Hi!
My daughter turned 17 this week, which still feels a little hard to believe.
When I think about what the world looked like when I was 17, it was completely different. Kids and teens today are growing up with technology changing so much of how they learn, communicate, and now even how they may experience healthcare.
Ethan Goh shared a post this week about something I think is becoming really important: AI needs to keep moving forward, but the evidence needs to move with it.
There is so much potential to improve access and quality of care, but we are still learning what actually works.
It makes me wonder what healthcare will look like by the time my daughter is my age. I feel hopeful about what is coming, but also a little uneasy about how much we still don’t know.
Let’s dive into today’s issue.
🤖AIBytes
Two clinical studies that deserve a closer look.
Can AI Automate Cervical Cancer Triage?
A positive HPV test often needs another step to decide who should have further evaluation.
Researchers tested an AI system that reads p16/Ki-67 dual-stain slides, a test used to help identify people at higher risk of cervical cancer.
🔬 Methods
Cytoreader-Fusion was trained on 1,919 cases.
It was tested on 9,006 cases across four blinded groups.
The strongest test included 4,996 cases evaluated only after the model was fully locked.
Performance was compared with biopsy-confirmed CIN 3 or worse, not simply agreement with another reader.
📊 Results
In the independent group, the AI reached 88.3% sensitivity and 54.4% specificity.
Manual dual-stain reading reached 78.1% sensitivity and 63.2% specificity.
Pap cytology reached 74.5% sensitivity and 45.9% specificity.
A preset designed for higher specificity reached 85.8% sensitivity and 61.7% specificity.
🔑 Key Takeaways
The standard AI setting detected more CIN 3+ cases than manual dual-stain reading, but with lower specificity.
Another preset came close to manual specificity while still detecting more CIN 3+ cases.
The system is already running alongside the routine Kaiser Permanente laboratory workflow, but it does not make clinical decisions.
The study used archived slides. Prospective studies are still needed to see how it affects referrals and patient care.
🔗 Lahrmann B, Keil A, Miranda Ruiz F, et al. Closing the Automation Gap in HPV-Based Cervical Cancer Screening: Independent External Validation of an AI Model for Dual-Stain Triage. NEJM AI. 2026;3(10). doi:10.1056/AIoa2600270
Do Junior Clinicians Recognize AI Hallucinations?
Many clinical AI systems still depend on a physician reviewing the output before it is used in patient care. In this paper, the researchers tested how well junior clinicians could identify hallucinations in AI-generated clinical responses.
🔬 Methods
Multicenter cross-sectional study across four hospitals in China.
137 junior clinicians reviewed GPT-4o responses.
The assessment included 37 questions from 16 simulated clinical cases.
Clinicians marked each AI response as agree, disagree, or uncertain.
📊 Results
Junior clinicians identified only 15.8% of the hallucinations.
13.1% did not identify any hallucinations.
Detection did not improve as clinical risk increased.
Differences between clinicians explained more of the variation in detection than differences between clinical scenarios.
🔑 Key Takeaways
Human review alone may not reliably catch AI errors.
Higher-risk errors were not more likely to be recognized.
Detection ability varied substantially between clinicians.
🔗 Zhang Y, Zhou J, Zhang Z, et al. A multicenter assessment of human oversight of generative AI outputs in simulated clinical decision making. npj Digital Medicine. 2026. doi:10.1038/s41746-026-03294-x
🧬AIMedily Snaps
Fast updates clinicians should not miss.
OpenAI introduced Dots, AI agents that can keep working on a task and take actions across connected apps. (Link)
Blue Cross says AI coding tools may be adding more secondary diagnoses to hospital bills without a similar change in the care patients receive. (Link)
An OpenAI agent accessed an Australian Medicare statistics site without authorization, raising new questions about what happens when AI systems can act on their own. (Link)
Doximity released an open benchmark with 500 physician-reviewed cases to test clinical AI on safety, calculations, hallucinations, and other common failure points. (Link)
OpenEvidence is adding research from more than 300 Sage journals, giving physicians more peer-reviewed sources directly inside the platform. (Link)
Heidi launched new AI agents that can help with pre-charting, referrals, follow-up tasks, and other work around the visit, with clinicians reviewing what goes out. (Link)
🧪Research Signals
New papers worth your time.
npj: AI-assisted preoperative communication reduced anxiety, physician workload, and communication time in a randomized prostate cancer trial. (Paper)
BMJ: Patients described what they need to trust AI decision support used during ophthalmology care. (Paper)
JAMIA: A real-time model used EHR data during cardiac surgery to update the risk of complications as the operation progressed. (Paper)
JAMIA: An AI system trained on more than 700,000 primary-care visits generated diagnostic and testing recommendations for older adults. (Paper)
npj: A locally hosted LLM retrieved information from longitudinal medical records with up to 95.3% accuracy. (Paper)
JAMA Network Open: An LLM was used to detect changes in primary outcomes across registered clinical trials. (Paper)
🦾TechTools
AI medical tools
Kardi AI (Link)
Uses a wearable ECG chest strap and AI to monitor heart rhythm over longer periods and flag possible abnormalities for clinical review.
Canary Ambient (Link)
Works with Dragon Copilot to analyze the patient conversation for cognitive and behavioral health signals while the visit is being documented.
📈 Productivity AI tool of the week:
Lindy (Link)
Lets you create AI assistants that handle recurring tasks across email, calendars, Slack, and other work tools.
That’s all for today.
Thank you for being part of AIMedily!
If it helped you stay up to date this week, feel free to share it with a colleague who might enjoy it too.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.
P.S. Enjoying AIMedily? 👉 Write a review here (it takes less than a minute).