Hi!
Remember last week when I mentioned that my son was bitten by a dog? Thankfully, he’s doing much better!
This week, it was our kitty’s turn. She spent two days at the veterinary hospital, but we’re happy to have her back home.
Since I’m not a veterinarian, I have to confess that I turned to AI to better understand what was going on. I had no idea how much veterinary research was out there! Claude Science was helpful in exploring the evidence and understanding her condition.
Hopefully, it will be a long time before I need to look up medical evidence for another family member or pet!
Now, let's get into this week's updates.
🧬AIMedily Snaps
Fast updates clinicians should not miss.
athenahealth launched a new connection that lets patients securely share their medical records with Claude and ask questions about their health information. (Link).
Google published a new framework for evaluating health AI, focusing on accuracy, safety, and usefulness through automated testing and expert review. (Link).
Stanford researchers revisited 50 published findings about LLMs and found that 44% no longer held in newer models, while another 34% showed weaker effects. (Link).
Nature Health announced a global commission to study the evidence, safety, and oversight of AI tools used for mental health and well-being. (Link).
ARISE Network is inviting researchers to submit 2026 clinical AI studies for its State of Clinical AI Report 2027. Deadline: December 11. (Submit here).
OpenEvidence launched Learn, a free AI-powered medical education platform with over 50,000 board-style questions, simulated patient cases, flashcards, and CME tools. (Link).
🤖AIBytes
Two clinical studies that deserve a closer look.
Can Simulation Training Improve AI Psychiatric Intake?
Researchers from Google and collaborating institutions tested whether training AI through simulated psychiatric interviews could improve clinical reasoning, communication, and safety.
Methods
They used reinforcement learning to train an AI model through simulated conversations. Then compared it with the original model using:
120 text-based consultations with psychiatrists role-playing patients.
127 psychiatrist-validated simulated intake scenarios.
256 challenging safety scenarios.
Results
The trained model received higher overall ratings in 76.7% of matched psychiatrist-evaluated cases.
Patient-centered communication: 82.0% vs. 65.2%.
Appropriate differential diagnoses: 68.3% vs. 33.3%.
Under-triage of acute risks: 12.9% vs. 30.5% in safety testing.
However, the trained model still missed at least one critical question in 60.6% of simulated intake encounters.
Schmidgall et al., arXiv, 2026. Source
Key Takeaways
Simulation training improved the AI's clinical reasoning, communication, and ability to assess psychiatric risks.
However, the model still missed important questions during intake and sometimes failed to recognize when patients needed urgent care.
The results are promising, but the study did not involve real patients. Further clinical testing is needed to understand how these systems could safely support physicians during psychiatric intake.
🔗 Schmidgall S, Strother T, Bijamov A, et al. Training language models in simulation for psychiatric intake. arXiv. 2026. doi:10.48550/arXiv.2610.40476.
AI Improves Detection of Cancer Immunotherapy Side Effects
Researchers from Harvard Medical School, Brigham and Women’s Hospital, and Dana-Farber Cancer Institute developed and tested an AI system to identify immune-related adverse events in clinical notes and evaluate whether it could help clinical research staff review medical records more quickly and accurately.
Methods
They developed an AI system to detect and assess six types of adverse events, including myocarditis, pneumonitis, and colitis.
Three phases were evaluated:
Retrospective testing using 263 clinical notes.
Prospective testing using 884 clinical notes over three months.
A randomized crossover study with 17 clinical research staff comparing AI-assisted and manual review.
Results
The AI performed well in retrospective detection, but performance decreased during prospective testing.
AI assistance reduced median review time by 43.5%.
Complete-match accuracy improved from 19.4% to 25.9%.
Agreement between reviewers also improved with AI assistance.
Gallifant et al., Nat Cancer, 2026. Figure 1.
Key Takeaways
AI helped clinical research staff review adverse events faster and more consistently.
Identifying events was easier than accurately grading their severity.
The study was conducted within one healthcare system. Further testing is needed to determine whether these tools can improve patient care and outcomes.
🔗 Gallifant J, Chen S, Shin KY, et al. An agentic AI system enhances clinical detection of immune-related adverse events in a multi-phase validation study. Nat Cancer. 2026. doi:10.1038/s43018-026-01264-6.
🦾TechTools
AI medical tools
An FDA-approved imaging system that combines optical coherence tomography with AI to help surgeons identify suspicious tissue at the margins of excised breast tumors during surgery.
An FDA-cleared AI tool that analyzes standard 12-lead ECGs for signs associated with pulmonary hypertension, helping identify patients who may need further evaluation.
📈 Productivity AI tool of the week:
Fathom
An AI meeting assistant that records, transcribes, and summarizes conversations, highlighting key decisions and action items.
🧪Research Signals
New papers worth your time.
Nature: Researchers tested an AI model on over 1.3 million medical images across six specialties. It performed well on diagnostic tasks and highlighted findings to support its predictions. (Link).
Nature: A study of 284,002 women found that AI-assisted cytology combined with HPV testing detected 96.37% of cervical precancerous lesions. (Link).
JAMA: AI-generated drafts were linked to 29.5% faster responses across 983,154 patient messages, but did not reduce overall after-hours messaging. (Link).
npj: AI reviewed about 25,000 clinical notes weekly across seven clinics and helped identify 225 missed specialist referrals. It detected 48.8% of true cases. (Link).
Sci Rep: AI helped 15 physicians interpret 69 ECGs after cardiac arrest, improving diagnostic performance but not significantly reducing overall errors. (Link).
Sci Rep: AI analyzed speech from 57 people with aphasia, identifying differences between aphasia types and changes over time. (Link).
And that wraps up this week's AIMedily!
Thank you for being here and taking the time to read.
If you enjoyed this issue, please share it with a colleague who might find it helpful too.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.