What does TRIPOD+AI ask you to report about a prediction model evaluation?

Why

Collins and colleagues published TRIPOD+AI in the BMJ in 2024 to harmonize reporting whether a model uses regression or machine learning, and it supersedes the 2015 TRIPOD checklist, which should no longer be used. It has 27 main items and a 13-item checklist for abstracts. For an evaluation, the important reporting points include the data sources and dates, how participants were selected, how outcomes and predictors were defined and measured, and how the sample size was decided.

It also asks how missing data were handled, how performance was measured, including both discrimination and calibration with uncertainty, and how the model performed in relevant subgroups. A notable emphasis is fairness. The authors embed it throughout the checklist so reports state whether and how performance was checked across groups such as sex, age or ethnicity.

For a team, the practical value is that writing to the checklist early exposes gaps, such as no calibration plot or an unclear outcome definition, while they can still be fixed. TRIPOD+AI is a reporting guideline, not a quality score. A well-reported study can still be poorly designed, and appraisal of bias needs a tool such as PROBAST.