Hi!
Have you noticed how quickly AI is moving from answering questions to becoming part of the care pathway itself?
In a new Nature Health Perspective, Dr. Eric Topol and colleagues describe this shift as AI connects with medical records, scheduling, prescriptions, payments, and clinical workflows.
The conversation is becoming much bigger than model performance or whether AI will replace doctors. It is also about who controls the pathway and how these systems shape what happens next in care.
Where do you think AI can add the most value? Reply to this email—I’d love to hear your thoughts!
Now, let’s dive into today’s issue.
🤖AIBytes
Two clinical studies that deserve a closer look.
Will Physicians Keep Using AI in the ED?
In this paper, researchers tested SHAKED, a clinical decision-support system built with multiple LLMs, during real care in a tertiary emergency department.
Methods
This 4-week prospective controlled pilot included 1,138 adults who needed specialty consultation. One ED wing had access to SHAKED, which summarized records and suggested diagnoses and workup. The other provided usual care.
Results
Physicians used SHAKED in 44.3% of eligible encounters, but use fell from 67.9% in week 1 to 29.7% in week 4.
ED length of stay was unchanged: 4.9 hours in both groups.
Consultation time was 9.4 minutes shorter, but the difference was not statistically significant (P = 0.077).
No adverse events were detected. Expert review rated 99 of 100 sampled differential diagnoses as appropriate.
Key Takeaways
The AI performed well on expert review, but physicians used it less over time, especially as shifts progressed. Workflow and sustained use may matter as much as model performance.
This was a short, single-center pilot. It was not designed to prove clinical efficacy, was not powered for rare harms, and the authors say the findings do not yet justify clinical deployment.
🔗 Leibovitch L, et al. Nature Medicine. 2026. https://doi.org/10.1038/s41591-026-04601-5
Can AI Catch Liver Malignancies Missed on CT?
In this paper, researchers tested LiON, an AI system designed to detect and classify liver malignancies on contrast-enhanced CT.
Methods
Researchers trained LiON on 6,443 patients and retrospectively validated it in 22,251 patients across multiple centers and real-world cohorts.
They then tested it as an additional AI reader during routine care in a prospective single-arm trial of 10,333 patients.
Results
LiON showed strong ability to distinguish malignant from non-malignant cases, with an AUC of 0.952.
AI-human review found 51 previously overlooked lesions, including 15 small metastases.
This led to 37 amended radiology reports and 22 multidisciplinary team reviews.
Four patients had immediate treatment changes after the AI findings showed it more advanced than initially recognized.
No AI-attributable adverse events were reported.
Key Takeaways
AI detected liver lesions that had been missed during routine radiology review.
In some cases, those findings changed patient management.
The AI worked as a second reader with mandatory human review, not independently.
The prospective study was single-center and had no control group. Validation in other healthcare systems is still needed.
🔗 Zhang X, Li C, Han X, et al. Large-scale AI-guided liver malignancy diagnosis: multicenter study and a single-arm trial. Nature Medicine. 2026. https://doi.org/10.1038/s41591-026-04589-y
🧬AIMedily Snaps
Fast updates clinicians should not miss.
72% of U.S. adults want providers to tell them when AI is used in their healthcare (Link).
OpenEvidence launches Patient Take-Homes, so physicians can share selected answers with patients as education materials (Link).
Oracle adds chart review, dictation, and coding to its Clinical AI Agent (Link).
Google Research introduces GlucoFM, a foundation model built to learn from continuous glucose monitoring data (Link).
HIMSS is developing an AI Outcomes Framework to measure the real-world impact and value of AI across health systems (Link).
Epic shares a major AI expansion including Ergo Visit, Agent Factory, and Curiosity for predicting future patient outcomes (Link).
🧪Research Signals
New papers worth your time.
Can language models help families make treatment decisions that reflect a patient’s values? (Paper).
Nature Health: Patient factors considered in medical AI research, including trust, safety, and patient involvement (Paper).
JAMA Surgery: Deep learning for detecting surgical site infections from postoperative wound images (Paper).
npj: Diagnostic performance of multimodal models in neuroradiology using open-ended questions vs answer choices (Paper).
npj: HealthFlow, a multi-agent system for automating complex EHR analysis (Paper).
Lessons from deploying the ChatEHR system at Stanford Medicine (Paper).
🦾TechTools
AI medical tools
Limbic (Link)
An AI tool used in mental health services to help with intake, assessment, and triage before patients are seen by the clinical team.
Neko Health (Link)
A preventive health scan that combines imaging, cardiovascular measurements, body composition, and blood tests, followed by a physician review of the results.
📈 Productivity AI Tool of the Week:
Sunsama (Link)
A daily planning tool that brings your calendar and tasks together and helps you organize your day without overloading it.
That’s all for today. 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! I’ll be eternally thankful 😊.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.