David JH Wu, MD
@davidjhwu
Followers
3K
Following
8K
Media
570
Statuses
3K
@stanfordradonc resident, previously @scvmed intern. research in healthcare + ai intersection
Cupertino, CA
Joined April 2012
full text here: https://t.co/rgc0BQ1JRn My thoughts on this are still evolving. In oncology, our 100m dash is the Kaplan-Meier overall survival curve (IMHO). That's the best way to get docs to align. Recently gave a talk on this, I will share with the greater internet soon(2/2)
0
0
1
Wanted to share some more personal reflections on what it's like to build, evaluate, and research some of these frontier models with real-world, de-identified patient cases. I think it's in these messy, non-polished cases where you can really see intelligence shine (or not) (1/2)
1
0
2
It was great to learn about what @drethangoh, @AdamRodmanMD & @jonc101x are building at ARiSE. My latest for @Forbes ✍️🚀 @Stanford @StanfordDeptMed @Harvard @erichorvitz, @andrewparsonsMD, @andrewolsonmd, @vishnuravi @LiamGMcCoy @davidjhwu & many more! https://t.co/W4E8A2Qs6I
0
5
8
Great thread by one of the most talented triple-threat physician-coder-catlover I have ever had the pleasure of working with!! Terrific work on some much needed eval of clinical reasoning in LLMs
Are language models overconfident in their clinical reasoning? Have reasoning optimizations made this problem worse? Our latest in NEJM AI - we adapt a human-validated, trivially scalable automated benchmark to get to the heart of clinical reasoning in AI systems
0
0
3
Cancer decision making is inherently multimodal but there is a lag in how we have adopted imaging into medical AI vs text
0
0
0
1st talk: Great talk by the legendary @mattlungrenMD about where we are today and some unique insight about Microsoft’s role in medical AI. Fun fact: 95% of EHR users use Microsoft. Could be an easy point of entry into healthcare, eg tumor boards held on Teams
1
0
1
Big day today! Welcome to the Symposium on AI in Medicine hosted by the @StanfordRadOnc department! Check out our awesome speaker lineup! Will tweet some highlights throughout the day 🌊
0
0
4
Much thanks to my amazing collaborators @drethangoh @vishnuravi @jonc101x @LiamGMcCoy @fatemenateghi
1
0
5
This small pilot (n=40) opens the door for more rigorous, large-scale real-world evaluation of medical AI. Instead of just testing on clean vignettes, we can assess how AI performs on the actual messy complexity of clinical practice in a scalable way. https://t.co/EUjwdxuKdy
medrxiv.org
Specialist consults in primary care and inpatient settings typically address complex clinical questions beyond standard guidelines. eConsults have been developed as a way for specialist physicians to...
1
1
3
Both compared AI answers to actual specialist responses. Key finding: LLM-as-judge matched human physician performance when assessing clinical concordance. This suggests we can scalably benchmark AI for complex medical tasks that is more considerate of physician reviewers.
1
1
5
The challenge? How do you evaluate AI performance on these real-world consults at scale? We can't have human physicians manually review every AI response so we tested two automated evaluation methods: LLM-as-judge Decompose-then-verify: Breaking down responses into atomic facts
1
0
2
Real-world cases are messy, often incomplete. At @Stanford we have collected a unique dataset: thousands of real physician-to-physician eConsult cases. Think actual specialists answering complex questions from primary care docs—not polished, textbook scenarios. (2/6)
1
0
4
Excited to share some of my recent work! Y'all have probably seen that there's been a lot of hype around LLMs acing the USMLE and medical vignettes. But here's the thing: standardized tests ≠ real-world medicine! How does AI do in the real-world and how can we eval it? Read on
1
6
25
Buying API credits off of Anthropic really has changed my perspective on what $5 is. I’ll never be able to look at a cup of coffee the same any more. I’ll take the $5 API credit over a latte ANY DAY
2
1
7
Wrote up some of my thoughts and reflections from this past year’s #AIMI25. Enjoy! https://t.co/eOEyuqQ9re
0
0
2
Great panel on publishing health in Health AI at #AIMI25 featuring @charlottehaug (NEJM AI), @yhswen (JAMA+AI), and Chris Paton (BMJ) hosted by @drnigam “What do editors want?” - Prefer something prospective, reproducible, and useful today. “Do science the standard way”
0
1
5
Some interesting quotes: “These past few months we’ve seen the same amount of progress in model developments as the past few years” “Fine-tuning on medical tasks usually causes loss of other skills. Training pipeline is delicate. Risk of catastrophic forgetting.”
1
0
1
Packed house at #AIMI25 for this star-studded panel featuring @Emily_Alsentzer (Stanford CS), @thekaransinghal (OpenAI), @KhaledSaab11 (DeepMind) and @marinkazitnik (Harvard Biomedical Informatics) and @drethangoh 🎯: foundational model roadmap for health AI
1
7
55