@article{li_investigation_2019,
	title = {An investigation of single-domain and multidomain medication and adverse drug event relation extraction from electronic health record notes using advanced deep learning models},
	volume = {26},
	url = {https://academic.oup.com/jamia/article/26/7/646/5426087},
	doi = {10.1093/jamia/ocz018},
	abstract = {AbstractObjective.  We aim to evaluate the effectiveness of advanced deep learning models (eg, capsule network [CapNet], adversarial training [ADV]) for single-},
	language = {en},
	number = {7},
	urldate = {2019-12-09},
	journal = {Journal of the American Medical Informatics Association},
	author = {Li, Fei and Yu, Hong},
	month = jul,
	year = {2019},
	pages = {646--654}
}

AbstractObjective. We aim to evaluate the effectiveness of advanced deep learning models (eg, capsule network [CapNet], adversarial training [ADV]) for single-

@article{chen_detecting_2019,
	title = {Detecting {Hypoglycemia} {Incidents} {Reported} in {Patients}’ {Secure} {Messages}: {Using} {Cost}-{Sensitive} {Learning} and {Oversampling} to {Reduce} {Data} {Imbalance}},
	volume = {21},
	copyright = {Unless stated otherwise, all articles are open-access distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work (},
	shorttitle = {Detecting {Hypoglycemia} {Incidents} {Reported} in {Patients}’ {Secure} {Messages}},
	url = {https://www.jmir.org/2019/3/e11990/},
	doi = {10.2196/11990},
	abstract = {Background: Improper dosing of medications such as insulin can cause hypoglycemic episodes, which may lead to severe morbidity or even death. Although secure messaging was designed for exchanging nonurgent messages, patients sometimes report hypoglycemia events through secure messaging. Detecting these patient-reported adverse events may help alert clinical teams and enable early corrective actions to improve patient safety. Objective: We aimed to develop a natural language processing system, called HypoDetect (Hypoglycemia Detector), to automatically identify hypoglycemia incidents reported in patients’ secure messages. Methods: An expert in public health annotated 3000 secure message threads between patients with diabetes and US Department of Veterans Affairs clinical teams as containing patient-reported hypoglycemia incidents or not. A physician independently annotated 100 threads randomly selected from this dataset to determine interannotator agreement. We used this dataset to develop and evaluate HypoDetect. HypoDetect incorporates 3 machine learning algorithms widely used for text classification: linear support vector machines, random forest, and logistic regression. We explored different learning features, including new knowledge-driven features. Because only 114 (3.80\%) messages were annotated as positive, we investigated cost-sensitive learning and oversampling methods to mitigate the challenge of imbalanced data. Results: The interannotator agreement was Cohen kappa=.976. Using cross-validation, logistic regression with cost-sensitive learning achieved the best performance (area under the receiver operating characteristic curve=0.954, sensitivity=0.693, specificity 0.974, F1 score=0.590). Cost-sensitive learning and the ensembled synthetic minority oversampling technique improved the sensitivity of the baseline systems substantially (by 0.123 to 0.728 absolute gains). Our results show that a variety of features contributed to the best performance of HypoDetect. Conclusions: Despite the challenge of data imbalance, HypoDetect achieved promising results for the task of detecting hypoglycemia incidents from secure messages. The system has a great potential to facilitate early detection and treatment of hypoglycemia.  [J Med Internet Res 2019;21(3):e11990]},
	language = {en},
	number = {3},
	urldate = {2019-04-10},
	journal = {Journal of Medical Internet Research},
	author = {Chen, Jinying and Lalor, John and Liu, Weisong and Druhl, Emily and Granillo, Edgard and Vimalananda, Varsha G. and Yu, Hong},
	year = {2019},
	pages = {e11990}
}

Background: Improper dosing of medications such as insulin can cause hypoglycemic episodes, which may lead to severe morbidity or even death. Although secure messaging was designed for exchanging nonurgent messages, patients sometimes report hypoglycemia events through secure messaging. Detecting these patient-reported adverse events may help alert clinical teams and enable early corrective actions to improve patient safety. Objective: We aimed to develop a natural language processing system, called HypoDetect (Hypoglycemia Detector), to automatically identify hypoglycemia incidents reported in patients’ secure messages. Methods: An expert in public health annotated 3000 secure message threads between patients with diabetes and US Department of Veterans Affairs clinical teams as containing patient-reported hypoglycemia incidents or not. A physician independently annotated 100 threads randomly selected from this dataset to determine interannotator agreement. We used this dataset to develop and evaluate HypoDetect. HypoDetect incorporates 3 machine learning algorithms widely used for text classification: linear support vector machines, random forest, and logistic regression. We explored different learning features, including new knowledge-driven features. Because only 114 (3.80%) messages were annotated as positive, we investigated cost-sensitive learning and oversampling methods to mitigate the challenge of imbalanced data. Results: The interannotator agreement was Cohen kappa=.976. Using cross-validation, logistic regression with cost-sensitive learning achieved the best performance (area under the receiver operating characteristic curve=0.954, sensitivity=0.693, specificity 0.974, F1 score=0.590). Cost-sensitive learning and the ensembled synthetic minority oversampling technique improved the sensitivity of the baseline systems substantially (by 0.123 to 0.728 absolute gains). Our results show that a variety of features contributed to the best performance of HypoDetect. Conclusions: Despite the challenge of data imbalance, HypoDetect achieved promising results for the task of detecting hypoglycemia incidents from secure messages. The system has a great potential to facilitate early detection and treatment of hypoglycemia. [J Med Internet Res 2019;21(3):e11990]

@inproceedings{lalor_learning_2019,
	address = {Hong Kong, China},
	title = {Learning {Latent} {Parameters} without {Human} {Response} {Patterns}: {Item} {Response} {Theory} with {Artificial} {Crowds}},
	shorttitle = {Learning {Latent} {Parameters} without {Human} {Response} {Patterns}},
	url = {https://www.aclweb.org/anthology/D19-1434},
	doi = {10.18653/v1/D19-1434},
	abstract = {Incorporating Item Response Theory (IRT) into NLP tasks can provide valuable information about model performance and behavior. Traditionally, IRT models are learned using human response pattern (RP) data, presenting a significant bottleneck for large data sets like those required for training deep neural networks (DNNs). In this work we propose learning IRT models using RPs generated from artificial crowds of DNN models. We demonstrate the effectiveness of learning IRT models using DNN-generated data through quantitative and qualitative analyses for two NLP tasks. Parameters learned from human and machine RPs for natural language inference and sentiment analysis exhibit medium to large positive correlations. We demonstrate a use-case for latent difficulty item parameters, namely training set filtering, and show that using difficulty to sample training data outperforms baseline methods. Finally, we highlight cases where human expectation about item difficulty does not match difficulty as estimated from the machine RPs.},
	urldate = {2019-11-11},
	booktitle = {Proceedings of the 2019 {Conference} on {Empirical} {Methods} in {Natural} {Language} {Processing} and the 9th {International} {Joint} {Conference} on {Natural} {Language} {Processing} ({EMNLP}-{IJCNLP})},
	publisher = {Association for Computational Linguistics},
	author = {Lalor, John P. and Wu, Hao and Yu, Hong},
	month = nov,
	year = {2019},
	pages = {4240--4250}
}

Incorporating Item Response Theory (IRT) into NLP tasks can provide valuable information about model performance and behavior. Traditionally, IRT models are learned using human response pattern (RP) data, presenting a significant bottleneck for large data sets like those required for training deep neural networks (DNNs). In this work we propose learning IRT models using RPs generated from artificial crowds of DNN models. We demonstrate the effectiveness of learning IRT models using DNN-generated data through quantitative and qualitative analyses for two NLP tasks. Parameters learned from human and machine RPs for natural language inference and sentiment analysis exhibit medium to large positive correlations. We demonstrate a use-case for latent difficulty item parameters, namely training set filtering, and show that using difficulty to sample training data outperforms baseline methods. Finally, we highlight cases where human expectation about item difficulty does not match difficulty as estimated from the machine RPs.

@article{jagannatha_overview_2019,
	title = {Overview of the {First} {Natural} {Language} {Processing} {Challenge} for {Extracting} {Medication}, {Indication}, and {Adverse} {Drug} {Events} from {Electronic} {Health} {Record} {Notes} ({MADE} 1.0)},
	issn = {1179-1942},
	doi = {10.1007/s40264-018-0762-z},
	abstract = {INTRODUCTION: This work describes the Medication and Adverse Drug Events from Electronic Health Records (MADE 1.0) corpus and provides an overview of the MADE 1.0 2018 challenge for extracting medication, indication, and adverse drug events (ADEs) from electronic health record (EHR) notes.
OBJECTIVE: The goal of MADE is to provide a set of common evaluation tasks to assess the state of the art for natural language processing (NLP) systems applied to EHRs supporting drug safety surveillance and pharmacovigilance. We also provide benchmarks on the MADE dataset using the system submissions received in the MADE 2018 challenge.
METHODS: The MADE 1.0 challenge has released an expert-annotated cohort of medication and ADE information comprising 1089 fully de-identified longitudinal EHR notes from 21 randomly selected patients with cancer at the University of Massachusetts Memorial Hospital. Using this cohort as a benchmark, the MADE 1.0 challenge designed three shared NLP tasks. The named entity recognition (NER) task identifies medications and their attributes (dosage, route, duration, and frequency), indications, ADEs, and severity. The relation identification (RI) task identifies relations between the named entities: medication-indication, medication-ADE, and attribute relations. The third shared task (NER-RI) evaluates NLP models that perform the NER and RI tasks jointly. In total, 11 teams from four countries participated in at least one of the three shared tasks, and 41 system submissions were received in total.
RESULTS: The best systems F1 scores for NER, RI, and NER-RI were 0.82, 0.86, and 0.61, respectively. Ensemble classifiers using the team submissions improved the performance further, with an F1 score of 0.85, 0.87, and 0.66 for the three tasks, respectively.
CONCLUSION: MADE results show that recent progress in NLP has led to remarkable improvements in NER and RI tasks for the clinical domain. However, some room for improvement remains, particularly in the NER-RI task.},
	language = {eng},
	journal = {Drug Safety},
	author = {Jagannatha, Abhyuday and Liu, Feifan and Liu, Weisong and Yu, Hong},
	month = jan,
	year = {2019},
	pmid = {30649735}
}

INTRODUCTION: This work describes the Medication and Adverse Drug Events from Electronic Health Records (MADE 1.0) corpus and provides an overview of the MADE 1.0 2018 challenge for extracting medication, indication, and adverse drug events (ADEs) from electronic health record (EHR) notes. OBJECTIVE: The goal of MADE is to provide a set of common evaluation tasks to assess the state of the art for natural language processing (NLP) systems applied to EHRs supporting drug safety surveillance and pharmacovigilance. We also provide benchmarks on the MADE dataset using the system submissions received in the MADE 2018 challenge. METHODS: The MADE 1.0 challenge has released an expert-annotated cohort of medication and ADE information comprising 1089 fully de-identified longitudinal EHR notes from 21 randomly selected patients with cancer at the University of Massachusetts Memorial Hospital. Using this cohort as a benchmark, the MADE 1.0 challenge designed three shared NLP tasks. The named entity recognition (NER) task identifies medications and their attributes (dosage, route, duration, and frequency), indications, ADEs, and severity. The relation identification (RI) task identifies relations between the named entities: medication-indication, medication-ADE, and attribute relations. The third shared task (NER-RI) evaluates NLP models that perform the NER and RI tasks jointly. In total, 11 teams from four countries participated in at least one of the three shared tasks, and 41 system submissions were received in total. RESULTS: The best systems F1 scores for NER, RI, and NER-RI were 0.82, 0.86, and 0.61, respectively. Ensemble classifiers using the team submissions improved the performance further, with an F1 score of 0.85, 0.87, and 0.66 for the three tasks, respectively. CONCLUSION: MADE results show that recent progress in NLP has led to remarkable improvements in NER and RI tasks for the clinical domain. However, some room for improvement remains, particularly in the NER-RI task.

@article{zheng_quiklite_2019,
	title = {{QuikLitE}, a {Framework} for {Quick} {Literacy} {Evaluation} in {Medicine}: {Development} and {Validation}},
	volume = {21},
	copyright = {Unless stated otherwise, all articles are open-access distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work (},
	shorttitle = {{QuikLitE}, a {Framework} for {Quick} {Literacy} {Evaluation} in {Medicine}},
	url = {https://www.jmir.org/2019/2/e12525/},
	doi = {10.2196/jmir.12525},
	abstract = {Background: A plethora of health literacy instruments was developed over the decades. They usually start with experts curating passages of text or word lists, followed by psychometric validation and revision based on test results obtained from a sample population. This process is costly and it is difficult to customize for new usage scenarios. Objective: This study aimed to develop and evaluate a framework for dynamically creating test instruments that can provide a focused assessment of patients’ health literacy. Methods: A health literacy framework and scoring method were extended from the vocabulary knowledge test to accommodate a wide range of item difficulties and various degrees of uncertainty in the participant’s answer. Web-based tests from Amazon Mechanical Turk users were used to assess reliability and validity. Results: Parallel forms of our tests showed high reliability (correlation=.78; 95\% CI 0.69-0.85). Validity measured as correlation with an electronic health record comprehension instrument was higher (.47-.61 among 3 groups) than 2 existing tools (Short Assessment of Health Literacy-English, .38-.43; Short Test of Functional Health Literacy in Adults, .34-.46). Our framework is able to distinguish higher literacy levels that are often not measured by other instruments. It is also flexible, allowing customizations to the test the designer’s focus on a particular interest in a subject matter or domain. The framework is among the fastest health literacy instrument to administer. Conclusions: We proposed a valid and highly reliable framework to dynamically create health literacy instruments, alleviating the need to repeat a time-consuming process when a new use scenario arises. This framework can be customized to a specific need on demand and can measure skills beyond the basic level.  [J Med Internet Res 2019;21(2):e12525]},
	language = {en},
	number = {2},
	urldate = {2019-02-22},
	journal = {Journal of Medical Internet Research},
	author = {Zheng, Jiaping and Yu, Hong},
	year = {2019},
	pages = {e12525}
}

Background: A plethora of health literacy instruments was developed over the decades. They usually start with experts curating passages of text or word lists, followed by psychometric validation and revision based on test results obtained from a sample population. This process is costly and it is difficult to customize for new usage scenarios. Objective: This study aimed to develop and evaluate a framework for dynamically creating test instruments that can provide a focused assessment of patients’ health literacy. Methods: A health literacy framework and scoring method were extended from the vocabulary knowledge test to accommodate a wide range of item difficulties and various degrees of uncertainty in the participant’s answer. Web-based tests from Amazon Mechanical Turk users were used to assess reliability and validity. Results: Parallel forms of our tests showed high reliability (correlation=.78; 95% CI 0.69-0.85). Validity measured as correlation with an electronic health record comprehension instrument was higher (.47-.61 among 3 groups) than 2 existing tools (Short Assessment of Health Literacy-English, .38-.43; Short Test of Functional Health Literacy in Adults, .34-.46). Our framework is able to distinguish higher literacy levels that are often not measured by other instruments. It is also flexible, allowing customizations to the test the designer’s focus on a particular interest in a subject matter or domain. The framework is among the fastest health literacy instrument to administer. Conclusions: We proposed a valid and highly reliable framework to dynamically create health literacy instruments, alleviating the need to repeat a time-consuming process when a new use scenario arises. This framework can be customized to a specific need on demand and can measure skills beyond the basic level. [J Med Internet Res 2019;21(2):e12525]

@article{liu_towards_2019,
	title = {Towards {Drug} {Safety} {Surveillance} and {Pharmacovigilance}: {Current} {Progress} in {Detecting} {Medication} and {Adverse} {Drug} {Events} from {Electronic} {Health} {Records}},
	issn = {1179-1942},
	shorttitle = {Towards {Drug} {Safety} {Surveillance} and {Pharmacovigilance}},
	url = {https://doi.org/10.1007/s40264-018-0766-8},
	doi = {10.1007/s40264-018-0766-8},
	language = {en},
	urldate = {2019-01-31},
	journal = {Drug Safety},
	author = {Liu, Feifan and Jagannatha, Abhyuday and Yu, Hong},
	month = jan,
	year = {2019}
}

@inproceedings{yang_generating_2019,
	address = {Hong Kong, China},
	title = {Generating {Classical} {Chinese} {Poems} from {Vernacular} {Chinese}},
	url = {https://www.aclweb.org/anthology/D19-1637},
	doi = {10.18653/v1/D19-1637},
	abstract = {Classical Chinese poetry is a jewel in the treasure house of Chinese culture. Previous poem generation models only allow users to employ keywords to interfere the meaning of generated poems, leaving the dominion of generation to the model. In this paper, we propose a novel task of generating classical Chinese poems from vernacular, which allows users to have more control over the semantic of generated poems. We adapt the approach of unsupervised machine translation (UMT) to our task. We use segmentation-based padding and reinforcement learning to address under-translation and over-translation respectively. According to experiments, our approach significantly improve the perplexity and BLEU compared with typical UMT models. Furthermore, we explored guidelines on how to write the input vernacular to generate better poems. Human evaluation showed our approach can generate high-quality poems which are comparable to amateur poems.},
	urldate = {2019-11-11},
	booktitle = {Proceedings of the 2019 {Conference} on {Empirical} {Methods} in {Natural} {Language} {Processing} and the 9th {International} {Joint} {Conference} on {Natural} {Language} {Processing} ({EMNLP}-{IJCNLP})},
	publisher = {Association for Computational Linguistics},
	author = {Yang, Zhichao and Cai, Pengshan and Feng, Yansong and Li, Fei and Feng, Weijiang and Chiu, Elena Suet-Ying and yu, hong},
	month = nov,
	year = {2019},
	pages = {6156--6165}
}

Classical Chinese poetry is a jewel in the treasure house of Chinese culture. Previous poem generation models only allow users to employ keywords to interfere the meaning of generated poems, leaving the dominion of generation to the model. In this paper, we propose a novel task of generating classical Chinese poems from vernacular, which allows users to have more control over the semantic of generated poems. We adapt the approach of unsupervised machine translation (UMT) to our task. We use segmentation-based padding and reinforcement learning to address under-translation and over-translation respectively. According to experiments, our approach significantly improve the perplexity and BLEU compared with typical UMT models. Furthermore, we explored guidelines on how to write the input vernacular to generate better poems. Human evaluation showed our approach can generate high-quality poems which are comparable to amateur poems.

@article{jin_automatic_2019,
	title = {Automatic {Detection} of {Hypoglycemic} {Events} {From} the {Electronic} {Health} {Record} {Notes} of {Diabetes} {Patients}: {Empirical} {Study}},
	volume = {7},
	copyright = {Unless stated otherwise, all articles are open-access distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work (},
	shorttitle = {Automatic {Detection} of {Hypoglycemic} {Events} {From} the {Electronic} {Health} {Record} {Notes} of {Diabetes} {Patients}},
	url = {https://medinform.jmir.org/2019/4/e14340/},
	doi = {10.2196/14340},
	abstract = {Background:  Hypoglycemic events are common and potentially dangerous conditions among patients being treated for diabetes. Automatic detection of such events could improve patient care and is valuable in population studies. Electronic health records (EHRs) are valuable resources for the detection of such events.
 Objective:  In this study, we aim to develop a deep-learning–based natural language processing (NLP) system to automatically detect hypoglycemic events from EHR notes. Our model is called the High-Performing System for Automatically Detecting Hypoglycemic Events (HYPE).
 Methods:  Domain experts reviewed 500 EHR notes of diabetes patients to determine whether each sentence contained a hypoglycemic event or not. We used this annotated corpus to train and evaluate HYPE, the high-performance NLP system for hypoglycemia detection. We built and evaluated both a classical machine learning model (ie, support vector machines [SVMs]) and state-of-the-art neural network models.
 Results:  We found that neural network models outperformed the SVM model. The convolutional neural network (CNN) model yielded the highest performance in a 10-fold cross-validation setting: mean precision=0.96 (SD 0.03), mean recall=0.86 (SD 0.03), and mean F1=0.91 (SD 0.03).
 Conclusions:  Despite the challenges posed by small and highly imbalanced data, our CNN-based HYPE system still achieved a high performance for hypoglycemia detection. HYPE can be used for EHR-based hypoglycemia surveillance and population studies in diabetes patients.
 [JMIR Med Inform 2019;7(4):e14340]},
	language = {en},
	number = {4},
	urldate = {2019-11-10},
	journal = {JMIR Medical Informatics},
	author = {Jin, Yonghao and Li, Fei and Vimalananda, Varsha G. and Yu, Hong},
	year = {2019},
	pages = {e14340}
}

Background: Hypoglycemic events are common and potentially dangerous conditions among patients being treated for diabetes. Automatic detection of such events could improve patient care and is valuable in population studies. Electronic health records (EHRs) are valuable resources for the detection of such events. Objective: In this study, we aim to develop a deep-learning–based natural language processing (NLP) system to automatically detect hypoglycemic events from EHR notes. Our model is called the High-Performing System for Automatically Detecting Hypoglycemic Events (HYPE). Methods: Domain experts reviewed 500 EHR notes of diabetes patients to determine whether each sentence contained a hypoglycemic event or not. We used this annotated corpus to train and evaluate HYPE, the high-performance NLP system for hypoglycemia detection. We built and evaluated both a classical machine learning model (ie, support vector machines [SVMs]) and state-of-the-art neural network models. Results: We found that neural network models outperformed the SVM model. The convolutional neural network (CNN) model yielded the highest performance in a 10-fold cross-validation setting: mean precision=0.96 (SD 0.03), mean recall=0.86 (SD 0.03), and mean F1=0.91 (SD 0.03). Conclusions: Despite the challenges posed by small and highly imbalanced data, our CNN-based HYPE system still achieved a high performance for hypoglycemia detection. HYPE can be used for EHR-based hypoglycemia surveillance and population studies in diabetes patients. [JMIR Med Inform 2019;7(4):e14340]

@article{rose_anticoagulant_2019,
	title = {Anticoagulant prescribing for non-valvular atrial fibrillation in the {Veterans} {Health} {Administration}},
	journal = {Journal of the American Heart Association},
	author = {Rose, AJ and Goldberg, R and McManus, DD and Kapoor, A and Wang, V and Liu, W and Yu, H},
	year = {2019}
}

@inproceedings{singh_clinical_2019,
	title = {Clinical {Question} {Answering} from {Electronic} {Health} {Records}. {In} the {MLHC} 2019 research track proceedings},
	booktitle = {The {MLHC} 2019 research track proceedings},
	author = {Singh, Bhanu and Li, Fei and Yu, Hong},
	year = {2019}
}

@article{lalor_comparing_2019,
	title = {Comparing {Human} and {DNN}-{Ensemble} {Response} {Patterns} for {Item} {Response} {Theory} {Model} {Fitting}},
	journal = {2019 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)The Workshop on Cognitive Modeling and Computational Linguistics (CMCL)},
	author = {Lalor, John and Wu, Hao and Yu, Hong},
	year = {2019}
}

@article{rawat_naranjo_2019,
	title = {Naranjo {Question} {Answering} using {End}-to-{End} {Multi}-task {Learning} {Model}},
	journal = {25th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
	author = {Rawat, Bhanu P and Li, Fei and Yu, Hong},
	year = {2019}
}

@patent{yu_method_2019,
	title = {Method for {Meta}-{Level} {Continual} {Learning}},
	url = {https://patents.google.com/patent/US20190034798A1/en},
	abstract = {Classification of an input task data set by meta level continual learning includes analyzing first and second training data sets in a task space to generate first and second meta weights and a slow weight value, and comparing an input task data set to the slow weight to generate a fast weight. The first and second meta weights are parameterized with the fast weight value to update the slow weight value, whereby a value is associated with the input task data set, thereby classifying the input task data set by meta level continual learning.},
	nationality = {US},
	assignee = {University Of Massachusetts Medical School},
	number = {US20190034798A1},
	urldate = {2019-04-10},
	author = {Yu, Hong and Munkhdalai, Tsendsuren},
	month = jan,
	year = {2019},
	keywords = {loss, meta, slow, task, weight}
}

Classification of an input task data set by meta level continual learning includes analyzing first and second training data sets in a task space to generate first and second meta weights and a slow weight value, and comparing an input task data set to the slow weight to generate a fast weight. The first and second meta weights are parameterized with the fast weight value to update the slow weight value, whereby a value is associated with the input task data set, thereby classifying the input task data set by meta level continual learning.

@incollection{liu_advancing_2019,
	address = {Cham},
	series = {Health {Informatics}},
	title = {Advancing {Clinical} {Research} {Through} {Natural} {Language} {Processing} on {Electronic} {Health} {Records}: {Traditional} {Machine} {Learning} {Meets} {Deep} {Learning}},
	isbn = {978-3-319-98779-8},
	shorttitle = {Advancing {Clinical} {Research} {Through} {Natural} {Language} {Processing} on {Electronic} {Health} {Records}},
	url = {https://doi.org/10.1007/978-3-319-98779-8_17},
	abstract = {Electronic health records (EHR) capture “real-world” disease and care processes and hence offer richer and more generalizable data for comparative effectiveness research than traditional randomized clinical trial studies. With the increasingly broadening adoption of EHR worldwide, there is a growing need to widen the use of EHR data to support clinical research. A big barrier to this goal is that much of the information in EHR is still narrative. This chapter describes the foundation of biomedical language processing and explains how traditional machine learning and the state-of-the-art deep learning techniques can be employed in the context of extracting and transforming narrative information in EHR to support clinical research.},
	language = {en},
	urldate = {2019-04-09},
	booktitle = {Clinical {Research} {Informatics}},
	publisher = {Springer International Publishing},
	author = {Liu, Feifan and Weng, Chunhua and Yu, Hong},
	editor = {Richesson, Rachel L. and Andrews, James E.},
	year = {2019},
	doi = {10.1007/978-3-319-98779-8_17},
	keywords = {Biomedical natural language processing, Clinical research, Deep learning, Electronic health records, Machine learning, Rule-based approach},
	pages = {357--378}
}

Electronic health records (EHR) capture “real-world” disease and care processes and hence offer richer and more generalizable data for comparative effectiveness research than traditional randomized clinical trial studies. With the increasingly broadening adoption of EHR worldwide, there is a growing need to widen the use of EHR data to support clinical research. A big barrier to this goal is that much of the information in EHR is still narrative. This chapter describes the foundation of biomedical language processing and explains how traditional machine learning and the state-of-the-art deep learning techniques can be employed in the context of extracting and transforming narrative information in EHR to support clinical research.

@article{lalor_improving_2019,
	title = {Improving {Electronic} {Health} {Record} {Note} {Comprehension} {With} {NoteAid}: {Randomized} {Trial} of {Electronic} {Health} {Record} {Note} {Comprehension} {Interventions} {With} {Crowdsourced} {Workers}},
	volume = {21},
	copyright = {Unless stated otherwise, all articles are open-access distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work (},
	shorttitle = {Improving {Electronic} {Health} {Record} {Note} {Comprehension} {With} {NoteAid}},
	url = {https://www.jmir.org/2019/1/e10793/},
	doi = {10.2196/jmir.10793},
	abstract = {Background: Patient portals are becoming more common, and with them, the ability of patients to access their personal electronic health records (EHRs). EHRs, in particular the free-text EHR notes, often contain medical jargon and terms that are difficult for laypersons to understand. There are many Web-based resources for learning more about particular diseases or conditions, including systems that directly link to lay definitions or educational materials for medical concepts. Objective: Our goal is to determine whether use of one such tool, NoteAid, leads to higher EHR note comprehension ability. We use a new EHR note comprehension assessment tool instead of patient self-reported scores. Methods: In this work, we compare a passive, self-service educational resource (MedlinePlus) with an active resource (NoteAid) where definitions are provided to the user for medical concepts that the system identifies. We use Amazon Mechanical Turk (AMT) to recruit individuals to complete ComprehENotes, a new test of EHR note comprehension. Results: Mean scores for individuals with access to NoteAid are significantly higher than the mean baseline scores, both for raw scores (P=.008) and estimated ability (P=.02). Conclusions: In our experiments, we show that the active intervention leads to significantly higher scores on the comprehension test as compared with a baseline group with no resources provided. In contrast, there is no significant difference between the group that was provided with the passive intervention and the baseline group. Finally, we analyze the demographics of the individuals who participated in our AMT task and show differences between groups that align with the current understanding of health literacy between populations. This is the first work to show improvements in comprehension using tools such as NoteAid as measured by an EHR note comprehension assessment tool as opposed to patient self-reported scores.  [J Med Internet Res 2019;21(1):e10793]},
	language = {en},
	number = {1},
	urldate = {2019-01-31},
	journal = {Journal of Medical Internet Research},
	author = {Lalor, John P. and Woolf, Beverly and Yu, Hong},
	year = {2019},
	pages = {e10793}
}

Background: Patient portals are becoming more common, and with them, the ability of patients to access their personal electronic health records (EHRs). EHRs, in particular the free-text EHR notes, often contain medical jargon and terms that are difficult for laypersons to understand. There are many Web-based resources for learning more about particular diseases or conditions, including systems that directly link to lay definitions or educational materials for medical concepts. Objective: Our goal is to determine whether use of one such tool, NoteAid, leads to higher EHR note comprehension ability. We use a new EHR note comprehension assessment tool instead of patient self-reported scores. Methods: In this work, we compare a passive, self-service educational resource (MedlinePlus) with an active resource (NoteAid) where definitions are provided to the user for medical concepts that the system identifies. We use Amazon Mechanical Turk (AMT) to recruit individuals to complete ComprehENotes, a new test of EHR note comprehension. Results: Mean scores for individuals with access to NoteAid are significantly higher than the mean baseline scores, both for raw scores (P=.008) and estimated ability (P=.02). Conclusions: In our experiments, we show that the active intervention leads to significantly higher scores on the comprehension test as compared with a baseline group with no resources provided. In contrast, there is no significant difference between the group that was provided with the passive intervention and the baseline group. Finally, we analyze the demographics of the individuals who participated in our AMT task and show differences between groups that align with the current understanding of health literacy between populations. This is the first work to show improvements in comprehension using tools such as NoteAid as measured by an EHR note comprehension assessment tool as opposed to patient self-reported scores. [J Med Internet Res 2019;21(1):e10793]

@article{pradhan_automatic_2019,
	title = {Automatic extraction of quantitative data from {ClinicalTrials}.gov to conduct meta-analyses},
	volume = {105},
	issn = {1878-5921},
	doi = {10.1016/j.jclinepi.2018.08.023},
	abstract = {OBJECTIVES: Systematic reviews and meta-analyses are labor-intensive and time-consuming. Automated extraction of quantitative data from primary studies can accelerate this process. ClinicalTrials.gov, launched in 2000, is the world's largest trial repository of results data from clinical trials; it has been used as a source instead of journal articles. We have developed a Web application called EXACT (EXtracting Accurate efficacy and safety information from ClinicalTrials.gov) that allows users without advanced programming skills to automatically extract data from ClinicalTrials.gov in analysis-ready format. We have also used the automatically extracted data to examine the reproducibility of meta-analyses in three published systematic reviews.
STUDY DESIGN AND SETTING: We developed a Python-based software application (EXACT) that automatically extracts data required for meta-analysis from the ClinicalTrials.gov database in a spreadsheet format. We confirmed the accuracy of the extracted data and then used those data to repeat meta-analyses in three published systematic reviews. To ensure that we used the same statistical methods and outcomes as the published systematic reviews, we repeated the meta-analyses using data manually extracted from the relevant journal articles. For the outcomes whose results we were able to reproduce using those journal article data, we examined the usability of ClinicalTrials.gov data.
RESULTS: EXACT extracted data at ClincalTrials.gov with 100\% accuracy, and it required 60\% less time than the usual practice of manually extracting data from journal articles. We found that 87\% of the data elements extracted using EXACT matched those extracted manually from the journal articles. We were able to reproduce 24 of 28 outcomes using the journal article data. Of these 24 outcomes, we were able to reproduce 83.3\% of the published estimates using data at ClinicalTrials.gov.
CONCLUSION: EXACT (http://bio-nlp.org/EXACT) automatically and accurately extracted data elements from ClinicalTrials.gov and thus reduced time in data extraction. The ClinicalTrials.gov data reproduced most meta-analysis results in our study, but this conclusion needs further validation.},
	language = {eng},
	journal = {Journal of Clinical Epidemiology},
	author = {Pradhan, Richeek and Hoaglin, David C. and Cornell, Matthew and Liu, Weisong and Wang, Victoria and Yu, Hong},
	month = jan,
	year = {2019},
	pmid = {30257185},
	keywords = {Automatic data extraction, ClinicalTrials.gov, Meta-analysis, Reproducibility, Simeprevir, Systematic review, Trametinib, Vortioxetine},
	pages = {92--100}
}

OBJECTIVES: Systematic reviews and meta-analyses are labor-intensive and time-consuming. Automated extraction of quantitative data from primary studies can accelerate this process. ClinicalTrials.gov, launched in 2000, is the world's largest trial repository of results data from clinical trials; it has been used as a source instead of journal articles. We have developed a Web application called EXACT (EXtracting Accurate efficacy and safety information from ClinicalTrials.gov) that allows users without advanced programming skills to automatically extract data from ClinicalTrials.gov in analysis-ready format. We have also used the automatically extracted data to examine the reproducibility of meta-analyses in three published systematic reviews. STUDY DESIGN AND SETTING: We developed a Python-based software application (EXACT) that automatically extracts data required for meta-analysis from the ClinicalTrials.gov database in a spreadsheet format. We confirmed the accuracy of the extracted data and then used those data to repeat meta-analyses in three published systematic reviews. To ensure that we used the same statistical methods and outcomes as the published systematic reviews, we repeated the meta-analyses using data manually extracted from the relevant journal articles. For the outcomes whose results we were able to reproduce using those journal article data, we examined the usability of ClinicalTrials.gov data. RESULTS: EXACT extracted data at ClincalTrials.gov with 100% accuracy, and it required 60% less time than the usual practice of manually extracting data from journal articles. We found that 87% of the data elements extracted using EXACT matched those extracted manually from the journal articles. We were able to reproduce 24 of 28 outcomes using the journal article data. Of these 24 outcomes, we were able to reproduce 83.3% of the published estimates using data at ClinicalTrials.gov. CONCLUSION: EXACT (http://bio-nlp.org/EXACT) automatically and accurately extracted data elements from ClinicalTrials.gov and thus reduced time in data extraction. The ClinicalTrials.gov data reproduced most meta-analysis results in our study, but this conclusion needs further validation.

@article{chen_natural_2018,
	title = {A {Natural} {Language} {Processing} {System} {That} {Links} {Medical} {Terms} in {Electronic} {Health} {Record} {Notes} to {Lay} {Definitions}: {System} {Development} {Using} {Physician} {Reviews}},
	volume = {20},
	issn = {1438-8871},
	shorttitle = {A {Natural} {Language} {Processing} {System} {That} {Links} {Medical} {Terms} in {Electronic} {Health} {Record} {Notes} to {Lay} {Definitions}},
	doi = {10.2196/jmir.8669},
	abstract = {BACKGROUND: Many health care systems now allow patients to access their electronic health record (EHR) notes online through patient portals. Medical jargon in EHR notes can confuse patients, which may interfere with potential benefits of patient access to EHR notes.
OBJECTIVE: The aim of this study was to develop and evaluate the usability and content quality of NoteAid, a Web-based natural language processing system that links medical terms in EHR notes to lay definitions, that is, definitions easily understood by lay people.
METHODS: NoteAid incorporates two core components: CoDeMed, a lexical resource of lay definitions for medical terms, and MedLink, a computational unit that links medical terms to lay definitions. We developed innovative computational methods, including an adapted distant supervision algorithm to prioritize medical terms important for EHR comprehension to facilitate the effort of building CoDeMed. Ten physician domain experts evaluated the user interface and content quality of NoteAid. The evaluation protocol included a cognitive walkthrough session and a postsession questionnaire. Physician feedback sessions were audio-recorded. We used standard content analysis methods to analyze qualitative data from these sessions.
RESULTS: Physician feedback was mixed. Positive feedback on NoteAid included (1) Easy to use, (2) Good visual display, (3) Satisfactory system speed, and (4) Adequate lay definitions. Opportunities for improvement arising from evaluation sessions and feedback included (1) improving the display of definitions for partially matched terms, (2) including more medical terms in CoDeMed, (3) improving the handling of terms whose definitions vary depending on different contexts, and (4) standardizing the scope of definitions for medicines. On the basis of these results, we have improved NoteAid's user interface and a number of definitions, and added 4502 more definitions in CoDeMed.
CONCLUSIONS: Physician evaluation yielded useful feedback for content validation and refinement of this innovative tool that has the potential to improve patient EHR comprehension and experience using patient portals. Future ongoing work will develop algorithms to handle ambiguous medical terms and test and evaluate NoteAid with patients.},
	language = {eng},
	number = {1},
	journal = {Journal of Medical Internet Research},
	author = {Chen, Jinying and Druhl, Emily and Polepalli Ramesh, Balaji and Houston, Thomas K. and Brandt, Cynthia A. and Zulman, Donna M. and Vimalananda, Varsha G. and Malkani, Samir and Yu, Hong},
	month = jan,
	year = {2018},
	pmid = {29358159},
	pmcid = {PMC5799720},
	keywords = {computer software, consumer health informatics, electronic health records, natural language processing, usability testing},
	pages = {e26}
}

BACKGROUND: Many health care systems now allow patients to access their electronic health record (EHR) notes online through patient portals. Medical jargon in EHR notes can confuse patients, which may interfere with potential benefits of patient access to EHR notes. OBJECTIVE: The aim of this study was to develop and evaluate the usability and content quality of NoteAid, a Web-based natural language processing system that links medical terms in EHR notes to lay definitions, that is, definitions easily understood by lay people. METHODS: NoteAid incorporates two core components: CoDeMed, a lexical resource of lay definitions for medical terms, and MedLink, a computational unit that links medical terms to lay definitions. We developed innovative computational methods, including an adapted distant supervision algorithm to prioritize medical terms important for EHR comprehension to facilitate the effort of building CoDeMed. Ten physician domain experts evaluated the user interface and content quality of NoteAid. The evaluation protocol included a cognitive walkthrough session and a postsession questionnaire. Physician feedback sessions were audio-recorded. We used standard content analysis methods to analyze qualitative data from these sessions. RESULTS: Physician feedback was mixed. Positive feedback on NoteAid included (1) Easy to use, (2) Good visual display, (3) Satisfactory system speed, and (4) Adequate lay definitions. Opportunities for improvement arising from evaluation sessions and feedback included (1) improving the display of definitions for partially matched terms, (2) including more medical terms in CoDeMed, (3) improving the handling of terms whose definitions vary depending on different contexts, and (4) standardizing the scope of definitions for medicines. On the basis of these results, we have improved NoteAid's user interface and a number of definitions, and added 4502 more definitions in CoDeMed. CONCLUSIONS: Physician evaluation yielded useful feedback for content validation and refinement of this innovative tool that has the potential to improve patient EHR comprehension and experience using patient portals. Future ongoing work will develop algorithms to handle ambiguous medical terms and test and evaluate NoteAid with patients.

@article{munkhdalai_clinical_2018,
	title = {Clinical {Relation} {Extraction} {Toward} {Drug} {Safety} {Surveillance} {Using} {Electronic} {Health} {Record} {Narratives}: {Classical} {Learning} {Versus} {Deep} {Learning}},
	volume = {4},
	issn = {2369-2960},
	shorttitle = {Clinical {Relation} {Extraction} {Toward} {Drug} {Safety} {Surveillance} {Using} {Electronic} {Health} {Record} {Narratives}},
	doi = {10.2196/publichealth.9361},
	abstract = {BACKGROUND: Medication and adverse drug event (ADE) information extracted from electronic health record (EHR) notes can be a rich resource for drug safety surveillance. Existing observational studies have mainly relied on structured EHR data to obtain ADE information; however, ADEs are often buried in the EHR narratives and not recorded in structured data.
OBJECTIVE: To unlock ADE-related information from EHR narratives, there is a need to extract relevant entities and identify relations among them. In this study, we focus on relation identification. This study aimed to evaluate natural language processing and machine learning approaches using the expert-annotated medical entities and relations in the context of drug safety surveillance, and investigate how different learning approaches perform under different configurations.
METHODS: We have manually annotated 791 EHR notes with 9 named entities (eg, medication, indication, severity, and ADEs) and 7 different types of relations (eg, medication-dosage, medication-ADE, and severity-ADE). Then, we explored 3 supervised machine learning systems for relation identification: (1) a support vector machines (SVM) system, (2) an end-to-end deep neural network system, and (3) a supervised descriptive rule induction baseline system. For the neural network system, we exploited the state-of-the-art recurrent neural network (RNN) and attention models. We report the performance by macro-averaged precision, recall, and F1-score across the relation types.
RESULTS: Our results show that the SVM model achieved the best average F1-score of 89.1\% on test data, outperforming the long short-term memory (LSTM) model with attention (F1-score of 65.72\%) as well as the rule induction baseline system (F1-score of 7.47\%) by a large margin. The bidirectional LSTM model with attention achieved the best performance among different RNN models. With the inclusion of additional features in the LSTM model, its performance can be boosted to an average F1-score of 77.35\%.
CONCLUSIONS: It shows that classical learning models (SVM) remains advantageous over deep learning models (RNN variants) for clinical relation identification, especially for long-distance intersentential relations. However, RNNs demonstrate a great potential of significant improvement if more training data become available. Our work is an important step toward mining EHRs to improve the efficacy of drug safety surveillance. Most importantly, the annotated data used in this study will be made publicly available, which will further promote drug safety research in the community.},
	language = {eng},
	number = {2},
	journal = {JMIR public health and surveillance},
	author = {Munkhdalai, Tsendsuren and Liu, Feifan and Yu, Hong},
	month = apr,
	year = {2018},
	pmid = {29695376},
	pmcid = {PMC5943628},
	keywords = {drug-related side effects and adverse reactions, electronic health records, medical informatics applications, natural language processing, neural networks},
	pages = {e29}
}

BACKGROUND: Medication and adverse drug event (ADE) information extracted from electronic health record (EHR) notes can be a rich resource for drug safety surveillance. Existing observational studies have mainly relied on structured EHR data to obtain ADE information; however, ADEs are often buried in the EHR narratives and not recorded in structured data. OBJECTIVE: To unlock ADE-related information from EHR narratives, there is a need to extract relevant entities and identify relations among them. In this study, we focus on relation identification. This study aimed to evaluate natural language processing and machine learning approaches using the expert-annotated medical entities and relations in the context of drug safety surveillance, and investigate how different learning approaches perform under different configurations. METHODS: We have manually annotated 791 EHR notes with 9 named entities (eg, medication, indication, severity, and ADEs) and 7 different types of relations (eg, medication-dosage, medication-ADE, and severity-ADE). Then, we explored 3 supervised machine learning systems for relation identification: (1) a support vector machines (SVM) system, (2) an end-to-end deep neural network system, and (3) a supervised descriptive rule induction baseline system. For the neural network system, we exploited the state-of-the-art recurrent neural network (RNN) and attention models. We report the performance by macro-averaged precision, recall, and F1-score across the relation types. RESULTS: Our results show that the SVM model achieved the best average F1-score of 89.1% on test data, outperforming the long short-term memory (LSTM) model with attention (F1-score of 65.72%) as well as the rule induction baseline system (F1-score of 7.47%) by a large margin. The bidirectional LSTM model with attention achieved the best performance among different RNN models. With the inclusion of additional features in the LSTM model, its performance can be boosted to an average F1-score of 77.35%. CONCLUSIONS: It shows that classical learning models (SVM) remains advantageous over deep learning models (RNN variants) for clinical relation identification, especially for long-distance intersentential relations. However, RNNs demonstrate a great potential of significant improvement if more training data become available. Our work is an important step toward mining EHRs to improve the efficacy of drug safety surveillance. Most importantly, the annotated data used in this study will be made publicly available, which will further promote drug safety research in the community.

@inproceedings{chen_detecting_2018,
	title = {Detecting {Hypoglycemia} {Incidence} from {Patients}’ {Secure} {Messages}},
	author = {Chen, J and Yu, H},
	year = {2018}
}

@article{li_extraction_2018,
	title = {Extraction of {Information} {Related} to {Adverse} {Drug} {Events} from {Electronic} {Health} {Record} {Notes}: {Design} of an {End}-to-{End} {Model} {Based} on {Deep} {Learning}},
	volume = {6},
	issn = {2291-9694},
	shorttitle = {Extraction of {Information} {Related} to {Adverse} {Drug} {Events} from {Electronic} {Health} {Record} {Notes}},
	doi = {10.2196/12159},
	abstract = {BACKGROUND: Pharmacovigilance and drug-safety surveillance are crucial for monitoring adverse drug events (ADEs), but the main ADE-reporting systems such as Food and Drug Administration Adverse Event Reporting System face challenges such as underreporting. Therefore, as complementary surveillance, data on ADEs are extracted from electronic health record (EHR) notes via natural language processing (NLP). As NLP develops, many up-to-date machine-learning techniques are introduced in this field, such as deep learning and multi-task learning (MTL). However, only a few studies have focused on employing such techniques to extract ADEs.
OBJECTIVE: We aimed to design a deep learning model for extracting ADEs and related information such as medications and indications. Since extraction of ADE-related information includes two steps-named entity recognition and relation extraction-our second objective was to improve the deep learning model using multi-task learning between the two steps.
METHODS: We employed the dataset from the Medication, Indication and Adverse Drug Events (MADE) 1.0 challenge to train and test our models. This dataset consists of 1089 EHR notes of cancer patients and includes 9 entity types such as Medication, Indication, and ADE and 7 types of relations between these entities. To extract information from the dataset, we proposed a deep-learning model that uses a bidirectional long short-term memory (BiLSTM) conditional random field network to recognize entities and a BiLSTM-Attention network to extract relations. To further improve the deep-learning model, we employed three typical MTL methods, namely, hard parameter sharing, parameter regularization, and task relation learning, to build three MTL models, called HardMTL, RegMTL, and LearnMTL, respectively.
RESULTS: Since extraction of ADE-related information is a two-step task, the result of the second step (ie, relation extraction) was used to compare all models. We used microaveraged precision, recall, and F1 as evaluation metrics. Our deep learning model achieved state-of-the-art results (F1=65.9\%), which is significantly higher than that (F1=61.7\%) of the best system in the MADE1.0 challenge. HardMTL further improved the F1 by 0.8\%, boosting the F1 to 66.7\%, whereas RegMTL and LearnMTL failed to boost the performance.
CONCLUSIONS: Deep learning models can significantly improve the performance of ADE-related information extraction. MTL may be effective for named entity recognition and relation extraction, but it depends on the methods, data, and other factors. Our results can facilitate research on ADE detection, NLP, and machine learning.},
	language = {eng},
	number = {4},
	journal = {JMIR medical informatics},
	author = {Li, Fei and Liu, Weisong and Yu, Hong},
	month = nov,
	year = {2018},
	pmid = {30478023},
	keywords = {adverse drug event, deep learning, multi-task learning, named entity recognition, natural language processing, relation extraction},
	pages = {e12159}
}

BACKGROUND: Pharmacovigilance and drug-safety surveillance are crucial for monitoring adverse drug events (ADEs), but the main ADE-reporting systems such as Food and Drug Administration Adverse Event Reporting System face challenges such as underreporting. Therefore, as complementary surveillance, data on ADEs are extracted from electronic health record (EHR) notes via natural language processing (NLP). As NLP develops, many up-to-date machine-learning techniques are introduced in this field, such as deep learning and multi-task learning (MTL). However, only a few studies have focused on employing such techniques to extract ADEs. OBJECTIVE: We aimed to design a deep learning model for extracting ADEs and related information such as medications and indications. Since extraction of ADE-related information includes two steps-named entity recognition and relation extraction-our second objective was to improve the deep learning model using multi-task learning between the two steps. METHODS: We employed the dataset from the Medication, Indication and Adverse Drug Events (MADE) 1.0 challenge to train and test our models. This dataset consists of 1089 EHR notes of cancer patients and includes 9 entity types such as Medication, Indication, and ADE and 7 types of relations between these entities. To extract information from the dataset, we proposed a deep-learning model that uses a bidirectional long short-term memory (BiLSTM) conditional random field network to recognize entities and a BiLSTM-Attention network to extract relations. To further improve the deep-learning model, we employed three typical MTL methods, namely, hard parameter sharing, parameter regularization, and task relation learning, to build three MTL models, called HardMTL, RegMTL, and LearnMTL, respectively. RESULTS: Since extraction of ADE-related information is a two-step task, the result of the second step (ie, relation extraction) was used to compare all models. We used microaveraged precision, recall, and F1 as evaluation metrics. Our deep learning model achieved state-of-the-art results (F1=65.9%), which is significantly higher than that (F1=61.7%) of the best system in the MADE1.0 challenge. HardMTL further improved the F1 by 0.8%, boosting the F1 to 66.7%, whereas RegMTL and LearnMTL failed to boost the performance. CONCLUSIONS: Deep learning models can significantly improve the performance of ADE-related information extraction. MTL may be effective for named entity recognition and relation extraction, but it depends on the methods, data, and other factors. Our results can facilitate research on ADE detection, NLP, and machine learning.

@article{pradhan_inadequate_2018,
	title = {Inadequate diversity of information resources searched in {US}-affiliated systematic reviews and meta-analyses: 2005-2016},
	volume = {102},
	issn = {1878-5921},
	shorttitle = {Inadequate diversity of information resources searched in {US}-affiliated systematic reviews and meta-analyses},
	doi = {10.1016/j.jclinepi.2018.05.024},
	abstract = {OBJECTIVE: Systematic reviews and meta-analyses (SRMAs) rely upon comprehensive searches into diverse resources that catalog primary studies. However, since what constitutes a comprehensive search is unclear, we examined trends in databases searched from 2005-2016, surrounding the publication of search guidelines in 2013, and associations between resources searched and evidence of publication bias in SRMAs involving human subjects.
STUDY DESIGN: To ensure comparability of included SRMAs over the 12 years in the face of a near 100-fold increase of international SRMAs (mainly genetic studies from China) during this period, we focused on USA-affiliated SRMAs, manually reviewing 100 randomly selected SRMAs from those published in each year. After excluding articles (mainly for inadequate detail or out-of-scope methods), we identified factors associated with the databases searched, used network analysis to see which resources were simultaneously searched, and used logistic regression to link information sources searched with a lower chance of finding publication bias.
RESULTS: Among 817 SRMA articles studied, the common resources used were Medline (95\%), EMBASE (44\%), and Cochrane (41\%). Methods journal SRMAs were most likely to use registries and grey literature resources. We found substantial co-searching of resources with only published materials, and not complemented by searches of registries and the grey literature. The 2013 guideline did not substantially increase searching of registries and grey literature resources to retrieve primary studies for the SRMAs. When used to augment Medline, Scopus (in all SRMAs) and ClinicalTrials.gov (in SRMAs with safety outcomes) were negatively associated with publication bias.
CONCLUSIONS: Even SRMAs that search multiple sources tend to search similar resources. Our study supports searching Scopus and CTG in addition to Medline to reduce the chance of publication bias.},
	language = {eng},
	journal = {Journal of Clinical Epidemiology},
	author = {Pradhan, Richeek and Garnick, Kyle and Barkondaj, Bikramjit and Jordan, Harmon S. and Ash, Arlene and Yu, Hong},
	month = oct,
	year = {2018},
	pmid = {29879464},
	pmcid = {PMC6250602},
	keywords = {Evidence synthesis, Grey literature, Literature databases, Meta-analysis, Publication bias, Systematic review, Trial registries},
	pages = {50--62}
}

OBJECTIVE: Systematic reviews and meta-analyses (SRMAs) rely upon comprehensive searches into diverse resources that catalog primary studies. However, since what constitutes a comprehensive search is unclear, we examined trends in databases searched from 2005-2016, surrounding the publication of search guidelines in 2013, and associations between resources searched and evidence of publication bias in SRMAs involving human subjects. STUDY DESIGN: To ensure comparability of included SRMAs over the 12 years in the face of a near 100-fold increase of international SRMAs (mainly genetic studies from China) during this period, we focused on USA-affiliated SRMAs, manually reviewing 100 randomly selected SRMAs from those published in each year. After excluding articles (mainly for inadequate detail or out-of-scope methods), we identified factors associated with the databases searched, used network analysis to see which resources were simultaneously searched, and used logistic regression to link information sources searched with a lower chance of finding publication bias. RESULTS: Among 817 SRMA articles studied, the common resources used were Medline (95%), EMBASE (44%), and Cochrane (41%). Methods journal SRMAs were most likely to use registries and grey literature resources. We found substantial co-searching of resources with only published materials, and not complemented by searches of registries and the grey literature. The 2013 guideline did not substantially increase searching of registries and grey literature resources to retrieve primary studies for the SRMAs. When used to augment Medline, Scopus (in all SRMAs) and ClinicalTrials.gov (in SRMAs with safety outcomes) were negatively associated with publication bias. CONCLUSIONS: Even SRMAs that search multiple sources tend to search similar resources. Our study supports searching Scopus and CTG in addition to Medline to reduce the chance of publication bias.

@inproceedings{vu_sentence_2018,
	title = {Sentence {Simplification} with {Memory}-{Augmented} {Neural} {Networks}},
	booktitle = {North {American} {Chapter} of the {Association} for {Computational} {Linguistics}: {Human} {Language} {Technologies}},
	author = {Vu, Tu and Hu, Baotian and Munkhdalai, Tsendsuren and Yu, Hong},
	year = {2018}
}

@article{zheng_assessing_2018,
	title = {Assessing the {Readability} of {Medical} {Documents}: {A} {Ranking} {Approach}},
	volume = {6},
	issn = {2291-9694},
	shorttitle = {Assessing the {Readability} of {Medical} {Documents}},
	doi = {10.2196/medinform.8611},
	abstract = {BACKGROUND: The use of electronic health record (EHR) systems with patient engagement capabilities, including viewing, downloading, and transmitting health information, has recently grown tremendously. However, using these resources to engage patients in managing their own health remains challenging due to the complex and technical nature of the EHR narratives.
OBJECTIVE: Our objective was to develop a machine learning-based system to assess readability levels of complex documents such as EHR notes.
METHODS: We collected difficulty ratings of EHR notes and Wikipedia articles using crowdsourcing from 90 readers. We built a supervised model to assess readability based on relative orders of text difficulty using both surface text features and word embeddings. We evaluated system performance using the Kendall coefficient of concordance against human ratings.
RESULTS: Our system achieved significantly higher concordance (.734) with human annotators than did a baseline using the Flesch-Kincaid Grade Level, a widely adopted readability formula (.531). The improvement was also consistent across different disease topics. This method's concordance with an individual human user's ratings was also higher than the concordance between different human annotators (.658).
CONCLUSIONS: We explored methods to automatically assess the readability levels of clinical narratives. Our ranking-based system using simple textual features and easy-to-learn word embeddings outperformed a widely used readability formula. Our ranking-based method can predict relative difficulties of medical documents. It is not constrained to a predefined set of readability levels, a common design in many machine learning-based systems. Furthermore, the feature set does not rely on complex processing of the documents. One potential application of our readability ranking is personalization, allowing patients to better accommodate their own background knowledge.},
	language = {eng},
	number = {1},
	journal = {JMIR medical informatics},
	author = {Zheng, Jiaping and Yu, Hong},
	month = mar,
	year = {2018},
	pmid = {29572199},
	pmcid = {PMC5889493},
	keywords = {comprehension, electronic health records, machine learning, readability},
	pages = {e17}
}

BACKGROUND: The use of electronic health record (EHR) systems with patient engagement capabilities, including viewing, downloading, and transmitting health information, has recently grown tremendously. However, using these resources to engage patients in managing their own health remains challenging due to the complex and technical nature of the EHR narratives. OBJECTIVE: Our objective was to develop a machine learning-based system to assess readability levels of complex documents such as EHR notes. METHODS: We collected difficulty ratings of EHR notes and Wikipedia articles using crowdsourcing from 90 readers. We built a supervised model to assess readability based on relative orders of text difficulty using both surface text features and word embeddings. We evaluated system performance using the Kendall coefficient of concordance against human ratings. RESULTS: Our system achieved significantly higher concordance (.734) with human annotators than did a baseline using the Flesch-Kincaid Grade Level, a widely adopted readability formula (.531). The improvement was also consistent across different disease topics. This method's concordance with an individual human user's ratings was also higher than the concordance between different human annotators (.658). CONCLUSIONS: We explored methods to automatically assess the readability levels of clinical narratives. Our ranking-based system using simple textual features and easy-to-learn word embeddings outperformed a widely used readability formula. Our ranking-based method can predict relative difficulties of medical documents. It is not constrained to a predefined set of readability levels, a common design in many machine learning-based systems. Furthermore, the feature set does not rely on complex processing of the documents. One potential application of our readability ranking is personalization, allowing patients to better accommodate their own background knowledge.

@article{lalor_comprehenotes_2018,
	title = {{ComprehENotes}, an {Instrument} to {Assess} {Patient} {Reading} {Comprehension} of {Electronic} {Health} {Record} {Notes}: {Development} and {Validation}},
	volume = {20},
	issn = {1438-8871},
	shorttitle = {{ComprehENotes}, an {Instrument} to {Assess} {Patient} {Reading} {Comprehension} of {Electronic} {Health} {Record} {Notes}},
	doi = {10.2196/jmir.9380},
	abstract = {BACKGROUND: Patient portals are widely adopted in the United States and allow millions of patients access to their electronic health records (EHRs), including their EHR clinical notes. A patient's ability to understand the information in the EHR is dependent on their overall health literacy. Although many tests of health literacy exist, none specifically focuses on EHR note comprehension.
OBJECTIVE: The aim of this paper was to develop an instrument to assess patients' EHR note comprehension.
METHODS: We identified 6 common diseases or conditions (heart failure, diabetes, cancer, hypertension, chronic obstructive pulmonary disease, and liver failure) and selected 5 representative EHR notes for each disease or condition. One note that did not contain natural language text was removed. Questions were generated from these notes using Sentence Verification Technique and were analyzed using item response theory (IRT) to identify a set of questions that represent a good test of ability for EHR note comprehension.
RESULTS: Using Sentence Verification Technique, 154 questions were generated from the 29 EHR notes initially obtained. Of these, 83 were manually selected for inclusion in the Amazon Mechanical Turk crowdsourcing tasks and 55 were ultimately retained following IRT analysis. A follow-up validation with a second Amazon Mechanical Turk task and IRT analysis confirmed that the 55 questions test a latent ability dimension for EHR note comprehension. A short test of 14 items was created along with the 55-item test.
CONCLUSIONS: We developed ComprehENotes, an instrument for assessing EHR note comprehension from existing EHR notes, gathered responses using crowdsourcing, and used IRT to analyze those responses, thus resulting in a set of questions to measure EHR note comprehension. Crowdsourced responses from Amazon Mechanical Turk can be used to estimate item parameters and select a subset of items for inclusion in the test set using IRT. The final set of questions is the first test of EHR note comprehension.},
	language = {eng},
	number = {4},
	journal = {Journal of Medical Internet Research},
	author = {Lalor, John P. and Wu, Hao and Chen, Li and Mazor, Kathleen M. and Yu, Hong},
	month = apr,
	year = {2018},
	pmid = {29695372},
	pmcid = {PMC5943623},
	keywords = {crowdsourcing, electronic health records, health literacy, psychometrics},
	pages = {e139}
}

BACKGROUND: Patient portals are widely adopted in the United States and allow millions of patients access to their electronic health records (EHRs), including their EHR clinical notes. A patient's ability to understand the information in the EHR is dependent on their overall health literacy. Although many tests of health literacy exist, none specifically focuses on EHR note comprehension. OBJECTIVE: The aim of this paper was to develop an instrument to assess patients' EHR note comprehension. METHODS: We identified 6 common diseases or conditions (heart failure, diabetes, cancer, hypertension, chronic obstructive pulmonary disease, and liver failure) and selected 5 representative EHR notes for each disease or condition. One note that did not contain natural language text was removed. Questions were generated from these notes using Sentence Verification Technique and were analyzed using item response theory (IRT) to identify a set of questions that represent a good test of ability for EHR note comprehension. RESULTS: Using Sentence Verification Technique, 154 questions were generated from the 29 EHR notes initially obtained. Of these, 83 were manually selected for inclusion in the Amazon Mechanical Turk crowdsourcing tasks and 55 were ultimately retained following IRT analysis. A follow-up validation with a second Amazon Mechanical Turk task and IRT analysis confirmed that the 55 questions test a latent ability dimension for EHR note comprehension. A short test of 14 items was created along with the 55-item test. CONCLUSIONS: We developed ComprehENotes, an instrument for assessing EHR note comprehension from existing EHR notes, gathered responses using crowdsourcing, and used IRT to analyze those responses, thus resulting in a set of questions to measure EHR note comprehension. Crowdsourced responses from Amazon Mechanical Turk can be used to estimate item parameters and select a subset of items for inclusion in the test set using IRT. The final set of questions is the first test of EHR note comprehension.

@article{kundu_recent_2018,
	title = {Recent {Trends} in {Oral} {Anticoagulant} {Use} and {Post}-{Discharge} {Complications} {Among} {Atrial} {Fibrillation} {Patients} with {Acute} {Myocardial} {Infarction}},
	volume = {10},
	issn = {1941-6911},
	doi = {10.4022/jafib.1749},
	abstract = {Background: Atrial fibrillation (AF) is a common complication of acute myocardial infarction (AMI).The CHA2DS2VAScand CHADS2risk scoresare used to identifypatients with AF at risk for strokeand to guide oral anticoagulants (OAC) use, including patients with AMI. However, the epidemiology of AF, further stratifiedaccording to patients' risk of stroke, has not been wellcharacterized among those hospitalized for AMI.
Methods: We examined trends in the frequency of AF, rates of discharge OAC use, and post-discharge outcomes among 6,627 residents of the Worcester, Massachusetts area who survived hospitalization for AMI at 11 medical centers between 1997 and 2011.
Results: A total of 1,050AMI patients had AF (16\%) andthe majority (91\%)had a CHA2DS2VAScscore {\textgreater}2.AF rates were highest among patients in the highest stroke risk group.In comparison to patients without AF, patients with AMI and AF in the highest stroke risk category had higher rates of post-discharge complications, including higher 30-day re-hospitalization [27 \% vs. 17 \%], 30-day post-discharge death [10 \% vs. 5\%], and 1-year post-discharge death [46 \% vs. 18 \%] (p {\textless} 0.001 for all). Notably, fewerthan half of guideline-eligible AF patientsreceived an OACprescription at discharge. Usage rates for other evidence-based therapiessuch as statins and beta-blockers,lagged in comparison to AMI patients free from AF.
Conclusions: Our findings highlight the need to enhance efforts towards stroke prevention among AMI survivors with AF.},
	language = {eng},
	number = {5},
	journal = {Journal of Atrial Fibrillation},
	author = {Kundu, Amartya and Day, Kevin O. and Lessard, Darleen M. and Gore, Joel M. and Lubitz, Steven A. and Yu, Hong and Akhter, Mohammed W. and Fisher, Daniel Z. and Hayward, Robert M. and Henninger, Nils and Saczynski, Jane S. and Walkey, Allan J. and Kapoor, Alok and Yarzebski, Jorge and Goldberg, Robert J. and McManus, David D.},
	month = feb,
	year = {2018},
	pmid = {29988239},
	pmcid = {PMC6006973},
	keywords = {Anticoagulation, Atrial Fibrillation, Epidemiology, Outcomes},
	pages = {1749}
}

Background: Atrial fibrillation (AF) is a common complication of acute myocardial infarction (AMI).The CHA2DS2VAScand CHADS2risk scoresare used to identifypatients with AF at risk for strokeand to guide oral anticoagulants (OAC) use, including patients with AMI. However, the epidemiology of AF, further stratifiedaccording to patients' risk of stroke, has not been wellcharacterized among those hospitalized for AMI. Methods: We examined trends in the frequency of AF, rates of discharge OAC use, and post-discharge outcomes among 6,627 residents of the Worcester, Massachusetts area who survived hospitalization for AMI at 11 medical centers between 1997 and 2011. Results: A total of 1,050AMI patients had AF (16%) andthe majority (91%)had a CHA2DS2VAScscore \textgreater2.AF rates were highest among patients in the highest stroke risk group.In comparison to patients without AF, patients with AMI and AF in the highest stroke risk category had higher rates of post-discharge complications, including higher 30-day re-hospitalization [27 % vs. 17 %], 30-day post-discharge death [10 % vs. 5%], and 1-year post-discharge death [46 % vs. 18 %] (p \textless 0.001 for all). Notably, fewerthan half of guideline-eligible AF patientsreceived an OACprescription at discharge. Usage rates for other evidence-based therapiessuch as statins and beta-blockers,lagged in comparison to AMI patients free from AF. Conclusions: Our findings highlight the need to enhance efforts towards stroke prevention among AMI survivors with AF.

@inproceedings{wang_accuracy_2018,
	title = {Accuracy of {International} {Classification} of {Disease} {Clinical} {Modification} {Codes} for {Detecting} {Bleeding} {Events} in {Electronic} {Health} {Records} and {When} to {Use} {Them}},
	author = {Wang, V and McManus, D and Ash, A and Hoaglin, D and Yu, H},
	year = {2018}
}

@inproceedings{lalor_understanding_2018,
	title = {Understanding {Deep} {Learning} {Performance} through an {Examination} of {Test} {Set} {Difficulty}: {A} {Psychometric} {Case} {Study}},
	booktitle = {{EMNLP}},
	author = {Lalor, John and Wu, Hao and Munkhdalai, Tsendsuren and Yu, Hong},
	year = {2018}
}

@inproceedings{lalor_soft_2018,
	title = {Soft {Label} {Memorization}-{Generalization} for {Natural} {Language} {Inference}.},
	author = {Lalor, John and Wu, Hao and Yu, Hong},
	year = {2018}
}

@inproceedings{amartya_kundu_recent_2018,
	title = {Recent {Trends} {In} {Oral} {Anticoagulant} {Use} and {Post}-{Discharge} {Complications} {Among} {Atrial} {Fibrillation} {Patients} {With} {Acute} {Myocardial} {Infarction}},
	publisher = {Journal of Atrial Fibrillation},
	author = {{Amartya Kundu} and {Kevin O ’Day} and {Darleen M. Lessard} and {Joel M. Gore1} and {Steven A. Lubitz} and {Hong Yu} and {Mohammed W. Akhter} and {Daniel Z. Fisher} and {Robert M. Hayward Jr.} and {Nils Henninger} and {Jane S. Saczynski} and {Allan J. Walkey} and {Alok Kapoor} and {Jorge Yarzebski} and {Robert J. Goldberg} and {David D. McManus}},
	year = {2018}
}

@inproceedings{abhyuday_jagannatha_frontiers_2018,
	title = {Frontiers of {Clinical} {Information} {Extraction}: {Current} {Progress} in {Medication} and {Adverse} {Drug} {Event} {Detection} from {Electronic} {Health} {Records}},
	url = {https://informaticssummit2018.zerista.com/event/member/470512},
	booktitle = {9th {Annual} {Pharmacy} {Informatics} {Conference}},
	author = {{Abhyuday Jagannatha} and {Feifan Liu} and {Weisong Liu} and {Hong Yu}},
	year = {2018}
}

@inproceedings{j._sun_panel_2018,
	title = {Panel – {Deep} {Learning} for {Healthcare} - {Hype} or the {Real} {Thing}?},
	booktitle = {{AMIA} 2018 {Informatics} {Summit}},
	author = {{J. Sun} and {B. Westover} and {H. Yu} and {D. Sontag} and {M. Ghassemi}},
	year = {2018}
}

@inproceedings{feifan_liu_panel_2018,
	title = {Panel - {Frontiers} of {Clinical} {Information} {Extraction}: {Current} {Progress} in {Medication} and {Adverse} {Drug} {Event} {Detection} from {Electronic} {Health} {Records}},
	booktitle = {{AMIA} 2018 {Informatics} {Summit}},
	author = {{Feifan Liu} and {Abhyuday Jagannatha} and {Hong Yu}},
	year = {2018}
}

@inproceedings{celena_b_peters_reference_2018,
	title = {Reference {Standard} {Development} to {Train} {Natural} {Language} {Processing} {Algorithms} to {Detect} {Problematic} {Buprenorphine}-{Naloxone} {Therapy}},
	booktitle = {{VA} {Pharmacy} {Informatics} {Conference} 2018},
	author = {{Celena B Peters} and {Fran Cunningham} and {Adam Gordon} and {Hong Yu} and {Cedric Salone} and {Jessica Zacher} and {Ronald Carico} and {Jianwei Leng} and {Nikolh Durley} and {Weisong Liu} and {Chao-Chin Lu} and {Emily Druhl} and {Feifan Liu} and {Brian C Sauer}},
	year = {2018}
}

@article{noauthor_zheng_2018,
	title = {Zheng {JP} and {Yu} {H}. 2018. {Assessing} {Readability} of {Medical} {Documents}: {A} {Ranking} {Approach}. {The} {Journal} of {Medical} {Internet} {Research} {Medical} {Informatics}.  {Accepted}.},
	year = {2018}
}

@article{noauthor_lalor_2018,
	title = {Lalor {JP}, {Chen} {L}, {Mazor} {K}, and {Yu} {H}. 2018. {ComprehENotes}: {An} {Instrument} to {Assess} {Patient} {EHR} {Note} {Reading} {Comprehension}. {The} {Journal} of {Medical} {Internet} {Research}. {Accepted}.},
	year = {2018}
}

@inproceedings{munkhdalai_neural_2017,
	title = {Neural {Semantic} {Encoders}},
	url = {https://arxiv.org/pdf/1607.04315v2.pdf},
	booktitle = {European {Chapter} of the {Association} for {Computational} {Linguistics} 2017 ({EACL})},
	author = {Munkhdalai, T and Yu, Hong},
	year = {2017}
}

@article{zheng_readability_2017,
	title = {Readability {Formulas} and {User} {Perceptions} of {Electronic} {Health} {Records} {Difficulty}: {A} {Corpus} {Study}},
	volume = {19},
	copyright = {Unless stated otherwise, all articles are open-access distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work (},
	shorttitle = {Readability {Formulas} and {User} {Perceptions} of {Electronic} {Health} {Records} {Difficulty}},
	url = {https://www.jmir.org/2017/3/e59/},
	doi = {10.2196/jmir.6962},
	abstract = {Background: Electronic health records (EHRs) are a rich resource for developing applications to engage patients and foster patient activation, thus holding a strong potential to enhance patient-centered care. Studies have shown that providing patients with access to their own EHR notes may improve the understanding of their own clinical conditions and treatments, leading to improved health care outcomes. However, the highly technical language in EHR notes impedes patients’ comprehension. Numerous studies have evaluated the difficulty of health-related text using readability formulas such as Flesch-Kincaid Grade Level (FKGL), Simple Measure of Gobbledygook (SMOG), and Gunning-Fog Index (GFI). They conclude that the materials are often written at a grade level higher than common recommendations. Objective: The objective of our study was to explore the relationship between the aforementioned readability formulas and the laypeople’s perceived difficulty on 2 genres of text: general health information and EHR notes. We also validated the formulas’ appropriateness and generalizability on predicting difficulty levels of highly complex technical documents. Methods: We collected 140 Wikipedia articles on diabetes and 242 EHR notes with diabetes International Classification of Diseases, Ninth Revision code. We recruited 15 Amazon Mechanical Turk (AMT) users to rate difficulty levels of the documents. Correlations between laypeople’s perceived difficulty levels and readability formula scores were measured, and their difference was tested. We also compared word usage and the impact of medical concepts of the 2 genres of text. Results: The distributions of both readability formulas’ scores (P{\textless}.001) and laypeople’s perceptions (P=.002) on the 2 genres were different. Correlations of readability predictions and laypeople’s perceptions were weak. Furthermore, despite being graded at similar levels, documents of different genres were still perceived with different difficulty (P{\textless}.001). Word usage in the 2 related genres still differed significantly (P{\textless}.001). Conclusions: Our findings suggested that the readability formulas’ predictions did not align with perceived difficulty in either text genre. The widely used readability formulas were highly correlated with each other but did not show adequate correlation with readers’ perceived difficulty. Therefore, they were not appropriate to assess the readability of EHR notes.  [J Med Internet Res 2017;19(3):e59]},
	language = {en},
	number = {3},
	urldate = {2017-03-06},
	journal = {Journal of Medical Internet Research},
	author = {Zheng, Jiaping and Yu, Hong},
	year = {2017},
	pages = {e59}
}

Background: Electronic health records (EHRs) are a rich resource for developing applications to engage patients and foster patient activation, thus holding a strong potential to enhance patient-centered care. Studies have shown that providing patients with access to their own EHR notes may improve the understanding of their own clinical conditions and treatments, leading to improved health care outcomes. However, the highly technical language in EHR notes impedes patients’ comprehension. Numerous studies have evaluated the difficulty of health-related text using readability formulas such as Flesch-Kincaid Grade Level (FKGL), Simple Measure of Gobbledygook (SMOG), and Gunning-Fog Index (GFI). They conclude that the materials are often written at a grade level higher than common recommendations. Objective: The objective of our study was to explore the relationship between the aforementioned readability formulas and the laypeople’s perceived difficulty on 2 genres of text: general health information and EHR notes. We also validated the formulas’ appropriateness and generalizability on predicting difficulty levels of highly complex technical documents. Methods: We collected 140 Wikipedia articles on diabetes and 242 EHR notes with diabetes International Classification of Diseases, Ninth Revision code. We recruited 15 Amazon Mechanical Turk (AMT) users to rate difficulty levels of the documents. Correlations between laypeople’s perceived difficulty levels and readability formula scores were measured, and their difference was tested. We also compared word usage and the impact of medical concepts of the 2 genres of text. Results: The distributions of both readability formulas’ scores (P\textless.001) and laypeople’s perceptions (P=.002) on the 2 genres were different. Correlations of readability predictions and laypeople’s perceptions were weak. Furthermore, despite being graded at similar levels, documents of different genres were still perceived with different difficulty (P\textless.001). Word usage in the 2 related genres still differed significantly (P\textless.001). Conclusions: Our findings suggested that the readability formulas’ predictions did not align with perceived difficulty in either text genre. The widely used readability formulas were highly correlated with each other but did not show adequate correlation with readers’ perceived difficulty. Therefore, they were not appropriate to assess the readability of EHR notes. [J Med Internet Res 2017;19(3):e59]

@article{munkhdalai_reasoning_2017,
	title = {Reasoning with memory augmented neural networks for language  comprehension.},
	url = {https://arxiv.org/abs/1610.06454},
	urldate = {2017-06-02},
	journal = {5th International Conference on Learning Representations (ICLR)},
	author = {Munkhdalai, Tsendsuren and Yu, Hong},
	year = {2017}
}

@article{lingeman_detecting_2017,
	title = {Detecting {Opioid}-{Related} {Aberrant} {Behavior} using {Natural} {Language} {Processing}},
	volume = {2017},
	issn = {1942-597X},
	abstract = {The United States is in the midst of a prescription opioid epidemic, with the number of yearly opioid-related overdose deaths increasing almost fourfold since 20001. To more effectively prevent unintentional opioid overdoses, the medical profession requires robust surveillance tools that can effectively identify at-risk patients. Drug-related aberrant behaviors observed in the clinical context may be important indicators of patients at risk for or actively abusing opioids. In this paper, we describe a natural language processing (NLP) method for automatic surveillance of aberrant behavior in medical notes relying only on the text of the notes. This allows for a robust and generalizable system that can be used for high volume analysis of electronic medical records for potential predictors of opioid abuse.},
	language = {eng},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Lingeman, Jesse M. and Wang, Priscilla and Becker, William and Yu, Hong},
	year = {2017},
	pmid = {29854186},
	pmcid = {PMC5977697},
	pages = {1179--1185}
}

The United States is in the midst of a prescription opioid epidemic, with the number of yearly opioid-related overdose deaths increasing almost fourfold since 20001. To more effectively prevent unintentional opioid overdoses, the medical profession requires robust surveillance tools that can effectively identify at-risk patients. Drug-related aberrant behaviors observed in the clinical context may be important indicators of patients at risk for or actively abusing opioids. In this paper, we describe a natural language processing (NLP) method for automatic surveillance of aberrant behavior in medical notes relying only on the text of the notes. This allows for a robust and generalizable system that can be used for high volume analysis of electronic medical records for potential predictors of opioid abuse.

@article{lalor_analysis_2017,
	title = {An {Analysis} of {Machine} {Learning} {Intelligence}},
	url = {http://arxiv.org/abs/1702.04811},
	abstract = {Deep neural networks (DNNs) have set state of the art results in many machine learning and NLP tasks. However, we do not have a strong understanding of what DNN models learn. In this paper, we examine learning in DNNs through analysis of their outputs. We compare DNN performance directly to a human population, and use characteristics of individual data points such as difficulty to see how well models perform on easy and hard examples. We investigate how training size and the incorporation of noise affect a DNN's ability to generalize and learn. Our experiments show that unlike traditional machine learning models (e.g., Naive Bayes, Decision Trees), DNNs exhibit human-like learning properties. As they are trained with more data, they are more able to distinguish between easy and difficult items, and performance on easy items improves at a higher rate than difficult items. We find that different DNN models exhibit different strengths in learning and are robust to noise in training data.},
	urldate = {2017-02-26},
	journal = {arXiv:1702.04811 [cs]},
	author = {Lalor, John P. and Wu, Hao and Munkhdalai, Tsendsuren and Yu, Hong},
	month = feb,
	year = {2017},
	note = {arXiv: 1702.04811},
	keywords = {Computer Science - Computation and Language}
}

Deep neural networks (DNNs) have set state of the art results in many machine learning and NLP tasks. However, we do not have a strong understanding of what DNN models learn. In this paper, we examine learning in DNNs through analysis of their outputs. We compare DNN performance directly to a human population, and use characteristics of individual data points such as difficulty to see how well models perform on easy and hard examples. We investigate how training size and the incorporation of noise affect a DNN's ability to generalize and learn. Our experiments show that unlike traditional machine learning models (e.g., Naive Bayes, Decision Trees), DNNs exhibit human-like learning properties. As they are trained with more data, they are more able to distinguish between easy and difficult items, and performance on easy items improves at a higher rate than difficult items. We find that different DNN models exhibit different strengths in learning and are robust to noise in training data.

@article{lalor_cift:_2017,
	title = {{CIFT}: {Crowd}-{Informed} {Fine}-{Tuning} to {Improve} {Machine} {Learning} {Ability}},
	shorttitle = {{CIFT}},
	journal = {arXiv preprint arXiv:1702.08563},
	author = {Lalor, John P. and Wu, Hao and Yu, Hong},
	year = {2017}
}

@article{chen_ranking_2017,
	title = {Ranking {Medical} {Terms} to {Support} {Expansion} of {Lay} {Language} {Resources} for {Patient} {Comprehension} of {Electronic} {Health} {Record} {Notes}: {Adapted} {Distant} {Supervision} {Approach}},
	volume = {5},
	issn = {2291-9694},
	shorttitle = {Ranking {Medical} {Terms} to {Support} {Expansion} of {Lay} {Language} {Resources} for {Patient} {Comprehension} of {Electronic} {Health} {Record} {Notes}},
	doi = {10.2196/medinform.8531},
	abstract = {BACKGROUND: Medical terms are a major obstacle for patients to comprehend their electronic health record (EHR) notes. Clinical natural language processing (NLP) systems that link EHR terms to lay terms or definitions allow patients to easily access helpful information when reading through their EHR notes, and have shown to improve patient EHR comprehension. However, high-quality lay language resources for EHR terms are very limited in the public domain. Because expanding and curating such a resource is a costly process, it is beneficial and even necessary to identify terms important for patient EHR comprehension first.
OBJECTIVE: We aimed to develop an NLP system, called adapted distant supervision (ADS), to rank candidate terms mined from EHR corpora. We will give EHR terms ranked as high by ADS a higher priority for lay language annotation-that is, creating lay definitions for these terms.
METHODS: Adapted distant supervision uses distant supervision from consumer health vocabulary and transfer learning to adapt itself to solve the problem of ranking EHR terms in the target domain. We investigated 2 state-of-the-art transfer learning algorithms (ie, feature space augmentation and supervised distant supervision) and designed 5 types of learning features, including distributed word representations learned from large EHR data for ADS. For evaluating ADS, we asked domain experts to annotate 6038 candidate terms as important or nonimportant for EHR comprehension. We then randomly divided these data into the target-domain training data (1000 examples) and the evaluation data (5038 examples). We compared ADS with 2 strong baselines, including standard supervised learning, on the evaluation data.
RESULTS: The ADS system using feature space augmentation achieved the best average precision, 0.850, on the evaluation set when using 1000 target-domain training examples. The ADS system using supervised distant supervision achieved the best average precision, 0.819, on the evaluation set when using only 100 target-domain training examples. The 2 ADS systems both performed significantly better than the baseline systems (P{\textless}.001 for all measures and all conditions). Using a rich set of learning features contributed to ADS's performance substantially.
CONCLUSIONS: ADS can effectively rank terms mined from EHRs. Transfer learning improved ADS's performance even with a small number of target-domain training examples. EHR terms prioritized by ADS were used to expand a lay language resource that supports patient EHR comprehension. The top 10,000 EHR terms ranked by ADS are available upon request.},
	language = {eng},
	number = {4},
	journal = {JMIR medical informatics},
	author = {Chen, Jinying and Jagannatha, Abhyuday N. and Fodeh, Samah J. and Yu, Hong},
	month = oct,
	year = {2017},
	pmid = {29089288},
	pmcid = {PMC5686421},
	keywords = {Information extraction, electronic health records, lexical entry selection, natural language processing, transfer learning},
	pages = {e42}
}

BACKGROUND: Medical terms are a major obstacle for patients to comprehend their electronic health record (EHR) notes. Clinical natural language processing (NLP) systems that link EHR terms to lay terms or definitions allow patients to easily access helpful information when reading through their EHR notes, and have shown to improve patient EHR comprehension. However, high-quality lay language resources for EHR terms are very limited in the public domain. Because expanding and curating such a resource is a costly process, it is beneficial and even necessary to identify terms important for patient EHR comprehension first. OBJECTIVE: We aimed to develop an NLP system, called adapted distant supervision (ADS), to rank candidate terms mined from EHR corpora. We will give EHR terms ranked as high by ADS a higher priority for lay language annotation-that is, creating lay definitions for these terms. METHODS: Adapted distant supervision uses distant supervision from consumer health vocabulary and transfer learning to adapt itself to solve the problem of ranking EHR terms in the target domain. We investigated 2 state-of-the-art transfer learning algorithms (ie, feature space augmentation and supervised distant supervision) and designed 5 types of learning features, including distributed word representations learned from large EHR data for ADS. For evaluating ADS, we asked domain experts to annotate 6038 candidate terms as important or nonimportant for EHR comprehension. We then randomly divided these data into the target-domain training data (1000 examples) and the evaluation data (5038 examples). We compared ADS with 2 strong baselines, including standard supervised learning, on the evaluation data. RESULTS: The ADS system using feature space augmentation achieved the best average precision, 0.850, on the evaluation set when using 1000 target-domain training examples. The ADS system using supervised distant supervision achieved the best average precision, 0.819, on the evaluation set when using only 100 target-domain training examples. The 2 ADS systems both performed significantly better than the baseline systems (P\textless.001 for all measures and all conditions). Using a rich set of learning features contributed to ADS's performance substantially. CONCLUSIONS: ADS can effectively rank terms mined from EHRs. Transfer learning improved ADS's performance even with a small number of target-domain training examples. EHR terms prioritized by ADS were used to expand a lay language resource that supports patient EHR comprehension. The top 10,000 EHR terms ranked by ADS are available upon request.

@inproceedings{noauthor_j._2017,
	title = {J. {Lalor}, {H}. {Wu}, {L}. {Chen}, {K}. {Mazor}, {H}. {Yu}. {Generating} a {Test} of {Electronic} {Health} {Record} {Narrative} {Comprehension} with {Item} {Response} {Theory}. {American} {Medical} {Informatics} {Association} ({AMIA}) {Annual} {Symposium} {Podium} {Abstract}.},
	year = {2017}
}

@inproceedings{noauthor_j._2017-1,
	title = {J. {Lalor}, {H}. {Wu}, {H}. {Yu}. {CIFT}: {Crowd}-{Informed} {Fine}-{Tuning} to {Improve} {Machine} {Learning} {Ability}. {Human} {Computation} and {Crowdsourcing} ({HCOMP}) {Works}-in-{Progress}.},
	year = {2017}
}

@inproceedings{noauthor_lingeman_2017,
	title = {Lingeman {J} and {Yu} {H}. 2017. {Detecting} {Opioid}-{Related} {Aberrant} {Behavior} using {Natural} {Language} {Processing}. {AMIA} {Fall} {Symposium}.},
	year = {2017}
}

@inproceedings{noauthor_li_2017,
	title = {Li {R} and {Yu} {H}. 2017. {A} {Hybrid} {Neural} {Network} {Model} for {Joint} {Prediction} of {Medical} {Presence} and {Period} {Assertions} in {Clinical} {Notes}. {AMIA} {Fall} {Symposium}.},
	year = {2017}
}

@inproceedings{noauthor_zheng_2017,
	title = {Zheng {J} and {Yu} {H}. 2017. {Assessing} {Electronic} {Health} {Record} {Readability}. {AMIA} {Fall} {Symposium}.},
	year = {2017}
}

@inproceedings{munkhdalai_meta_2017,
	address = {Sydney, Australia},
	title = {Meta {Networks}},
	booktitle = {{ICML}},
	author = {Munkhdalai, Tsendsuren and Yu, Hong},
	year = {2017}
}

@inproceedings{lalor_improving_2017,
	title = {Improving {Machine} {Learning} {Ability} with {Fine}-{Tuning}},
	booktitle = {{ICML}},
	author = {Lalor, John and Wu, Hao and Yu, Hong},
	year = {2017}
}

@article{prakash_condensed_2016,
	title = {Condensed {Memory} {Networks} for {Clinical} {Diagnostic} {Inferencing}},
	url = {http://arxiv.org/abs/1612.01848},
	abstract = {Diagnosis of a clinical condition is a challenging task, which often requires significant medical investigation. Previous work related to diagnostic inferencing problems mostly consider multivariate observational data (e.g. physiological signals, lab tests etc.). In contrast, we explore the problem using free-text medical notes recorded in an electronic health record (EHR). Complex tasks like these can benefit from structured knowledge bases, but those are not scalable. We instead exploit raw text from Wikipedia as a knowledge source. Memory networks have been demonstrated to be effective in tasks which require comprehension of free-form text. They use the final iteration of the learned representation to predict probable classes. We introduce condensed memory neural networks (C-MemNNs), a novel model with iterative condensation of memory representations that preserves the hierarchy of features in the memory. Experiments on the MIMIC-III dataset show that the proposed model outperforms other variants of memory networks to predict the most probable diagnoses given a complex clinical scenario.},
	urldate = {2017-01-12},
	journal = {arXiv:1612.01848 [cs]},
	author = {Prakash, Aaditya and Zhao, Siyuan and Hasan, Sadid A. and Datla, Vivek and Lee, Kathy and Qadir, Ashequl and Liu, Joey and Farri, Oladimeji},
	month = dec,
	year = {2016},
	note = {arXiv: 1612.01848},
	keywords = {Computer Science - Computation and Language}
}

Diagnosis of a clinical condition is a challenging task, which often requires significant medical investigation. Previous work related to diagnostic inferencing problems mostly consider multivariate observational data (e.g. physiological signals, lab tests etc.). In contrast, we explore the problem using free-text medical notes recorded in an electronic health record (EHR). Complex tasks like these can benefit from structured knowledge bases, but those are not scalable. We instead exploit raw text from Wikipedia as a knowledge source. Memory networks have been demonstrated to be effective in tasks which require comprehension of free-form text. They use the final iteration of the learned representation to predict probable classes. We introduce condensed memory neural networks (C-MemNNs), a novel model with iterative condensation of memory representations that preserves the hierarchy of features in the memory. Experiments on the MIMIC-III dataset show that the proposed model outperforms other variants of memory networks to predict the most probable diagnoses given a complex clinical scenario.

@article{chen_finding_2016,
	title = {Finding {Important} {Terms} for {Patients} in {Their} {Electronic} {Health} {Records}: {A} {Learning}-to-{Rank} {Approach} {Using} {Expert} {Annotations}},
	volume = {4},
	shorttitle = {Finding {Important} {Terms} for {Patients} in {Their} {Electronic} {Health} {Records}},
	doi = {10.2196/medinform.6373},
	abstract = {BACKGROUND: Many health organizations allow patients to access their own electronic health record (EHR) notes through online patient portals as a way to enhance patient-centered care. However, EHR notes are typically long and contain abundant medical jargon that can be difficult for patients to understand. In addition, many medical terms in patients' notes are not directly related to their health care needs. One way to help patients better comprehend their own notes is to reduce information overload and help them focus on medical terms that matter most to them. Interventions can then be developed by giving them targeted education to improve their EHR comprehension and the quality of care.
OBJECTIVE: We aimed to develop a supervised natural language processing (NLP) system called Finding impOrtant medical Concepts most Useful to patientS (FOCUS) that automatically identifies and ranks medical terms in EHR notes based on their importance to the patients.
METHODS: First, we built an expert-annotated corpus. For each EHR note, 2 physicians independently identified medical terms important to the patient. Using the physicians' agreement as the gold standard, we developed and evaluated FOCUS. FOCUS first identifies candidate terms from each EHR note using MetaMap and then ranks the terms using a support vector machine-based learn-to-rank algorithm. We explored rich learning features, including distributed word representation, Unified Medical Language System semantic type, topic features, and features derived from consumer health vocabulary. We compared FOCUS with 2 strong baseline NLP systems.
RESULTS: Physicians annotated 90 EHR notes and identified a mean of 9 (SD 5) important terms per note. The Cohen's kappa annotation agreement was .51. The 10-fold cross-validation results show that FOCUS achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.940 for ranking candidate terms from EHR notes to identify important terms. When including term identification, the performance of FOCUS for identifying important terms from EHR notes was 0.866 AUC-ROC. Both performance scores significantly exceeded the corresponding baseline system scores (P{\textless}.001). Rich learning features contributed to FOCUS's performance substantially.
CONCLUSIONS: FOCUS can automatically rank terms from EHR notes based on their importance to patients. It may help develop future interventions that improve quality of care.},
	language = {eng},
	number = {4},
	journal = {JMIR medical informatics},
	author = {Chen, Jinying and Zheng, Jiaping and Yu, Hong},
	month = nov,
	year = {2016},
	pmid = {27903489},
	pmcid = {PMC5156821},
	keywords = {Information extraction, Learning to rank, Supervised learning, electronic health records, natural language processing},
	pages = {e40}
}

BACKGROUND: Many health organizations allow patients to access their own electronic health record (EHR) notes through online patient portals as a way to enhance patient-centered care. However, EHR notes are typically long and contain abundant medical jargon that can be difficult for patients to understand. In addition, many medical terms in patients' notes are not directly related to their health care needs. One way to help patients better comprehend their own notes is to reduce information overload and help them focus on medical terms that matter most to them. Interventions can then be developed by giving them targeted education to improve their EHR comprehension and the quality of care. OBJECTIVE: We aimed to develop a supervised natural language processing (NLP) system called Finding impOrtant medical Concepts most Useful to patientS (FOCUS) that automatically identifies and ranks medical terms in EHR notes based on their importance to the patients. METHODS: First, we built an expert-annotated corpus. For each EHR note, 2 physicians independently identified medical terms important to the patient. Using the physicians' agreement as the gold standard, we developed and evaluated FOCUS. FOCUS first identifies candidate terms from each EHR note using MetaMap and then ranks the terms using a support vector machine-based learn-to-rank algorithm. We explored rich learning features, including distributed word representation, Unified Medical Language System semantic type, topic features, and features derived from consumer health vocabulary. We compared FOCUS with 2 strong baseline NLP systems. RESULTS: Physicians annotated 90 EHR notes and identified a mean of 9 (SD 5) important terms per note. The Cohen's kappa annotation agreement was .51. The 10-fold cross-validation results show that FOCUS achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.940 for ranking candidate terms from EHR notes to identify important terms. When including term identification, the performance of FOCUS for identifying important terms from EHR notes was 0.866 AUC-ROC. Both performance scores significantly exceeded the corresponding baseline system scores (P\textless.001). Rich learning features contributed to FOCUS's performance substantially. CONCLUSIONS: FOCUS can automatically rank terms from EHR notes based on their importance to patients. It may help develop future interventions that improve quality of care.

@inproceedings{choi_retain:_2016,
	title = {{RETAIN}: {An} {Interpretable} {Predictive} {Model} for {Healthcare} using {Reverse} {Time} {Attention} {Mechanism}},
	shorttitle = {{RETAIN}},
	url = {http://papers.nips.cc/paper/6321-retain-an-interpretable-predictive-model-for-healthcare-using-reverse-time-attention-mechanism},
	urldate = {2017-01-12},
	booktitle = {Advances in {Neural} {Information} {Processing} {Systems}},
	author = {Choi, Edward and Bahadori, Mohammad Taha and Sun, Jimeng and Kulas, Joshua and Schuetz, Andy and Stewart, Walter},
	year = {2016},
	pages = {3504--3512}
}

@inproceedings{jagannatha_structured_2016,
	title = {Structured prediction models for {RNN} based sequence labeling in clinical text},
	volume = {2016},
	abstract = {Sequence labeling is a widely used method for named entity recognition and information extraction from unstructured natural language data. In clinical domain one major application of sequence labeling involves extraction of medical entities such as medication, indication, and side-effects from Electronic Health Record narratives. Sequence labeling in this domain, presents its own set of challenges and objectives. In this work we experimented with various CRF based structured learning models with Recurrent Neural Networks. We extend the previously studied LSTM-CRF models with explicit modeling of pairwise potentials. We also propose an approximate version of skip-chain CRF inference with RNN potentials. We use these methodologies for structured prediction in order to improve the exact phrase detection of various medical entities.},
	language = {eng},
	booktitle = {Proceedings of the {Conference} on {Empirical} {Methods} in {Natural} {Language} {Processing}},
	author = {Jagannatha, Abhyuday N. and Yu, Hong},
	month = nov,
	year = {2016},
	pmid = {28004040},
	pmcid = {PMC5167535},
	keywords = {Computer Science - Computation and Language},
	pages = {856--865}
}

Sequence labeling is a widely used method for named entity recognition and information extraction from unstructured natural language data. In clinical domain one major application of sequence labeling involves extraction of medical entities such as medication, indication, and side-effects from Electronic Health Record narratives. Sequence labeling in this domain, presents its own set of challenges and objectives. In this work we experimented with various CRF based structured learning models with Recurrent Neural Networks. We extend the previously studied LSTM-CRF models with explicit modeling of pairwise potentials. We also propose an approximate version of skip-chain CRF inference with RNN potentials. We use these methodologies for structured prediction in order to improve the exact phrase detection of various medical entities.

@inproceedings{yu_ehr_2016,
	title = {{EHR} {Note} {Paraphrasing} for {NoteAid} {Evaluation}.},
	booktitle = {{SBM}},
	author = {Yu, Hong},
	year = {2016}
}

@inproceedings{yu_mismatch_2016,
	title = {Mismatch between {Patient} {Information}-{Seeking} and {Physician} {Expectation} at a {Diabetes} {Outpatient} {Clinic}.},
	booktitle = {{SBM}},
	author = {Yu, Hong and Makkapati, Shreya and Maranda, Louis and Malkani, Samir},
	year = {2016}
}

@article{lingeman_learning_2016,
	title = {Learning to {Rank} {Scientific} {Documents} from the {Crowd}},
	url = {https://arxiv.org/pdf/1611.01400v1.pdf},
	journal = {arXiv:1611.01400},
	author = {Lingeman, Jesse M and Yu, Hong},
	month = nov,
	year = {2016}
}

@article{liu_learning_2016,
	title = {Learning for {Biomedical} {Information} {Extraction}: {Methodological} {Review} of {Recent} {Advances}},
	url = {https://arxiv.org/ftp/arxiv/papers/1606/1606.07993.pdf},
	journal = {arXiv:1606.07993},
	author = {Liu, Feifan and Chen, Jinying and Jagannatha, Abhyuday and Yu, Hong},
	month = jun,
	year = {2016}
}

@inproceedings{munkhdalai_citation_2016,
	address = {Austin, TX},
	title = {Citation {Analysis} with {Neural} {Attention} {Models}},
	url = {http://www.aclweb.org/anthology/W/W16/W16-6109.pdf},
	booktitle = {Proceedings of the {Seventh} {International} {Workshop} on {Health} {Text} {Mining} and {Information} {Analysis} ({LOUHI}) ,},
	publisher = {Association for Computational Linguistics},
	author = {Munkhdalai, M and Lalor, J and Yu, H},
	month = nov,
	year = {2016},
	pages = {69--77}
}

@article{lalor_building_2016,
	title = {Building an {Evaluation} {Scale} using {Item} {Response} {Theory}},
	url = {http://arxiv.org/abs/1605.08889},
	abstract = {Evaluation of NLP methods requires testing against a previously vetted gold-standard test set and reporting standard metrics (accuracy/precision/recall/F1). The current assumption is that all items in a given test set are equal with regards to difficulty and discriminating power. We propose Item Response Theory (IRT) from psychometrics as an alternative means for gold-standard test-set generation and NLP system evaluation. IRT is able to describe characteristics of individual items - their difficulty and discriminating power - and can account for these characteristics in its estimation of human intelligence or ability for an NLP task. In this paper, we demonstrate IRT by generating a gold-standard test set for Recognizing Textual Entailment. By collecting a large number of human responses and fitting our IRT model, we show that our IRT model compares NLP systems with the performance in a human population and is able to provide more insight into system performance than standard evaluation metrics. We show that a high accuracy score does not always imply a high IRT score, which depends on the item characteristics and the response pattern.},
	urldate = {2016-09-26},
	journal = {arXiv:1605.08889 [cs]},
	author = {Lalor, John P. and Wu, Hao and Yu, Hong},
	month = may,
	year = {2016},
	note = {arXiv: 1605.08889},
	keywords = {Computer Science - Computation and Language}
}

Evaluation of NLP methods requires testing against a previously vetted gold-standard test set and reporting standard metrics (accuracy/precision/recall/F1). The current assumption is that all items in a given test set are equal with regards to difficulty and discriminating power. We propose Item Response Theory (IRT) from psychometrics as an alternative means for gold-standard test-set generation and NLP system evaluation. IRT is able to describe characteristics of individual items - their difficulty and discriminating power - and can account for these characteristics in its estimation of human intelligence or ability for an NLP task. In this paper, we demonstrate IRT by generating a gold-standard test set for Recognizing Textual Entailment. By collecting a large number of human responses and fitting our IRT model, we show that our IRT model compares NLP systems with the performance in a human population and is able to provide more insight into system performance than standard evaluation metrics. We show that a high accuracy score does not always imply a high IRT score, which depends on the item characteristics and the response pattern.

@article{jagannatha_bidirectional_2016,
	title = {Bidirectional {Recurrent} {Neural} {Networks} for {Medical} {Event} {Detection} in {Electronic} {Health} {Records}},
	url = {http://arxiv.org/abs/1606.07953},
	abstract = {Sequence labeling for extraction of medical events and their attributes from unstructured text in Electronic Health Record (EHR) notes is a key step towards semantic understanding of EHRs. It has important applications in health informatics including pharmacovigilance and drug surveillance. The state of the art supervised machine learning models in this domain are based on Conditional Random Fields (CRFs) with features calculated from fixed context windows. In this application, we explored various recurrent neural network frameworks and show that they significantly outperformed the CRF models.},
	urldate = {2016-08-19},
	journal = {arXiv:1606.07953 [cs]},
	author = {Jagannatha, Abhyuday and Yu, Hong},
	month = jun,
	year = {2016},
	note = {arXiv: 1606.07953},
	keywords = {Computer Science - Computation and Language, Computer Science - Learning, Computer Science - Neural and Evolutionary Computing}
}

Sequence labeling for extraction of medical events and their attributes from unstructured text in Electronic Health Record (EHR) notes is a key step towards semantic understanding of EHRs. It has important applications in health informatics including pharmacovigilance and drug surveillance. The state of the art supervised machine learning models in this domain are based on Conditional Random Fields (CRFs) with features calculated from fixed context windows. In this application, we explored various recurrent neural network frameworks and show that they significantly outperformed the CRF models.

@article{zheng_methods_2015,
	title = {Methods for {Linking} {EHR} {Notes} to {Education} {Materials}},
	volume = {2015},
	issn = {2153-4063},
	abstract = {It has been shown that providing patients with access to their own electronic health records (EHR) can enhance their medical understanding and provide clinically relevant benefits. However, languages that are difficult for non-medical professionals to comprehend are prevalent in the EHR notes, including medical terms, abbreviations, and domain-specific language patterns. Furthermore, limited average health literacy forms a barrier for patients to understand their health condition, impeding their ability to actively participate in managing their health. Therefore, we are developing a system to retrieve EHR note-tailored online consumer-oriented health education materials to improve patients' health knowledge of their own clinical conditions. Our experiments show that queries combining key concepts and other medical concepts present in the EHR notes significantly outperform (more than doubled) a baseline system of using the phrases from topic models.},
	language = {eng},
	journal = {AMIA Joint Summits on Translational Science proceedings AMIA Summit on Translational Science},
	author = {Zheng, Jiaping and Yu, Hong},
	year = {2015},
	pmid = {26306273},
	pmcid = {PMC4525231},
	pages = {209--215}
}

It has been shown that providing patients with access to their own electronic health records (EHR) can enhance their medical understanding and provide clinically relevant benefits. However, languages that are difficult for non-medical professionals to comprehend are prevalent in the EHR notes, including medical terms, abbreviations, and domain-specific language patterns. Furthermore, limited average health literacy forms a barrier for patients to understand their health condition, impeding their ability to actively participate in managing their health. Therefore, we are developing a system to retrieve EHR note-tailored online consumer-oriented health education materials to improve patients' health knowledge of their own clinical conditions. Our experiments show that queries combining key concepts and other medical concepts present in the EHR notes significantly outperform (more than doubled) a baseline system of using the phrases from topic models.

@article{polepalli_ramesh_figure-associated_2015,
	title = {Figure-{Associated} {Text} {Summarization} and {Evaluation}},
	volume = {10},
	issn = {1932-6203},
	url = {http://dx.plos.org/10.1371/journal.pone.0115671},
	doi = {10.1371/journal.pone.0115671},
	language = {en},
	number = {2},
	urldate = {2015-02-26},
	journal = {PLOS ONE},
	author = {Polepalli Ramesh, Balaji and Sethi, Ricky J. and Yu, Hong},
	editor = {Sarkar, Indra Neil},
	month = feb,
	year = {2015},
	pages = {e0115671}
}

@inproceedings{zheng_identifying_2015,
	title = {Identifying {Key} {Concepts} from {EHR} {Notes} {Using} {Domain} {Adaptation}},
	url = {http://www.anthology.aclweb.org/W/W15/W15-26.pdf#page=127},
	urldate = {2017-02-23},
	booktitle = {{SIXTH} {INTERNATIONAL} {WORKSHOP} {ON} {HEALTH} {TEXT} {MINING} {AND} {INFORMATION} {ANALYSIS} ({LOUHI})},
	author = {Zheng, Jiaping and Yu, Hong and Bedford, M. A.},
	year = {2015},
	pages = {115}
}

@inproceedings{zheng_improving_2015,
	address = {Lisboa, Portugal},
	title = {Improving {Concept} {Identification} for linking {EHR} notes to education materials.},
	booktitle = {Empirical {Methods} in {Natural} {Language} {Processing}},
	author = {Zheng, J and Yu, Hong},
	year = {2015}
}

@inproceedings{yu_towards_2015,
	title = {Towards {Mining} {Electronic} {Health} {Records} for {Opioid} {ADE} {Surveillance}},
	url = {http://www.hsrd.research.va.gov/meetings/2015/abstract-display.cfm?RecordID=200},
	abstract = {Objectives:
Prescription opioids are commonly used to treat acute and cancer-related pain, and, over the last two decades, have increasingly been used in the management of chronic non-cancer pain. Patients taking opioids can experience a wide range of adverse drug events (ADEs), including constipation, nausea/vomiting, pruritus, drowsiness and dizziness, hormonal dysfunction, depression, oversedation, falls, fractures, addiction, overdose, respiratory depression, sleep-disordered breathing, and death. Since such ADEs are frequently described in the unstructured electronic health record (EHR) notes, we are developing natural language processing (NLP) system to automatically extract opioid and ADEs from EHRs. The purpose of this study was to test out the feasibility of mining EHR notes for ADE detection using NLP approaches.
 
Methods:
We developed an annotation guideline using an interactive process during which physicians and linguists worked together to define rules and resolve discrepancy. Following the guideline, two annotators annotated 150 discharge summaries (or 8,672 sentences comprising 102,807 word tokens). The overall pairwise annotation agreement was 88\%. The total number of annotated ADEs and medications were 103 and 3,290. Using this annotated corpus, we developed a NLP system to detect medication and ADE information. Our NLP system is trained on the supervised machine learning model Conditional Random Fields. We compared our NLP system with the state-of-the-art NLP system the MetaMap for ADE detection.
 
Results:
NLP performed well on discharge summaries on certain named entities, including frequency (92\% F1), route (89\% F1), dosage (87\% F1), and medication (84\% F1). Because the number of ADE instances is small, NLP performed poorly on ADE (24\% F1). MetaMap performed on average 62\% F1 for medication and 4\% F1 for ADE.
 
Implications:
Our NLP system outperformed MetaMap for EHR notes ADE detection. NLP generally performs well with a sufficient size of annotated data. While the performance of ADE detection is low, more annotated data yielding a higher prevalence of ADEs would likely improve opioid ADE detection. Use of larger datasets is underway.
 
Impacts:
NLP has the potential to improve understanding of the nature and prevalence of opioid ADEs and, ultimately, advance the field of medication safety.},
	booktitle = {The 2015 {HSR}\&{D}/{QUERI} {National} {Conference}},
	author = {Yu, H and Brandt, C and Becker, W and Kem, R},
	year = {2015}
}

Objectives: Prescription opioids are commonly used to treat acute and cancer-related pain, and, over the last two decades, have increasingly been used in the management of chronic non-cancer pain. Patients taking opioids can experience a wide range of adverse drug events (ADEs), including constipation, nausea/vomiting, pruritus, drowsiness and dizziness, hormonal dysfunction, depression, oversedation, falls, fractures, addiction, overdose, respiratory depression, sleep-disordered breathing, and death. Since such ADEs are frequently described in the unstructured electronic health record (EHR) notes, we are developing natural language processing (NLP) system to automatically extract opioid and ADEs from EHRs. The purpose of this study was to test out the feasibility of mining EHR notes for ADE detection using NLP approaches. Methods: We developed an annotation guideline using an interactive process during which physicians and linguists worked together to define rules and resolve discrepancy. Following the guideline, two annotators annotated 150 discharge summaries (or 8,672 sentences comprising 102,807 word tokens). The overall pairwise annotation agreement was 88%. The total number of annotated ADEs and medications were 103 and 3,290. Using this annotated corpus, we developed a NLP system to detect medication and ADE information. Our NLP system is trained on the supervised machine learning model Conditional Random Fields. We compared our NLP system with the state-of-the-art NLP system the MetaMap for ADE detection. Results: NLP performed well on discharge summaries on certain named entities, including frequency (92% F1), route (89% F1), dosage (87% F1), and medication (84% F1). Because the number of ADE instances is small, NLP performed poorly on ADE (24% F1). MetaMap performed on average 62% F1 for medication and 4% F1 for ADE. Implications: Our NLP system outperformed MetaMap for EHR notes ADE detection. NLP generally performs well with a sufficient size of annotated data. While the performance of ADE detection is low, more annotated data yielding a higher prevalence of ADEs would likely improve opioid ADE detection. Use of larger datasets is underway. Impacts: NLP has the potential to improve understanding of the nature and prevalence of opioid ADEs and, ultimately, advance the field of medication safety.

@inproceedings{liu_translating_2015,
	address = {Bei Jing, China},
	title = {Translating {Electronic} {Health} {Record} {Notes} from {English} to {Spanish}: {A} {Preliminary} {Study}},
	url = {http://aclweb.org/anthology/W/W15/W15-3816.pdf},
	booktitle = {{ACL}-{IJCNLP}},
	author = {Liu, Weisong and Cai, Shu and Balaji, Ramesh and Chiriboga, German and Knight, Kevin and Yu, Hong},
	month = jul,
	year = {2015},
	pages = {134}
}

@inproceedings{lingerman_learning_2015,
	title = {Learning to rank scientific articles.},
	url = {https://arxiv.org/pdf/1611.01400.pdf},
	booktitle = {{AMIA} {Fall} {Symposium}},
	author = {Lingerman, J and Hong, Yu},
	year = {2015}
}

@inproceedings{yu_systems_2015,
	title = {Systems for helping {Veterans} {Comprehend} their own {EHR} notes.},
	url = {http://www.hsrd.research.va.gov/meetings/2015/abstract-display.cfm?RecordID=149},
	abstract = {Objectives:
In January 2013, the VHA began to make clinical notes in the electronic health record (EHR) available through the My HealtheVet portal. However, studies have shown that patients are confused by EHR notes, especially patients in vulnerable groups (e.g., low literacy, low income). Such confusion may result in unintended increases in service utilization and lead to changes in perceptions that may disrupt patient-provider relationships. We are developing NoteAid, a multi-component, natural language processing (NLP) system that translates medical jargon into lay terms and provides definitions and links to related educational material from trusted resources. Here we report the development and evaluation of NoteAid.
 
Methods:
NoteAid has two main components: a knowledge resource comprised of patient education materials and the NLP system that links EHR notes to the knowledge resource. The knowledge resource integrates the medical jargon-lay term-definitions from the Unified Medical Language System and the MedlinePlus. The NLP system links the complex medical jargon that appear in the note to simple consumer oriented definitions and explanations from the knowledge resource. We evaluated NoteAid using 40 de-identified EHR notes (20 progress notes and 20 discharge summaries). We recruited 64 subjects. Each subject reads an assigned EHR note before and after NoteAid and self-scores comprehension (on a scale of 1 to 5, with 1 the poorest and 5 the best comprehension). Each subject completes the evaluation of either 20 PGNs or 20 DSs. We used Flesch-Kincaid grade level (FKGL) to score readability of each note.
 
Results:
Our results show a negative association between FKGL of a note and its subject's self-reported comprehension score (Spearman rho coefficient, -0.807; p {\textless} 0.0001). NoteAid improves self-reported EHR comprehension in both PGNs and DSs and the improvement is statistically significant (Wilcoxon signed-rank test, p {\textless} 0.05).
 
Implications:
NoteAid improves EHR note comprehension.
 
Impacts:
We are preparing to implement and evaluate NoteAid within the VA. NoteAid has the potential to facilitate translation of clinical records to patients, which could enhance patient engagement and lead to improved self-management and clinical outcomes.},
	booktitle = {2015 {HSR}\&{D}/{QUERI} {National} {Conference}},
	author = {Yu, H and Brandt, C and Houston, T},
	year = {2015}
}

Objectives: In January 2013, the VHA began to make clinical notes in the electronic health record (EHR) available through the My HealtheVet portal. However, studies have shown that patients are confused by EHR notes, especially patients in vulnerable groups (e.g., low literacy, low income). Such confusion may result in unintended increases in service utilization and lead to changes in perceptions that may disrupt patient-provider relationships. We are developing NoteAid, a multi-component, natural language processing (NLP) system that translates medical jargon into lay terms and provides definitions and links to related educational material from trusted resources. Here we report the development and evaluation of NoteAid. Methods: NoteAid has two main components: a knowledge resource comprised of patient education materials and the NLP system that links EHR notes to the knowledge resource. The knowledge resource integrates the medical jargon-lay term-definitions from the Unified Medical Language System and the MedlinePlus. The NLP system links the complex medical jargon that appear in the note to simple consumer oriented definitions and explanations from the knowledge resource. We evaluated NoteAid using 40 de-identified EHR notes (20 progress notes and 20 discharge summaries). We recruited 64 subjects. Each subject reads an assigned EHR note before and after NoteAid and self-scores comprehension (on a scale of 1 to 5, with 1 the poorest and 5 the best comprehension). Each subject completes the evaluation of either 20 PGNs or 20 DSs. We used Flesch-Kincaid grade level (FKGL) to score readability of each note. Results: Our results show a negative association between FKGL of a note and its subject's self-reported comprehension score (Spearman rho coefficient, -0.807; p \textless 0.0001). NoteAid improves self-reported EHR comprehension in both PGNs and DSs and the improvement is statistically significant (Wilcoxon signed-rank test, p \textless 0.05). Implications: NoteAid improves EHR note comprehension. Impacts: We are preparing to implement and evaluate NoteAid within the VA. NoteAid has the potential to facilitate translation of clinical records to patients, which could enhance patient engagement and lead to improved self-management and clinical outcomes.

@inproceedings{zheng_learning_2015,
	title = {Learning {Distributed} {Word} {Representations} and {Applications} in {Biomedial} {Natural} {Language} {Processing}},
	url = {http://medinfo2015.com/wp-content/uploads/2015/06/Learning-Distributed-Word-Representations-and-Applications-in-Biomedial-Natural-Language-Processing.pdf},
	booktitle = {{MEDINFO} 2015: {EHealth}-enabled {Health}: {Proceedings} of the 15th {World} {Congress} on {Health} and {Biomedical} {Informatics}},
	author = {Zheng, Jiaping and Yu, Hong},
	year = {2015}
}

@article{yin_detext:_2015,
	title = {{DeTEXT}: {A} {Database} for {Evaluating} {Text} {Extraction} from {Biomedical} {Literature} {Figures}},
	volume = {10},
	issn = {1932-6203},
	shorttitle = {{DeTEXT}},
	url = {http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4423993/},
	doi = {10.1371/journal.pone.0126200},
	abstract = {Hundreds of millions of figures are available in biomedical literature, representing important biomedical experimental evidence. Since text is a rich source of information in figures, automatically extracting such text may assist in the task of mining figure information. A high-quality ground truth standard can greatly facilitate the development of an automated system. This article describes DeTEXT: A database for evaluating text extraction from biomedical literature figures. It is the first publicly available, human-annotated, high quality, and large-scale figure-text dataset with 288 full-text articles, 500 biomedical figures, and 9308 text regions. This article describes how figures were selected from open-access full-text biomedical articles and how annotation guidelines and annotation tools were developed. We also discuss the inter-annotator agreement and the reliability of the annotations. We summarize the statistics of the DeTEXT data and make available evaluation protocols for DeTEXT. Finally we lay out challenges we observed in the automated detection and recognition of figure text and discuss research directions in this area. DeTEXT is publicly available for downloading at http://prir.ustb.edu.cn/DeTEXT/.},
	number = {5},
	urldate = {2015-06-03},
	journal = {PLoS ONE},
	author = {Yin, Xu-Cheng and Yang, Chun and Pei, Wei-Yi and Man, Haixia and Zhang, Jun and Learned-Miller, Erik and Yu, Hong},
	month = may,
	year = {2015}
}

Hundreds of millions of figures are available in biomedical literature, representing important biomedical experimental evidence. Since text is a rich source of information in figures, automatically extracting such text may assist in the task of mining figure information. A high-quality ground truth standard can greatly facilitate the development of an automated system. This article describes DeTEXT: A database for evaluating text extraction from biomedical literature figures. It is the first publicly available, human-annotated, high quality, and large-scale figure-text dataset with 288 full-text articles, 500 biomedical figures, and 9308 text regions. This article describes how figures were selected from open-access full-text biomedical articles and how annotation guidelines and annotation tools were developed. We also discuss the inter-annotator agreement and the reliability of the annotations. We summarize the statistics of the DeTEXT data and make available evaluation protocols for DeTEXT. Finally we lay out challenges we observed in the automated detection and recognition of figure text and discuss research directions in this area. DeTEXT is publicly available for downloading at http://prir.ustb.edu.cn/DeTEXT/.

@article{li_robust_2014,
	title = {A robust data-driven approach for gene ontology annotation},
	volume = {2014},
	issn = {1758-0463},
	url = {http://database.oxfordjournals.org/cgi/doi/10.1093/database/bau113},
	doi = {10.1093/database/bau113},
	abstract = {Gene ontology (GO) and GO annotation are important resources for biological information management and knowledge discovery, but the speed of manual annotation became a major bottleneck of database curation. BioCreative IV GO annotation task aims to evaluate the performance of system that automatically assigns GO terms to genes based on the narrative sentences in biomedical literature. This article presents our work in this task as well as the experimental results after the competition. For the evidence sentence extraction subtask, we built a binary classifier to identify evidence sentences using reference distance estimator (RDE), a recently proposed semi-supervised learning method that learns new features from around 10 million unlabeled sentences, achieving an F1 of 19.3\% in exact match and 32.5\% in relaxed match. In the post-submission experiment, we obtained 22.1\% and 35.7\% F1 performance by incorporating bigram features in RDE learning. In both development and test sets, RDE-based method achieved over 20\% relative improvement on F1 and AUC performance against classical supervised learning methods, e.g. support vector machine and logistic regression. For the GO term prediction subtask, we developed an information retrieval-based method to retrieve the GO term most relevant to each evidence sentence using a ranking function that combined cosine similarity and the frequency of GO terms in documents, and a filtering method based on high-level GO classes. The best performance of our submitted runs was 7.8\% F1 and 22.2\% hierarchy F1. We found that the incorporation of frequency information and hierarchy filtering substantially improved the performance. In the post-submission evaluation, we obtained a 10.6\% F1 using a simpler setting. Overall, the experimental analysis showed our approaches were robust in both the two tasks.},
	language = {eng},
	journal = {Database: The Journal of Biological Databases and Curation},
	author = {Li, Yanpeng and Yu, Hong},
	year = {2014},
	pmid = {25425037},
	pmcid = {PMC4243380},
	note = {00000 },
	pages = {bau113}
}

Gene ontology (GO) and GO annotation are important resources for biological information management and knowledge discovery, but the speed of manual annotation became a major bottleneck of database curation. BioCreative IV GO annotation task aims to evaluate the performance of system that automatically assigns GO terms to genes based on the narrative sentences in biomedical literature. This article presents our work in this task as well as the experimental results after the competition. For the evidence sentence extraction subtask, we built a binary classifier to identify evidence sentences using reference distance estimator (RDE), a recently proposed semi-supervised learning method that learns new features from around 10 million unlabeled sentences, achieving an F1 of 19.3% in exact match and 32.5% in relaxed match. In the post-submission experiment, we obtained 22.1% and 35.7% F1 performance by incorporating bigram features in RDE learning. In both development and test sets, RDE-based method achieved over 20% relative improvement on F1 and AUC performance against classical supervised learning methods, e.g. support vector machine and logistic regression. For the GO term prediction subtask, we developed an information retrieval-based method to retrieve the GO term most relevant to each evidence sentence using a ranking function that combined cosine similarity and the frequency of GO terms in documents, and a filtering method based on high-level GO classes. The best performance of our submitted runs was 7.8% F1 and 22.2% hierarchy F1. We found that the incorporation of frequency information and hierarchy filtering substantially improved the performance. In the post-submission evaluation, we obtained a 10.6% F1 using a simpler setting. Overall, the experimental analysis showed our approaches were robust in both the two tasks.

@article{polepalli_ramesh_automatically_2014,
	title = {Automatically {Recognizing} {Medication} and {Adverse} {Event} {Information} {From} {Food} and {Drug} {Administration}’s {Adverse} {Event} {Reporting} {System} {Narratives}},
	volume = {2},
	issn = {2291-9694},
	url = {http://medinform.jmir.org/2014/1/e10/},
	doi = {10.2196/medinform.3022},
	language = {en},
	number = {1},
	urldate = {2015-05-02},
	journal = {JMIR Medical Informatics},
	author = {Polepalli Ramesh, Balaji and Belknap, Steven M and Li, Zuofeng and Frid, Nadya and West, Dennis P and Yu, Hong},
	month = jun,
	year = {2014},
	pages = {e10}
}

@article{liu_learning_2014,
	title = {Learning to {Rank} {Figures} within a {Biomedical} {Article}},
	volume = {9},
	issn = {1932-6203},
	url = {http://dx.plos.org/10.1371/journal.pone.0061567},
	doi = {10.1371/journal.pone.0061567},
	language = {en},
	number = {3},
	urldate = {2015-02-26},
	journal = {PLoS ONE},
	author = {Liu, Feifan and Yu, Hong},
	editor = {Preis, Tobias},
	month = mar,
	year = {2014},
	pages = {e61567}
}

@article{schulman_treatment_2014,
	title = {Treatment of acute venous thromboembolism with dabigatran or warfarin and pooled analysis},
	volume = {129},
	issn = {1524-4539},
	doi = {10.1161/CIRCULATIONAHA.113.004450},
	abstract = {BACKGROUND: Dabigatran and warfarin have been compared for the treatment of acute venous thromboembolism (VTE) in a previous trial. We undertook this study to extend those findings.
METHODS AND RESULTS: In a randomized, double-blind, double-dummy trial of 2589 patients with acute VTE treated with low-molecular-weight or unfractionated heparin for 5 to 11 days, we compared dabigatran 150 mg twice daily with warfarin. The primary outcome, recurrent symptomatic, objectively confirmed VTE and related deaths during 6 months of treatment occurred in 30 of the 1279 dabigatran patients (2.3\%) compared with 28 of the 1289 warfarin patients (2.2\%; hazard ratio, 1.08; 95\% confidence interval [CI], 0.64-1.80; absolute risk difference, 0.2\%; 95\% CI, -1.0 to 1.3; P{\textless}0.001 for the prespecified noninferiority margin for both criteria). The safety end point, major bleeding, occurred in 15 patients receiving dabigatran (1.2\%) and in 22 receiving warfarin (1.7\%; hazard ratio, 0.69; 95\% CI, 0.36-1.32). Any bleeding occurred in 200 dabigatran (15.6\%) and 285 warfarin (22.1\%; hazard ratio, 0.67; 95\% CI, 0.56-0.81) patients. Deaths, adverse events, and acute coronary syndromes were similar in both groups. Pooled analysis of this study RE-COVER II and the RE-COVER trial gave hazard ratios for recurrent VTE of 1.09 (95\% CI, 0.76-1.57), for major bleeding of 0.73 (95\% CI, 0.48-1.11), and for any bleeding of 0.70 (95\% CI, 0.61-0.79).
CONCLUSION: Dabigatran has similar effects on VTE recurrence and a lower risk of bleeding compared with warfarin for the treatment of acute VTE.
CLINICAL TRIAL REGISTRATION URL: www.clinicaltrials.gov. Unique identifiers: NCT00680186 and NCT00291330.},
	language = {eng},
	number = {7},
	journal = {Circulation},
	author = {Schulman, Sam and Kakkar, Ajay K. and Goldhaber, Samuel Z. and Schellong, Sebastian and Eriksson, Henry and Mismetti, Patrick and Christiansen, Anita Vedel and Friedman, Jeffrey and Le Maulf, Florence and Peter, Nuala and Kearon, Clive and {RE-COVER II Trial Investigators}},
	month = feb,
	year = {2014},
	pmid = {24344086},
	keywords = {Acute Disease, Adolescent, Adult, Aged, Anticoagulants, Antithrombins, Benzimidazoles, Dabigatran, Double-Blind Method, Female, Follow-Up Studies, Hemorrhage, Heparin, Heparin, Low-Molecular-Weight, Humans, Male, Middle Aged, Recurrence, Risk Factors, Venous thromboembolism, Warfarin, Young Adult, antagonists \& inhibitors, beta-Alanine, thrombin},
	pages = {764--772}
}

BACKGROUND: Dabigatran and warfarin have been compared for the treatment of acute venous thromboembolism (VTE) in a previous trial. We undertook this study to extend those findings. METHODS AND RESULTS: In a randomized, double-blind, double-dummy trial of 2589 patients with acute VTE treated with low-molecular-weight or unfractionated heparin for 5 to 11 days, we compared dabigatran 150 mg twice daily with warfarin. The primary outcome, recurrent symptomatic, objectively confirmed VTE and related deaths during 6 months of treatment occurred in 30 of the 1279 dabigatran patients (2.3%) compared with 28 of the 1289 warfarin patients (2.2%; hazard ratio, 1.08; 95% confidence interval [CI], 0.64-1.80; absolute risk difference, 0.2%; 95% CI, -1.0 to 1.3; P\textless0.001 for the prespecified noninferiority margin for both criteria). The safety end point, major bleeding, occurred in 15 patients receiving dabigatran (1.2%) and in 22 receiving warfarin (1.7%; hazard ratio, 0.69; 95% CI, 0.36-1.32). Any bleeding occurred in 200 dabigatran (15.6%) and 285 warfarin (22.1%; hazard ratio, 0.67; 95% CI, 0.56-0.81) patients. Deaths, adverse events, and acute coronary syndromes were similar in both groups. Pooled analysis of this study RE-COVER II and the RE-COVER trial gave hazard ratios for recurrent VTE of 1.09 (95% CI, 0.76-1.57), for major bleeding of 0.73 (95% CI, 0.48-1.11), and for any bleeding of 0.70 (95% CI, 0.61-0.79). CONCLUSION: Dabigatran has similar effects on VTE recurrence and a lower risk of bleeding compared with warfarin for the treatment of acute VTE. CLINICAL TRIAL REGISTRATION URL: www.clinicaltrials.gov. Unique identifiers: NCT00680186 and NCT00291330.

@article{zhang_computational_2014,
	title = {Computational {Approaches} for {Predicting} {Biomedical} {Research} {Collaborations}},
	volume = {9},
	issn = {1932-6203},
	url = {http://dx.plos.org/10.1371/journal.pone.0111795},
	doi = {10.1371/journal.pone.0111795},
	language = {en},
	number = {11},
	urldate = {2015-02-26},
	journal = {PLoS ONE},
	author = {Zhang, Qing and Yu, Hong},
	editor = {Smalheiser, Neil R.},
	month = nov,
	year = {2014},
	pages = {e111795}
}

@inproceedings{polepalli_ramesh_systems_2013,
	title = {Systems for {Improving} {Electronic} {Health} {Record} {Note} {Comprehension}},
	booktitle = {{ACM} {SIGIR} {Workshop} on {Health} {Search} \& {Discovery}},
	author = {Polepalli Ramesh, Balaji and Yu, Hong},
	year = {2013}
}

@article{qing_citegraph:_2013,
	title = {{CiteGraph}: {A} {Citation} {Network} {System} for {MEDLINE} {Articles} and {Analysis}},
	copyright = {©2013 \© IMIA and IOS Press.},
	issn = {0926-9630},
	shorttitle = {{CiteGraph}},
	url = {http://www.medra.org/servlet/aliasResolver?alias=iospressISSNISBN&issn=0926-9630&volume=192&spage=832},
	doi = {10.3233/978-1-61499-289-9-832},
	abstract = {This paper details the development and implementation of CiteGraph, a system for constructing large-scale citation and co-authorship networks from full-text biomedical articles. CiteGraph represents articles and authors by uniquely identified nodes, and connects those nodes through citation and co-authorship relations. CiteGraph network encompasses over 1.65 million full-text articles and 6.35 million citations by 1.37 million unique authors from the Elsevier full-text articles. Our evaluation shows 98\% 99\% F1-score for mapping a citation to the corresponding article and identifying MEDLINE articles. We further analyzed the characteristics of CiteGraph and found that they are consistent with assumptions made using small-scale bibliometric analysis. We also developed several novel network-based methods for analyzing publication, citation and collaboration patterns. This is the first work to develop a completely automated system for the creation of a large-scale citation network in the biomedical domain, and also to introduce novel findings in researcher publication histories. CiteGraph can be a useful resource to both the biomedical community, and bibliometric research.},
	urldate = {2016-11-30},
	journal = {Studies in Health Technology and Informatics},
	author = {Qing, Zhang and Hong, Yu},
	year = {2013},
	pages = {832--836}
}

This paper details the development and implementation of CiteGraph, a system for constructing large-scale citation and co-authorship networks from full-text biomedical articles. CiteGraph represents articles and authors by uniquely identified nodes, and connects those nodes through citation and co-authorship relations. CiteGraph network encompasses over 1.65 million full-text articles and 6.35 million citations by 1.37 million unique authors from the Elsevier full-text articles. Our evaluation shows 98% 99% F1-score for mapping a citation to the corresponding article and identifying MEDLINE articles. We further analyzed the characteristics of CiteGraph and found that they are consistent with assumptions made using small-scale bibliometric analysis. We also developed several novel network-based methods for analyzing publication, citation and collaboration patterns. This is the first work to develop a completely automated system for the creation of a large-scale citation network in the biomedical domain, and also to introduce novel findings in researcher publication histories. CiteGraph can be a useful resource to both the biomedical community, and bibliometric research.

@article{ramesh_automatic_2012,
	title = {Automatic discourse connective detection in biomedical text},
	volume = {19},
	issn = {1527-974X},
	doi = {10.1136/amiajnl-2011-000775},
	abstract = {OBJECTIVE

Relation extraction in biomedical text mining systems has largely focused on identifying clause-level relations, but increasing sophistication demands the recognition of relations at discourse level. A first step in identifying discourse relations involves the detection of discourse connectives: words or phrases used in text to express discourse relations. In this study supervised machine-learning approaches were developed and evaluated for automatically identifying discourse connectives in biomedical text.


MATERIALS AND METHODS

Two supervised machine-learning models (support vector machines and conditional random fields) were explored for identifying discourse connectives in biomedical literature. In-domain supervised machine-learning classifiers were trained on the Biomedical Discourse Relation Bank, an annotated corpus of discourse relations over 24 full-text biomedical articles ({\textasciitilde}112,000 word tokens), a subset of the GENIA corpus. Novel domain adaptation techniques were also explored to leverage the larger open-domain Penn Discourse Treebank ({\textasciitilde}1 million word tokens). The models were evaluated using the standard evaluation metrics of precision, recall and F1 scores.


RESULTS AND CONCLUSION

Supervised machine-learning approaches can automatically identify discourse connectives in biomedical text, and the novel domain adaptation techniques yielded the best performance: 0.761 F1 score. A demonstration version of the fully implemented classifier BioConn is available at: http://bioconn.askhermes.org.},
	number = {5},
	journal = {Journal of the American Medical Informatics Association: JAMIA},
	author = {Ramesh, Balaji Polepalli and Prasad, Rashmi and Miller, Tim and Harrington, Brian and Yu, Hong},
	month = oct,
	year = {2012},
	pmid = {22744958},
	keywords = {Knowledge Bases, NLP, analysis, automated learning, controlled terminologies and vocabularies, discovery, display, image representation, knowledge acquisition and knowledge management, knowledge representations, natural language processing, ontologies, processing, text and data mining methods},
	pages = {800--808}
}

OBJECTIVE Relation extraction in biomedical text mining systems has largely focused on identifying clause-level relations, but increasing sophistication demands the recognition of relations at discourse level. A first step in identifying discourse relations involves the detection of discourse connectives: words or phrases used in text to express discourse relations. In this study supervised machine-learning approaches were developed and evaluated for automatically identifying discourse connectives in biomedical text. MATERIALS AND METHODS Two supervised machine-learning models (support vector machines and conditional random fields) were explored for identifying discourse connectives in biomedical literature. In-domain supervised machine-learning classifiers were trained on the Biomedical Discourse Relation Bank, an annotated corpus of discourse relations over 24 full-text biomedical articles (\textasciitilde112,000 word tokens), a subset of the GENIA corpus. Novel domain adaptation techniques were also explored to leverage the larger open-domain Penn Discourse Treebank (\textasciitilde1 million word tokens). The models were evaluated using the standard evaluation metrics of precision, recall and F1 scores. RESULTS AND CONCLUSION Supervised machine-learning approaches can automatically identify discourse connectives in biomedical text, and the novel domain adaptation techniques yielded the best performance: 0.761 F1 score. A demonstration version of the fully implemented classifier BioConn is available at: http://bioconn.askhermes.org.

@article{bockhorst_beyond_2012,
	title = {Beyond {Captions}: {Linking} {Figures} with {Abstract} {Sentences} in {Biomedical} {Articles}},
	volume = {7},
	issn = {1932-6203},
	shorttitle = {Beyond {Captions}},
	url = {http://dx.plos.org/10.1371/journal.pone.0039618},
	doi = {10.1371/journal.pone.0039618},
	language = {en},
	number = {7},
	urldate = {2016-11-30},
	journal = {PLoS ONE},
	author = {Bockhorst, Joseph P. and Conroy, John M. and Agarwal, Shashank and O’Leary, Dianne P. and Yu, Hong},
	editor = {Ouzounis, Christos A.},
	month = jul,
	year = {2012},
	pages = {e39618}
}

@incollection{liu_natural_2012,
	title = {Natural {Language} {Processing}, {Electronic} {Health} {Records}, and {Clinical} {Research}},
	booktitle = {Clinical {Research} {Informatics}},
	publisher = {Springer London},
	author = {Liu, Feifan and Weng, Chunhua and Yu, Hong},
	year = {2012},
	pages = {293--310}
}

@article{liu_medtxting:_2012,
	title = {{MedTxting}: learning based and knowledge rich {SMS}-style medical text contraction},
	volume = {2012},
	issn = {1942-597X},
	shorttitle = {{MedTxting}},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3540574/},
	abstract = {In mobile health (M-health), Short Message Service (SMS) has shown to improve disease related self-management and health service outcomes, leading to enhanced patient care. However, the hard limit on character size for each message limits the full value of exploring SMS communication in health care practices. To overcome this problem and improve the efficiency of clinical workflow, we developed an innovative system, MedTxting (available at http://medtxting.askhermes.org), which is a learning-based but knowledge-rich system that compresses medical texts in a SMS style. Evaluations on clinical questions and discharge summary narratives show that MedTxting can effectively compress medical texts with reasonable readability and noticeable size reduction. Findings in this work reveal potentials of MedTxting to the clinical settings, allowing for real-time and cost-effective communication, such as patient condition reporting, medication consulting, physicians connecting to share expertise to improve point of care.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Liu, Feifan and Moosavinasab, Soheil and Houston, Thomas K. and Yu, Hong},
	year = {2012},
	pmid = {23304328},
	pmcid = {PMC3540574},
	keywords = {Humans, Pilot Projects, Telemedicine, Text Messaging},
	pages = {558--567}
}

In mobile health (M-health), Short Message Service (SMS) has shown to improve disease related self-management and health service outcomes, leading to enhanced patient care. However, the hard limit on character size for each message limits the full value of exploring SMS communication in health care practices. To overcome this problem and improve the efficiency of clinical workflow, we developed an innovative system, MedTxting (available at http://medtxting.askhermes.org), which is a learning-based but knowledge-rich system that compresses medical texts in a SMS style. Evaluations on clinical questions and discharge summary narratives show that MedTxting can effectively compress medical texts with reasonable readability and noticeable size reduction. Findings in this work reveal potentials of MedTxting to the clinical settings, allowing for real-time and cost-effective communication, such as patient condition reporting, medication consulting, physicians connecting to share expertise to improve point of care.

@article{miller_investigation_2011,
	title = {An investigation into the feasibility of spoken clinical question answering},
	volume = {2011},
	issn = {1942-597X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3243288/},
	abstract = {Spoken question answering for clinical decision support is a potentially revolutionary technology for improving the efficiency and quality of health care delivery. This application involves many technologies currently being researched, including automatic speech recognition (ASR), information retrieval (IR), and summarization, all in the biomedical domain. In certain domains, the problem of spoken document retrieval has been declared solved because of the robustness of IR to ASR errors. This study investigates the extent to which spoken medical question answering benefits from that same robustness. We used the best results from previous speech recognition experiments as inputs to a clinical question answering system, and had physicians perform blind evaluations of results generated both by ASR transcripts of questions and gold standard transcripts of the same questions. Our results suggest that the medical domain differs enough from the open domain to require additional work in automatic speech recognition adapted for the biomedical domain.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Miller, Tim and Ravvaz, Kourosh and Cimino, James J. and Yu, Hong},
	year = {2011},
	pmid = {22195154},
	pmcid = {PMC3243288},
	keywords = {Decision Support Systems, Clinical, Feasibility Studies, Humans, Information Storage and Retrieval, Speech Recognition Software, natural language processing},
	pages = {954--959}
}

Spoken question answering for clinical decision support is a potentially revolutionary technology for improving the efficiency and quality of health care delivery. This application involves many technologies currently being researched, including automatic speech recognition (ASR), information retrieval (IR), and summarization, all in the biomedical domain. In certain domains, the problem of spoken document retrieval has been declared solved because of the robustness of IR to ASR errors. This study investigates the extent to which spoken medical question answering benefits from that same robustness. We used the best results from previous speech recognition experiments as inputs to a clinical question answering system, and had physicians perform blind evaluations of results generated both by ASR transcripts of questions and gold standard transcripts of the same questions. Our results suggest that the medical domain differs enough from the open domain to require additional work in automatic speech recognition adapted for the biomedical domain.

@article{granger_apixaban_2011,
	title = {Apixaban versus warfarin in patients with atrial fibrillation},
	volume = {365},
	issn = {1533-4406},
	url = {http://www.nejm.org/doi/full/10.1056/NEJMoa1107039},
	doi = {10.1056/NEJMoa1107039},
	abstract = {BACKGROUND: Vitamin K antagonists are highly effective in preventing stroke in patients with atrial fibrillation but have several limitations. Apixaban is a novel oral direct factor Xa inhibitor that has been shown to reduce the risk of stroke in a similar population in comparison with aspirin.
METHODS: In this randomized, double-blind trial, we compared apixaban (at a dose of 5 mg twice daily) with warfarin (target international normalized ratio, 2.0 to 3.0) in 18,201 patients with atrial fibrillation and at least one additional risk factor for stroke. The primary outcome was ischemic or hemorrhagic stroke or systemic embolism. The trial was designed to test for noninferiority, with key secondary objectives of testing for superiority with respect to the primary outcome and to the rates of major bleeding and death from any cause.
RESULTS: The median duration of follow-up was 1.8 years. The rate of the primary outcome was 1.27\% per year in the apixaban group, as compared with 1.60\% per year in the warfarin group (hazard ratio with apixaban, 0.79; 95\% confidence interval [CI], 0.66 to 0.95; P{\textless}0.001 for noninferiority; P=0.01 for superiority). The rate of major bleeding was 2.13\% per year in the apixaban group, as compared with 3.09\% per year in the warfarin group (hazard ratio, 0.69; 95\% CI, 0.60 to 0.80; P{\textless}0.001), and the rates of death from any cause were 3.52\% and 3.94\%, respectively (hazard ratio, 0.89; 95\% CI, 0.80 to 0.99; P=0.047). The rate of hemorrhagic stroke was 0.24\% per year in the apixaban group, as compared with 0.47\% per year in the warfarin group (hazard ratio, 0.51; 95\% CI, 0.35 to 0.75; P{\textless}0.001), and the rate of ischemic or uncertain type of stroke was 0.97\% per year in the apixaban group and 1.05\% per year in the warfarin group (hazard ratio, 0.92; 95\% CI, 0.74 to 1.13; P=0.42).
CONCLUSIONS: In patients with atrial fibrillation, apixaban was superior to warfarin in preventing stroke or systemic embolism, caused less bleeding, and resulted in lower mortality. (Funded by Bristol-Myers Squibb and Pfizer; ARISTOTLE ClinicalTrials.gov number, NCT00412984.).},
	language = {eng},
	number = {11},
	journal = {The New England Journal of Medicine},
	author = {Granger, Christopher B. and Alexander, John H. and McMurray, John J. V. and Lopes, Renato D. and Hylek, Elaine M. and Hanna, Michael and Al-Khalidi, Hussein R. and Ansell, Jack and Atar, Dan and Avezum, Alvaro and Bahit, M. Cecilia and Diaz, Rafael and Easton, J. Donald and Ezekowitz, Justin A. and Flaker, Greg and Garcia, David and Geraldes, Margarida and Gersh, Bernard J. and Golitsyn, Sergey and Goto, Shinya and Hermosillo, Antonio G. and Hohnloser, Stefan H. and Horowitz, John and Mohan, Puneet and Jansky, Petr and Lewis, Basil S. and Lopez-Sendon, Jose Luis and Pais, Prem and Parkhomenko, Alexander and Verheugt, Freek W. A. and Zhu, Jun and Wallentin, Lars and {ARISTOTLE Committees and Investigators}},
	month = sep,
	year = {2011},
	pmid = {21870978},
	keywords = {Aged, Anticoagulants, Atrial Fibrillation, Double-Blind Method, Factor Xa Inhibitors, Female, Follow-Up Studies, Hemorrhage, Humans, International Normalized Ratio, Kaplan-Meier Estimate, Male, Middle Aged, Pyrazoles, Pyridones, Stroke, Thromboembolism, Treatment Outcome, Warfarin},
	pages = {981--992}
}

BACKGROUND: Vitamin K antagonists are highly effective in preventing stroke in patients with atrial fibrillation but have several limitations. Apixaban is a novel oral direct factor Xa inhibitor that has been shown to reduce the risk of stroke in a similar population in comparison with aspirin. METHODS: In this randomized, double-blind trial, we compared apixaban (at a dose of 5 mg twice daily) with warfarin (target international normalized ratio, 2.0 to 3.0) in 18,201 patients with atrial fibrillation and at least one additional risk factor for stroke. The primary outcome was ischemic or hemorrhagic stroke or systemic embolism. The trial was designed to test for noninferiority, with key secondary objectives of testing for superiority with respect to the primary outcome and to the rates of major bleeding and death from any cause. RESULTS: The median duration of follow-up was 1.8 years. The rate of the primary outcome was 1.27% per year in the apixaban group, as compared with 1.60% per year in the warfarin group (hazard ratio with apixaban, 0.79; 95% confidence interval [CI], 0.66 to 0.95; P\textless0.001 for noninferiority; P=0.01 for superiority). The rate of major bleeding was 2.13% per year in the apixaban group, as compared with 3.09% per year in the warfarin group (hazard ratio, 0.69; 95% CI, 0.60 to 0.80; P\textless0.001), and the rates of death from any cause were 3.52% and 3.94%, respectively (hazard ratio, 0.89; 95% CI, 0.80 to 0.99; P=0.047). The rate of hemorrhagic stroke was 0.24% per year in the apixaban group, as compared with 0.47% per year in the warfarin group (hazard ratio, 0.51; 95% CI, 0.35 to 0.75; P\textless0.001), and the rate of ischemic or uncertain type of stroke was 0.97% per year in the apixaban group and 1.05% per year in the warfarin group (hazard ratio, 0.92; 95% CI, 0.74 to 1.13; P=0.42). CONCLUSIONS: In patients with atrial fibrillation, apixaban was superior to warfarin in preventing stroke or systemic embolism, caused less bleeding, and resulted in lower mortality. (Funded by Bristol-Myers Squibb and Pfizer; ARISTOTLE ClinicalTrials.gov number, NCT00412984.).

@article{cao_askhermes:_2011,
	title = {{AskHERMES}: {An} online question answering system for complex clinical questions},
	volume = {44},
	issn = {1532-0480},
	shorttitle = {{AskHERMES}},
	url = {http://www.ncbi.nlm.nih.gov/pubmed/21256977},
	doi = {10.1016/j.jbi.2011.01.004},
	abstract = {{\textless}AbstractText Label="OBJECTIVE" NlmCategory="OBJECTIVE"{\textgreater}Clinical questions are often long and complex and take many forms. We have built a clinical question answering system named AskHERMES to perform robust semantic analysis on complex clinical questions and output question-focused extractive summaries as answers.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="DESIGN" NlmCategory="METHODS"{\textgreater}This paper describes the system architecture and a preliminary evaluation of AskHERMES, which implements innovative approaches in question analysis, summarization, and answer presentation. Five types of resources were indexed in this system: MEDLINE abstracts, PubMed Central full-text articles, eMedicine documents, clinical guidelines and Wikipedia articles.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="MEASUREMENT" NlmCategory="METHODS"{\textgreater}We compared the AskHERMES system with Google (Google and Google Scholar) and UpToDate and asked physicians to score the three systems by ease of use, quality of answer, time spent, and overall performance.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="RESULTS" NlmCategory="RESULTS"{\textgreater}AskHERMES allows physicians to enter a question in a natural way with minimal query formulation and allows physicians to efficiently navigate among all the answer sentences to quickly meet their information needs. In contrast, physicians need to formulate queries to search for information in Google and UpToDate. The development of the AskHERMES system is still at an early stage, and the knowledge resource is limited compared with Google or UpToDate. Nevertheless, the evaluation results show that AskHERMES' performance is comparable to the other systems. In particular, when answering complex clinical questions, it demonstrates the potential to outperform both Google and UpToDate systems.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="CONCLUSIONS" NlmCategory="CONCLUSIONS"{\textgreater}AskHERMES, available at http://www.AskHERMES.org, has the potential to help physicians practice evidence-based medicine and improve the quality of patient care.{\textless}/AbstractText{\textgreater}},
	number = {2},
	urldate = {2011-03-25},
	journal = {Journal of Biomedical Informatics},
	author = {Cao, Yonggang and Liu, Feifan and Simpson, Pippa and Antieau, Lamont and Bennett, Andrew and Cimino, James J and Ely, John and Yu, Hong},
	month = apr,
	year = {2011},
	pmid = {21256977},
	keywords = {Algorithms, Clinical Medicine, Databases, Factual, Information Storage and Retrieval, Online Systems, Software, expert systems, natural language processing},
	pages = {277--288}
}

\textlessAbstractText Label="OBJECTIVE" NlmCategory="OBJECTIVE"\textgreaterClinical questions are often long and complex and take many forms. We have built a clinical question answering system named AskHERMES to perform robust semantic analysis on complex clinical questions and output question-focused extractive summaries as answers.\textless/AbstractText\textgreater \textlessAbstractText Label="DESIGN" NlmCategory="METHODS"\textgreaterThis paper describes the system architecture and a preliminary evaluation of AskHERMES, which implements innovative approaches in question analysis, summarization, and answer presentation. Five types of resources were indexed in this system: MEDLINE abstracts, PubMed Central full-text articles, eMedicine documents, clinical guidelines and Wikipedia articles.\textless/AbstractText\textgreater \textlessAbstractText Label="MEASUREMENT" NlmCategory="METHODS"\textgreaterWe compared the AskHERMES system with Google (Google and Google Scholar) and UpToDate and asked physicians to score the three systems by ease of use, quality of answer, time spent, and overall performance.\textless/AbstractText\textgreater \textlessAbstractText Label="RESULTS" NlmCategory="RESULTS"\textgreaterAskHERMES allows physicians to enter a question in a natural way with minimal query formulation and allows physicians to efficiently navigate among all the answer sentences to quickly meet their information needs. In contrast, physicians need to formulate queries to search for information in Google and UpToDate. The development of the AskHERMES system is still at an early stage, and the knowledge resource is limited compared with Google or UpToDate. Nevertheless, the evaluation results show that AskHERMES' performance is comparable to the other systems. In particular, when answering complex clinical questions, it demonstrates the potential to outperform both Google and UpToDate systems.\textless/AbstractText\textgreater \textlessAbstractText Label="CONCLUSIONS" NlmCategory="CONCLUSIONS"\textgreaterAskHERMES, available at http://www.AskHERMES.org, has the potential to help physicians practice evidence-based medicine and improve the quality of patient care.\textless/AbstractText\textgreater

@article{kim_automatic_2011,
	title = {Automatic figure classification in bioscience literature},
	volume = {44},
	issn = {15320464},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046411000943},
	doi = {10.1016/j.jbi.2011.05.003},
	language = {en},
	number = {5},
	urldate = {2016-11-30},
	journal = {Journal of Biomedical Informatics},
	author = {Kim, Daehyun and Ramesh, Balaji Polepalli and Yu, Hong},
	month = oct,
	year = {2011},
	pages = {848--858}
}

@article{agarwal_figure_2011,
	title = {Figure summarizer browser extensions for {PubMed} {Central}},
	volume = {27},
	issn = {1367-4803, 1460-2059},
	url = {https://academic.oup.com/bioinformatics/article-lookup/doi/10.1093/bioinformatics/btr194},
	doi = {10.1093/bioinformatics/btr194},
	language = {en},
	number = {12},
	urldate = {2016-11-30},
	journal = {Bioinformatics},
	author = {Agarwal, S. and Yu, H.},
	month = jun,
	year = {2011},
	pages = {1723--1724}
}

@article{kim_figure_2011,
	title = {Figure {Text} {Extraction} in {Biomedical} {Literature}},
	volume = {6},
	issn = {1932-6203},
	url = {http://dx.plos.org/10.1371/journal.pone.0015338},
	doi = {10.1371/journal.pone.0015338},
	language = {en},
	number = {1},
	urldate = {2016-11-30},
	journal = {PLoS ONE},
	author = {Kim, Daehyun and Yu, Hong},
	editor = {Uversky, Vladimir N.},
	month = jan,
	year = {2011},
	pages = {e15338}
}

@article{zhang_parsing_2011,
	title = {Parsing citations in biomedical articles using conditional random fields},
	volume = {41},
	issn = {00104825},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S0010482511000291},
	doi = {10.1016/j.compbiomed.2011.02.005},
	language = {en},
	number = {4},
	urldate = {2016-11-30},
	journal = {Computers in Biology and Medicine},
	author = {Zhang, Qing and Cao, Yong-Gang and Yu, Hong},
	month = apr,
	year = {2011},
	pages = {190--194}
}

@article{prasad_biomedical_2011,
	title = {The biomedical discourse relation bank},
	volume = {12},
	copyright = {2011 Prasad et al; licensee BioMed Central Ltd.},
	issn = {1471-2105},
	url = {http://www.biomedcentral.com/1471-2105/12/188/abstract},
	doi = {10.1186/1471-2105-12-188},
	abstract = {Identification of discourse relations, such as causal and contrastive relations, between situations mentioned in text is an important task for biomedical text-mining. A biomedical text corpus annotated with discourse relations would be very useful for developing and evaluating methods for biomedical discourse processing. However, little effort has been made to develop such an annotated resource.},
	language = {en},
	number = {1},
	urldate = {2013-05-23},
	journal = {BMC Bioinformatics},
	author = {Prasad, Rashmi and McRoy, Susan and Frid, Nadya and Joshi, Aravind and Yu, Hong},
	month = may,
	year = {2011},
	pmid = {21605399},
	pages = {188}
}

Identification of discourse relations, such as causal and contrastive relations, between situations mentioned in text is an important task for biomedical text-mining. A biomedical text corpus annotated with discourse relations would be very useful for developing and evaluating methods for biomedical discourse processing. However, little effort has been made to develop such an annotated resource.

@article{liu_toward_2011,
	title = {Toward automated consumer question answering: {Automatically} separating consumer questions from professional questions in the healthcare domain},
	volume = {44},
	issn = {15320464},
	shorttitle = {Toward automated consumer question answering},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046411001353},
	doi = {10.1016/j.jbi.2011.08.008},
	language = {en},
	number = {6},
	urldate = {2016-11-30},
	journal = {Journal of Biomedical Informatics},
	author = {Liu, Feifan and Antieau, Lamont D. and Yu, Hong},
	month = dec,
	year = {2011},
	keywords = {Artificial Intelligence, Consumer Participation, Databases, Factual, Delivery of Health Care, Humans, Information Dissemination, Information Storage and Retrieval, Internet, Point-of-Care Systems, Semantics, natural language processing},
	pages = {1032--1038}
}

@article{liu_towards_2011,
	title = {Towards spoken clinical-question answering: evaluating and adapting automatic speech-recognition systems for spoken clinical questions},
	volume = {18},
	issn = {1527-974X},
	shorttitle = {Towards spoken clinical-question answering},
	url = {http://www.ncbi.nlm.nih.gov/pubmed/21705457},
	doi = {10.1136/amiajnl-2010-000071},
	abstract = {OBJECTIVE

To evaluate existing automatic speech-recognition (ASR) systems to measure their performance in interpreting spoken clinical questions and to adapt one ASR system to improve its performance on this task.


DESIGN AND MEASUREMENTS

The authors evaluated two well-known ASR systems on spoken clinical questions: Nuance Dragon (both generic and medical versions: Nuance Gen and Nuance Med) and the SRI Decipher (the generic version SRI Gen). The authors also explored language model adaptation using more than 4000 clinical questions to improve the SRI system's performance, and profile training to improve the performance of the Nuance Med system. The authors reported the results with the NIST standard word error rate (WER) and further analyzed error patterns at the semantic level.


RESULTS

Nuance Gen and Med systems resulted in a WER of 68.1\% and 67.4\% respectively. The SRI Gen system performed better, attaining a WER of 41.5\%. After domain adaptation with a language model, the performance of the SRI system improved 36\% to a final WER of 26.7\%.


CONCLUSION

Without modification, two well-known ASR systems do not perform well in interpreting spoken clinical questions. With a simple domain adaptation, one of the ASR systems improved significantly on the clinical question task, indicating the importance of developing domain/genre-specific ASR systems.},
	number = {5},
	urldate = {2011-12-13},
	journal = {Journal of the American Medical Informatics Association: JAMIA},
	author = {Liu, Feifan and Tur, Gokhan and Hakkani-Tür, Dilek and Yu, Hong},
	month = oct,
	year = {2011},
	pmid = {21705457},
	pages = {625--630}
}

OBJECTIVE To evaluate existing automatic speech-recognition (ASR) systems to measure their performance in interpreting spoken clinical questions and to adapt one ASR system to improve its performance on this task. DESIGN AND MEASUREMENTS The authors evaluated two well-known ASR systems on spoken clinical questions: Nuance Dragon (both generic and medical versions: Nuance Gen and Nuance Med) and the SRI Decipher (the generic version SRI Gen). The authors also explored language model adaptation using more than 4000 clinical questions to improve the SRI system's performance, and profile training to improve the performance of the Nuance Med system. The authors reported the results with the NIST standard word error rate (WER) and further analyzed error patterns at the semantic level. RESULTS Nuance Gen and Med systems resulted in a WER of 68.1% and 67.4% respectively. The SRI Gen system performed better, attaining a WER of 41.5%. After domain adaptation with a language model, the performance of the SRI system improved 36% to a final WER of 26.7%. CONCLUSION Without modification, two well-known ASR systems do not perform well in interpreting spoken clinical questions. With a simple domain adaptation, one of the ASR systems improved significantly on the clinical question task, indicating the importance of developing domain/genre-specific ASR systems.

@article{agarwal_simple_2011,
	title = {Simple and efficient machine learning frameworks for identifying protein-protein interaction relevant articles and experimental methods used to study the interactions},
	volume = {12},
	issn = {1471-2105},
	url = {http://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-12-S8-S10},
	doi = {10.1186/1471-2105-12-S8-S10},
	language = {en},
	number = {Suppl 8},
	urldate = {2016-11-30},
	journal = {BMC Bioinformatics},
	author = {Agarwal, Shashank and Liu, Feifan and Yu, Hong},
	year = {2011},
	pages = {S10}
}

@article{agarwal_biont:_2011,
	title = {{BioN}∅{T}: {A} searchable database of biomedical negated sentences},
	volume = {12},
	issn = {1471-2105},
	shorttitle = {{BioN}∅{T}},
	url = {http://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-12-420},
	doi = {10.1186/1471-2105-12-420},
	language = {en},
	number = {1},
	urldate = {2016-11-30},
	journal = {BMC Bioinformatics},
	author = {Agarwal, Shashank and Yu, Hong and Kohane, Issac},
	year = {2011},
	pages = {420}
}

@article{cao_ir-aided_2010,
	title = {An {IR}-aided machine learning framework for the {BioCreative} {II}.5 {Challenge}},
	volume = {7},
	issn = {1557-9964},
	url = {http://www.ncbi.nlm.nih.gov/pubmed/20671317},
	doi = {10.1109/TCBB.2010.56},
	abstract = {The team at the University of Wisconsin-Milwaukee developed an information retrieval and machine learning framework. Our framework requires only the standardized training data and depends upon minimal external knowledge resources and minimal parsing. Within the framework, we built our text mining systems and participated for the first time in all three BioCreative II.5 Challenge tasks. The results show that our systems performed among the top five teams for raw F1 scores in all three tasks and came in third place for the homonym ortholog F1 scores for the INT task. The results demonstrated that our IR-based framework is efficient, robust, and potentially scalable.},
	number = {3},
	urldate = {2010-09-21},
	journal = {IEEE/ACM Transactions on Computational Biology and Bioinformatics / IEEE, ACM},
	author = {Cao, Yonggang and Li, Zuofeng and Liu, Feifan and Agarwal, Shashank and Zhang, Qing and Yu, Hong},
	month = sep,
	year = {2010},
	pmid = {20671317},
	keywords = {Text mining, bioinformatics (genome or protein) databases, information search and retrieval, systems and software},
	pages = {454--461}
}

The team at the University of Wisconsin-Milwaukee developed an information retrieval and machine learning framework. Our framework requires only the standardized training data and depends upon minimal external knowledge resources and minimal parsing. Within the framework, we built our text mining systems and participated for the first time in all three BioCreative II.5 Challenge tasks. The results show that our systems performed among the top five teams for raw F1 scores in all three tasks and came in third place for the homonym ortholog F1 scores for the INT task. The results demonstrated that our IR-based framework is efficient, robust, and potentially scalable.

@article{yu_automatic_2010,
	title = {Automatic {Figure} {Ranking} and {User} {Interfacing} for {Intelligent} {Figure} {Search}},
	volume = {5},
	issn = {1932-6203},
	url = {http://dx.plos.org/10.1371/journal.pone.0012983},
	doi = {10.1371/journal.pone.0012983},
	language = {en},
	number = {10},
	urldate = {2016-11-30},
	journal = {PLoS ONE},
	author = {Yu, Hong and Liu, Feifan and Ramesh, Balaji Polepalli},
	editor = {Wong, Kelvin Kian Loong},
	month = oct,
	year = {2010},
	keywords = {balaji-ramesh, bioinformatics, feifan-liu, hong-yu},
	pages = {e12983}
}

@article{cao_automatically_2010,
	title = {Automatically extracting information needs from complex clinical questions. {Best} {Paper} in {International} {Medical} {Informatics} {Association} ({IMIA}) {Yearbook} 2011},
	issn = {1532-0480},
	url = {http://www.ncbi.nlm.nih.gov/pubmed/20670693},
	doi = {10.1016/j.jbi.2010.07.007},
	abstract = {OBJECTIVE: Clinicians pose complex clinical questions when seeing patients, and identifying the answers to those questions in a timely manner helps improve the quality of patient care. We report here on two natural language processing models, namely, automatic topic assignment and keyword identification, that together automatically and effectively extract information needs from ad hoc clinical questions. Our study is motivated in the context of developing the larger clinical question answering system AskHERMES (Help clinicians to Extract and aRrticulate Multimedia information for answering clinical quEstionS). DESIGN AND MEASUREMENTS: We developed supervised machine-learning systems to automatically assign predefined general categories (e.g. etiology, procedure, and diagnosis) to a question. We also explored both supervised and unsupervised systems to automatically identify keywords that capture the main content of the question. RESULTS: We evaluated our systems on 4654 annotated clinical questions that were collected in practice. We achieved an F1 score of 76.0\% for the task of general topic classification and 58.0\% for keyword extraction. Our systems have been implemented into the larger question answering system AskHERMES. Our error analyses suggested that inconsistent annotation in our training data have hurt both question analysis tasks. CONCLUSION: Our systems, available at http://www.askhermes.org, can automatically extract information needs from both short (the number of word tokens {\textless}20) and long questions (the number of word tokens {\textgreater}20), and from both well-structured and ill-formed questions. We speculate that the performance of general topic classification and keyword extraction can be further improved if consistently annotated data are made available.},
	urldate = {2010-09-21},
	journal = {Journal of Biomedical Informatics},
	author = {Cao, Yong-Gang and Cimino, James J and Ely, John and Yu, Hong},
	month = jul,
	year = {2010},
	pmid = {20670693},
	keywords = {Keyword extraction, Question analysis, Question answering, natural language processing}
}

OBJECTIVE: Clinicians pose complex clinical questions when seeing patients, and identifying the answers to those questions in a timely manner helps improve the quality of patient care. We report here on two natural language processing models, namely, automatic topic assignment and keyword identification, that together automatically and effectively extract information needs from ad hoc clinical questions. Our study is motivated in the context of developing the larger clinical question answering system AskHERMES (Help clinicians to Extract and aRrticulate Multimedia information for answering clinical quEstionS). DESIGN AND MEASUREMENTS: We developed supervised machine-learning systems to automatically assign predefined general categories (e.g. etiology, procedure, and diagnosis) to a question. We also explored both supervised and unsupervised systems to automatically identify keywords that capture the main content of the question. RESULTS: We evaluated our systems on 4654 annotated clinical questions that were collected in practice. We achieved an F1 score of 76.0% for the task of general topic classification and 58.0% for keyword extraction. Our systems have been implemented into the larger question answering system AskHERMES. Our error analyses suggested that inconsistent annotation in our training data have hurt both question analysis tasks. CONCLUSION: Our systems, available at http://www.askhermes.org, can automatically extract information needs from both short (the number of word tokens \textless20) and long questions (the number of word tokens \textgreater20), and from both well-structured and ill-formed questions. We speculate that the performance of general topic classification and keyword extraction can be further improved if consistently annotated data are made available.

@article{agarwal_biomedical_2010,
	title = {Biomedical negation scope detection with conditional random fields},
	volume = {17},
	issn = {1527-974X},
	url = {http://www.ncbi.nlm.nih.gov/pubmed/20962133},
	doi = {10.1136/jamia.2010.003228},
	abstract = {{\textless}AbstractText Label="OBJECTIVE" NlmCategory="OBJECTIVE"{\textgreater}Negation is a linguistic phenomenon that marks the absence of an entity or event. Negated events are frequently reported in both biological literature and clinical notes. Text mining applications benefit from the detection of negation and its scope. However, due to the complexity of language, identifying the scope of negation in a sentence is not a trivial task.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="DESIGN" NlmCategory="METHODS"{\textgreater}Conditional random fields (CRF), a supervised machine-learning algorithm, were used to train models to detect negation cue phrases and their scope in both biological literature and clinical notes. The models were trained on the publicly available BioScope corpus.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="MEASUREMENT" NlmCategory="METHODS"{\textgreater}The performance of the CRF models was evaluated on identifying the negation cue phrases and their scope by calculating recall, precision and F1-score. The models were compared with four competitive baseline systems.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="RESULTS" NlmCategory="RESULTS"{\textgreater}The best CRF-based model performed statistically better than all baseline systems and NegEx, achieving an F1-score of 98\% and 95\% on detecting negation cue phrases and their scope in clinical notes, and an F1-score of 97\% and 85\% on detecting negation cue phrases and their scope in biological literature.{\textless}/AbstractText{\textgreater}
{\textless}AbstractText Label="CONCLUSIONS" NlmCategory="CONCLUSIONS"{\textgreater}This approach is robust, as it can identify negation scope in both biological and clinical text. To benefit text mining applications, the system is publicly available as a Java API and as an online application at http://negscope.askhermes.org.{\textless}/AbstractText{\textgreater}},
	number = {6},
	urldate = {2011-03-25},
	journal = {Journal of the American Medical Informatics Association: JAMIA},
	author = {Agarwal, Shashank and Yu, Hong},
	month = nov,
	year = {2010},
	note = {00033},
	keywords = {Humans, natural language processing},
	pages = {696--701}
}

\textlessAbstractText Label="OBJECTIVE" NlmCategory="OBJECTIVE"\textgreaterNegation is a linguistic phenomenon that marks the absence of an entity or event. Negated events are frequently reported in both biological literature and clinical notes. Text mining applications benefit from the detection of negation and its scope. However, due to the complexity of language, identifying the scope of negation in a sentence is not a trivial task.\textless/AbstractText\textgreater \textlessAbstractText Label="DESIGN" NlmCategory="METHODS"\textgreaterConditional random fields (CRF), a supervised machine-learning algorithm, were used to train models to detect negation cue phrases and their scope in both biological literature and clinical notes. The models were trained on the publicly available BioScope corpus.\textless/AbstractText\textgreater \textlessAbstractText Label="MEASUREMENT" NlmCategory="METHODS"\textgreaterThe performance of the CRF models was evaluated on identifying the negation cue phrases and their scope by calculating recall, precision and F1-score. The models were compared with four competitive baseline systems.\textless/AbstractText\textgreater \textlessAbstractText Label="RESULTS" NlmCategory="RESULTS"\textgreaterThe best CRF-based model performed statistically better than all baseline systems and NegEx, achieving an F1-score of 98% and 95% on detecting negation cue phrases and their scope in clinical notes, and an F1-score of 97% and 85% on detecting negation cue phrases and their scope in biological literature.\textless/AbstractText\textgreater \textlessAbstractText Label="CONCLUSIONS" NlmCategory="CONCLUSIONS"\textgreaterThis approach is robust, as it can identify negation scope in both biological and clinical text. To benefit text mining applications, the system is publicly available as a Java API and as an online application at http://negscope.askhermes.org.\textless/AbstractText\textgreater

@article{agarwal_detecting_2010,
	title = {Detecting hedge cues and their scope in biomedical text with conditional random fields},
	volume = {43},
	issn = {1532-0480},
	doi = {10.1016/j.jbi.2010.08.003},
	abstract = {OBJECTIVE: Hedging is frequently used in both the biological literature and clinical notes to denote uncertainty or speculation. It is important for text-mining applications to detect hedge cues and their scope; otherwise, uncertain events are incorrectly identified as factual events. However, due to the complexity of language, identifying hedge cues and their scope in a sentence is not a trivial task. Our objective was to develop an algorithm that would automatically detect hedge cues and their scope in biomedical literature.
METHODOLOGY: We used conditional random fields (CRFs), a supervised machine-learning algorithm, to train models to detect hedge cue phrases and their scope in biomedical literature. The models were trained on the publicly available BioScope corpus. We evaluated the performance of the CRF models in identifying hedge cue phrases and their scope by calculating recall, precision and F1-score. We compared our models with three competitive baseline systems.
RESULTS: Our best CRF-based model performed statistically better than the baseline systems, achieving an F1-score of 88\% and 86\% in detecting hedge cue phrases and their scope in biological literature and an F1-score of 93\% and 90\% in detecting hedge cue phrases and their scope in clinical notes.
CONCLUSIONS: Our approach is robust, as it can identify hedge cues and their scope in both biological and clinical text. To benefit text-mining applications, our system is publicly available as a Java API and as an online application at http://hedgescope.askhermes.org. To our knowledge, this is the first publicly available system to detect hedge cues and their scope in biomedical literature.},
	language = {eng},
	number = {6},
	journal = {Journal of Biomedical Informatics},
	author = {Agarwal, Shashank and Yu, Hong},
	month = dec,
	year = {2010},
	pmid = {20709188},
	pmcid = {PMC2991497},
	keywords = {Algorithms, Artificial Intelligence, Data Mining, Natural Language Processing, Pattern Recognition, Automated, Vocabulary, Controlled},
	pages = {953--961}
}

OBJECTIVE: Hedging is frequently used in both the biological literature and clinical notes to denote uncertainty or speculation. It is important for text-mining applications to detect hedge cues and their scope; otherwise, uncertain events are incorrectly identified as factual events. However, due to the complexity of language, identifying hedge cues and their scope in a sentence is not a trivial task. Our objective was to develop an algorithm that would automatically detect hedge cues and their scope in biomedical literature. METHODOLOGY: We used conditional random fields (CRFs), a supervised machine-learning algorithm, to train models to detect hedge cue phrases and their scope in biomedical literature. The models were trained on the publicly available BioScope corpus. We evaluated the performance of the CRF models in identifying hedge cue phrases and their scope by calculating recall, precision and F1-score. We compared our models with three competitive baseline systems. RESULTS: Our best CRF-based model performed statistically better than the baseline systems, achieving an F1-score of 88% and 86% in detecting hedge cue phrases and their scope in biological literature and an F1-score of 93% and 90% in detecting hedge cue phrases and their scope in clinical notes. CONCLUSIONS: Our approach is robust, as it can identify hedge cues and their scope in both biological and clinical text. To benefit text-mining applications, our system is publicly available as a Java API and as an online application at http://hedgescope.askhermes.org. To our knowledge, this is the first publicly available system to detect hedge cues and their scope in biomedical literature.

@article{ramesh_identifying_2010,
	title = {Identifying discourse connectives in biomedical text},
	volume = {2010},
	issn = {1942-597X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3041460/},
	abstract = {Discourse connectives are words or phrases that connect or relate two coherent sentences or phrases and indicate the presence of discourse relations. Automatic recognition of discourse connectives may benefit many natural language processing applications. In this pilot study, we report the development of the supervised machine-learning classifiers with conditional random fields (CRFs) for automatically identifying discourse connectives in full-text biomedical articles. Our first classifier was trained on the open-domain 1 million token Penn Discourse Tree Bank (PDTB). We performed cross validation on biomedical articles (approximately 100K word tokens) that we annotated. The results show that the classifier trained on PDTB data attained a 0.55 F1-score for identifying discourse connectives in biomedical text, while the cross-validation results in the biomedical text attained a 0.69 F1-score, a much better performance despite a much smaller training size. Our preliminary analysis suggests the existence of domain-specific features, and we speculate that domain-adaption approaches may further improve performance.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Ramesh, Balaji Polepalli and Yu, Hong},
	month = nov,
	year = {2010},
	pmid = {21347060},
	pmcid = {PMC3041460},
	keywords = {Algorithms, Artificial Intelligence, Databases, Factual, Humans, Pilot Projects, Supervised Machine Learning, natural language processing},
	pages = {657--661}
}

Discourse connectives are words or phrases that connect or relate two coherent sentences or phrases and indicate the presence of discourse relations. Automatic recognition of discourse connectives may benefit many natural language processing applications. In this pilot study, we report the development of the supervised machine-learning classifiers with conditional random fields (CRFs) for automatically identifying discourse connectives in full-text biomedical articles. Our first classifier was trained on the open-domain 1 million token Penn Discourse Tree Bank (PDTB). We performed cross validation on biomedical articles (approximately 100K word tokens) that we annotated. The results show that the classifier trained on PDTB data attained a 0.55 F1-score for identifying discourse connectives in biomedical text, while the cross-validation results in the biomedical text attained a 0.69 F1-score, a much better performance despite a much smaller training size. Our preliminary analysis suggests the existence of domain-specific features, and we speculate that domain-adaption approaches may further improve performance.

@article{li_lancet:_2010,
	title = {Lancet: a high precision medication event extraction system for clinical text},
	volume = {17},
	issn = {1527-974X},
	shorttitle = {Lancet},
	url = {http://www.ncbi.nlm.nih.gov/pubmed/20819865},
	doi = {10.1136/jamia.2010.004077},
	abstract = {OBJECTIVE: This paper presents Lancet, a supervised machine-learning system that automatically extracts medication events consisting of medication names and information pertaining to their prescribed use (dosage, mode, frequency, duration and reason) from lists or narrative text in medical discharge summaries. DESIGN: Lancet incorporates three supervised machine-learning models: a conditional random fields model for tagging individual medication names and associated fields, an AdaBoost model with decision stump algorithm for determining which medication names and fields belong to a single medication event, and a support vector machines disambiguation model for identifying the context style (narrative or list). MEASUREMENTS: The authors, from the University of Wisconsin-Milwaukee, participated in the third i2b2 shared-task for challenges in natural language processing for clinical data: medication extraction challenge. With the performance metrics provided by the i2b2 challenge, the micro F1 (precision/recall) scores are reported for both the horizontal and vertical level. RESULTS: Among the top 10 teams, Lancet achieved the highest precision at 90.4\% with an overall F1 score of 76.4\% (horizontal system level with exact match), a gain of 11.2\% and 12\%, respectively, compared with the rule-based baseline system jMerki. By combining the two systems, the hybrid system further increased the F1 score by 3.4\% from 76.4\% to 79.0\%. CONCLUSIONS: Supervised machine-learning systems with minimal external knowledge resources can achieve a high precision with a competitive overall F1 score.Lancet based on this learning framework does not rely on expensive manually curated rules. The system is available online at http://code.google.com/p/lancet/.},
	number = {5},
	urldate = {2010-09-21},
	journal = {Journal of the American Medical Informatics Association: JAMIA},
	author = {Li, Zuofeng and Liu, Feifan and Antieau, Lamont and Cao, Yonggang and Yu, Hong},
	month = oct,
	year = {2010},
	pmid = {20819865},
	pages = {563--567}
}

OBJECTIVE: This paper presents Lancet, a supervised machine-learning system that automatically extracts medication events consisting of medication names and information pertaining to their prescribed use (dosage, mode, frequency, duration and reason) from lists or narrative text in medical discharge summaries. DESIGN: Lancet incorporates three supervised machine-learning models: a conditional random fields model for tagging individual medication names and associated fields, an AdaBoost model with decision stump algorithm for determining which medication names and fields belong to a single medication event, and a support vector machines disambiguation model for identifying the context style (narrative or list). MEASUREMENTS: The authors, from the University of Wisconsin-Milwaukee, participated in the third i2b2 shared-task for challenges in natural language processing for clinical data: medication extraction challenge. With the performance metrics provided by the i2b2 challenge, the micro F1 (precision/recall) scores are reported for both the horizontal and vertical level. RESULTS: Among the top 10 teams, Lancet achieved the highest precision at 90.4% with an overall F1 score of 76.4% (horizontal system level with exact match), a gain of 11.2% and 12%, respectively, compared with the rule-based baseline system jMerki. By combining the two systems, the hybrid system further increased the F1 score by 3.4% from 76.4% to 79.0%. CONCLUSIONS: Supervised machine-learning systems with minimal external knowledge resources can achieve a high precision with a competitive overall F1 score.Lancet based on this learning framework does not rely on expensive manually curated rules. The system is available online at http://code.google.com/p/lancet/.

@article{agarwal_automatically_2010,
	title = {Automatically classifying the role of citations in biomedical articles},
	volume = {2010},
	issn = {1942-597X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3041379/},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Agarwal, Shashank and Choubey, Lisha and Yu, Hong},
	month = nov,
	year = {2010},
	pmid = {21346931},
	pmcid = {PMC3041379},
	keywords = {Algorithms, Artificial Intelligence, Humans, Supervised Machine Learning},
	pages = {11--15}
}

@article{yu_are_2009,
	title = {Are figure legends sufficient? {Evaluating} the contribution of associated text to biomedical figure comprehension},
	volume = {4},
	issn = {1747-5333},
	shorttitle = {Are figure legends sufficient?},
	url = {http://www.j-biomed-discovery.com/content/4/1/1},
	doi = {10.1186/1747-5333-4-1},
	abstract = {BACKGROUND:Biomedical scientists need to access figures to validate research facts and to formulate or to test novel research hypotheses. However, figures are difficult to comprehend without associated text (e.g., figure legend and other reference text). We are developing automated systems to extract the relevant explanatory information along with figures extracted from full text articles. Such systems could be very useful in improving figure retrieval and in reducing the workload of biomedical scientists, who otherwise have to retrieve and read the entire full-text journal article to determine which figures are relevant to their research. As a crucial step, we studied the importance of associated text in biomedical figure comprehension.METHODS:Twenty subjects evaluated three figure-text combinations: figure+legend, figure+legend+title+abstract, and figure+full-text. Using a Likert scale, each subject scored each figure+text according to the extent to which the subject thought he/she understood the meaning of the figure and the confidence in providing the assigned score. Additionally, each subject entered a free text summary for each figure-text. We identified missing information using indicator words present within the text summaries. Both the Likert scores and the missing information were statistically analyzed for differences among the figure-text types. We also evaluated the quality of text summaries with the text-summarization evaluation method the ROUGE score.RESULTS:Our results showed statistically significant differences in figure comprehension when varying levels of text were provided. When the full-text article is not available, presenting just the figure+legend left biomedical researchers lacking 39-68\% of the information about a figure as compared to having complete figure comprehension; adding the title and abstract improved the situation, but still left biomedical researchers missing 30\% of the information. When the full-text article is available, figure comprehension increased to 86-97\%; this indicates that researchers felt that only 3-14\% of the necessary information for full figure comprehension was missing when full text was available to them. Clearly there is information in the abstract and in the full text that biomedical scientists deem important for understanding the figures that appear in full-text biomedical articles.CONCLUSION:We conclude that the texts that appear in full-text biomedical articles are useful for understanding the meaning of a figure, and an effective figure-mining system needs to unlock the information beyond figure legend. Our work provides important guidance to the figure mining systems that extract information only from figure and figure legend.},
	number = {1},
	urldate = {2009-03-03},
	journal = {Journal of Biomedical Discovery and Collaboration},
	author = {Yu, Hong and Agarwal, Shashank and Johnston, Mark and Cohen, Aaron},
	year = {2009},
	pages = {1}
}

BACKGROUND:Biomedical scientists need to access figures to validate research facts and to formulate or to test novel research hypotheses. However, figures are difficult to comprehend without associated text (e.g., figure legend and other reference text). We are developing automated systems to extract the relevant explanatory information along with figures extracted from full text articles. Such systems could be very useful in improving figure retrieval and in reducing the workload of biomedical scientists, who otherwise have to retrieve and read the entire full-text journal article to determine which figures are relevant to their research. As a crucial step, we studied the importance of associated text in biomedical figure comprehension.METHODS:Twenty subjects evaluated three figure-text combinations: figure+legend, figure+legend+title+abstract, and figure+full-text. Using a Likert scale, each subject scored each figure+text according to the extent to which the subject thought he/she understood the meaning of the figure and the confidence in providing the assigned score. Additionally, each subject entered a free text summary for each figure-text. We identified missing information using indicator words present within the text summaries. Both the Likert scores and the missing information were statistically analyzed for differences among the figure-text types. We also evaluated the quality of text summaries with the text-summarization evaluation method the ROUGE score.RESULTS:Our results showed statistically significant differences in figure comprehension when varying levels of text were provided. When the full-text article is not available, presenting just the figure+legend left biomedical researchers lacking 39-68% of the information about a figure as compared to having complete figure comprehension; adding the title and abstract improved the situation, but still left biomedical researchers missing 30% of the information. When the full-text article is available, figure comprehension increased to 86-97%; this indicates that researchers felt that only 3-14% of the necessary information for full figure comprehension was missing when full text was available to them. Clearly there is information in the abstract and in the full text that biomedical scientists deem important for understanding the figures that appear in full-text biomedical articles.CONCLUSION:We conclude that the texts that appear in full-text biomedical articles are useful for understanding the meaning of a figure, and an effective figure-mining system needs to unlock the information beyond figure legend. Our work provides important guidance to the figure mining systems that extract information only from figure and figure legend.

@article{agarwal_automatically_2009,
	title = {Automatically classifying sentences in full-text biomedical articles into {Introduction}, {Methods}, {Results} and {Discussion}},
	volume = {25},
	issn = {1367-4803, 1460-2059},
	url = {https://academic.oup.com/bioinformatics/article-lookup/doi/10.1093/bioinformatics/btp548},
	doi = {10.1093/bioinformatics/btp548},
	language = {en},
	number = {23},
	urldate = {2016-11-30},
	journal = {Bioinformatics},
	author = {Agarwal, S. and Yu, H.},
	month = dec,
	year = {2009},
	pages = {3174--3180}
}

@inproceedings{cao_evaluation_2009,
	title = {Evaluation of the clinical question answering presentation},
	isbn = {978-1-932432-30-5},
	url = {http://portal.acm.org/citation.cfm?doid=1572364.1572388},
	doi = {10.3115/1572364.1572388},
	language = {en},
	urldate = {2016-11-30},
	booktitle = {Proceedings of the {Workshop} on {Current} {Trends} in {Biomedical} {Natural} {Language} {Processing}},
	publisher = {Association for Computational Linguistics},
	author = {Cao, Yong-Gang and Ely, John and Antieau, Lamont and Yu, Hong},
	year = {2009},
	pages = {171}
}

@article{agarwal_figsum:_2009,
	title = {{FigSum}: automatically generating structured text summaries for figures in biomedical literature},
	volume = {2009},
	issn = {1942-597X},
	shorttitle = {{FigSum}},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2815407/},
	abstract = {Figures are frequently used in biomedical articles to support research findings; however, they are often difficult to comprehend based on their legends alone and information from the full-text articles is required to fully understand them. Previously, we found that the information associated with a single figure is distributed throughout the full-text article the figure appears in. Here, we develop and evaluate a figure summarization system - FigSum, which aggregates this scattered information to improve figure comprehension. For each figure in an article, FigSum generates a structured text summary comprising one sentence from each of the four rhetorical categories - Introduction, Methods, Results and Discussion (IMRaD). The IMRaD category of sentences is predicted by an automated machine learning classifier. Our evaluation shows that FigSum captures 53\% of the sentences in the gold standard summaries annotated by biomedical scientists and achieves an average ROUGE-1 score of 0.70, which is higher than a baseline system.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Agarwal, Shashank and Yu, Hong},
	month = nov,
	year = {2009},
	pmid = {20351812},
	pmcid = {PMC2815407},
	keywords = {Algorithms, Artificial Intelligence, Medical Illustration, Periodicals as Topic},
	pages = {6--10}
}

Figures are frequently used in biomedical articles to support research findings; however, they are often difficult to comprehend based on their legends alone and information from the full-text articles is required to fully understand them. Previously, we found that the information associated with a single figure is distributed throughout the full-text article the figure appears in. Here, we develop and evaluate a figure summarization system - FigSum, which aggregates this scattered information to improve figure comprehension. For each figure in an article, FigSum generates a structured text summary comprising one sentence from each of the four rhetorical categories - Introduction, Methods, Results and Discussion (IMRaD). The IMRaD category of sentences is predicted by an automated machine learning classifier. Our evaluation shows that FigSum captures 53% of the sentences in the gold standard summaries annotated by biomedical scientists and achieves an average ROUGE-1 score of 0.70, which is higher than a baseline system.

@article{kim_hierarchical_2009,
	title = {Hierarchical image classification in the bioscience literature},
	volume = {2009},
	issn = {1942-597X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2815366/},
	abstract = {Our previous work has shown that images appearing in bioscience articles can be classified into five types: Gel-Image, Image-of-Thing, Graph, Model, and Mix. For this paper, we explored and analyzed features strongly associated with each image type and developed a hierarchical image classification approach for classifying an image into one of the five types. First, we applied texture features to separate images into two groups: 1) a texture group comprising Gel Image, Image-of-Thing, and Mix, and 2) a non-texture group comprising Graph and Model. We then applied entropy, skewness, and uniformity for the first group, and edge difference, uniformity, and smoothness for the second group to classify images into specific types. Our results show that hierarchical image classification accurately divided images into the two groups during the initial classification and that the overall accuracy of the image classification was higher than that of our previous approach. In particular, the recall of hierarchical image classification was greatly improved due to the high accuracy of the initial classification.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Kim, Daehyun and Yu, Hong},
	month = nov,
	year = {2009},
	pmid = {20351874},
	pmcid = {PMC2815366},
	keywords = {Classification, Information Storage and Retrieval, Medical Illustration, Pattern Recognition, Automated},
	pages = {327--331}
}

Our previous work has shown that images appearing in bioscience articles can be classified into five types: Gel-Image, Image-of-Thing, Graph, Model, and Mix. For this paper, we explored and analyzed features strongly associated with each image type and developed a hierarchical image classification approach for classifying an image into one of the five types. First, we applied texture features to separate images into two groups: 1) a texture group comprising Gel Image, Image-of-Thing, and Mix, and 2) a non-texture group comprising Graph and Model. We then applied entropy, skewness, and uniformity for the first group, and edge difference, uniformity, and smoothness for the second group to classify images into specific types. Our results show that hierarchical image classification accurately divided images into the two groups during the initial classification and that the overall accuracy of the image classification was higher than that of our previous approach. In particular, the recall of hierarchical image classification was greatly improved due to the high accuracy of the initial classification.

@inproceedings{yu_investigating_2009,
	title = {Investigating and annotating the role of citation in biomedical full-text articles},
	isbn = {978-1-4244-5121-0},
	url = {http://ieeexplore.ieee.org/document/5332080/},
	doi = {10.1109/BIBMW.2009.5332080},
	urldate = {2016-11-30},
	booktitle = {Bioinformatics and {Biomedicine} {Workshop}},
	publisher = {IEEE},
	author = {Yu, Hong and Agarwal, Shashank and Frid, Nadya},
	month = nov,
	year = {2009},
	pages = {308--313}
}

@article{yu_using_2009,
	title = {Using the {Weighted} {Keyword} {Model} to {Improve} {Information} {Retrieval} for {Answering} {Biomedical} {Questions}},
	volume = {2009},
	journal = {Summit on translational bioinformatics},
	author = {Yu, Hong and Cao, Yong-Gang},
	year = {2009},
	pages = {143}
}

@inproceedings{cao_evaluating_2009,
	title = {Evaluating the weighted-keyword model to improve clinical question answering},
	isbn = {978-1-4244-5121-0},
	url = {http://ieeexplore.ieee.org/document/5332084/},
	doi = {10.1109/BIBMW.2009.5332084},
	urldate = {2016-11-30},
	booktitle = {Bioinformatics and {Biomedicine} {Workshop}},
	publisher = {IEEE},
	author = {Cao, Yong-Gang and Ely, John and Yu, Hong},
	month = nov,
	year = {2009},
	pages = {331--335}
}

@article{yu_survey_2009,
	title = {A {Survey} of {Artificial} {Immune} {Theory}-based {Computer} {Defense} {Technology}},
	volume = {3},
	journal = {Journal of University of South China (Science and Technology)},
	author = {YU, Hong-xiao and TAN, Min-sheng},
	year = {2009},
	pages = {011}
}

@inproceedings{yu_pilot_2008,
	title = {A pilot annotation to investigate discourse connectivity in biomedical text},
	url = {https://www.aclweb.org/anthology/W/W08/W08-0614.pdf},
	booktitle = {Proceedings of the {Workshop} on {Current} {Trends} in {Biomedical} {Natural} {Language} {Processing}},
	publisher = {Association for Computational Linguistics},
	author = {Yu, Hong and Frid, Nadya and McRoy, Susan and Prasad, Rashmi and Lee, Alan and Joshi, Aravind},
	year = {2008},
	pages = {92--93}
}

@article{yu_automatically_2008,
	title = {Automatically extracting information needs from {Ad} {Hoc} clinical questions},
	issn = {1942-597X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2655957/},
	abstract = {Automatically extracting information needs from ad hoc clinical questions is an important step towards medical question answering. In this work, we first explored supervised machine-learning approaches to automatically classify an ad hoc clinical question into general topics. We then evaluated different methods for automatically extracting keywords from an ad hoc clinical question. Our methods were evaluated on the 4,654 clinical questions maintained by the National Library of Medicine. Our best systems or methods showed F-score of 76\% for the task of question-topic classification and an average F-score of 56\% for extracting keywords from ad hoc clinical questions.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Yu, Hong and Cao, Yong-Gang},
	month = nov,
	year = {2008},
	pmid = {18999100},
	pmcid = {PMC2655957},
	keywords = {Algorithms, Artificial Intelligence, Communication, Decision Support Systems, Clinical, Information Dissemination, Internet, Pattern Recognition, Automated, Point-of-Care Systems, Remote Consultation, User-Computer Interface, natural language processing},
	pages = {96--100}
}

Automatically extracting information needs from ad hoc clinical questions is an important step towards medical question answering. In this work, we first explored supervised machine-learning approaches to automatically classify an ad hoc clinical question into general topics. We then evaluated different methods for automatically extracting keywords from an ad hoc clinical question. Our methods were evaluated on the 4,654 clinical questions maintained by the National Library of Medicine. Our best systems or methods showed F-score of 76% for the task of question-topic classification and an average F-score of 56% for extracting keywords from ad hoc clinical questions.

@inproceedings{cohen_translating_2008,
	title = {Translating biology: text mining tools that work},
	volume = {13},
	url = {http://psb.stanford.edu/psb-online/proceedings/psb08/textmining.pdf},
	booktitle = {Pacific {Symposium} on {Biocomputing}. {Pacific} {Symposium} on {Biocomputing}},
	author = {Cohen, K Bretonnel and Yu, Hong and Bourne, Philip E and Hirschman, Lynette},
	year = {2008},
	pages = {551}
}

@article{yu_cognitive_2007,
	title = {A cognitive evaluation of four online search engines for answering definitional questions posed by physicians},
	issn = {2335-6936},
	url = {http://psb.stanford.edu/psb-online/proceedings/psb07/yu.pdf},
	abstract = {The Internet is having a profound impact on physicians' medical decision making. One recent survey of 277 physicians showed that 72\% of physicians regularly used the Internet to research medical information and 51\% admitted that information from web sites influenced their clinical decisions. This paper describes the first cognitive evaluation of four state-of-the-art Internet search engines: Google (i.e., Google and Scholar.Google), MedQA, Onelook, and PubMed for answering definitional questions (i.e., questions with the format of "What is X?") posed by physicians. Onelook is a portal for online definitions, and MedQA is a question answering system that automatically generates short texts to answer specific biomedical questions. Our evaluation criteria include quality of answer, ease of use, time spent, and number of actions taken. Our results show that MedQA outperforms Onelook and PubMed in most of the criteria, and that MedQA surpasses Google in time spent and number of actions, two important efficiency criteria. Our results show that Google is the best system for quality of answer and ease of use. We conclude that Google is an effective search engine for medical definitions, and that MedQA exceeds the other search engines in that it provides users direct answers to their questions; while the users of the other search engines have to visit several sites before finding all of the pertinent information.},
	language = {ENG},
	journal = {Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing},
	author = {Yu, Hong and Kaufman, David},
	year = {2007},
	pmid = {17990503},
	keywords = {Adult, Computational Biology, Female, Humans, Internet, Male, Middle Aged, Physicians, PubMed},
	pages = {328--339}
}

The Internet is having a profound impact on physicians' medical decision making. One recent survey of 277 physicians showed that 72% of physicians regularly used the Internet to research medical information and 51% admitted that information from web sites influenced their clinical decisions. This paper describes the first cognitive evaluation of four state-of-the-art Internet search engines: Google (i.e., Google and Scholar.Google), MedQA, Onelook, and PubMed for answering definitional questions (i.e., questions with the format of "What is X?") posed by physicians. Onelook is a portal for online definitions, and MedQA is a question answering system that automatically generates short texts to answer specific biomedical questions. Our evaluation criteria include quality of answer, ease of use, time spent, and number of actions taken. Our results show that MedQA outperforms Onelook and PubMed in most of the criteria, and that MedQA surpasses Google in time spent and number of actions, two important efficiency criteria. Our results show that Google is the best system for quality of answer and ease of use. We conclude that Google is an effective search engine for medical definitions, and that MedQA exceeds the other search engines in that it provides users direct answers to their questions; while the users of the other search engines have to visit several sites before finding all of the pertinent information.

@article{yu_development_2007,
	title = {Development, implementation, and a cognitive evaluation of a definitional question answering system for physicians},
	volume = {40},
	issn = {15320464},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046407000202},
	doi = {10.1016/j.jbi.2007.03.002},
	language = {en},
	number = {3},
	urldate = {2016-11-30},
	journal = {Journal of Biomedical Informatics},
	author = {Yu, Hong and Lee, Minsuk and Kaufman, David and Ely, John and Osheroff, Jerome A. and Hripcsak, George and Cimino, James},
	month = jun,
	year = {2007},
	keywords = {Algorithms Attitude of Health Personnel Attitude to Computers *Cognition Databases, Bibliographic Databases, Factual *Decision Support Techniques Humans Information Storage and Retrieval Information Systems Internet Logical Observation Identifiers Names and Codes Online Systems *Physicians PubMed Research Design Software},
	pages = {236--251}
}

@article{zweigenbaum_frontiers_2007,
	title = {Frontiers of biomedical text mining: current progress},
	volume = {8},
	issn = {1477-4054},
	shorttitle = {Frontiers of biomedical text mining},
	url = {http://www.ncbi.nlm.nih.gov/pubmed/17977867},
	doi = {10.1093/bib/bbm045},
	abstract = {It is now almost 15 years since the publication of the first paper on text mining in the genomics domain, and decades since the first paper on text mining in the medical domain. Enormous progress has been made in the areas of information retrieval, evaluation methodologies and resource construction. Some problems, such as abbreviation-handling, can essentially be considered solved problems, and others, such as identification of gene mentions in text, seem likely to be solved soon. However, a number of problems at the frontiers of biomedical text mining continue to present interesting challenges and opportunities for great improvements and interesting research. In this article we review the current state of the art in biomedical text mining or 'BioNLP' in general, focusing primarily on papers published within the past year.},
	number = {5},
	urldate = {2009-03-03},
	journal = {Briefings in Bioinformatics},
	author = {Zweigenbaum, Pierre and Demner-Fushman, Dina and Yu, Hong and Cohen, Kevin B},
	month = sep,
	year = {2007},
	pmid = {17977867},
	keywords = {Abstracting and Indexing as Topic, Biology, Databases, Bibliographic, Forecasting, Periodicals as Topic, Vocabulary, Controlled},
	pages = {358--75}
}

It is now almost 15 years since the publication of the first paper on text mining in the genomics domain, and decades since the first paper on text mining in the medical domain. Enormous progress has been made in the areas of information retrieval, evaluation methodologies and resource construction. Some problems, such as abbreviation-handling, can essentially be considered solved problems, and others, such as identification of gene mentions in text, seem likely to be solved soon. However, a number of problems at the frontiers of biomedical text mining continue to present interesting challenges and opportunities for great improvements and interesting research. In this article we review the current state of the art in biomedical text mining or 'BioNLP' in general, focusing primarily on papers published within the past year.

@article{yu_using_2007,
	title = {Using {MEDLINE} as a knowledge source for disambiguating abbreviations and acronyms in full-text biomedical journal articles},
	volume = {40},
	issn = {15320464},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046406000621},
	doi = {10.1016/j.jbi.2006.06.001},
	language = {en},
	number = {2},
	urldate = {2016-11-30},
	journal = {Journal of Biomedical Informatics},
	author = {Yu, Hong and Kim, Won and Hatzivassiloglou, Vasileios and Wilbur, W. John},
	month = apr,
	year = {2007},
	keywords = {*Artificial Intelligence Database Management Systems Information Storage and Retrieval/*methods *Medline *Natural Language Processing Pattern Recognition, Automated/*methods *Periodicals *Terminology},
	pages = {150--159}
}

@article{lassen_efficacy_2007,
	title = {The efficacy and safety of apixaban, an oral, direct factor {Xa} inhibitor, as thromboprophylaxis in patients following total knee replacement},
	volume = {5},
	issn = {15387933, 15387836},
	url = {http://doi.wiley.com/10.1111/j.1538-7836.2007.02764.x},
	doi = {10.1111/j.1538-7836.2007.02764.x},
	language = {en},
	number = {12},
	urldate = {2016-11-30},
	journal = {Journal of Thrombosis and Haemostasis},
	author = {Lassen, M. R. and Davidson, B. L. and Gallus, A. and Pineo, G. and Ansell, J. and Deitchman, D.},
	month = dec,
	year = {2007},
	pages = {2368--2375}
}

@article{yu_towards_2006,
	title = {Towards answering biological questions with experimental evidence: automatically identifying text that summarize image content in full-text articles},
	issn = {1942-597X},
	shorttitle = {Towards answering biological questions with experimental evidence},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1839512/},
	abstract = {Images (i.e., figures) are important experimental evidence that are typically reported in bioscience full-text articles. Biologists need to access images to validate research facts and to formulate or to test novel research hypotheses. We propose to build a biological question answering system that provides experimental evidences as answers in response to biological questions. As a first step, we develop natural language processing techniques to identify sentences that summarize image content.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Yu, Hong},
	year = {2006},
	pmid = {17238458},
	pmcid = {PMC1839512},
	keywords = {Algorithms, Biological Science Disciplines, Information Storage and Retrieval, Medical Illustration, natural language processing},
	pages = {834--838}
}

Images (i.e., figures) are important experimental evidence that are typically reported in bioscience full-text articles. Biologists need to access images to validate research facts and to formulate or to test novel research hypotheses. We propose to build a biological question answering system that provides experimental evidences as answers in response to biological questions. As a first step, we develop natural language processing techniques to identify sentences that summarize image content.

@article{yu_accessing_2006,
	title = {Accessing bioscience images from abstract sentences},
	volume = {22},
	issn = {1367-4803, 1460-2059},
	url = {http://bioinformatics.oxfordjournals.org/content/22/14/e547},
	doi = {10.1093/bioinformatics/btl261},
	abstract = {Images (e.g., figures) are important experimental results that are typically reported in bioscience full-text articles. Biologists need to access images to validate research facts and to formulate or to test novel research hypotheses. On the other hand, biologists live in an age of information explosion. As thousands of biomedical articles are published every day, systems that help biologists efficiently access images in literature would greatly facilitate biomedical research. We hypothesize that much of image content reported in a full-text article can be summarized by the sentences in the abstract of the article. In our study, more than one hundred biologists had tested this hypothesis and more than 40 biologists had evaluated a novel user-interface BioEx that allows biologists to access images directly from abstract sentences. Our results show that 87.8\% biologists were in favor of BioEx over two other baseline user-interfaces. We further developed systems that explored hierarchical clustering algorithms to automatically identify abstract sentences that summarize the images. One of the systems achieves a precision of 100\% that corresponds to a recall of 4.6\%.
Contact:hongyu@uwm.edu or hy52@columbia.edu},
	language = {en},
	number = {14},
	urldate = {2013-12-31},
	journal = {Bioinformatics},
	author = {Yu, Hong and Lee, Minsuk},
	month = jul,
	year = {2006},
	note = {00039 PMID: 16873519},
	pages = {e547--e556}
}

Images (e.g., figures) are important experimental results that are typically reported in bioscience full-text articles. Biologists need to access images to validate research facts and to formulate or to test novel research hypotheses. On the other hand, biologists live in an age of information explosion. As thousands of biomedical articles are published every day, systems that help biologists efficiently access images in literature would greatly facilitate biomedical research. We hypothesize that much of image content reported in a full-text article can be summarized by the sentences in the abstract of the article. In our study, more than one hundred biologists had tested this hypothesis and more than 40 biologists had evaluated a novel user-interface BioEx that allows biologists to access images directly from abstract sentences. Our results show that 87.8% biologists were in favor of BioEx over two other baseline user-interfaces. We further developed systems that explored hierarchical clustering algorithms to automatically identify abstract sentences that summarize the images. One of the systems achieves a precision of 100% that corresponds to a recall of 4.6%. Contact:hongyu@uwm.edu or hy52@columbia.edu

@article{lee_beyond_2006,
	title = {Beyond information retrieval--medical question answering},
	issn = {1942-597X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1839371/},
	abstract = {Physicians have many questions when caring for patients, and frequently need to seek answers for their questions. Information retrieval systems (e.g., PubMed) typically return a list of documents in response to a user's query. Frequently the number of returned documents is large and makes physicians' information seeking "practical only 'after hours' and not in the clinical settings". Question answering techniques are based on automatically analyzing thousands of electronic documents to generate short-text answers in response to clinical questions that are posed by physicians. The authors address physicians' information needs and described the design, implementation, and evaluation of the medical question answering system (MedQA). Although our long term goal is to enable MedQA to answer all types of medical questions, currently, we implemented MedQA to integrate information retrieval, extraction, and summarization techniques to automatically generate paragraph-level text for definitional questions (i.e., "What is X?"). MedQA can be accessed at http://www.dbmi.columbia.edu/{\textasciitilde}yuh9001/research/MedQA.html.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Lee, Minsuk and Cimino, James and Zhu, Hai R. and Sable, Carl and Shanker, Vijay and Ely, John and Yu, Hong},
	year = {2006},
	pmid = {17238385},
	pmcid = {PMC1839371},
	keywords = {Decision Support Techniques, Humans, Information Storage and Retrieval, Internet, MEDLINE, Physicians, Pilot Projects, expert systems, natural language processing},
	pages = {469--473}
}

Physicians have many questions when caring for patients, and frequently need to seek answers for their questions. Information retrieval systems (e.g., PubMed) typically return a list of documents in response to a user's query. Frequently the number of returned documents is large and makes physicians' information seeking "practical only 'after hours' and not in the clinical settings". Question answering techniques are based on automatically analyzing thousands of electronic documents to generate short-text answers in response to clinical questions that are posed by physicians. The authors address physicians' information needs and described the design, implementation, and evaluation of the medical question answering system (MedQA). Although our long term goal is to enable MedQA to answer all types of medical questions, currently, we implemented MedQA to integrate information retrieval, extraction, and summarization techniques to automatically generate paragraph-level text for definitional questions (i.e., "What is X?"). MedQA can be accessed at http://www.dbmi.columbia.edu/\textasciitildeyuh9001/research/MedQA.html.

@inproceedings{yu_bioex:_2006,
	title = {{BioEx}: a novel user-interface that accesses images from abstract sentences},
	url = {http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=EC9B0313E8DBD1F5C07A06DA15ED2944?doi=10.1.1.95.5862&rep=rep1&type=pdf},
	booktitle = {Proceedings of the {Human} {Language} {Technology} {Conference} of the {NAACL}, {Companion} {Volume}: {Short} {Papers}},
	publisher = {Association for Computational Linguistics},
	author = {Yu, Hong and Lee, Minsuk},
	year = {2006},
	pages = {189--192}
}

@article{lee_exploring_2006,
	title = {Exploring supervised and unsupervised methods to detect topics in biomedical text},
	volume = {7},
	issn = {1471-2105},
	url = {http://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-7-140},
	doi = {10.1186/1471-2105-7-140},
	abstract = {BACKGROUND: Topic detection is a task that automatically identifies topics (e.g., "biochemistry" and "protein structure") in scientific articles based on information content. Topic detection will benefit many other natural language processing tasks including information retrieval, text summarization and question answering; and is a necessary step towards the building of an information system that provides an efficient way for biologists to seek information from an ocean of literature.
RESULTS: We have explored the methods of Topic Spotting, a task of text categorization that applies the supervised machine-learning technique naïve Bayes to assign automatically a document into one or more predefined topics; and Topic Clustering, which apply unsupervised hierarchical clustering algorithms to aggregate documents into clusters such that each cluster represents a topic. We have applied our methods to detect topics of more than fifteen thousand of articles that represent over sixteen thousand entries in the Online Mendelian Inheritance in Man (OMIM) database. We have explored bag of words as the features. Additionally, we have explored semantic features; namely, the Medical Subject Headings (MeSH) that are assigned to the MEDLINE records, and the Unified Medical Language System (UMLS) semantic types that correspond to the MeSH terms, in addition to bag of words, to facilitate the tasks of topic detection. Our results indicate that incorporating the MeSH terms and the UMLS semantic types as additional features enhances the performance of topic detection and the naïve Bayes has the highest accuracy, 66.4\%, for predicting the topic of an OMIM article as one of the total twenty-five topics.
CONCLUSION: Our results indicate that the supervised topic spotting methods outperformed the unsupervised topic clustering; on the other hand, the unsupervised topic clustering methods have the advantages of being robust and applicable in real world settings.},
	language = {ENG},
	journal = {BMC bioinformatics},
	author = {Lee, Minsuk and Wang, Weiqing and Yu, Hong},
	month = mar,
	year = {2006},
	pmid = {16539745},
	pmcid = {PMC1472693},
	keywords = {Abstracting and Indexing as Topic, Abstracting and Indexing/methods *Artificial Intelligence Humans *Medline *Natural Language Processing Pattern Recognition, Artificial Intelligence, Automated/*methods *Periodicals *Terminology *Vocabulary, Controlled, Humans, MEDLINE, Pattern Recognition, Automated, Periodicals as Topic, Terminology as Topic, Vocabulary, Controlled, natural language processing},
	pages = {140}
}

BACKGROUND: Topic detection is a task that automatically identifies topics (e.g., "biochemistry" and "protein structure") in scientific articles based on information content. Topic detection will benefit many other natural language processing tasks including information retrieval, text summarization and question answering; and is a necessary step towards the building of an information system that provides an efficient way for biologists to seek information from an ocean of literature. RESULTS: We have explored the methods of Topic Spotting, a task of text categorization that applies the supervised machine-learning technique naïve Bayes to assign automatically a document into one or more predefined topics; and Topic Clustering, which apply unsupervised hierarchical clustering algorithms to aggregate documents into clusters such that each cluster represents a topic. We have applied our methods to detect topics of more than fifteen thousand of articles that represent over sixteen thousand entries in the Online Mendelian Inheritance in Man (OMIM) database. We have explored bag of words as the features. Additionally, we have explored semantic features; namely, the Medical Subject Headings (MeSH) that are assigned to the MEDLINE records, and the Unified Medical Language System (UMLS) semantic types that correspond to the MeSH terms, in addition to bag of words, to facilitate the tasks of topic detection. Our results indicate that incorporating the MeSH terms and the UMLS semantic types as additional features enhances the performance of topic detection and the naïve Bayes has the highest accuracy, 66.4%, for predicting the topic of an OMIM article as one of the total twenty-five topics. CONCLUSION: Our results indicate that the supervised topic spotting methods outperformed the unsupervised topic clustering; on the other hand, the unsupervised topic clustering methods have the advantages of being robust and applicable in real world settings.

@inproceedings{rafkind_exploring_2006,
	title = {Exploring text and image features to classify images in bioscience literature},
	url = {http://portal.acm.org/citation.cfm?doid=1654415.1654428},
	doi = {10.3115/1654415.1654428},
	language = {en},
	urldate = {2016-11-30},
	booktitle = {Proceedings of the {HLT}-{NAACL} {BioNLP} {Workshop} on {Linking} {Natural} {Language} and {Biology}},
	publisher = {Association for Computational Linguistics},
	author = {Rafkind, Barry and Lee, Minsuk and Chang, Shih-Fu and Yu, Hong},
	year = {2006},
	pages = {73}
}

@inproceedings{yu_semantics_2006,
	address = {New York, USA},
	title = {The semantics of a definiendum constrains both the lexical semantics and the lexicosyntactic patterns in the definiens},
	booktitle = {Proceedings of the {BioNLP} {Workshop} on {Linking} {Natural} {Language} {Processing} and {Biology} at {HLT}-{NAACL}},
	author = {Yu, H. and Wei, Y.},
	year = {2006},
	pages = {1--8}
}

@book{yu_system_2006,
	title = {System and methods for automatically identifying answerable questions},
	author = {Yu, Hong},
	month = jun,
	year = {2006},
	note = {00012 
US Patent App. 11/479,645}
}

@article{sable_question_2005,
	title = {Question analysis for biomedical question answering},
	issn = {1942-597X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1560855/},
	abstract = {We are developing a biomedical question answering system. This paper describes our system's architecture and our question analysis component. Specifically, we have explored the use of various supervised machine learning approaches to filter out unanswerable questions based on physicians' annotations.},
	language = {ENG},
	journal = {AMIA ... Annual Symposium proceedings. AMIA Symposium},
	author = {Sable, Carl and Lee, Minsuk and Zhu, Hai Ran and Yu, Hong},
	year = {2005},
	pmid = {16779389},
	pmcid = {PMC1560855},
	keywords = {Artificial Intelligence, Information Storage and Retrieval, medical informatics applications},
	pages = {1102}
}

We are developing a biomedical question answering system. This paper describes our system's architecture and our question analysis component. Specifically, we have explored the use of various supervised machine learning approaches to filter out unanswerable questions based on physicians' annotations.

@article{rzhetsky_geneways:_2004,
	title = {{GeneWays}: a system for extracting, analyzing, visualizing, and integrating molecular pathway data},
	volume = {37},
	issn = {1532-0464},
	shorttitle = {{GeneWays}},
	doi = {10.1016/j.jbi.2003.10.001},
	abstract = {The immense growth in the volume of research literature and experimental data in the field of molecular biology calls for efficient automatic methods to capture and store information. In recent years, several groups have worked on specific problems in this area, such as automated selection of articles pertinent to molecular biology, or automated extraction of information using natural-language processing, information visualization, and generation of specialized knowledge bases for molecular biology. GeneWays is an integrated system that combines several such subtasks. It analyzes interactions between molecular substances, drawing on multiple sources of information to infer a consensus view of molecular networks. GeneWays is designed as an open platform, allowing researchers to query, review, and critique stored information.},
	language = {eng},
	number = {1},
	journal = {Journal of Biomedical Informatics},
	author = {Rzhetsky, Andrey and Iossifov, Ivan and Koike, Tomohiro and Krauthammer, Michael and Kra, Pauline and Morris, Mitzi and Yu, Hong and Duboué, Pablo Ariel and Weng, Wubin and Wilbur, W. John and Hatzivassiloglou, Vasileios and Friedman, Carol},
	month = feb,
	year = {2004},
	pmid = {15016385},
	note = {00251 },
	keywords = {*Artificial Intelligence Comparative Study Computer Graphics Database Management Systems Databases, Artificial Intelligence, Bioinformatics, Computer Graphics, Controlled, Database, Database Management Systems, Databases, Factual, Documentation, Factual Documentation/methods Gene Expression Regulation/physiology Information Storage and Retrieval/*methods Internet Metabolism/*physiology *Natural Language Processing *Periodicals Research Support, Gene Expression Regulation, Information Storage and Retrieval, Information extraction, Internet, Knowledge engineering, Metabolism, Molecular interactions, Molecular networks, Non-P.H.S. Research Support, P.H.S. Signal Transduction/physiology *Software *User-Computer Interface Vocabulary, Periodicals as Topic, Signal Transduction, Software, Text mining, U.S. Gov't, User-Computer Interface, Vocabulary, Controlled, natural language processing},
	pages = {43--53}
}

The immense growth in the volume of research literature and experimental data in the field of molecular biology calls for efficient automatic methods to capture and store information. In recent years, several groups have worked on specific problems in this area, such as automated selection of articles pertinent to molecular biology, or automated extraction of information using natural-language processing, information visualization, and generation of specialized knowledge bases for molecular biology. GeneWays is an integrated system that combines several such subtasks. It analyzes interactions between molecular substances, drawing on multiple sources of information to infer a consensus view of molecular networks. GeneWays is designed as an open platform, allowing researchers to query, review, and critique stored information.

@inproceedings{yu_using_2004,
	title = {Using {MEDLINE} as a knowledge source for disambiguating abbreviations in full-text biomedical journal articles},
	booktitle = {Computer-{Based} {Medical} {Systems}, 2004. {CBMS} 2004. {Proceedings}. 17th {IEEE} {Symposium} on},
	publisher = {IEEE},
	author = {Yu, Hong and Kim, Won and Hatzivassiloglou, Vasileios and John Wilbur, W},
	year = {2004},
	pages = {27--32}
}

@article{bethard_automatic_2004,
	title = {Automatic extraction of opinion propositions and their holders},
	url = {http://web.stanford.edu/~jurafsky/SS404BethardS.pdf},
	journal = {2004 AAAI Spring Symposium on Exploring Attitude and Affect in Text},
	author = {Bethard, Steven and Yu, Hong and Thornton, Ashley and Hatzivassiloglou, Vasileios and Jurafsky, Dan},
	year = {2004},
	pages = {2224}
}

@book{yu_methods_2004,
	title = {Methods and systems for extracting synonymous gene and protein terms from biological literature},
	abstract = {00003 
US Patent App. 10/915,168},
	author = {Yu, Hong and Rzhetsky, Andrey},
	month = aug,
	year = {2004}
}

00003 US Patent App. 10/915,168

@article{yu_extracting_2003,
	title = {Extracting synonymous gene and protein terms from biological literature},
	volume = {19},
	issn = {1367-4803, 1460-2059},
	url = {https://academic.oup.com/bioinformatics/article-lookup/doi/10.1093/bioinformatics/btg1047},
	doi = {10.1093/bioinformatics/btg1047},
	language = {en},
	number = {Suppl 1},
	urldate = {2016-11-30},
	journal = {Bioinformatics},
	author = {Yu, H. and Agichtein, E.},
	month = jul,
	year = {2003},
	keywords = {Abstracting and Indexing/*methods/standards Acetaminophen Algorithms Biology/methods/standards Computational Biology/methods/standards Database Management Systems *Databases, Bibliographic Documentation *Genes Information Storage and Retrieval/methods/standards *Natural Language Processing *Periodicals *Proteins Research Support, Controlled, Non-P.H.S. *Terminology Vocabulary, U.S. Gov't},
	pages = {i340--i349}
}