• 2019-05-08.jpg
  • 12877.jpg
  • HongYuGroup.2018.IMG_2037.JPG
  • IMG_0782.JPG
  • IMG_1675.jpeg
  • IMG_1703.jpg
  • IMG_1850.jpeg
  • IMG_1855_copy.jpg
  • MayHiking_11.jpg
  • MayHiking_12.jpg
  • MIT_UML_NLP_0.jpg
  • MIT_UML_NLP_1.jpg
  • WechatIMG95.jpeg
  • WechatIMG98.jpeg
  • YHMessageIMG.jpg
  • YHPengshan_Defense.jpg
  • YH_Amherst.jpg
  • YH_HongYuGroup_2023.jpg

Abstract: An adverse event is an injury to a patient and an adverse drug event is an injury to a patient resulting from a medical intervention related to pharmacotherapy. Adverse drug events complicate two million hospital stays annually, are associated with a prolonged hospital stay, account for upwards of two thirds of post-discharge complications, and are a significant contributor to escalating health care costs. The Office of Disease Prevention and Health Promotion has identified adverse drug event prevention as a patient safety priority. Electronic health records (EHRs) contain important adverse drug event-related information and manual chart review is prohibitively expensive. In contrast, biomedical natural language processing (NLP) provides automated tools that facilitate chart review and can improve patient drug safety surveillance and post-marketing pharmacovigilance through enhanced cost efficiencies and provision of real-time information. In this talk, I will first introduce an expert-annotated EHR corpus we developed. I will then describe several new deep neural network models (e.g., LSTM-CRF and memory-augmented NNs) we developed to build the state-of-the-art NLP systems for automated medication and adverse drug event detection from EHR narratives. I will also describe Item Response Theory (IRT) as a new evaluation metrics for NLP systems. Unlike the traditional evaluation metrics of recall/precision/F-score, IRT models characteristics of individual data points (called “items”) such as difficulty and discriminatory ability to estimate ability as a function of the characteristics of correctly answered items. Based on our IRT analysis, we found that deep neural network models exhibit human-like learning process and intelligence capabilities. Our work is an important step towards ADE surveillance and pharmacovigilance.

 

Bio: Professor Yu is an elected fellow of the American College of Medical Informatics and a tenured full professor in the Department of Computer Science, University of Massachusetts, Lowell. She is also an adjunct professor in the Department of Medicine at the University of Massachusetts Medical School, and in the College of Information and Computer Science at the University of Massachusetts, Amherst. She is a Research Health Scientist at the Bedford VA Medical Center.