Events
Foundation Models for Electronic Health Record Data Abstract
The Institute of Artificial Intelligence for Digital Health invites you to attend its monthly seminar series, featuring Matthew McDermott, Assistant Professor at Columbia Department of Biomedical Informatics.
Artificial Intelligence today is undergoing a “foundation model” revolution. From tools such as ChatGPT, to Sora, to AlphaFold, foundation models now dominate the popular press and academic directions of many areas of AI. In the high-impact, high-risk domain of health and biomedicine, the appeal of foundation models is especially poignant, but also faces some of the greatest risks and challenges. In this talk, Dr. McDermott will give a holistic overview on foundation models, their importance in health in biomedicine, and why they are so hard to build in our field — and what we can do about it. This talk will span challenges including reproducibility and effective science, the emergence of new, EHR foundation models like Ethos or COMET, and some of the major technical challenges we face in trying to build better models more aligned to healthcare.
Speaker Bio
Matthew McDermott is an Assistant Professor at Columbia Department of Biomedical Informatics. He received his PhD in Computer Science from MIT and served as a Berkowitz Postdoctoral Fellow at Harvard Medical School His research focuses on high-capacity representation learning for health and biomedicine, in particular examining the question of how to build “foundation models” for electronic health record data and other forms of health and biomedical data. Some of his key prior works include Clinical BERT, Structure-inducing Pre-training, and the emerging MEDS framework for reproducible, effective high-capacity AI over EHR data.