Hi!
AI is moving quickly from pilot projects into everyday care. But the evidence that it improves patient outcomes is not moving nearly as fast.
A new PLOS Digital Health analysis found that among 1,357 AI medical devices cleared or approved by the FDA, only three were linked to studies evaluating patient outcomes.
But this is no longer only about how well the technology performs. It is also about who benefits, who makes the decisions, and how AI will impact people.
Bill Gates argues that these decisions should not be left to AI companies alone, but shaped by policymakers, educators, health workers, and communities—with stronger support for those most affected.
How do we move quickly enough to benefit from AI without getting ahead of the evidence—or the people it will affect?
Let’s dive into today’s issue.
🤖AIBytes
Two clinical studies that deserve a closer look.
Which AI-Drafted Messages Take Physicians More Time?
This study examined physician edits to AI-drafted patient messages and their link to response time.
Methods
This quality improvement study identified 14,350 AI-assisted replies from 1,131 physicians at UC San Diego Health.
An expert-validated language model grouped the changes into 15 overlapping categories. The adjusted analysis included 11,344 messages from 281 physicians. Response time served as a proxy for workload.
Results
Scheduling was the most common edit, followed by lifestyle advice and empathy or emotional support.
All 15 edit types were linked to longer response times. The largest increases involved radiology-result interpretation (70.1%), diagnostic clarification (63.9%), and lab-result interpretation (60.8%).
Scheduling was linked to 18.2% more time per message. Together, these edits accounted for the most additional physician time.
Most edits changed clinical meaning without changing the care recommendation.
Key Takeaways
Edits requiring more clinical judgment were linked to more time. Yet common, shorter edits also added up across the health system.
Because the study did not compare AI-assisted replies with replies written from scratch, it cannot show whether AI saved time overall.
It was observational, came from one health system, and measured only one part of workload.
🔗 Poursoltan L, Cao J, Chen W, et al. Physician edits to AI-drafted patient messages and their impact on clinical workload. NEJM AI. Published online August 27, 2026. doi:10.1056/AIoa2501034
Can Novices Use AI-Guided Ultrasound to Screen for Aortic Stenosis?
This study asked if novices could use AI-guided handheld ultrasound to help find moderate or greater aortic stenosis.
Methods
Researchers developed and validated a deep learning model.
They then tested it prospectively in 1,302 adults referred for echocardiogram. Nine research staff had no clinical or ultrasound experience. They received four hours of training, then used AI guidance to collect handheld heart images.
Results
AI could assess 96.6% of scans. Sensitivity was 93.5%. Specificity was 96.4%. Positive predictive value was 49.4%, and negative predictive value was 99.7%.
Experts reviewed about 10% of scans. This raised positive predictive value to 91.1%, but lowered sensitivity to 85.4%.
The hybrid workflow missed 7 of 48 cases. It would have referred 4.8% of participants for a comprehensive echocardiogram.
Key Takeaways
With AI guidance, novices collected scans suitable for automated assessment in most patients. This approach may help with triage where access to echocardiography is limited.
But the prospective group was clinically referred at one Mayo Clinic center. It was not a community screening group.
The study did not test patient outcomes, resource use, or costs. Comprehensive Doppler echocardiography is still needed for diagnosis.
🔗 Lee E, Naser JA, Kane CJ, et al. Artificial intelligence–enabled acquisition and interpretation for screening aortic stenosis. JAMA Cardiol. Published online August 28, 2026. doi:10.1001/jamacardio.2026.3829
🧬AIMedily Snaps
Fast updates clinicians should not miss.
If your organization uses Epic, ChatGPT for Healthcare can now review authorized patient information, but it still cannot write back into the EHR (Link).
AI is moving beyond the screen: Anthropic’s new standard allows agents to operate laboratory equipment such as microscopes and liquid handlers (Link).
Your patients may soon be able to ask questions about their health records through Emmie, Epic’s AI assistant inside MyChart (Link).
Hospitals using Bayesian Health’s AI sepsis monitor may receive additional Medicare reimbursement beginning October 1 (Link).
Owkin: Boehringer Ingelheim will use its AI Scientist to help develop new cancer and immunology treatments (Link).
WHO new recommendations offer practical guidance for building stronger governance into the health system (Link).
🧪Research Signals
New papers worth your time.
NEJM AI: Agentic AI Teammates in Medical Research — From Tools to Collaborators — and the Accelerating Digital Divide (Paper).
npj: A locally deployed open-source LLM extracted structured kidney-cancer data from more than 130,000 clinical notes, with 97.5% agreement with a manually curated database (Paper).
npj: Generative deep learning reconstructed deep-brain signals from cortical recordings in people with movement disorders, a possible step toward more reliable closed-loop brain stimulation (Paper).
The Lancet: A contrastive learning foundation model (ECG‑CLIP) improved ECG‑based prediction of cardiovascular diseases and outcomes (Paper).
JAMA: In a study of nearly 40,000 students, 21.1% used generative AI for emotional support, and this was associated with more emotional problems (Paper).
The Lancet: A Comment examines what patients may not be told when ambient AI records and processes their clinical conversations (Paper).
🦾TechTools
Medical AI tools:
Anumana ECG‑AI Pulmonary Hypertension (Link)
AI algorithm that analyzes a standard 12‑lead ECG to estimate the likelihood of pulmonary hypertension, helping flag at‑risk patients earlier for confirmatory testing.
Lunit INSIGHT MMG (Link)
AI software for mammography that helps radiologists detect breast cancer earlier and reduce false positives, recently cleared by the FDA.
📈 Productivity AI tool of the week:
Undermind (Link)
AI research assistant that builds living literature reviews, continuously updating summaries as new papers are published in your topic area.