iiinova — where your digital self keeps evolving

Valentin Liévin

@valentinlievin.bsky.social

14 followers 11 following 9 posts

ML research at Google DeepMind Better LLMs for healthcare and science.

Posts Replies Media

Reposted by Valentin Liévin

Mike Schaekermann @mike-shake.bsky.social · 07/03/2025

🚨 New paper 🚨 Excited to share new capabilities for AMIE our conversational medical AI @googIe.blog We advance AMIE beyond diagnosis 🩺 towards treatment 💊 over time w/ guidelines 📃📃📃 - Dive into 🔗 research.google/blog/from-di... with my amazing co-author @adamrodmanmd.bsky.social!

research.google

From diagnosis to treatment: Advancing AMIE for longitudinal disease management

132

Valentin Liévin @valentinlievin.bsky.social · 07/03/2025

Our latest AMIE paper is out!

000

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

Thank you @mike-shake.bsky.social, Anil Palepu, Tao Tu, Alan Kharti, Adam Rodman, Vivek Natarajan, Wei-Hung, and many others!

010

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

[7/n] This is a major update for AMIE. While further research is needed, our study hints that AI could become a powerful tool for improving clinical decisions and healthcare access! Congratulations to the team for this incredible piece of work!

100

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

[6/n] We also challenged AMIE with RxQA, a new medication reasoning benchmark requiring precise pharmacological knowledge. Again, AMIE outperformed PCPs.

100

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

[5/n] We assessed AMIE’s clinical skills via a multi-visit Objective Structured Clinical Evaluation (OSCE), a common tool to evaluate medical professionals. AMIE scored (slightly) higher than human doctors (PCPs) in both management quality and alignment with guidelines!

100

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

[4/n] In management reasoning, there is generally not a single ground truth, but rather a space of acceptable solutions. Capturing a reliable signal to gauge plan quality was a major challenge. Gemini-as-a-judge provided a robust signal, which we hill-climbed to improve our agent

100

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

[3/n] Management reasoning is hard. AMIE can in-context retrieve and reason about specific recommendations from the guidelines, and compile these findings into a personalized plan with citations. Structured generation was a key ingredient to elicit long and controllable reasoning

100

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

[2/n] We upgraded AMIE’s original dialogue agent with a reasoning partner. The new reasoning agent taps into Gemini’s long-context reasoning and retrieval capabilities to process 100+ pages of evidence-based clinical guidelines from @NICEComms and @BMJBestPractice, in real time!

110

Valentin Liévin @valentinlievin.bsky.social · 06/03/2025

[1/n] Happy to share a big step forward for AMIE, the Articulate Medical Intelligence Explorer! We explored the frontier of AI for disease management. How well can AI manage patients over time? 🔗 research.google/blog/from-di...

100