Hello!
I'm writing this week's newsletter from Mexico City.
I had the opportunity to present some of the work we're doing at the University of Michigan's Neurobionics Lab at a rehabilitation research congress held at the National Institute of Rehabilitation.
I really enjoyed visiting the institute and meeting so many wonderful people. One of my favorite things about these events is getting to know people, sharing ideas, and making new connections.
Now, let's get into this week's updates!
🧬AIMedily Snaps
Fast updates clinicians should not miss.
OpenAI introduced new AI agents that can work across electronic health records and insurance systems to help with referrals, prior authorizations, and other administrative tasks. (Link)
Google DeepMind introduced AlphaGenome 2, a new AI model that predicts how genetic changes may affect different types of cells, helping study their role in disease. (Link)
Apple launched a study with OpenEvidence to understand how people use AI-generated health information. Up to 200,000 U.S. adults may participate. (Link)
Mayo Clinic developed an AI model that identified kidney stone composition using urine tests, with 73% accuracy compared with 33% using a conventional method. (Link)
BMJ Digital Health partnered with Google Cloud to give clinicians easier access to evidence-based medical information through AI tools. (Link)
🤖AIBytes
Two clinical studies that deserve a closer look.
When AI Gets It Wrong, Do Clinicians Notice?
As AI becomes more common in clinical practice, physicians are expected to recognize when its recommendations are wrong. But can we always tell when AI is wrong?
This study tested whether clinicians could recognize when an AI system became less reliable while interpreting heart sounds.
Methods
A randomized study included 48 participants, from final-year medical students to experienced physicians.
Each participant classified 40 heart sound recordings with AI assistance.
AI-First: Received the AI recommendation before making a diagnosis.
AI-Second: Made a diagnosis before seeing the AI recommendation.
The AI's accuracy was gradually reduced from 90% to 0% across four stages.
Results
Diagnostic accuracy dropped from 81.5% to 58.3%.
Making a diagnosis before reviewing the AI recommendation did not improve overall performance.
Only 12.5% of participants spontaneously reported noticing that the AI was becoming less reliable.
When directly asked, 37.5% recognized the decline.

Key Takeaways
Most participants did not notice when the AI became less reliable.
Making a diagnosis before seeing the AI recommendation did not prevent errors.
The findings show why physician review alone may not be enough to catch AI errors.
The study used simulated tasks, so we still need to understand how this happens in real clinical settings.
Schlicker N, et al. Challenges for human oversight in detecting system malfunctions in AI-assisted medical diagnostics. npj Digital Medicine. 2026. doi:10.1038/s41746-026-03410-x
AI Uses Routine Chest X-Rays to Predict Fracture Risk
Researchers tested whether AI could identify patients at higher risk of fractures using routine chest X-rays.
Methods
This retrospective study developed an AI model using chest X-rays from 53,109 adults.
The model was tested in 85,284 patients from two other hospitals in South Korea.
All participants had undergone chest X-rays and bone density testing (DXA).
Then, they compared AI with FRAX, a tool that estimates fracture risk using clinical information and bone density.
Results
AI was better than FRAX with bone density at identifying patients who later had a major osteoporotic fracture.
This advantage was seen when predicting fractures over 2, 3, 5, and 10 years.
The difference was greater in 35,358 patients with osteopenia.
Adding FRAX to the AI model did not meaningfully improve its performance.

Key Takeaways
AI could help identify patients who need further fracture-risk evaluation.
AI showed better performance than FRAX with bone density, particularly in patients with osteopenia.
All patients had already been referred for bone density testing, so we still need to know how well this approach works in patients who have not.
Kim Y, Park CM, Ko H, et al. Opportunistic Risk Assessment of Major Osteoporotic Fractures from Routine Chest Radiographs Using Deep Learning. NEJM AI. 2026. doi:10.1056/AIoa2600590
🦾TechTools
AI medical tools
AI-enabled EEG monitoring that helps clinicians screen for and monitor delirium in critically ill adults, alongside clinical assessments.
FDA-cleared AI software that helps radiologists detect and assess brain aneurysms on CT angiography.
📈 Productivity AI tool of the week:
An AI tool to help researchers screen studies, organize references, and prioritize relevant papers for systematic and literature reviews.
🧪Research Signals
New papers worth your time.
JAMA: AI-generated discharge summaries were more likely to receive physician approval with no or minor edits than clinician-written summaries (92.2% vs. 70.0%, adjusted estimates). (Paper)
Nature: AI identified a biomarker linked to longer progression-free survival in lung cancer patients receiving datopotamab deruxtecan (7.2 vs. 4.0 months). (Paper)
npj: AI used retinal images and clinical data from 8,616 pregnancies to predict preeclampsia. (Paper)
npj: AI analyzed 2.7 million lab results from 3,905 cancer patients to predict complications recorded within the next two years. (Paper)
Nature Medicine: OpenScientist, an AI research agent, completed five biomedical research analyses, including Alzheimer's and cancer research. (Paper)
That’s it for today.
Thank you for taking the time to read.
If AIMedily helped you stay up to date, please share it with a colleague who would find it useful.
Itzel Fer, MD
Forwarded this email? Subscribe free.
