📖Read more: arxiv.org/abs/2506.10157
🧠Thanks to my amazing team: Ben Reis, @adamrodmanmd.bsky.social, Tianxi Cai, Noa Dagan, @ranbalicer.bsky.social, Joseph Loscalzo, Isaac Kohane, @marinkazitnik.bsky.social @harvardmed.bsky.social @hsph.harvard.edu @broadinstitute.org @bostonchildrens.bsky.social
Models must not only generalize across typical care settings but also adapt to the unique needs of specific users, institutions, geographies, diseases, & populations.
✍️We outline concrete strategies on three contextual vignettes to develop and evaluate context-switching models.
Medical AI must operate as ✨adaptable systems✨ that respond to variation as it arises.
🫀AI trained on adult cardiology should recognize patterns relevant to pediatric cases & adjust reasoning
🧑⚕️Documentation assistant should shift tones when addressing patients vs clinicians
We argue that ✨ #contextswitching ✨ is a core paradigm in medical AI.
It's a model’s ability to adjust its reasoning & outputs in real-time based on shifts in clinical specialty, patient population, local practices, etc.
❌ Acquire new data
✅ Apply existing knowledge flexibly
Medical AI operates via pattern recognition. So outputs resemble rote memorization rather than flexible, context-aware reasoning.
This leads to contextual error❌ where outputs appear appropriate based on the medical record but are flawed due to missing or misunderstood context.
🚨Medical AI’s fatal flaw exposed?🚨
Medical AI has shown promise in streamlining clinical workflows & decision-making. But they don't reason like human clinical domain experts
Are we killing patients with “correct” diagnoses applied to the wrong contexts?
🧵 #contextswitching
You think you can destroy science? Well, many scientists and I are just doubling our efforts to train the next generation.