Pranav Rajpurkar

The work toward it

I think of the path in four parts. For each, one result we have shown and one question that is still open. The research happens in the Rajpurkar Lab at Harvard.

1. Reason from evidence

   An AI doctor has to read everything a patient has, from scans to notes, and tie each answer to the evidence behind it.

   Shown

   Radiologists rated reports from MedVersa, our generalist imaging model, as good as human-written ones in 64% of cases. NEJM AI, 2026

   Open

   How should a model combine a patient’s scans, notes, labs, and history into one picture that changes over time?

   

2. Test before we trust

   Exams and fixed benchmarks overstate readiness. Before an AI meets a patient, it has to be tested the way it will be used: across hospitals, in conversation, and alongside clinicians.

   Shown

   GPT-4’s diagnostic accuracy fell from 82% on written cases to 63% when it had to take the history itself. Nature Medicine, 2025

   Open

   How can a simulated hospital respond plausibly to an AI’s decisions over time?

   

3. Learn to assist

   Much of care is physical work. I want robots that learn from expert demonstration and keep the clinician in control.

   Shown

   Trained on under an hour of demonstrations, a low-cost robot arm followed 23 of 25 stitches alongside a surgeon. Machine Learning for Healthcare, 2026

   Open

   Which forms of robotic assistance are useful, safe, and robust enough to evaluate clinically?

   

4. Reach patients

   Research only helps patients once it reaches the clinic. I co-founded a2z Radiology AI to take this work there.

   Shown

   a2z’s two FDA-cleared devices, a2z-Abdo-Triage and a2z-Unified-Triage, flag suspected cases across 17 findings on adult abdomen and pelvis CT for worklist prioritization. a2z-Abdo-Triage clearance · a2z-Unified-Triage clearance

   Open

   What evidence should let AI take on more responsibility in care?

5. Expert medical care for everyone.