Software Verification in Medical Device AI
Holding
Validation asks whether the product is clinically useful for its intended purpose. Verification asks whether the software was built correctly against its specifications. Medical AI needs both, and the distinction should not be collapsed into model performance.
Authority
EU medical device rules reference software verification and validation as part of technical documentation. FDA's quality system regulation incorporates ISO 13485 principles for design and development controls. IMDRF frames SaMD as software that can itself perform a medical purpose.
Issue
A model can pass retrospective validation while the application mishandles units, missing values, permissions, time windows, or threshold display. Those are software failures, not model failures, and they require verification evidence.
Resolution
The system should define requirements for data ingestion, preprocessing, inference, display, logging, access, error handling, and model version selection. Verification should test each requirement, including boundary cases and failure modes, before clinical validation is used to assess intended use.
Evidence Package
The package should include software requirements, architecture, traceability matrix, test protocols, test results, defect history, release approval, and configuration records. The practical rule is that no clinical claim should rest on software behavior that has not been verified.