Hi!
This week, Dr. Robert Wachter shared a thoughtful perspective on the debate about slowing AI development. His point is clear: concerns about future, more powerful models should not hold back the careful evaluation and use of today’s AI tools.
You see the same tension in Dario Amodei’s (Claude co-founder) recent essay, where he calls for “pacing the frontier”—slowing the race toward ever-more-capable AI while safety and alignment catch up.
To me, these views fit together. In healthcare, we should move carefully with the tools we already have, even as people debate what comes next in AI. Our priority stays the same: protecting patients and improving outcomes.
This week’s studies show where AI may help and where closer physician review is still needed.
Let’s dive in.
🤖AIBytes
Two clinical studies that deserve a closer look.
Can AI Spot Fetal Brain Malformations on Ultrasound?
Prenatal ultrasound depends heavily on operator experience. Researchers tested whether real-time AI support could help sonographers detect fetal brain malformations more accurately.
Methods
The randomized trial included 1,584 ultrasound scans from high-risk pregnancies across five centers in China.
Sonographers evaluated each scan with and without an AI system called PAICS.
An expert panel reviewed the ultrasound videos as the reference standard.
The system was evaluated on 10 fetal brain malformations it was designed to recognize.
Results
With AI assistance, sensitivity improved:
From 78.6% to 87.3% when evaluated by fetus.
From 69.6% to 81.4% when evaluated by individual malformation.
Specificity remained similar, meaning the improvement in detection did not meaningfully increase false-positive findings.
No adverse events related to the ultrasound procedure or AI assistance were reported.
🔑 Key Takeaways
AI helped sonographers detect more of the 10 targeted fetal brain malformations.
The improvement came without a meaningful reduction in specificity.
The trial involved high-risk pregnancies and experienced sonographers in China, so the findings may not apply to routine prenatal screening or other clinical settings.
Lei T, Liao J, Gao J, et al. Evaluating AI-assisted detection of fetal intracranial malformations in prenatal ultrasound practice: a multicentre, self-crossover, randomised controlled trial in China. Lancet Digit Health. 2026. doi:10.1016/j.landig.2026.101040
Medical AI That Runs Inside Hospitals and Shows Which Cases Need Human Review
Hospitals may want to run medical AI locally so patient data remain within their own systems. Researchers tested whether locally hosted AI agents could approach the performance of a cloud model—and whether repeated answers could help identify unreliable diagnoses.
Methods
The AI agents completed simulated clinical encounters using three datasets.
The main dataset included 551 cases covering seven diagnoses.
The AI interviewed a simulated patient, used clinical tools, and generated a diagnosis.
Each case was repeated five times to measure whether the system consistently reached the same diagnosis.
Results
On the primary dataset:
The best locally run model was correct in 90.0% of cases, compared with 90.7% for the cloud-based GPT-5.2 model.
Consistency across repeated runs was better than the model’s own confidence score at separating correct from incorrect answers.
When researchers kept only cases with very high consistency, the agent retained 49.4% of cases and was correct in 98.9% of that selected group.
🔑 Key Takeaways
A locally hosted medical AI agent approached the cloud model’s performance on the primary benchmark.
Checking if the AI gives the same answer each time may help spot cases that are less reliable and should be reviewed more closely.
This was a retrospective simulation—not a clinical deployment. Repeated runs also required approximately five times more computing.
Zhang L, Wölflein G, Ferber D, et al. On-premise medical AI agents for reliable clinical decision-making. Nat Med. 2026. doi:10.1038/s41591-026-04609-x
🧬AIMedily Snaps
Fast updates clinicians should not miss.
ARPA-H is funding a Duke-led consortium to test AI agents that could help monitor patients with heart failure. (Link)
How should AI earn your trust? A UK commission recommends approval in stages and ongoing safety monitoring of AI medical devices. (Link)
Penn researchers are exploring potential antimicrobial molecules, with ChatGPT and Codex supporting early discovery work. (Link)
Oracle’s new AI agent for nurses offers patient summaries, voice-based chart searches, and EHR documentation. (Link)
Teams from 25 health systems are building Epic AI agents for medication reviews and more. (Link)
GE HealthCare’s new AI tool is designed to predict hospital bottlenecks up to 72 hours ahead. (Link)
🧪Research Signals
New papers worth your time.
Nature Medicine: Prospective AMIE studies revealed challenges involving clinical workflows, patient trust, and integration (Paper).
npj Digital Medicine: Researchers prospectively compared machine learning models with clinical scores to predict Gram-negative bloodstream infections in ICU patients. (Paper)
Nature: Paper2Agent turns papers, code, and data into AI agents that can reproduce analyses and answer questions (Paper).
npj: A review examined LLM-supported patient education and found that evidence remains limited (Paper).
JAMA: An LLM tracked primary-outcome changes across more than 15,000 registered clinical trials (Paper).
Nature Human Behaviour: A Perspective proposes designing health AI around comprehension, agency, accountability, and risk (Paper).
🦾TechTools
AI medical tools
Isabel Pro (Link)
A clinical decision-support tool that generates a differential diagnosis from a patient’s age, sex, symptoms, signs, and test results.
AlphaGenome Atlas by Google (Link)
A free, searchable map predicting the regulatory effects of all 9 billion possible single-letter DNA changes, helping physicians prioritize variants tied to disease.
📈 Productivity AI tool of the week:
Mantle Chat (Link)
Brings team chat, multiple AI models, shared agents, and connected tools into one collaborative workspace.
That’s it for today.
Thank you for being part of AIMedily.
You’re already ahead of the curve in medical AI—if you know someone who’d appreciate it, feel free to forward AIMedily their way.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.
P.S. If AIMedily has been helpful, would you consider leaving a quick note about what you’ve found useful? 👉 It takes less than a minute. Thank you!