Pranav Rajpurkar

Pranav Rajpurkar works on a question medicine cannot avoid: how should general-purpose AI enter a system where evidence, workflow and accountability matter as much as benchmark accuracy? His research spans diagnostic imaging, multimodal foundation models and new environments for testing clinical agents over time.

Pranav Rajpurkar is an associate professor of biomedical informatics at Harvard Medical School, where his laboratory develops and evaluates AI for medicine. His earlier first-author work on CheXNet helped establish deep learning as a serious method for chest-radiograph interpretation. More recently, he has argued that medical AI must move beyond isolated prediction towards systems that can integrate modalities, specialists and sequential decisions. MedVersa, published in NEJM AI in 2026, embodies that direction: a language-model-orchestrated generalist system for multiple medical-imaging tasks. In a retrospective reader study reported by his laboratory, radiologists judged generated chest-X-ray reports clinically equivalent to human reports in 64% of cases overall and 91% of normal cases. Those findings do not establish patient benefit or clinical approval. Rajpurkar also co-authored a 2026 Nature Medicine perspective proposing a clinical-environment simulator with hospital and patient engines for dynamic evaluation of AI agents. Companion work in Nature Biomedical Engineering set out a generalist–specialist collaboration framework and examined large reasoning models as potential thinking machines for medicine. Across these projects, his emphasis is consistent: assess AI not as a detached benchmark entrant, but as one component of a clinical system with changing information, expert roles and consequences.