RESEARCH AT THE RAJPURKAR LAB

Reason. Test. Assist.

We build AI that reasons from evidence, we test it before we trust it with patients, and we teach it to assist clinicians with its own hands.

PROGRAM 01

Reason

Reason from evidence

Can AI read every scan and record a patient has, and tell us what is coming?

Early medical AI matched specialists one task at a time. We build generalist models that read and write across images and text, find what no one asked them to look for, and tie each answer to the evidence behind it.

WHAT WE ARE BUILDING TOWARDAI that predicts each patient's future from every scan and every record, so clinicians can act before disease arrives.

THE RESEARCH ARC

  1. 01Read
  2. 02Ground
  3. 03Predict

QUESTIONS WE ARE WORKING ON

  1. Q1

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

  2. Q2

    Which routine data already carries warning signs of future disease that no one checks today?

  3. Q3

    How should language be tied to spatial evidence, so every finding points to where it came from?

SELECTED PROJECTS

Connected steps in the program, not a complete publication list.

  1. 01

    MedVersa

    A generalist model that reads and writes both images and text across seven kinds of medical imaging, with tasks specified on the fly.

    Radiologists rated its reports as good as human-written ones in 64% of cases.

    NEJM AI · 2026

  2. 02

    Heart risk from head CT

    Reads routine emergency head CTs, which are never checked for heart disease, to predict heart attack, stroke, and heart failure.

    Outperformed the standard PREVENT risk score (C-index 0.82 vs 0.75) and changed the risk group for 1 in 6 patients without an extra test.

    JACC · 2026

  3. 03

    ReXGroundingCT

    Connects free-text radiology findings to expert-verified, pixel-level evidence in three-dimensional chest CT.

    NEJM AI · 2026

  4. 04

    3DReasonKnee

    Tests whether models can localize relevant anatomy before producing step-by-step diagnostic reasoning and structured severity assessments.

    Pacific Symposium on Biocomputing · 2026

PEOPLE

Pranav RajpurkarSungeun KimEmma ChenOishi BanerjeeMohammed Baharoon

Open role: Postdoctoral fellow, Reason

PROGRAM 02

Test

Test before we trust

Is an AI ready for patients, and how would we know before it meets one?

Exams and fixed benchmarks overstate readiness. We build public scoreboards, AI patients that must be interviewed, studies across many hospitals, and experiments with clinicians. We bring them together in simulated hospitals, where every decision changes what happens next.

WHAT WE ARE BUILDING TOWARDA simulator of an entire hospital, with every department, bed, and handoff, so AI can be tested on a whole hospital's work before it meets a patient.

THE RESEARCH ARC

  1. 01Measure
  2. 02Interact
  3. 03Generalize
  4. 04Simulate

QUESTIONS WE ARE WORKING ON

  1. Q1

    Which errors matter clinically, and which evaluations reveal failures that averages hide?

  2. Q2

    How can simulated patients and hospitals respond plausibly to decisions over time?

  3. Q3

    How should work be divided between AI and clinicians, and how do we test each in its role?

SELECTED PROJECTS

Connected steps in the program, not a complete publication list.

  1. 01

    Clinical Environment Simulator

    Proposes a persistent digital hospital where agent decisions alter patients, resources, and subsequent choices.

    Nature Medicine · Perspective · 2026

  2. 02

    CRAFT-MD

    Evaluates clinical language models through interactive patient encounters rather than isolated question answering.

    GPT-4's diagnostic accuracy fell from 82% on written cases to 63% when it had to take the history itself, across 2,000 cases in 12 specialties.

    Nature Medicine · 2025

  3. 03

    AI across 23 hospitals

    Tests one AI across 23 hospitals in 12 countries through MAIDA, Medical AI Data for All.

    Error nearly tripled from one hospital to another; continual learning beat locally fine-tuned models at 21 of 22 hospitals.

    NEJM AI · 2026

  4. 04

    When AI helps clinicians

    Measures how AI assistance changes the performance of 140 radiologists across 15 chest X-ray tasks.

    Experience, specialty, and familiarity with AI did not predict who AI would help, and inaccurate AI pulled radiologists down.

    Nature Medicine · 2024

  5. 05

    ReXrank

    Provides a common, held-out evaluation for comparing report-generation systems across diverse clinical settings.

    Evaluates report generation on 10,000 studies from 67 U.S. sites, the largest test set of its kind.

    AAAI Bridge Program · 2025

  6. 06

    Medical sycophancy

    Shows how a model's clinical position can degrade when a user repeatedly challenges it across a conversation.

    HeaLing at ACL · 2026

PEOPLE

Pranav RajpurkarSungeun KimAaditya Ura

Open role: Postdoctoral fellow, Test

PROGRAM 03

Assist

Learn to assist

Can AI learn the physical work of a procedure well enough to be a useful extra pair of hands?

Most medical AI works on a screen. We teach robots to assist clinicians in the operating room and at the bedside: learning from expert demonstration, understanding procedures from video, and keeping the clinician in control.

WHAT WE ARE BUILDING TOWARDA portable surgical assistant that rolls up to any operating table, and robotic imaging that brings expert scans to every patient, wherever they are.

THE RESEARCH ARC

  1. 01Perceive
  2. 02Learn from demonstration
  3. 03Assist

QUESTIONS WE ARE WORKING ON

  1. Q1

    Which representations capture anatomy, geometry, coverage, and procedural state from video?

  2. Q2

    How can robots learn expert technique from limited demonstrations?

  3. Q3

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

SELECTED PROJECTS

Connected steps in the program, not a complete publication list.

  1. 01

    Robot suture following

    The first evaluation of general-purpose imitation learning for a robot assisting a surgeon in open surgery.

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

    MLHC · 2026

  2. 02

    ReXSonoVQA

    Tests whether video-language models understand probe movement, image quality, anatomy, and procedural intent in ultrasound.

    The first benchmark for video-based procedural ultrasound understanding, built on 514 videos; models struggle most with fixing a poor image.

    CHIL · 2026

  3. 03

    ColonCrafter

    Estimates temporally consistent depth from colonoscopy video as a foundation for geometry-aware procedural analysis.

    Pacific Symposium on Biocomputing · 2026

  4. 04

    Seeing Down the Line

    Adds colon-aware geometry to three-dimensional reconstruction so procedural coverage can be measured directly.

    MIDL · 2026

PEOPLE

Pranav RajpurkarSungeun KimRomain HardyDavid WangZhizhou (Jason) Yang

Open role: Postdoctoral fellow, Assist

THE COMPLETE RECORD

View all publications