Building an AI Validation Dossier

Module 10 established that every safety database requires Computer System Validation (CSV) before it’s used for regulated data. AI models used inside a regulated PV workflow require the same underlying discipline, adapted for what makes AI systems specifically different from a conventional database: they learn patterns from data rather than executing fixed logic, which means their behaviour has to be characterised empirically, not just specified in advance.

The pre-deployment phase starts with training data assessment: defining the model’s intended use precisely, documenting the training data’s sources, size, and demographic coverage, assessing that data for bias, and setting concrete performance targets — precision, recall, and F1-score — for the specific use case. This is a joint responsibility between the data science team and the PV medical team, with the bias assessment specifically requiring independent review rather than being self-assessed by whoever built the model.

Test dataset construction follows, and the requirement here is genuinely strict: the test dataset must be constructed independently of the training data — cases the model has never seen — and it must include cases from the intended operational population, covering demographic subgroups, geographic regions, and languages the model will actually encounter once deployed, not just the population that happened to be convenient to train on. Where patient data is involved, de-identification documentation is required alongside the test dataset specification, and the dataset needs formal approval before validation testing can even begin.

Performance testing runs the model against that approved test dataset, calculating precision, recall, F1-score, and — critically — subgroup-level performance, not just an aggregate figure, compared against human expert performance on the same cases. The output is a validation test report with a gap analysis against the pre-defined targets and an explicit deploy/do-not-deploy decision, requiring independent statistical review and formal PV Quality sign-off before the model can go live. This entire framework is, in effect, Module 10’s system validation discipline extended to cover a category of system that didn’t exist when that discipline was first written — the underlying principle (nothing regulated goes live unvalidated) hasn’t changed at all.