Hong-Yu Zhou

(he/his/him) AI for Medicine | Now: Assistant Professor @Tsinghua_Uni | Prev.: PostDoc @HarvardDBMI with @pranavrajpurkar

[Nature Biomedical Engineering @natBME]

Large reasoning models as thinking machines for medicine https://t.co/QwqyLNHt7E

[Hong-Yu Zhou @HongYuZhou14]

Conventional medical AI finds patterns.
But medicine demands reasoning!

New Perspective @natBME : we propose Medical Reasoning AI (MRAI), systems that think through clinical problems, integrate real-world data, and continuously improve under clinician guidance.

The goal: AI as a transparent, auditable thinking partner in patient care.

Full text access: https://t.co/VKLDP2LlpT

With @EricTopol @AdamRodmanMD @pranavrajpurkar @Tony_Y_Hu @TienYinWong

[Hong-Yu Zhou @HongYuZhou14]

@LorenzoRighett7 Good luck, Lorenzo!

[Pranav Rajpurkar @pranavrajpurkar]

A pleasure to collaborate with Jung-Oh Lee, @HongYuZhou14, @tberzin, @DanielSodickson on rethinking the frontier of multimodal generative ai for complex medical data!

[UC Joint Computational Precision Health Program @UCJointCPH]

It's no trick! 🎃 Join us 10/31 for the CPH fall seminar series on medical #AI: @pranavrajpurkar PhD presents on "Human-AI Interaction in Radiology."
Thu, Oct. 31, 10-11 AM PDT

Register here: https://t.co/zBh4M2c5t3

[Bryan Johnson @bryan_johnson]

MRI saved my life. It may save yours too.

I’m excited to announce the world's best full-body MRI protocol in partnership with @ezrainc. This Blueprint MRI includes everything I’ve learned over the past four years becoming the most MRI-measured person in the world. This is me👇

[Philipp Schmid @_philschmid]

Can @AnthropicAI Claude 3.5 sonnet outperform @OpenAI o1 in reasoning? Combining Dynamic Chain of Thoughts, reflection, and verbal reinforcement, existing LLMs like Claude 3.5 Sonnet can be prompted to increase test-time compute and match reasoning strong models like OpenAI o1. 👀

TL;DR:
🧠 Combines Dynamic Chain of thoughts + reflection + verbal reinforcement prompting
📊 Benchmarked against tough academic tests (JEE Advanced, UPSC, IMO, Putnam)
🏆 Claude 3.5 Sonnet outperformes GPT-4 and matched O1 models
🔍 LLMs can create internal simulations and take 50+ reasoning steps for complex problems
📚 Works for smaller, open models like Llama 3.1 8B +10% (Llama 3.1 8B 33/48 vs GPT-4o 36/48)
❌ Didn’t benchmark like MMLU, MMLU pro, or GPQA due to computing and budget constraints
📈 High token usage - Claude Sonnet 3.5 used around 1 million tokens for just 7 questions

[Eric Topol @EricTopol]

I'm excited to let you know I have a new book coming out next spring.
https://t.co/3fodbmO27o @simonschuster

[Chad Byers @chadbyers]

Healthcare feels like a top 3 place to build over the next decade.

-- $4.5T US spend//20% of GDP --> huge market
-- up to 25% of the world's data is healthcare related --> huge data set --> big surface area to build
-- not available on the public internet --> not in foundational models --> less big co competition
-- multimodal --> AI particularly good at this
-- majority of data not used clinically --> latent value that can be captured by new companies

Often the challenge in healthcare is getting distribution.

Distribution --> data --> value/moat

[Hong-Yu Zhou @HongYuZhou14]

@AyooshPareekMD Thank you for delivering such an incredible talk to us!

[Isaac Kohane @zakkohane]

At least 50% of the variance in common diseases is due to environmental exposures. NEXUS is a first step to the Human Exposome Project that is essential to understanding and preventing disease.

[Hong-Yu Zhou @HongYuZhou14]

🚀 Catching Misses by Doctors! 🚀
a2z-1 launches: AI tackling one of the most common radiological exams - abdominal-pelvis CT scans.

21 conditions. One AI. Unlimited potential.
And it's just the beginning.

Ever wondered the vision and story behind? 👇
https://t.co/MvCdRI1lKZ

[Pranav Rajpurkar @pranavrajpurkar]

It's launch day! 🚀 Announcing a2z Radiology AI and our first product, a2z-1.

a2z-1 is an AI that analyzes abdominal-pelvis CT scans and reports to catch potential misses across 21 conditions.

Our mission is to create a comprehensive AI safety net for radiology, ensuring no disease goes undetected.

[Danielle Bitterman, MD @dbittermanmd]

AI models should be clinically validated to demonstrate benefit and implement effectively! Our clinical trial of AI-assisted contouring for breast radiation therapy is a case in point, showing no efficiency improvements despite promising pre-clinical evidence. @Dr_RayMak #ASTRO24

[Pranav Rajpurkar @pranavrajpurkar]

Looking forward to it. We’re at an inflection point for radiology Ai, and in my talk, I’ll make the case for the emerging generalist medical ai paradigm and how that will change the kinds of AI models we will deploy for radiologists in the future.

[Hong-Yu Zhou @HongYuZhou14]

We are transitioning from CXR to CT reporting!

[Pranav Rajpurkar @pranavrajpurkar]

📢 Introducing HeadCT-ONE: Our new paper addresses a major gap in AI evaluation for radiology—capturing semantic equivalence. Using ontologies, we standardize medical terms, making AI-generated head CT reports more accurately comparable, even when phrasing differs.🧠✨

[Michael Moor @Michael_D_Moor]

Arguably one of the most underrated skills in academic research (esp in CS/ML/AI world) is to create captivating visuals, see eg 👇

[Julián Nicolás Acosta @jn_acosta]

Are you a radiologist interested in shaping the future of AI in healthcare? At Rajpurkar Lab, we are looking for collaborators to join us in an exciting upcoming project. If interested, please reach out - DMs are open! @pranavrajpurkar

[Eric Topol @EricTopol]

Cancer in the young is on the rise worldwide, but the basis for this is not known. A new review covers it well, open-access
https://t.co/iS6yMuMUxu @Albert0Bardelli @GianMau1990 @CellRepMed

[Pranav Rajpurkar @pranavrajpurkar]

⭐️ Announcing ReXRank, a competition for radiology report generation from Chest X-Rays.

Featuring 16 existing competitive approaches, a large private test leaderboard of 10k cases indicative of more real-world performance, and a live arena where radiology experts can do a head-to-head test. 📈

Check out the live leaderboard.
https://t.co/CsU0K5htqK