Hi!
Welcome to AIMedily.
I’m writing this week from Mexico—grateful for the sunshine and the simple joy of not wearing layers (and more layers) to stay warm in Michigan.
There’s a lot of strong research and news this week, let’s get started.
🤖 AIBytes
Researchers tested whether a LLM could help general cardiologists manage complex cardiac cases that are usually referred to subspecialists.
🔬 Methods
Study design: Blinded, randomized controlled trial.
Patients: 107 real-world cardiology cases
Clinicians: General cardiologists
AI system: AMIE, built on Gemini 2.0 Flash, adapted using prompt engineering (no medical fine-tuning).
What clinicians did:
Reviewed real patient data (ECGs, echocardiograms, CMR, Holter, CPX, genetics).
Completed assessments with or without LLM support.
Evaluation:
Blinded subspecialist cardiologists rated each assessment for errors, missing information, reasoning quality, and bias
📊 Results
Clinician preference:
Subspecialists preferred LLM-assisted assessments overall
Clinical errors:
13.1% of LLM-assisted assessments had errors
24.3% of unassisted assessments had errors
11.2% absolute reduction in clinically significant errors (P = 0.033)
Missing information:
17.8% with LLM support vs 37.4% without
19.6% reduction in missing key content (P = 0.0021)
Clinical reasoning: Equivalent quality with and without LLM support.
Clinician experience:
AI helped in 57% of cases
Saved time in 50.5% of cases
No clinically significant hallucinations in 93.5% of cases
🔑 Key Takeaways
LLM support helped general cardiologists make fewer mistakes in complex cases.
The model reduced omissions without lowering clinical reasoning quality.
Time savings were common, but not universal.
Careful oversight is needed to avoid overreliance and automation bias.
🔗 Oskotsky B, Palepu A, Tu T, et al. A randomized controlled trial of a large language model to assist cardiologists in complex clinical assessment. Nature Medicine. 2025. doi:10.1038/s41591-025-04190-9
Researchers evaluated whether a real-time machine learning model could predict delayed cerebral ischemia (DCI) after subarachnoid hemorrhage (SAH) without affecting clinical decisions, using a “silent validation” approach in the ICU.
🔬 Methods
Setting: Neurocritical Care Unit, University Hospital Zurich.
Participants: 63 patients with aneurysmal SAH.
DCI rate: 17 patients (27%).
AI model: Previously validated model generating static, dynamic, and combined DCI risk scores (0 = lowest risk, 1 = highest risk).
Evaluation process:
Three independent research physicians (not involved in patient care) reviewed AI outputs daily from day 2 to day 14 after SAH.
They estimated DCI risk, adjusted neurologic monitoring frequency, and recommended additional imaging.
Their decisions were compared with actual bedside clinical decisions and confirmed DCI events.
📊 Results
Total assessments: 794
Neurologic monitoring frequency:
Increased monitoring recommended 466 times with AI support vs. 629 without AI.
25.9% reduction in intensified monitoring.
164 fewer false-positive alerts with AI support.
No increase in missed DCI cases.
Diagnostic imaging:
128 imaging recommendations with AI vs. 153 without.
16.3% reduction in additional imaging orders.
🔑 Key Takeaways
Real-time AI risk estimates can be generated without disrupting ICU workflows.
Silent validation enables safe evaluation of AI before clinical deployment.
AI support reduced unnecessary monitoring and false positives without compromising DCI detection.
This approach may lower clinician workload while maintaining diagnostic accuracy in neurocritical care.
🔗 Willms J, Schmid J, Inauen C, et al. Silent validation of a longitudinal model for predicting delayed cerebral ischemia in real time after subarachnoid hemorrhage. NEJM AI. 2026;3(2). doi:10.1056/AIoa2500749
🦾TechTools
Portable, handheld ultrasound device that connects to a smartphone or tablet.
Uses AI guidance to help capture cardiac and lung images at the bedside.
Designed for point-of-care imaging in clinics, hospitals, and emergency settings.
AI voice agents that call insurers and pharmacies on behalf of healthcare teams.
Automates tasks like prior authorizations, benefits verification, and prescription coordination.
Reduces hold times and manual phone work, returning structured call summaries to staff.
Voice-first AI productivity assistant that captures ideas, notes, and tasks from WhatsApp, Telegram, or email.
Automatically organizes content into structured notes, tasks, and summaries inside Notion and linked apps.
Supports voice transcription, reminders, search, and automations without needing a separate new app.
🧬AIMedily Snaps
University of Michigan: An AI model that can read and diagnose a brain MRI in seconds (Link).
Oracle Health adds order creation capabilities to their clinical AI gent to support accurate, complete records (Link).
MIT, Harvard University, and MGH: AI algorithm enables tracking of vital white matter pathways (Link).
Affordable microscope speeds up malaria diagnosis with AI (Link).
University of Liverpool-US collaboration to accelerate drug discovery using AI (Link).
From the founders of Fitbit™, Luffu is the app that proactively watches over your family’s health and safety (Link).
🧪 Research Signals
Consumer-facing AI can improve public health — but more evidence is needed (Link).
Advancing healthcare AI governance through a comprehensive maturity model based on systematic review (Link).
TEMPO: Experimenting with AI Sandboxes in the United States (Link).
The Paradoxical Challenge of High-Value Medical Artificial Intelligence (Link).
Comparative Evaluation of Machine Learning and Specialist Physicians in Breast Care Triaging: A Real-World Observational Study (Link).
Evaluating LLM–Generated Clinical Summaries Through a Dual-Perspective Framework (Link).
🧩TriviaRX
Which medical innovation was first tested on a 13-year-old boy in 1796 and went on to save millions of lives worldwide?
A. The first blood transfusion
B. The smallpox vaccine
C. Use of Anesthesia
D. The stethoscope
Now, the answer from last week TriviaRX: ✅ B) Bedside cardiac monitoring
Continuous bedside cardiac monitoring was introduced in the 1960s, reducing in-hospital mortality from acute myocardial infarction.
That’s all 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 Wednesday.
Itzel Fer, MD PM&R
Forwarded this email? Sign up here
