Abstract:
Learning from ordered sequences is central to modern language models and offers a useful starting point for studying longitudinal clinical records. In clinical data, the sequence and timing of diagnoses, treatments, and measurements provide information about how disease develops across years of care. Their interpretation also depends on how illness is observed and recorded. In this talk, I will examine the connection between sequence learning in AI and temporal phenotyping, with attention to the irregular timing and clinical context of electronic health records. I will present our work on the phylophenome, a directed, confidence-ranked map of temporal relationships among clinical phenotypes. Using records from 295,678 patients at Mass General Brigham, we identified 435,240 time-ordered disease associations and combined locally deployed medical language models with machine learning to rank them. We compared these rankings with independent genetic correlations from the UK Biobank to assess concordance with shared genetic architecture. Through ClaiD, an interactive browser for exploring the resulting map, I will illustrate how temporal clinical patterns and biological evidence can guide hypotheses about disease progression. The discussion will address the distinction between predicting a subsequent clinical event and explaining its occurrence, alongside opportunities for evaluation across populations.
Bio:
Hossein Estiri, PhD, is an Associate Professor of Medicine at Harvard Medical School, Director of the Clinical Augmented Intelligence group at Massachusetts General Hospital, and an Associate Member of the Broad Institute of MIT and Harvard. His research develops computational methods for studying longitudinal electronic health records, with a focus on temporal phenotyping, clinical machine learning, and disease progression. His group applies these methods to questions including Long COVID and dementia.
听
There will be a networking opportunity following this talk. If you would like to attend, please complete the registration form. Registration will close on Monday 18 January 2027 at 12.00.