What happens in SAIL 2025 stays in SAIL 2025 -- except for these anonymized hot takes! 🔥 Jotted down 17 de-identified quotes on AI and medicine from medical executives, journal editors, and academics in off-the-record discussions in Puerto Rico

irenechen.net/sail2025/

AI deployments in health are often understudied because they require time and careful analysis.⌛️🤔

We share thoughts in @ai.nejm.org about a recent AI tool for emergency dept triage that: 1) improves wait times and fairness (!), and 2) helps nurses unevenly based on triage ability

tl;dr Healthcare access disparities cascade through the entire ML pipeline.

Check out our working paper here: arxiv.org/pdf/2412.07712

Finding 3: Here's what helped: Adding patient self-reported data boosted model sensitivity by 11.2% for underserved patients. (Note adding the low/high access info did NOT help)

Finding 2: "Just train a better model" isn't enough. EHR reliability has large effects on balanced accuracy (5.8% drop) and sensitivity (9.4% drop).

Finding 1: For 78% of medical conditions we examined, data quality was worse for patients facing cost/time barriers to health. EHR reliability is defined compared against patient self-report

How do disparities in healthcare access affect ML models? 💰📉🧐 We found that low access to care -> worse EHR data quality -> worse ML performance in a dataset of 134k patients. Work with Anna Zink (on the faculty job market rn!) + Hongzhou Luan, presented at #ML4H2024

It's giving Best Paper at the ML for Health Symposium (co-located w NeurIPS)!! 🥳 Congrats to co-authors Emily, Jin, and many others 👏. Check out our work using LLMs to understand liver transplants, esp understudied social and economic factors 🏥💰🏠! #ml4h2024

arxiv.org/pdf/2412.07924

This year our CHEN lab holiday party featured cookie decorating! 🎄 Grateful to have such creative and inspiring students and collaborators. 🥰 Can you spot all of the ML-related cookies? 📈

Important caveats: 1) Very small sample size (6 medical cases) -> p=0.03 which is kinda sus, 2) human physicians in study had only 3 yrs of training, 3) no nuance of how to use LLMs for diag reasoning: clinical notes != clean cases; paper does not engage with this.

What do it mean to be a “low resourced” language? I’ve seen definitions for less training data to low number of speakers. Great to see this important clarifying work at #EMNLP2024 from @hellinanigatu.bsky.social et al

aclanthology.org/2024.emnlp-m...

If you trained 10 models and they had a huge variance on predictions for you, would you have any faith in the model? Enjoyed this paper defining self-consistency -- and showing enforcing that makes models more fair! Cool AAAI24 paper from A. Feder Cooper et al.

katelee168.github.io/pdfs/arbitra...

Giving a talk tomorrow 11:40am PT at the Simons Domain Adaptation Workshop. I'll be speaking about our recent paper on the Data Addition Dilemma! Catch the talk on live-stream or recorded afterwards

Paper: arxiv.org/pdf/2408.04154
Workshop: simons.berkeley.edu/workshops/do...

Creative AIES 2024 paper by andreawwenyi.bsky.social that uses NLP to help uncover gender bias for men/women defendants. Legal experts used NLP to build consensus and evidence on annotation rules. Could have relevant tie-ins to healthcare and bias in clinical notes

The CHEN Lab doesn't just work on cool ML+health problems! We also enjoy viewing cacti, making pasta, and climbing on chairs

irenechen.net/join-lab/

First post! I'm recruiting PhD students this PhD admission cycle who want to work on: a) impactful ML methods for healthcare 🤖, b) computational methods to improve health equity ⚖️, or c) AI for women's health or climate health 🤰🌎

Apply via UC Berkeley CPH or EECS (AI-H) 🌉.

irenechen.net/join-lab/