@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{liu_learning_2019,
	title = {Learning to detect and understand drug discontinuation events from clinical narratives},
	url = {https://academic.oup.com/jamia/advance-article/doi/10.1093/jamia/ocz048/5481540},
	doi = {10.1093/jamia/ocz048},
	abstract = {AbstractObjective.  Identifying drug discontinuation (DDC) events and understanding their reasons are important for medication management and drug safety survei},
	language = {en},
	urldate = {2019-04-29},
	journal = {Journal of the American Medical Informatics Association},
	author = {Liu, Feifan and Pradhan, Richeek and Druhl, Emily and Freund, Elaine and Liu, Weisong and Sauer, Brian C. and Cunningham, Fran and Gordon, Adam J. and Peters, Celena B. and Yu, Hong},
	year = {2019}
}

AbstractObjective. Identifying drug discontinuation (DDC) events and understanding their reasons are important for medication management and drug safety survei

@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}
}

@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]

@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{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{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{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}
}

@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{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{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.

@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.

@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.

@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.

@inproceedings{chen_detecting_2018,
	title = {Detecting {Hypoglycemia} {Incidence} from {Patients}’ {Secure} {Messages}},
	author = {Chen, Jinying and Lalor, John and Yu, Hong},
	year = {2018}
}

@article{pradhan_automatic_2018,
	title = {Automatic extraction of quantitative data from {ClinicalTrials}.gov to conduct meta-analyses},
	volume = {0},
	issn = {0895-4356, 1878-5921},
	url = {https://www.jclinepi.com/article/S0895-4356(17)31306-9/fulltext},
	doi = {10.1016/j.jclinepi.2018.08.023},
	abstract = {Objective
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 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
We developed a Python-based software application (EXACT, Extracting Accurate efficacy and safety information from ClinicalTrials.gov) 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 statistics 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 = {English},
	number = {0},
	urldate = {2018-10-08},
	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 = sep,
	year = {2018},
	pmid = {30257185},
	keywords = {Automatic data extraction, ClinicalTrials.gov, Meta-analysis, Reproducibility}
}

Objective 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 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 We developed a Python-based software application (EXACT, Extracting Accurate efficacy and safety information from ClinicalTrials.gov) 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 statistics 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.

@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_understanding_2018-1,
	title = {Understanding {Deep} {Learning} {Performance} through an {Examination} of {Test} {Set} {Difficulty}: {A} {Psychometric} {Case} {Study}},
	booktitle = {{UAI} {Workshop} on {Uncertainty} in {Deep} {Learning}},
	author = {Lalor, John and Wu, Hao and Munkhdalai, T and Yu, Hong},
	year = {2018}
}

@article{lalor_comprehenotes_2018-1,
	title = {{ComprehENotes}, an {Instrument} to {Assess} {Patient} {Reading} {Comprehension} of {Electronic} {Health} {Record} {Notes}: {Development} and {Validation}},
	volume = {20},
	shorttitle = {{ComprehENotes}, an {Instrument} to {Assess} {Patient} {Reading} {Comprehension} of {Electronic} {Health} {Record} {Notes}},
	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},
	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}
}

@article{munkhdalai_clinical_2018-1,
	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{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{pradhan_inadequate_2018-1,
	title = {Inadequate diversity of information resources searched in {USA}-affiliated systematic reviews and meta-analyses: 2005-2016},
	issn = {1878-5921},
	shorttitle = {Inadequate diversity of information resources searched in {USA}-affiliated systematic reviews and meta-analyses},
	doi = {10.1016/j.jclinepi.2018.05.024},
	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 = jun,
	year = {2018},
	pmid = {29879464}
}

@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}
}

@article{chen_natural_2018-1,
	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}},
	url = {http://www.jmir.org/2018/1/e26/},
	doi = {10.2196/jmir.8669},
	language = {en},
	number = {1},
	urldate = {2018-02-26},
	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},
	pages = {e26}
}

@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 {Ability} in {Deep} {Neural} {Networks}},
	journal = {arXiv preprint arXiv:1702.04811},
	author = {Lalor, John P. and Wu, Hao and Munkhdalai, Tsendsuren and Yu, Hong},
	year = {2017}
}

@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}
}

@article{zheng_readability_2017,
	title = {Readability {Formulas} and {User} {Perceptions} of {Electronic} {Health} {Records} {Difficulty}: {A} {Corpus} {Study}},
	volume = {19},
	issn = {1438-8871},
	shorttitle = {Readability {Formulas} and {User} {Perceptions} of {Electronic} {Health} {Records} {Difficulty}},
	url = {http://www.jmir.org/2017/3/e59/},
	doi = {10.2196/jmir.6962},
	language = {en},
	number = {3},
	urldate = {2018-02-26},
	journal = {Journal of Medical Internet Research},
	author = {Zheng, Jiaping and Yu, Hong},
	month = mar,
	year = {2017},
	pages = {e59}
}

@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{lalor_analysis_2017-1,
	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{zheng_readability_2017-1,
	title = {Readability {Formulas} and {User} {Perceptions} of {Electronic} {Health} {Records} {Difficulty}: {A} {Corpus} {Study}},
	journal = {Journal of Medical Internet Research},
	author = {Zheng, JP and Yu, H.},
	year = {2017}
}

@inproceedings{munkhdalai_neural_2017,
	title = {Neural  tree  indexers for  text  understanding},
	url = {https://arxiv.org/pdf/1607.04492v1.pdf},
	booktitle = {European {Chapter} of the {Association} for {Computational} {Linguistics} 2017 ({EACL})},
	author = {Munkhdalai, Tsendsuren and Yu, Hong},
	year = {2017}
}

@inproceedings{munkhdalai_neural_2017-1,
	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{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.

@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{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{jagannatha_structured_2016,
	title = {Structured prediction models for {RNN} based sequence labeling in clinical text},
	url = {http://arxiv.org/abs/1608.00612},
	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.},
	urldate = {2016-08-19},
	journal = {arXiv:1608.00612 [cs]},
	author = {Jagannatha, Abhyuday and Yu, Hong},
	month = aug,
	year = {2016},
	note = {arXiv: 1608.00612},
	keywords = {Computer Science - Computation and Language}
}

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.

@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},
	issn = {2291-9694},
	shorttitle = {Finding {Important} {Terms} for {Patients} in {Their} {Electronic} {Health} {Records}},
	url = {http://medinform.jmir.org/2016/4/e40/},
	doi = {10.2196/medinform.6373},
	language = {en},
	number = {4},
	urldate = {2016-11-30},
	journal = {JMIR Medical Informatics},
	author = {Chen, Jinying and Zheng, Jiaping and Yu, Hong},
	month = nov,
	year = {2016},
	pages = {e40}
}

@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}
}

@article{munkhdalai_reasoning_2016,
	title = {Reasoning with {Memory} {Augmented} {Neural} {Networks} for {Language} {Comprehension}},
	url = {https://arxiv.org/pdf/1610.06454v1.pdf},
	journal = {arXiv:1610.06454},
	author = {Munkhdalai, Tsendsuren and Yu, Hong},
	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.

@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_identifying_2015,
	title = {Identifying {Key} {Concepts} from {EHR} {Notes} {Using} {Domain} {Adaptation}},
	url = {https://aclweb.org/anthology/W/W15/W15-2615.pdf},
	booktitle = {Proceedings of the {Sixth} {International} {Workshop} on {Health} {Text} {Mining} and {Information} {Analysis}},
	author = {Zheng, J and Yu, Hong},
	year = {2015},
	pages = {115--119}
}

@inproceedings{zheng_methods_2015,
	title = {Methods for linking {EHR} notes to education materials},
	volume = {2015},
	url = {http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4525231/},
	booktitle = {{AMIA} {Jt} {Summits} {Transl} {Sci} {Proc}. 2015},
	author = {Zheng, J and Yu, H},
	year = {2015},
	pages = {209--215}
}

@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{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}
}

@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{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{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 = {2015-02-26},
	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{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},
	language = {en},
	number = {0},
	urldate = {2015-02-26},
	journal = {Database},
	author = {Li, Y. and Yu, H.},
	month = nov,
	year = {2014},
	pages = {bau113--bau113}
}

@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.

@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{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}
}

@article{polepalli_ramesh_automatic_2012,
	title = {Automatic discourse connective detection in biomedical text},
	volume = {19},
	issn = {1067-5027, 1527-974X},
	url = {http://jamia.oxfordjournals.org/lookup/doi/10.1136/amiajnl-2011-000775},
	doi = {10.1136/amiajnl-2011-000775},
	language = {en},
	number = {5},
	urldate = {2016-11-30},
	journal = {Journal of the American Medical Informatics Association},
	author = {Polepalli Ramesh, Balaji and Prasad, Rashmi and Miller, Tim and Harrington, Brian and Yu, Hong},
	month = sep,
	year = {2012},
	pages = {800--808}
}

@article{liu_towards_2011,
	title = {Towards spoken clinical-question answering: evaluating and adapting automatic speech-recognition systems for spoken clinical questions},
	volume = {18},
	issn = {1067-5027, 1527-974X},
	shorttitle = {Towards spoken clinical-question answering},
	url = {http://jamia.oxfordjournals.org/lookup/doi/10.1136/amiajnl-2010-000071},
	doi = {10.1136/amiajnl-2010-000071},
	language = {en},
	number = {5},
	urldate = {2016-11-30},
	journal = {Journal of the American Medical Informatics Association},
	author = {Liu, Feifan and Tur, Gokhan and Hakkani-Tür, Dilek and Yu, Hong},
	month = sep,
	year = {2011},
	pages = {625--630}
}

@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},
	pages = {1032--1038}
}

@article{prasad_biomedical_2011,
	title = {The biomedical discourse relation bank},
	volume = {12},
	issn = {1471-2105},
	url = {http://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-12-188},
	doi = {10.1186/1471-2105-12-188},
	language = {en},
	number = {1},
	urldate = {2016-11-30},
	journal = {BMC Bioinformatics},
	author = {Prasad, Rashmi and McRoy, Susan and Frid, Nadya and Joshi, Aravind and Yu, Hong},
	year = {2011},
	pages = {188}
}

@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{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{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{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{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{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{cao_askhermes:_2011,
	title = {{AskHERMES}: {An} online question answering system for complex clinical questions},
	volume = {44},
	issn = {15320464},
	shorttitle = {{AskHERMES}},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046411000062},
	doi = {10.1016/j.jbi.2011.01.004},
	language = {en},
	number = {2},
	urldate = {2016-11-30},
	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},
	pages = {277--288}
}

@article{granger_apixaban_2011,
	title = {Apixaban versus {Warfarin} in {Patients} with {Atrial} {Fibrillation}},
	volume = {365},
	issn = {0028-4793, 1533-4406},
	url = {http://www.nejm.org/doi/10.1056/NEJMoa1107039},
	doi = {10.1056/NEJMoa1107039},
	language = {en},
	number = {11},
	urldate = {2016-11-30},
	journal = {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},
	month = sep,
	year = {2011},
	pages = {981--992}
}

@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{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{li_lancet:_2010,
	title = {Lancet: a high precision medication event extraction system for clinical text},
	volume = {17},
	issn = {1067-5027, 1527-974X},
	shorttitle = {Lancet},
	url = {http://jamia.oxfordjournals.org/lookup/doi/10.1136/jamia.2010.004077},
	doi = {10.1136/jamia.2010.004077},
	language = {en},
	number = {5},
	urldate = {2016-11-30},
	journal = {Journal of the American Medical Informatics Association},
	author = {Li, Zuofeng and Liu, Feifan and Antieau, Lamont and Cao, Yonggang and Yu, Hong},
	month = sep,
	year = {2010},
	pages = {563--567}
}

@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{agarwal_detecting_2010,
	title = {Detecting hedge cues and their scope in biomedical text with conditional random fields},
	volume = {43},
	issn = {15320464},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046410001140},
	doi = {10.1016/j.jbi.2010.08.003},
	language = {en},
	number = {6},
	urldate = {2016-11-30},
	journal = {Journal of Biomedical Informatics},
	author = {Agarwal, Shashank and Yu, Hong},
	month = dec,
	year = {2010},
	pages = {953--961}
}

@article{agarwal_biomedical_2010,
	title = {Biomedical negation scope detection with conditional random fields},
	volume = {17},
	issn = {1067-5027, 1527-974X},
	url = {http://jamia.oxfordjournals.org/lookup/doi/10.1136/jamia.2010.003228},
	doi = {10.1136/jamia.2010.003228},
	language = {en},
	number = {6},
	urldate = {2016-11-30},
	journal = {Journal of the American Medical Informatics Association},
	author = {Agarwal, Shashank and Yu, Hong},
	month = nov,
	year = {2010},
	pages = {696--701}
}

@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},
	pages = {e12983}
}

@article{cao_automatically_2010,
	title = {Automatically extracting information needs from complex clinical questions},
	volume = {43},
	issn = {15320464},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046410001061},
	doi = {10.1016/j.jbi.2010.07.007},
	language = {en},
	number = {6},
	urldate = {2016-11-30},
	journal = {Journal of Biomedical Informatics},
	author = {Cao, Yong-gang and Cimino, James J. and Ely, John and Yu, Hong},
	month = dec,
	year = {2010},
	pages = {962--971}
}

@article{yong-gang_cao_ir-aided_2010,
	title = {An {IR}-{Aided} {Machine} {Learning} {Framework} for the {BioCreative} {II}.5 {Challenge}},
	volume = {7},
	issn = {1545-5963},
	url = {http://ieeexplore.ieee.org/document/5473215/},
	doi = {10.1109/TCBB.2010.56},
	number = {3},
	urldate = {2016-11-30},
	journal = {IEEE/ACM Transactions on Computational Biology and Bioinformatics},
	author = {{Yong-gang Cao} and {Zuofeng Li} and {Feifan Liu} and Agarwal, Shashank and {Qing Zhang} and {Hong Yu}},
	month = jul,
	year = {2010},
	pages = {454--461}
}

@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}
}

@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}
}

@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{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.

@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{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}
}

@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://j-biomed-discovery.biomedcentral.com/articles/10.1186/1747-5333-4-1},
	doi = {10.1186/1747-5333-4-1},
	language = {en},
	number = {1},
	urldate = {2016-11-30},
	journal = {Journal of Biomedical Discovery and Collaboration},
	author = {Yu, Hong and Agarwal, Shashank and Johnston, Mark and Cohen, Aaron},
	year = {2009},
	pages = {1}
}

@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}
}

@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}
}

@article{cohen_translating_2008,
	title = {Translating biology: text mining tools that work},
	volume = {13},
	url = {http://psb.stanford.edu/psb-online/proceedings/psb08/textmining.pdf},
	journal = {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_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{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_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},
	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{zweigenbaum_frontiers_2007,
	title = {Frontiers of biomedical text mining: current progress},
	volume = {8},
	issn = {1467-5463, 1477-4054},
	shorttitle = {Frontiers of biomedical text mining},
	url = {https://academic.oup.com/bib/article-lookup/doi/10.1093/bib/bbm045},
	doi = {10.1093/bib/bbm045},
	language = {en},
	number = {5},
	urldate = {2016-11-30},
	journal = {Briefings in Bioinformatics},
	author = {Zweigenbaum, P. and Demner-Fushman, D. and Yu, H. and Cohen, K. B.},
	month = jun,
	year = {2007},
	pages = {358--375}
}

@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},
	pages = {236--251}
}

@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.

@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,
	title = {The semantics of a definiendum constrains both the lexical semantics and the lexicosyntactic patterns in the definiens},
	booktitle = {Proceedings of the {Workshop} on {Linking} {Natural} {Language} {Processing} and {Biology}: {Towards} {Deeper} {Biological} {Literature} {Analysis}},
	publisher = {Association for Computational Linguistics},
	author = {Yu, Hong and Wei, Ying},
	year = {2006},
	pages = {1--8}
}

@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{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, Artificial Intelligence, 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{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_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.

@article{yu_accessing_2006,
	title = {Accessing bioscience images from abstract sentences},
	volume = {22},
	issn = {1367-4803, 1460-2059},
	url = {https://academic.oup.com/bioinformatics/article-lookup/doi/10.1093/bioinformatics/btl261},
	doi = {10.1093/bioinformatics/btl261},
	language = {en},
	number = {14},
	urldate = {2016-11-30},
	journal = {Bioinformatics},
	author = {Yu, H. and Lee, M.},
	month = jul,
	year = {2006},
	pages = {e547--e556}
}

@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.

@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{rzhetsky_geneways:_2004,
	title = {{GeneWays}: a system for extracting, analyzing, visualizing, and integrating molecular pathway data},
	volume = {37},
	issn = {15320464},
	shorttitle = {{GeneWays}},
	url = {http://linkinghub.elsevier.com/retrieve/pii/S1532046403001126},
	doi = {10.1016/j.jbi.2003.10.001},
	language = {en},
	number = {1},
	urldate = {2016-11-30},
	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},
	pages = {43--53}
}

@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

@inproceedings{yu_towards_2003,
	title = {Towards answering opinion questions: separating facts from opinions and identifying the polarity of opinion sentences},
	volume = {10},
	shorttitle = {Towards answering opinion questions},
	url = {http://portal.acm.org/citation.cfm?doid=1119355.1119372},
	doi = {10.3115/1119355.1119372},
	language = {en},
	urldate = {2016-11-30},
	booktitle = {Proceedings of the 2003 conference on {Empirical} methods in natural language processing},
	publisher = {Association for Computational Linguistics},
	author = {Yu, Hong and Hatzivassiloglou, Vasileios},
	year = {2003},
	pages = {129--136}
}

@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},
	pages = {i340--i349}
}

@phdthesis{yu_synonym_2002,
	title = {Synonym and homonym resolution of gene and protein names},
	school = {Columbia University},
	author = {Yu, Hong},
	year = {2002}
}

@article{yu_automatically_2002,
	title = {Automatically identifying gene/protein terms in {MEDLINE} abstracts},
	volume = {35},
	issn = {1532-0464},
	url = {http://www.sciencedirect.com/science/article/pii/S1532046403000327},
	abstract = {MOTIVATION: Natural language processing (NLP) techniques are used to extract information automatically from computer-readable literature. In biology, the identification of terms corresponding to biological substances (e.g., genes and proteins) is a necessary step that precedes the application of other NLP systems that extract biological information (e.g., protein-protein interactions, gene regulation events, and biochemical pathways). We have developed GPmarkup (for "gene/protein-full name mark up"), a software system that automatically identifies gene/protein terms (i.e., symbols or full names) in MEDLINE abstracts. As a part of marking up process, we also generated automatically a knowledge source of paired gene/protein symbols and full names (e.g., LARD for lymphocyte associated receptor of death) from MEDLINE. We found that many of the pairs in our knowledge source do not appear in the current GenBank database. Therefore our methods may also be used for automatic lexicon generation.
RESULTS: GPmarkup has 73\% recall and 93\% precision in identifying and marking up gene/protein terms in MEDLINE abstracts.
AVAILABILITY: A random sample of gene/protein symbols and full names and a sample set of marked up abstracts can be viewed at http://www.cpmc.columbia.edu/homepages/yuh9001/GPmarkup/. Contact. hy52ATcolumbia.edu. Voice: 212-939-7028; fax: 212-666-0140.},
	language = {ENG},
	number = {5-6},
	journal = {Journal of Biomedical Informatics},
	author = {Yu, Hong and Hatzivassiloglou, Vasileios and Rzhetsky, Andrey and Wilbur, W. John},
	month = dec,
	year = {2002},
	pmid = {12968781},
	keywords = {Abstracting and Indexing as Topic, Automation, Chromosome Mapping, Genes, MEDLINE, Proteins, Terminology as Topic},
	pages = {322--330}
}

MOTIVATION: Natural language processing (NLP) techniques are used to extract information automatically from computer-readable literature. In biology, the identification of terms corresponding to biological substances (e.g., genes and proteins) is a necessary step that precedes the application of other NLP systems that extract biological information (e.g., protein-protein interactions, gene regulation events, and biochemical pathways). We have developed GPmarkup (for "gene/protein-full name mark up"), a software system that automatically identifies gene/protein terms (i.e., symbols or full names) in MEDLINE abstracts. As a part of marking up process, we also generated automatically a knowledge source of paired gene/protein symbols and full names (e.g., LARD for lymphocyte associated receptor of death) from MEDLINE. We found that many of the pairs in our knowledge source do not appear in the current GenBank database. Therefore our methods may also be used for automatic lexicon generation. RESULTS: GPmarkup has 73% recall and 93% precision in identifying and marking up gene/protein terms in MEDLINE abstracts. AVAILABILITY: A random sample of gene/protein symbols and full names and a sample set of marked up abstracts can be viewed at http://www.cpmc.columbia.edu/homepages/yuh9001/GPmarkup/. Contact. hy52ATcolumbia.edu. Voice: 212-939-7028; fax: 212-666-0140.

@article{yu_mapping_2002,
	title = {Mapping {Abbreviations} to {Full} {Forms} in {Biomedical} {Articles}},
	volume = {9},
	issn = {10675027, 1527974X},
	url = {http://jamia.oxfordjournals.org/cgi/doi/10.1197/jamia.M0913},
	doi = {10.1197/jamia.M0913},
	number = {3},
	urldate = {2016-11-30},
	journal = {Journal of the American Medical Informatics Association},
	author = {Yu, H.},
	month = may,
	year = {2002},
	pages = {262--272}
}

@article{yu_automatic_2002,
	title = {Automatic extraction of gene and protein synonyms from {MEDLINE} and journal articles},
	issn = {1531-605X},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2244511/},
	abstract = {Genes and proteins are often associated with multiple names, and more names are added as new functional or structural information is discovered. Because authors often alternate between these synonyms, information retrieval and extraction benefits from identifying these synonymous names. We have developed a method to extract automatically synonymous gene and protein names from MEDLINE and journal articles. We first identified patterns authors use to list synonymous gene and protein names. We developed SGPE (for synonym extraction of gene and protein names), a software program that recognizes the patterns and extracts from MEDLINE abstracts and full-text journal articles candidate synonymous terms. SGPE then applies a sequence of filters that automatically screen out those terms that are not gene and protein names. We evaluated our method to have an overall precision of 71\% on both MEDLINE and journal articles, and 90\% precision on the more suitable full-text articles alone},
	language = {ENG},
	journal = {Proceedings. AMIA Symposium},
	author = {Yu, Hong and Hatzivassiloglou, Vasileios and Friedman, Carol and Rzhetsky, Andrey and Wilbur, W. John},
	year = {2002},
	pmid = {12463959},
	pmcid = {PMC2244511},
	keywords = {Automatic Data Processing, Genes, Information Storage and Retrieval, MEDLINE, Names, Pattern Recognition, Automated, Periodicals as Topic, Proteins, Software},
	pages = {919--923}
}

Genes and proteins are often associated with multiple names, and more names are added as new functional or structural information is discovered. Because authors often alternate between these synonyms, information retrieval and extraction benefits from identifying these synonymous names. We have developed a method to extract automatically synonymous gene and protein names from MEDLINE and journal articles. We first identified patterns authors use to list synonymous gene and protein names. We developed SGPE (for synonym extraction of gene and protein names), a software program that recognizes the patterns and extracts from MEDLINE abstracts and full-text journal articles candidate synonymous terms. SGPE then applies a sequence of filters that automatically screen out those terms that are not gene and protein names. We evaluated our method to have an overall precision of 71% on both MEDLINE and journal articles, and 90% precision on the more suitable full-text articles alone

@article{yu_knowledge-based_2001,
	title = {Knowledge-based disambiguation of abbreviations},
	url = {https://pantherfile.uwm.edu/hongyu/www/files/articles/D010001419.pdf},
	journal = {Proceedings of the AMIA Symposium},
	author = {Yu, Hong},
	year = {2001},
	pages = {1067}
}

@article{friedman_genies:_2001,
	title = {{GENIES}: a natural-language processing system for the extraction of molecular pathways from journal articles},
	volume = {17},
	issn = {1367-4803, 1460-2059},
	shorttitle = {{GENIES}},
	url = {https://academic.oup.com/bioinformatics/article-lookup/doi/10.1093/bioinformatics/17.suppl_1.S74},
	doi = {10.1093/bioinformatics/17.suppl_1.S74},
	language = {en},
	number = {Suppl 1},
	urldate = {2016-11-30},
	journal = {Bioinformatics},
	author = {Friedman, C. and Kra, P. and Yu, H. and Krauthammer, M. and Rzhetsky, A.},
	month = jun,
	year = {2001},
	pages = {S74--S82}
}

@article{yu_automatic_2001,
	title = {Automatic extraction from scientific abstracts of synonyms for proteins and genes},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2243300/pdf/procamiasymp00002-1105.pdf},
	journal = {Proceedings of the AMIA Symposium},
	author = {Yu, Hong},
	year = {2001},
	pages = {1066}
}

@article{yu_large_2000,
	title = {A large scale, cross-disease family health history data set},
	url = {https://pantherfile.uwm.edu/hongyu/www/files/articles/D200262.pdf},
	journal = {Proceedings of the AMIA Symposium},
	author = {Yu, Hong and Hripcsak, George},
	year = {2000},
	pages = {1162}
}

@article{yu_representing_1999,
	title = {Representing genomic knowledge in the {UMLS} semantic network.},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2232882/},
	journal = {Proceedings of the AMIA Symposium},
	author = {Yu, Hong and Friedman, Carol and Rhzetsky, Andrey and Kra, Pauline},
	year = {1999},
	pages = {181}
}

@article{li_investigation_nodate,
	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},
	url = {https://academic.oup.com/jamia/advance-article/doi/10.1093/jamia/ocz018/5426087},
	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},
	urldate = {2019-04-10},
	journal = {Journal of the American Medical Informatics Association},
	author = {Li, Fei and Yu, Hong}
}

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

@article{li_investigation_nodate-1,
	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},
	url = {https://academic.oup.com/jamia/advance-article/doi/10.1093/jamia/ocz018/5426087},
	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},
	urldate = {2019-04-05},
	journal = {Journal of the American Medical Informatics Association},
	author = {Li, Fei and Yu, Hong}
}

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

@inproceedings{vu_sentence_nodate,
	title = {Sentence {Simplification} with {Memory}-{Augmented} {Neural} {Networks}},
	author = {Vu, T and Hu, B and Munkhdalai, T and Yu, H}
}

@misc{noauthor_jmi-finding_nodate,
	title = {{JMI}-{Finding} {Important} {Terms} for {Patients} in {Their} {Electronic} {Health} {Records}: {A} {Learning}-to-{Rank} {Approach} {Using} {Expert} {Annotations} {\textbar} {Chen} {\textbar} {JMIR} {Medical} {Informatics}},
	url = {https://medinform.jmir.org/2016/4/e40/},
	urldate = {2017-02-26}
}

@article{yu_rule-based_nodate,
	title = {A rule-based approach for automatically identifying gene and protein names in {MEDLINE} abstracts},
	url = {http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.68.9363&rep=rep1&type=pdf},
	author = {Yu, Hong and Hatzivassiloglou, Vasileios and Friedman, Carol and Iossifov, Ivan H and Rzhetsky, Andrey and Wilbur, W John}
}

@article{yu_question_nodate,
	title = {Question {Answering} with {Biomedical} {Images}},
	author = {Yu, Hong and Cimino, James J and Chang, Shih-fu and {others}}
}