Research
Clinical NLP
Using LLMs to surface clinical decision patterns from unstructured medical narratives, with a focus on heart failure and heart transplant care.
A collaboration with NewYork-Presbyterian Hospital on clinical NLP for heart failure and heart transplant patients. The shared question across these projects: clinical reasoning lives in unstructured narratives — discharge summaries, exception requests, progress notes — and the patterns inside them are too numerous and too subtle to extract by hand at scale.
Heart-transplant exception requests
When a patient doesn’t fit the standard heart-transplant allocation criteria, their care team writes a free-text justification arguing for an exception. Thousands of these requests accumulate; manual review for policy design is infeasible. Our JHLT 2026 paper uses an LLM to perform thematic analysis of accepted exception requests, surfacing the clinical rationales that drive decisions and producing evidence directly relevant to allocation-policy design.
Heart-failure care
Beyond transplant, the broader heart-failure population generates dense longitudinal narratives across admissions, clinic visits, and device interrogations. We’re applying similar narrative-analysis methods to surface decision patterns and risk signals from this unstructured record.
Why this matters
Cardiology is a domain where the clinical reasoning is the data — and where the cost of overlooking signal in narratives is measured in patient outcomes and policy design. LLMs are finally good enough at reading long, technical, idiosyncratic clinical text to make this kind of analysis possible at scale.