Intended use and boundaries
Define one task, target user, permitted actions, and the conditions that require the agent to stop or hand off to a person.
Science and methods · reviewed 19 Aug 2026
How medical AI agents should be evaluated
We separate a product hypothesis from demonstrated medical benefit. Each agent needs a defined purpose, user, data boundary, failure set, stopping criteria and professional oversight.
Scientific principles
Evaluate before deployment
Technical accuracy alone does not demonstrate clinical benefit. Evaluation covers the data, model, interface, workflow and the human–AI team.
Pre-specified metrics
Set completeness, false positives, omissions, time, source quality and critical stopping errors before testing starts.
The human–AI team
Measure how professionals interpret, check, correct and use the output—not only the model response.
Versioning and generalizability
Record the model, prompt, knowledge base and dataset version; do not assume results transfer across populations or sites.
Evidence ladder
Five separate stages
Each move requires a new protocol, approvals and an explicit decision by accountable specialists.
- 01
Product hypothesis
Describe the user, problem, expected action and unacceptable harm.
- 02Current stage
Synthetic evaluation
Test fictional cases, sources, refusals, escalation and interface resilience.
- 03
Offline validation
Use authorized data to assess pre-specified metrics, subgroups, bias and external applicability.
- 04
Early clinical evaluation
Under a protocol, study safety, human factors and small-scale operation in a live workflow.
- 05
Comparative evaluation
Evaluate process or outcome effects against a comparator, including unintended consequences.
Research programme
What should be measured
Clinical-trial navigation
Search recall, criterion-matching errors, unknown information, coordinator review time and the quality of handoff to a physician.
OMS, VMP and self-pay oncology pathways
Currency of official sources, clarity of the next step, document completeness and absence of covert treatment or provider selection.
Clinical copilot
Claims with sources, unsupported conclusions, critical omissions, review time and the quality of professional checking.
Operational agent
Missed actions, handoff quality, manual corrections, cycle time and newly introduced risk points.
Primary and official sources
Methodological foundation
WHO · Regulatory considerations on artificial intelligence for health
Regulatory topics including intended use, transparency, validation, data and lifecycle oversight.
↗DECIDE-AI · Early-stage clinical evaluationReporting checklist for early evaluation of AI decision support in live clinical settings.
↗TRIPOD+AI · Reporting clinical prediction modelsTransparent reporting guidance for clinical prediction models, including machine learning.
↗FDA · Good Machine Learning PracticeQuality and safety principles for AI/ML medical devices across the product lifecycle.
↗NCI · TrialGPT clinical-trial matchingOfficial description of a research system for clinical-trial search and matching.
↗ClinicalTrials.gov · Data APIOfficial programmatic access to registry data; status and criteria require date-specific verification.
↗Evidence boundary
What this page does not demonstrate
A methodology and source list do not mean that the project has completed a clinical study, is a registered medical device or is ready to process real health data. Those claims require separate evidence and approvals.