Computer scientist · Medical artificial intelligence

Pranav Rajpurkar

Associate Professor of Biomedical Informatics
Harvard Medical School
Co-founder of a2z Radiology AI

Pranav Rajpurkar develops and evaluates artificial intelligence for medicine. He leads the Rajpurkar Lab at Harvard and co-founded a2z Radiology AI.

His work spans medical imaging, multimodal learning, clinical reasoning agents, and the study of how clinicians and AI make decisions together.

Computer scientist working on artificial intelligence for medicine

Position: Associate Professor of Biomedical Informatics, Harvard Medical School
Lab: Rajpurkar Lab at Harvard Medical School
Industry: Co-founder of a2z Radiology AI
Training: BS, MS, and PhD in computer science, Stanford University
Teaching: 90,000+ learners in AI for Medicine
Record: 201 research publications · 58,040 citations per Google Scholar

Scientific profile

From measuring machine comprehension to building medical AI for real clinical work.

Rajpurkar trained in computer science at Stanford, where his early work helped establish widely used benchmarks for machine comprehension and medical image interpretation. At Harvard Medical School, his group now studies the broader problem: how AI can interpret different forms of medical data, reason through clinical tasks, and work reliably with people.

The research connects foundational models and evaluation methods with clinical translation. The governing standard is practical: a capability is only useful if the evidence survives the conditions in which care actually happens.

Expert medical care should reach everyone.
Getting there will take AI doctors: systems that can carry real clinical work, not only answer questions about it.

Selected contributions

A scientific arc from language benchmarks to systems that reason, act, and enter practice.

Five chapters show how the work expanded from bounded benchmarks and medical perception to clinical reasoning, interactive simulation, robotics, and translation.

2016 · Machine comprehension

A benchmark for extractive question answering

SQuAD introduced more than 100,000 questions grounded in Wikipedia passages and made one bounded form of machine comprehension easier to measure. It was my first first-author paper, and it showed that a benchmark can focus a field without capturing the whole problem.

SQuAD paper

2017–2020 · Medical signals and images

Testing perception on medical data

Work on arrhythmias and chest radiographs compared deep-learning systems with physicians on retrospective perception tasks. CheXpert also made a large labeled chest X-ray dataset available for research. These projects made the difference between benchmark performance and usefulness in clinical practice impossible to ignore.

Arrhythmia detection CheXNet CheXpert

2021–2024 · Broader medical AI

Broadening the problem

The work expanded toward generalist medical AI, systems that can use several kinds of medical information, and studies of how clinicians and AI make decisions together. A central lesson was that combining a person and a model does not guarantee a better decision. These remain active questions in the lab today.

Generalist medical AI

2025–present · Simulation and embodied AI

Moving from static answers to interactive systems

Recent work extends evaluation from isolated model outputs to settings where decisions unfold over time. The Clinical Environment Simulator proposes testing clinical agents against evolving patients and hospital constraints; robotics studies evaluate assistance in open surgery and staged paths for embodied AI in operating rooms.

Clinical environment simulator Robot assistance in open surgery Embodied AI in the operating room

2025–present · Imaging AI translation

Bringing multi-finding imaging AI into clinical workflow

Rajpurkar co-founded a2z Radiology AI, whose first product, a2z-Unified-Triage, received FDA clearance to prioritize adult abdomen and pelvis CT studies with suspected findings across seven conditions. The product brings multi-finding imaging AI into the clinical worklist.

a2z Radiology AI FDA clearance

Research program

Four areas shape the lab’s current work.

The Rajpurkar Lab builds models, datasets, evaluation methods, and simulation environments for medical AI.

01Research area

Multimodal medical AI

A clinician combines images, language, video, laboratory results, and physiological signals. We are studying how a single system can reason across them without losing the context that connects them.

Can one system understand the whole medical picture?

02Research area

Clinical agents and evaluation

Care is a sequence of decisions, not a single answer. We build agents that gather information, revise their view, and work through clinical tasks—and environments that test those abilities dynamically.

Can clinical AI reason over time?

03Research area

Clinician–AI collaboration

Putting a clinician and an AI system together does not automatically improve a decision. We study when the combination helps, when it fails, and how responsibility should be shared.

When does AI actually help a clinician?

04Research area

Medical imaging and embodied AI

Interpreting an image is only one part of care. We are connecting medical perception to action, including procedural and robotic settings where mistakes have immediate consequences.

Can perception lead safely to action?

The standard: a capability is only useful if the evidence survives the conditions in which care actually happens.

Beyond the lab

Translation and teaching extend the scientific work.

The same emphasis on useful evidence carries into clinical software and education.

Clinical translation

a2z-Unified-Triage is FDA-cleared to prioritize adult abdomen and pelvis CT studies with suspected findings across seven conditions. Rajpurkar co-founded a2z Radiology AI.

Teaching

90,000+ learners have enrolled in AI for Medical Diagnosis, one course in the three-part AI for Medicine series. The courses teach the same standard the research holds to: measure what a model does, and say where it fails.

Recent work