Title: What, Why and How of Clinical Risk Prediction using Electronic Health Records
Date and time: 12 June 2025 (Thursday), 10 – 11 a.m.
Venue: IEOR Seminar Room
Speaker: Dr. Sandhya Tripathi, Washington University at St. Louis
Abstract: Clinical risk prediction using electronic health records (EHRs) presents challenges due to fragmented data sources, heterogeneous modalities, and concerns around model generalizability and fairness. This talk focuses on contrastive learning—a representation learning technique that brings semantically similar features or samples closer in latent space while pushing dissimilar ones apart—as a core method for aligning tabular datasets without metadata. We extend this approach to intermodality contrastive learning, leveraging relationships across structured data, clinical notes, and intraoperative time series to enhance integration and predictive accuracy. We evaluate the contributions of each modality, highlighting tradeoffs between data availability and performance. We also discuss how social and demographic factors influence outcomes and model behaviour, and briefly introduce a framework to surface model limitations for informed use in clinical settings.
Bio: Sandhya Tripathi is a staff scientist in the Department of Anesthesiology at Washington University School of Medicine, specialising in machine learning applications within healthcare. She is an alumna of Dept. of Statistics, Lady Shri Ram Collage and Dept. of IEOR, IIT Bombay. Her research interests span perioperative and critical care settings, focusing on the development of AI-based clinical decision support systems—from designing AI models to deploying them in electronic health record (EHR) systems—and exploring how these models impact different societal demographics.