🤖AIBytes
Two papers that deserve a closer look.
AI-Triggered Rapid Response Linked to Lower Mortality
Researchers assessed whether linking the Epic Deterioration Index to automated rapid response team alerts was associated with lower inpatient mortality.
Methods
This staggered, quasi-experimental study included 23,132 high-risk adult medical–surgical admissions across 11 hospitals from October 2022 to August 2024.
Patients were included if their Epic Deterioration Index score reached 60 or higher.
The intervention combined:
Risk scores updated every 15 minutes
Visible electronic health record alerts
Automated mobile notifications to rapid response teams
Staff education and standardized clinical workflows
The primary outcome was risk-adjusted in-hospital mortality.
Results
After implementation:
Rapid response activations increased: 25.3% to 37.5%
In-hospital mortality decreased: 23.1% to 18.6%
After adjustment, patients had 18% lower odds of dying in the hospital
Higher-level care transfers did not significantly increase
Mortality reductions were statistically significant in academic and community teaching hospitals, but not in nonteaching community hospitals.
Nahass et al. NEJM AI. 2026. doi:10.1056/AIoa2500973.
Key Takeaways
The AI score was linked directly to rapid response alerts, helping care teams act sooner when patients were at high risk.
The intervention was more than the model itself. It also required clear alerts, staff education, and integration into clinical workflows.
This suggests that how AI is implemented may be just as important as how well the model performs.
Because the study was not randomized, it shows an association but cannot prove that the intervention caused the lower mortality.
Nahass TA, Hanna JS, Liu N, et al. Implementation of an AI-triggered rapid response—association with mortality. NEJM AI. Published online July 29, 2026. doi:10.1056/AIoa2500973.
What Must Medical AI Prove Before Acting Alone?
This important paper brings together authors from academic medicine, major AI companies, and a U.S. federal health research agency to address a growing question:
What should AI prove before we allow it to act more independently in patient care?
Strong performance on medical tests does not necessarily mean that an AI system is ready for real clinical care. The authors propose the Medical AI Superintelligence Test, or MAST, to evaluate AI across clinically meaningful areas, including:
Diagnosis and treatment decisions
Safety and harm prevention
Multimodal interpretation
Long-term and complex care
Use of medical records
Decisions under uncertainty
The authors argue that medical superintelligence should not simply mean outperforming the average physician. AI should outperform teams of the best specialists across a range of meaningful clinical tasks.
How AI is tested also matters. A model can appear stronger or weaker depending on the information, tools, and type of interaction it is allowed to use.
For example, ChatGPT Health undertriaged 52% of emergencies after only one patient message. In a later, more natural conversation-based evaluation, accuracy improved from 48% to 80% in the scenario most associated with undertriage.
In MedAgentBench, the best-performing AI agent successfully completed only 69.67% of 300 clinical tasks in a simulated electronic health record.
Key Takeaways
High benchmark scores do not prove clinical readiness.
Performance depends on the task, setting, and evaluation design.
AI should be compared with expert clinical teams, not only average physicians.
Evaluations must move beyond exam-style questions to real clinical workflows and long-term care.
Real-world safety and patient outcomes should ultimately become the standard.
Goh E, Wu D, Walton C, et al. Toward a test of medical AI superintelligence. Nat Med. Published online 2026. DOI: s41591-026-04539-8
🧬AIMedily Snaps
Fast updates clinicians should not miss.
OpenAI begins rolling out Health in ChatGPT to eligible U.S. adults, with optional connections to supported medical records and Apple Health (Link).
The ZS Impact Institute surveyed nearly 10,000 healthcare consumers and providers across the U.S., Germany and China. The findings highlight how AI has shifted how patients interact with healthcare (Link).
NVIDIA releases open-source simulation framework allowing surgical robots to be designed and tested in virtual environments (Link).
The U.S. Department of Veterans Affairs awards Salesforce $1.6B contract to transform veteran care and services with AI (Link).
Samsung launches Health Assistant Beta — First Fully Integrated AI-Powered Personal Health Assistant (Link).
A Florida man sues OpenAI, alleging that ChatGPT’s medical advice delayed care before a life-threatening pulmonary embolism (Link).
🧪Research Signals
New papers worth your time.
Cell: Advancing cancer detection and treatment using longitudinal routine clinical data (Paper).
Nature: A multi-agent multimodal LLM framework for imaging-grounded treatment recommendations in acute ischemic stroke (Paper).
Nature: AI assistance for fetal ultrasound interpretation in a multi-reader study (Paper).
NEJM: After Clearance — Continuous Monitoring as the Foundation of Clinical AI Oversight (Paper).
Nature: Global analysis of country-level factors associated with chatbot usage for health (Paper).
Nature: Evaluating large language models for assessment of psychosis risk (Paper).
🦾TechTools
Useful tools
Analyzes brain CT and MRI images to support the detection of tumors, hemorrhage, and other neurological conditions. Designed as clinical decision support for neurology and neuroradiology teams.
Searches the TRIP medical database and gives structured, evidence-based answers to clinical questions. Helps clinicians and researchers find and compare relevant evidence faster.
📈 Productivity AI tool of the week:
Works across connected apps and files to research, organize information, and complete multi-step tasks. Designed to act on work.
Thank you for taking the time to read.
Know one clinician who should be reading AIMedily? Forward this issue to them.
Itzel Fer, MD PM&R
Forwarded this email? Subscribe free.