TRIPOD is the checklist for clinical prediction models — the risk scores and calculators authors increasingly submit. This walks the items that separate a usable model from an unverifiable one, with a self-check for each.
TRIPOD — Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis — is the reporting guideline ↗ for studies that develop, validate, or update a prediction model, maintained through the EQUATOR Network ↗ (with the 2024 TRIPOD+AI extension for models built with machine learning). Its companion, PROBAST, is the matching risk-of-bias tool reviewers apply. Prediction models are held to a high reporting bar because a published calculator can be put into clinical use — so the manuscript has to give a reader everything needed to check it and, ideally, to reproduce it.
Describe the source of data and study design (item 4) — cohort, registry, case-control — and the eligibility and setting (item 5). Define the outcome the model predicts (item 6), how and when it was assessed, and whether outcome assessment was blind to the predictors.
Self-check. Is the outcome ever ascertained at the same moment as a predictor? If an event and one of its own predictors are recorded in the same window with no guaranteed ordering, the model can look far more accurate than it is — reverse causation dressed as discrimination. That is the central finding in our PE risk-model case study.
Define all predictors (item 7) and how they were measured — again, blind to the outcome where possible. Justify the sample size (item 8); models fit on too few events per predictor overfit and validate poorly. State how missing data were handled (item 9) — complete-case analysis silently changes who the model is about, and the excluded count belongs in the main text, not buried in a supplement.
Self-check. How many events per candidate predictor did you have, and what did you do with the records missing a predictor? If the answer to the second is “dropped them,” say so and show a sensitivity analysis.
Report discrimination and calibration (item 16), not one without the other. An AUC (discrimination) says how well the model ranks patients; calibration says whether its predicted probabilities match observed rates. A model can rank well and still be systematically over- or under-confident. Specify the type of validation (item 10c): internal (resampling in the same data), temporal (a later slice of the same source), or external (an independent population). These are not interchangeable, and the strength of the claim must track which one you did.
Self-check. If your model is validated only internally or temporally, is any deployed calculator labeled investigational pending external validation? Deploying on internal validation alone is a claim that outruns the evidence — the pattern our PE risk-model case study rates as serious.
Present the full model (item 15) — every coefficient, the intercept or baseline hazard, and enough detail to compute a prediction for a new patient. A calculator readers cannot reconstruct is a claim they cannot check. State the limitations (item 18) and give a fair account of what the model can and cannot be used for (item 19), including the population and care setting it was built in.
Self-check. Does the paper let a reader reproduce a single prediction by hand from the published coefficients? If not, an item is missing — and reproducibility is precisely what makes a prediction model trustworthy.
Complete the official checklist — TRIPOD or TRIPOD+AI as appropriate — and read your study against PROBAST before submission. A structured pre-submission review does both alongside the statistics: RigorMD maps a prediction-model manuscript to TRIPOD item by item as one of six scored domains, recomputes reported discrimination, calibration, and reclassification figures deterministically, and has two independent engines assess whether the model's central claim is calibrated to its validation. It flags; it does not certify, and it does not replace peer review or a statistician's input on design.
See the full sample report on a TRIPOD prediction model →, read how the engine works, or review pricing — the pre-submission review is $25. For other designs, see the STROBE, CONSORT, and PRISMA walkthroughs.