Hello,
Welcome back!
This week I’m writing from Mexico, and the World Cup energy here is amazing.
It has also been a busy week in medical AI.
There was a lot of discussion after the 2 new papers on AI agents in medicine. One of the papers I shared last week compared clinical AI tools, including OpenEvidence and UpToDate Expert AI, with frontier LLMs. OpenEvidence responded quickly, which made the conversation even more interesting.
This week, I’m sharing another paper that looks at AI agents working inside simulated clinical workflows.
And another topic that got a lot of attention: Midjourney, best known for AI image generation, is now moving into health with a proposed full-body ultrasound scanner.
A lot to unpack.
Let’s dive into today’s issue.
🤖 AIBytes
Two clinical studies that deserve a closer look.
AI Agent Works Inside Simulated EHR
Researchers tested MIRA, an AI agent built to work inside a simulated electronic health record that could take a patient case from history to diagnosis, tests, treatment, and admission decisions.
Methods
The study used 574 real patient cases from MIMIC-IV.
Cases covered 8 emergency diagnoses.
MIRA could ask questions, order tests, read results, prescribe medications, request procedures, and make admission plans.
It was compared with 4 board-certified physicians and a second group of 6 physicians with mixed experience.
The study also tested safety, medication orders, guideline use, and patient-agent robustness.
Results
MIRA reached 88.9% diagnostic accuracy.
MIRA had higher diagnostic accuracy than board-certified physicians:
87.8% vs. 78.1%
Captured home medications with:
95.2% recall
99.6% precision
Matched more reference procedures than physicians:
53.5% vs. 38.3%
Showed stronger guideline adherence in many medication categories.
In a safety review of 56 cases, no high-severity drug interactions, renal dosing errors, allergy-medication mismatches, QT-risk prescribing, or unsafe opioid prescribing were found.
Three cases had therapeutic duplication.
Key Takeaways
This study moves beyond a chatbot. MIRA worked inside a simulated EHR and could take clinical actions.
MIRA followed a care workflow from history and testing to diagnosis, treatment, and admission planning.
The strongest signals were diagnostic accuracy, medication reconciliation, and guideline-aligned decisions.
The authors emphasize that medical AI agents should support clinicians, not replace them.
🔗 Ferber D, Hilgers L, Höper C, et al. Towards autonomous medical artificial intelligence agents. Nature. 2026. doi:10.1038/s41586-026-10675-5
AI Learns from Surgical Videos While Keeping Privacy
This study tested whether hospitals could train AI models on surgical videos without sharing the videos themselves.
Methods
Researchers used 397 laparoscopic surgery videos from six international centers. Most were appendectomy cases.
The models were trained to predict:
Appendix perforation
Laparoscopic severity
Inflammation grade from pathology reports
They compared three approaches:
Models trained at one hospital
Models trained with all hospitals data pooled together
Swarm Learning, where hospitals kept data locally and shared only model updates
Results
Swarm Learning performed similarly to models trained with all the data in one place.
It performed better than models trained by a single hospital alone.
The strongest results were seen for identifying perforated appendicitis from surgical video.
Performance was lower when predicting pathology-based inflammation grades, suggesting that some information may not be fully visible from video alone.
Key Takeaways
Hospitals may be able to collaborate on surgical AI without sharing raw surgical videos.
Weak supervision allowed models to learn from clinical labels in the patient record.
This could reduce manual labeling and make surgical video AI easier to scale.
This was still proof of concept. It did not show improved surgical outcomes or patient benefit yet.
🔗 Saldanha OL, Pfeiffer K, Bodenstedt S, et al. Privacy-Preserving Surgical Video Analysis with Swarm Learning - Results from a Multinational Appendectomy Cohort. NEJM AI. 2026. doi:10.1056/AIoa2501116.
🧬 AIMedily Snaps
Medical AI updates worth knowing this week.
Midjourney Health is working on a scanner that uses underwater ultrasound sensors to create fast body images (Read more).
The WHO has published: Artificial intelligence and evidence-informed policy – emerging challenges and opportunities (Link).
How Doximity Ask Answers Clinical Questions (Learn more).
OpenAI: Using AI to help physicians diagnose rare genetic diseases affecting children (Link).
Philips Future Health Index 2026: AI is already saving clinicians time and delivering measurable impact in healthcare (Read more).
GPT-5.5 Instant now matches frontier thinking models on health tasks and is available to all free ChatGPT users (Link).
🧪 Research Signals
New papers worth your time.
NEJM AI: SmartAlert helps reduce unnecessary inpatient labs (Paper).
Nature: Automated reanalysis of genomic data for rare disease diagnostics (Paper).
JMIR: A multicenter study to predict early hospital admission after stroke using machine learning (Paper).
JAMA: AI note summarization in the emergency department (Paper).
Nature: Privacy risks in medical AI may affect patients unequally(Paper).
JAMA: AI helps generate urine drug test sign-outs (Paper).
🦾 TechTools
AI Tools clinicians may want on their radar.
Is a clinical decision support platform for critical care monitoring.
It brings together real-time physiologic data and risk analytics to help ICU teams track patient status.
Useful for pediatric and adult critical care teams.
AI clinical documentation and insights platform.
It reviews data in the chart to recommend diagnoses and generate draft documentation at the point of care.
Useful for inpatient teams that want to reduce documentation burden and avoid missing important clinical context.
📈 Productivity AI tool of the week:
You can use Gumloop to build AI-powered workflows without code.
It helps connect steps like web research, document review, data extraction, and email drafts.
You can use its visual builder to automate repetitive work across different apps.
🧩 TriviaRX
A quick question to test your knowledge.
What was the original name given to ether anesthesia before physicians realized it was ether?
A) Letheon
B) Laudanum
C) Chloroform
D) Nitrous oxide
Now, the answer from last week’s TriviaRX: ✅ B) Tuberculosis
The 1948 streptomycin trial for pulmonary tuberculosis helped shape the way modern clinical trials are designed.
That’s it for today.
As always, thank you for taking the time to read.
You’re already ahead of the curve in medical AI — don’t keep it to yourself. Forward AIMedily to your colleagues who’d appreciate the insights.
Until next week.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.
P.S. Enjoying AIMedily? 👉 Write a review here (it takes less than a minute).
