Validation study overview — External prediction model validation

Validation question

Define the intended population and decision for which an existing model needs evaluation.

Identify frozen model

Obtain the exact frozen model, coefficients or executable version, preprocessing and baseline risk where required.

Select independent population

Select data independent of development and document setting, calendar period, case mix and ascertainment.

Map predictors and outcomes

Map each input and the outcome horizon to the original definition; record unavailable inputs and deviations.

Generate predictions

Generate original-model predictions without refitting and retain reproducible model and data versions.

Assess performance

Assess calibration, discrimination and utility with intervals and appropriate handling of censoring or missing outcomes.

Evaluate updating separately

Preserve the original-model evaluation; any recalibration or revision creates an updated model needing separate validation.

Report

Report independence, mapping, performance and any separately evaluated update using TRIPOD+AI.

Validation study overview

End flow

Protocol and decision register

Record the protocol, approvals and registration where applicable. Preserve planned decisions and amendments separately from results.

Protocol and rationale evidence

Attach the dated protocol, model or test specification, approvals and amendments. Identify any applicable study registration, registry and identifier explicitly. Diagnostic accuracy and prediction studies are not automatically interventional clinical trials; a disease registry is a data source, not protocol registration.