- Evaluation
Evaluating clinical AI before it touches patients
A 20-page report on how to judge a clinical AI tool before it reaches a patient.
PDF · 20 pages · Updated 30 September 2026
- Who it is for
- Physicians and health-system leaders who have to say yes or no to a clinical AI tool.
- Format
- Length
- 20 pages
- Updated
- 30 September 2026
- Topic
- Evaluation
Three findings inside
6%
of 516 studies of AI for diagnostic imaging performed external validation.
Kim et al., Korean Journal of Radiology, 2019
87%
of machine-learning prediction-model analyses were at high risk of bias.
Andaur Navarro et al., BMJ, 2021
0.63
external AUROC for a widely used sepsis model, which missed 67% of patients who had sepsis.
Wong et al., JAMA Internal Medicine, 2021
What is inside
- The gap between adoption and evidence
- What the number on the slide meansAUROC, positive predictive value, calibration, subgroups.
- Whose patients the evidence is about
- What FDA authorization actually establishes
- Europe: the clock was reset, not stopped
- Benchmarks, and the evidence hierarchy for language models
- The vendor demo, and the one-page checklist
- After go-live: evaluation becomes governance
Get it
A 20-page report on how to judge a clinical AI tool before it reaches a patient.
PDF · 20 pages. The download starts on this page and the file is also sent to your inbox.
Read the guides behind it
Questions
Is this clinical or regulatory advice?
No. It appraises published evidence and sources. Confirm anything that affects a patient or a purchase with your own governance team and counsel.
Do I have to join the briefing to download?
No. The download does not depend on the box. The box is only for the AIMOCS briefing, and nothing is sent until you confirm from the email we send you.
What do you do with my email address?
We use it to send you this file. The privacy notice says what else we do, and how to ask us to delete it.
Physicians and teams working on AI in healthcare
Meet them in the network.
Evaluating clinical AI before it touches patients
Get it