Pranav Rajpurkar PhD

Pranav Rajpurkar PhD.

@rajpurkar.bsky.social Biomedical Informatics Research Colloquia
“Beyond Assistance: Rethinking AI-Human Integration in Radiology”
Pranav Rajpurkar, PhD.

Thursday, April 17th, 2025
12:00 to 1:00 pm PST

stanford.zoom.us/j/9788759601...

Webinar ID: 978 8759 6012
Webinar Passcode: 420642

A pleasure to be included as part of the innovation driving our field forward!

Our @nature.com paper today on multimodal generative #AI for interpretation of medical images and generating reports across many specialties nature.com/articles/s41... @rajpurkar.bsky.social
free access rdcu.be/efcIG

When Doctors With A.I. Are Outperformed by A.I. Alone Interpreting Some Surprising Results

Our op-ed @nytopinion.nytimes.com addresses the surprising results of recent medical studies that showed A.I. alone outperformed physicians using A.I. With @rajpurkar.bsky.social
Here is a summary Table, an overview, and a gift link to the op-ed
erictopol.substack.com/p/when-docto...

Opinion | How Doctors Can Best Integrate A.I. Into Medical Care Research suggests physicians are undervaluing how artificial intelligence can improve medical care.

Our new essay in @nytimes.com: What if letting doctors and AI each focus on their strengths works better than constant collaboration? Some early & surprising evidence. w/ @erictopol.bsky.social

www.nytimes.com/2025/02/02/o...

An evaluation framework for clinical use of large language models in patient interaction tasks - Nature Medicine By simulating realistic doctor–patient conversations, a framework can be applied to large language models to investigate shortcomings and bias in patient interactions, providing insight before actual ...

How do LLMs really perform in diagnosis? There are claims of superhuman AI performance on clinical vignettes, which have much of the relevant information included. That's not how medicine works. That’s why @rajpurkar.bsky.social and I worked on CRAFT-MD: www.nature.com/articles/s41...

A clinical certification pathway for generalist medical AI systems There have been extensive evaluations of artificial intelligence (AI) systems for narrow medical tasks, but more work is needed to systematically evaluate and deploy AI systems that can perform a broad range of medical tasks. The medical training process itself might offer a template for addressing this challenge. Clinicians undergo rigorous education and training, progressing through stages of increasing responsibility and autonomy. Similarly, generalist medical AI systems could be subjected to a phased certification model before they are granted greater autonomy in patient care.

Our new essay @TheLancet on certifying generalist medical #AI systems, from a template of the training process in medicine
www.thelancet.com/journals/lan... @rajpurkar.bsky.social

Important #AI chatbot for medical diagnosis report today @naturemedicine.bsky.social by @rajpurkar.bsky.social and colleagues, and coverage @newscientist.bsky.social
www.nature.com/articles/s41...
www.newscientist.com/article/2462...

Fantastic co-authors including co-first author Jaehwan Jeong, with Benjamin Tran, Daniel Schlessinger, Shannon Wongvibulsin, Leandra Barnes, Hong-Yu Zhou, Zhuo Ran Cai, Eli Van Allen, and David Kim.

An evaluation framework for clinical use of large language models in patient interaction tasks Nature Medicine - By simulating realistic doctor–patient conversations, a framework can be applied to large language models to investigate shortcomings and bias in patient interactions,...

A big congratulations to star Harvard PhD student Shreya Johri, and what a delight to co-lead this work with
@roxanadaneshjou.bsky.social.

Read our paper here: rdcu.be/d5ovt

Our new Harvard/Stanford study out today in Nature Medicine! Our results show language models, while strong at medical exams, stumble when they need to gather patient information through natural conversation - a key finding as healthcare explores AI integration.

How to solve? Enter CRAFT-MD.

ReXrank: Radiology Report Generation Leaderboard Open-source leaderboard for comparing and benchmarking AI models in radiology report generation.

I encourage academic and commercial groups to join in the competition and showcase the best you have:
rexrank.ai

ReXrank: Radiology Report Generation Leaderboard Open-source leaderboard for comparing and benchmarking AI models in radiology report generation.

If developers are coming to you with a model claiming to perform great report generation on chest X-rays, ask them to do a fair comparison on ReXrank. Let's separate the hype from the reality.

ReXrank: Radiology Report Generation Leaderboard Open-source leaderboard for comparing and benchmarking AI models in radiology report generation.

We have now launched a competition called ReXrank to track progress.
With Gradient Health, we have collaborated to produce the now the largest and cleanest private dataset to evaluate models.

I don't think it's widely clear to the #RadiologyAI community just how poorly GPT-4V compares with the top report generation models on chest X-rays, like MedVersa or MAIRA-2.

It's clear we need a way to track progress.

Please welcome MedAutoCorrect = Spelling check for medical imaging reports.
"Model leverages image-conditioned data processing to enhance the accuracy of medical report generation."
It seems simple, yet at the same time, it is incredibly advanced.
arxiv.org/pdf/2412.02971
@rajpurkar.bsky.social

Biomedical Signal Interpretation - HMX | Harvard Medical School Primary Faculty More Information

Medical Image Interpretation course page: onlinelearning.hms.harvard.edu/hmx/hmx-pro/...
Biomedical Signal Interpretation course page: onlinelearning.hms.harvard.edu/hmx/hmx-pro/...

Includes a cameo by @mattlungrenmd.bsky.social , as the course covers an expert perspective on real-world AI in radiology.

Most importantly, it's curated to be very accessible and approachable, requiring no mathematical background. I walk you through the science. And it's short :) Here's a teaser for one of the courses – this is on Medical Image Interpretation.

I've curated lessons using real-world research examples, where I go through classic problem formulations and how they have been approached by world-class teams. For some case studies, I've used examples from my own research to give you insight into the back story surrounding cutting-edge research.

Delighted to launch 2 new short courses on AI for Medicine, now open for registration at Harvard Medical School online learning platform!

Medical Image Interpretation course page: onlinelearning.hms.harvard.edu/hmx/hmx-pro/...

“Automation neglect” may be responsible for suboptimal human ai collaboration @erictopol.bsky.social @rsnasky.bsky.social #rsna24 #radiology #ai

No FDA approved Ai for level 3 and 4 AI!

Opportunistic screening AI evidence! @erictopol.bsky.social @rsnasky.bsky.social

#rsna24

If a machine makes the same mistake again and again, it's precise. but that's not what we want. We also need accurate AI.

Talk from @erictopol.bsky.social at @rsnasky.bsky.social #RSNA

Excited for the keynote of @erictopol.bsky.social at #RSNA24.

Powerful plenary keynote by @curtlanglotz.bsky.social, painting the vision for a future with intelligent connections in radiology.

AI powered draft reporting being showcased on bone x-rays

title slide of presentation

The @rsnasky.bsky.social meeting is the largest medical meeting in the world. The radiology community is the pacesetter for medical #AI, that we can all learn from. I'm excited to give a presentation during tomorrow morning's plenary session
reg.meeting.rsna.org/flow/rsna/rs... #RSNA2024

Excited to be back at #RSNA2024! 🎉 It's going to be a packed one.

I'm delighted to showcase a2z-2: our AI model that can detect 140 findings in abdomen-pelvis CT scans, ranging from acute appendicitis to aortic aneurysms. It is currently available for research use and clinical trials.

140 Findings. The most comprehensive detection for abdomen-pelvis CTs. This is what it means to go from A to Z.