Sam Margolis

Research

What AI actually does in medicine, and where it helps

Evaluating medical AI

Agents in DrugTargetWorld get a simulated biobank with hidden causal structure; the best of 9 scored 39.98 of 100. In TriageBench, branded health AI over-triaged 28% of low-acuity cases, against 3% and 2% for other product categories.

Cardiovascular AI and genomics

As a Sarnoff Fellow in Euan Ashley's group, Sam uses large-scale patient data and AI models to split disease subtypes that are lumped together, then traces them to genomics and treatments.

Biomedical sensing and imaging

Learning models on new hardware with Tzung Hsiai's group: an on-skin sensor that detects early fatty liver disease, stretchable sensing fibers, and virtual immunohistochemistry slides.

Cancer survivorship and patient-reported outcomes

With Mina Sedrak's group at UCLA: how often patient-reported outcomes from the trials behind breast cancer drug approvals are reported and used, and the PROFFi trial of fisetin and exercise in breast cancer survivors.

Open clinical software

OpenScribe is an MIT-licensed, self-hostable AI scribe. Preclinical stress-tests healthcare agents with simulated adversarial patients.

All publications · Projects