Welcome back,
This week, ARISE opened applications for a one-year consortium that will bring physicians into research evaluating advanced AI systems in healthcare.
It reminded me of something we discussed at the RAISE Health Symposium at Stanford: physicians need to be part of these conversations, but not only physicians. Patients, researchers, developers, and the communities affected by these systems should also have a voice.
The UN’s first Global Dialogue on AI Governance made a similar point at a broader level: AI is too important to be shaped by only a few.
AI will keep moving quickly. The people who use it—and the people affected by it—should help shape where it goes.
Let’s dive into today’s issue.
🤖AIBytes
Two clinical studies that deserve a closer look.
Who is Better Predicting Hospital Discharge, AI or Case Managers?
This study compared discharge dates generated by AI with estimates made by case managers.
Methods
This single-center quality-improvement study included 22,349 hospitalizations involving 17,173 patients at Houston Methodist Hospital.
Researchers compared estimates from a commercial AI tool and case managers with the actual discharge date at three points:
At admission
48 hours before discharge
24 hours before discharge
The average prediction error was measured as the number of days between the estimated and actual discharge dates.
Results
At admission, the average error was similar:
AI: 4.20 days
Case managers: 4.27 days
However, case managers predicted the exact date more often than AI: 23.6% versus 15.3%.
As discharge approached, case managers became more accurate:
At 48 hours, the average error was 1.29 days for case managers and 1.59 days for AI.
At 24 hours, it was 0.98 days for case managers and 1.93 days for AI.
At 24 hours, 79.5% of case manager estimates were within one day of discharge, compared with 37.9% of AI estimates.
Key Takeaways
AI and case managers had similar average errors at admission, but case managers were more accurate as discharge approached, when reliable estimates may be most useful.
The AI performed better for patients hospitalized for 5 to 7 days, suggesting that its usefulness may vary by patient group and clinical setting.
This was a single-center study of prediction accuracy. It did not evaluate whether the AI improved hospital operations or patient outcomes. Case manager estimates were also visible to care teams and may have influenced discharge planning.
🔗 Kantheti HS, Dolan C, Dale J. Artificial intelligence-generated discharge dates and estimation accuracy in hospitalized patients. JAMA Netw Open. 2026;9(9). doi:10.1001/jamanetworkopen.2026.32033
A Predictive Mother-Child AI Agent
This Nature Medicine study developed MoChiAgent, an AI system that links maternal and infant health records to predict current diagnoses and future outcomes.
Methods
Researchers developed and tested the system using longitudinal health records from three hospitals in China:
Development data included 1,458,108 mothers and infants from two hospitals.
External testing included 90,234 mothers and infants from a third hospital.
A clinical prototype was evaluated by six perinatal specialists using 25 diagnostically enriched cases.
Results
In external testing:
The model’s ability to distinguish between patients who would and would not develop an outcome ranged from:
0.700 to 0.867 for mothers
0.801 to 0.908 for infants
AUROC measures accuracy: 0.5 is chance, while 1.0 is perfect.
Adding maternal pregnancy records improved the prediction of several infant conditions compared with using infant records alone.
Specialists gave MoChiAgent a mean diagnostic-accuracy score of 4.70 out of 5 in the 25-case evaluation.
Liu, S., Zheng, W., Kang, J. et al.
Key Takeaways
Connecting maternal and infant records may reveal health risks that could be missed when their records are analyzed separately.
The prediction model was externally validated, but the complete clinical prototype was evaluated in only 25 selected cases.
All data came from hospitals in China. Prospective studies are needed to test the system during clinical care.
🔗 Liu, S., Zheng, W., Kang, J. et al. Prediction of maternal and infant outcomes from longitudinal electronic health records with a Mother-Child AI agent. Nat Med (2026). https://doi.org/10.1038/s41591-026-04694-y
🧬AIMedily Snaps
Fast updates clinicians should not miss.
OpenEvidence introduced three models for different clinical needs: Osler for quick questions, Sackett for deeper evidence searches, and Snow for more complex reviews (Link).
OpenAI released GPT-6 Astra, its newest model, and reported stronger performance on HealthBench Professional (Link).
AI governance is going global: the UN’s first Global Dialogue put equal access, human rights, and broader participation at the center—including in health (Link).
Epic reports that patients sent 32% fewer follow-up questions when physicians used AI-assisted drafts to explain test results (Link).
The FDA is exploring how to identify medical devices that use foundation models in future updates to its device list (Link).
Microsoft Dragon Copilot can now create reusable templates for referral letters, patient instructions, work notes, and other clinical documents (Link).
🧪Research Signals
New papers worth your time.
npj: In a randomized trial across three countries, physicians performed better on clinical cases when using GPT-4o (Paper).
npj: AI used standard carotid ultrasound to identify vulnerable plaque and improve prediction of ischemic cerebrovascular events in 6,618 high-risk patients (Paper).
JAMA: GPT-5 identified adverse events in immunotherapy trial notes about as well as individual physician reviewers (Paper).
npj: Researchers reviewed 4,979 references returned by OpenEvidence across five specialties (Paper).
NEJM AI: the Centers for Medicare & Medicaid Services is removing an expedited Medicare payment pathway for FDA breakthrough devices. This could make it harder for independent clinical AI developers to compete with large EHR vendors (Paper).
Nature Medicine: Matching physicians is not enough. AI should be judged by whether physicians and AI working together improve patient outcomes (Paper).
🦾TechTools
AI medical tools
Eko SENSORA (Link) FDA-cleared AI that works with Eko digital stethoscopes to help identify signs of heart murmurs, atrial fibrillation, and low ejection fraction during an exam.
Brainomix 360 Stroke (Link) AI imaging software that analyzes brain scans to support faster assessment and treatment decisions in patients with suspected stroke.
📈 Productivity AI tool of the week:
Raycast (Link) A productivity tool for Mac that lets you open apps, search files, and use AI commands directly from your keyboard.
That’s all for today.
Thank you for taking the time to read and being part of this community.
You’re already ahead of the curve in medical AI—don’t keep it to yourself. Forward AIMedily to someone who’d find it useful. I’ll be eternally grateful.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.