Hi!
This has been a stressful week. My son was bitten by a dog, and we've spent the last few days going to medical appointments and dealing with a lot of uncertainty.
Even though I'm a physician, when it comes to my own child, I'm also a worried mom.
One thing that helped was having access to OpenEvidence. I used it to check the latest guidelines, which gave me some reassurance.
Being able to review the guidelines and discuss them with the doctors, while considering our specific situation, helped us make informed decisions.
Having access to reliable medical information really matters, but having a physician to discuss it with makes all the difference!
Let’s dive into today’s issue.
🧬AIMedily Snaps
Fast updates clinicians should not miss.
A Yale study finds that 71% of surveyed family medicine residents used AI at work, but only 36% of AI users reported recent training (Link).
HHS removes AI transparency and some risk-management requirements from federal health IT certification, raising concerns about how clinicians evaluate these tools (Link).
Cleveland Clinic reports that teams using AI for clinical trial matching enrolled 40% more patients, 47% faster on average (Link).
Stanford researchers use AI-powered digital twins to identify a potential drug combination for preventing preterm birth (Link).
OpenEvidence partners with the American Society of Addiction Medicine to make addiction treatment guidelines available through its platform (Link).
UCLA researchers use AI to screen 3,592 compounds, with laboratory testing identifying 23 promising combinations for prostate cancer research (Link).
🤖AIBytes
Two clinical studies that deserve a closer look.
What if AI could ask questions, order tests, and make a diagnosis step by step?
A new study in Science tested an AI system designed to do just that, with promising results.
Methods
Microsoft researchers developed Diagnostic Orchestrator (DxO), an AI system that gathers information and updates diagnoses as new findings become available.
They tested it using:
304 complex cases from NEJM.
325 emergency department cases from MIMIC-IV-ED.
Researchers compared diagnostic accuracy and estimated testing costs with those of practicing generalist physicians and other AI models.
Results
Complex cases: DxO achieved 83.2% accuracy, compared with 36.1% for physicians using reference tools.
Emergency cases: DxO achieved 80.3% accuracy, compared with 46.8% for physicians.
Testing costs: With budget limits, DxO achieved 80.9% accuracy at an estimated cost of $2,241 per complex case, compared with 36.1% accuracy and $4,725 for physicians using reference tools.
DxO generally asked more questions and ordered fewer tests than physicians.
Key Takeaways
AI showed strong diagnostic performance using a step-by-step approach.
Budget limits reduced estimated testing costs while maintaining high accuracy.
Results came from a controlled, text-based benchmark, not real clinical care.
More research is needed to assess performance and safety in clinical practice.
Nori H, Daswani M, Kelly C, et al. Sequential diagnosis with language models. Science. 2026;394(6822):423-428. doi:10.1126/science.aeh7930.
AI can help doctors make better decisions. But what happens when they trust incorrect advice or ignore advice that is right?
A new study in explored how physicians respond to AI advice when reading brain CT scans.
Methods
Researchers conducted a randomized crossover experiment with 125 participants from 10 countries, including 50 radiologists and 75 other physicians or medical trainees.
Participants reviewed 50 head CT cases under three conditions:
No AI advice.
Basic AI advice indicating whether intracranial hemorrhage was present.
AI advice with visual annotations.
AI advice was correct in 80% of cases. Researchers measured diagnostic accuracy, confidence, reading time, and how participants responded to the advice.
Results
Physicians sometimes followed incorrect AI advice and ignored correct advice.
Ignoring correct AI advice had a greater overall impact on diagnostic accuracy.
Correct AI advice with visual explanations improved diagnostic accuracy compared with no AI advice.
Visual explanations helped reduce missed intracranial hemorrhages, particularly among less experienced participants.
Incorrect AI advice reduced diagnostic accuracy, and visual explanations did not significantly reduce reading time.
Key Takeaways
AI can improve diagnostic accuracy, but incorrect advice can also lead to errors.
Ignoring correct AI advice can also affect diagnostic accuracy.
Visual explanations can help, but they don't prevent all errors.
We still need to learn how to use AI safely in everyday clinical practice.
Gaube S, et al. Examining reliance patterns on AI advice in medical imaging: a mixed-methods, randomised, crossover experiment. Lancet Digit Health. Published online October 23, 2026. doi:10.1016/j.landig.2026.101050.
🧪Research Signals
New papers worth your time.
Nature: AI diabetic eye screening reached completion rates above 98% in a study of 397 young people with diabetes, in both racial groups studied (Paper).
npj: AI slightly improved fracture detection in a prospective emergency department study, but performed worse than radiographers when used alone (Paper).
Nature: A multimodal AI model trained on 1.3 million samples showed promising results across 10 medical imaging modalities (Paper).
JAMIA: AI scribes were linked to more patient instructions, but the instructions were longer and harder to read (Paper).
JAMA: Researchers compared ChatGPT, Gemini, and Claude on clinical and image-based ophthalmology questions (Paper).
npj: A pilot study of 58 medical students found no significant difference in clinical skills improvement between AI and human standardized-patient training (Paper).
🦾TechTools
AI medical tools
📈 Productivity AI tool of the week:
Uses AI to analyze digital cognitive tests and help clinicians detect early signs of cognitive impairment.
An AI symptom checker that asks patients questions and suggests possible conditions and next steps for care.
Productivity Tool of the Week
An AI assistant that records and summarizes meetings and conversations, helping you keep track of ideas and important information.
That’s it for today.
Thank you for taking the time to read.
Know someone interested in AI in medicine? Share AIMedily with them so we can all keep learning together.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.