Hi!
We have a lot of interesting AI news and research this week! OpenAI, PerturbAI, and Scripps Research used thousands of AI agents to analyze data from millions of mouse brain cells in just 32 minutes.
The agents connected genetic changes with existing research and identified a possible link between cholesterol regulation and neurodegeneration.
These findings still need to be tested, but they show the great potential of AI when used well. Not just to speed up research, but to find connections and open new directions for discovery.
Let’s dive into today’s issue!
🧬AIMedily Snaps
Fast updates clinicians should not miss.
SimBioSys received FDA clearance for an AI tool that predicts breast cancer recurrence risk in certain early-stage patients. (Link)
Radical Numerics, a biotech company, reports using AI-designed bacteriophages to treat a patient with an antibiotic-resistant infection. (Link)
Doximity Ask will soon include UpToDate's clinical content alongside AI-generated answers. (Link)
Stanford researchers used AI models of the immune system during pregnancy to identify a drug combination that may help prevent premature birth. Human trials are planned. (Link)
The American Hospital Association is asking the FDA for clearer rules on generative AI medical devices, especially around safety, transparency, and monitoring. (Link)
JAMA announced a new journal, JAMA AI, covering AI research, safety, and its use in medicine. (Link)
🤖AIBytes
Two clinical studies that deserve a closer look.
Does AI Recommend More Expensive Care Than Physicians?
Researchers compared the cost of care recommended by AI systems with physicians' treatment plans.
Methods
Tested 24 AI systems using 200 real outpatient and emergency department cases.
Compared recommended tests, referrals, and medications with physicians' plans.
Estimated test and referral costs using 2026 Medicare rates.
Results
AI-recommended care averaged $314 per visit, compared with $97 for physicians, or 3.2 times more.
AI systems recommended more tests, referrals, and medications.
In 66.5% of visits, physicians ordered no tests, while AI often suggested additional testing.
In a review of 1,000 AI plans, 18% included guideline-recommended workup missing from physicians' plans.
Asking AI to recommend fewer tests reduced costs but also lowered agreement with physicians' diagnoses.
Key Takeaways
AI may increase healthcare costs through additional testing and referrals.
Some extra recommendations were appropriate and identified potentially missed care.
The study measured estimated costs, not actual spending or patient outcomes. Physician review remains essential.
Aran D, Goldhaber-Fiebert JD, Many I, et al. Comparative cost of care recommended by large language models versus treating physicians. NEJM AI. 2026. doi:10.1056/AIoa2600537
How Well Does AI Detect Pulmonary Embolism in Hospitals?
Researchers evaluated how two FDA-cleared AI systems performed when detecting pulmonary embolism (PE) in routine clinical practice.
Methods
Retrospective study across 17 facilities in one US academic health system.
Analyzed 30,678 CT pulmonary angiograms and 37,191 routine contrast-enhanced CT scans.
Tested two AI algorithms: one for detecting PE on dedicated scans and another for finding incidental PE on routine scans.
Compared AI findings with radiology reports.
Results
On dedicated PE scans, AI achieved 86.8% sensitivity and 99.1% specificity.
For incidental PE on routine CT scans, sensitivity was 73.5% and specificity was 99.8%.
Sensitivity was higher for large saddle emboli (99.3%) than for smaller subsegmental emboli (72.9%).
Both algorithms showed lower sensitivity than reported in their FDA-clearance evaluations.
Key Takeaways
AI performance in everyday clinical practice may differ from results reported during regulatory evaluation.
Smaller and non-acute emboli were more likely to be missed, making continued radiologist review essential.
The study relied on radiology reports rather than independent image review and did not assess patient outcomes.
Mansuri A, Chavoshi M, Dapamede T, et al. Multi-site real-world performance of commercial AI for pulmonary and incidental pulmonary embolism detection. npj Digit Med. 2026. doi:10.1038/s41746-026-03418-3
🦾TechTools
AI medical tools
An FDA-authorized AI tool that helps physicians spot polyps during colonoscopy by highlighting areas they might otherwise miss.
This AI tool helps you detect wheezing and other abnormal lung sounds using a digital stethoscope, making it easier to assess patients remotely.
📈 Productivity AI tool of the week:
An AI email assistant that organizes your inbox, drafts replies in your writing style, and keeps track of emails that need a response. Works with Gmail and Outlook.
🧪Research Signals
New papers worth your time.
NEJM AI: In a trial of 300 patients, AI-generated summaries before medical visits improved patient-rated communication and shortened consultations from 33 to 25 minutes. (Paper)
JAMA: An analysis of 271 million OpenEvidence queries found that physicians used AI more often for treatment and management questions than for diagnosis. (Paper)
JAMIA: A study of 220 primary care clinicians found that AI scribes reduced note-writing time by 16% and documentation time outside work hours by 13% after five months. (Paper)
BMJ: A review of 68 clinical trials of large language models found that 69% were not preregistered, raising concerns about research transparency. (Paper)
npj: In a study of neonatal X-rays, trainees accepted 52% of incorrect AI suggestions, while three experienced specialists overrode most AI errors. (Paper)
npj: A randomized trial found that AI-assisted lung biopsies required fewer punctures and reduced pneumothorax complications from 23.9% to 12.7%. (Paper)
That’s it for today.
Thank you for reading AIMedily! If you know someone who might enjoy it too, I’d love for you to share it with them.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.