TRIPOD+AI: reporting clinical prediction models

TRIPOD+AI, published in 2024, supersedes the 2015 TRIPOD statement. Its 27-item checklist covers clinical prediction models developed or evaluated using regression or machine learning. It includes reporting on fairness, open science, and intended clinical use. The official scope ↗ explains which studies it covers; the expanded checklist ↗ provides every item and its explanation. This guide selects a few requirements; it does not reproduce the full checklist.

Reporting completeness and methodological quality are different questions. PROBAST+AI ↗, published in 2025, is an updated tool for assessing quality, risk of bias, and applicability of prediction models using regression or AI methods.

Describe the data sources and collection dates (item 5) for development and evaluation. Explain the setting, eligibility criteria, and relevant treatments (item 6), then the data preparation and quality checks (item 7). Define the outcome and time horizon, how and when it was assessed, and any steps to blind its assessment (item 8).

Self-check. For a prognostic model, were all predictors available at the intended prediction time? A variable measured after that point can introduce information that would be unavailable when the model is used. Diagnostic models need a timing assessment appropriate to their intended use, too.

Define the predictors and how they were measured (item 9). Explain how the sample size was determined and why it is sufficient for development or evaluation (item 10). Describe missing-data handling and reasons for omitting data (item 11).

Self-check. Does the sample-size justification account for the planned model and outcome frequency? Explain which records were excluded, what was imputed, and the assumptions behind that choice. If data were split for training and testing, imputation and other preparation should avoid passing information from test data into model development.

Describe the analysis and validation methods (item 12), including data partitioning, model-building steps, tuning, internal validation (12c), clustering (12d), and the measures used to assess performance (12e). Explain differences between development and evaluation data (item 16). Report performance estimates with confidence intervals and relevant subgroup results (item 23).

An AUC describes how well a model ranks patients; calibration describes agreement between predicted probabilities and observed outcomes. Good discrimination does not establish good calibration. Internal validation and evaluation in a separate population answer different questions about performance.

Self-check. Can a reader tell which data produced each performance estimate, whether model-building steps were repeated during internal validation, and how clustering or repeated observations were handled?

Provide the full model or access details (item 22) so others can calculate predictions and evaluate it, including any restrictions. For a regression model, this may include the coefficients and intercept or baseline hazard; other models may require code or an accessible model object. Interpret the findings in context (item 25), explain limitations and their effects (item 26), and discuss practical use and further evaluation (item 27).

Self-check. Could another team calculate predictions from what you provide? Does the proposed use match the populations and settings in which the model was evaluated? State access restrictions and remaining evaluation needs directly.

Complete the current TRIPOD+AI checklist and assess quality, risk of bias, and applicability with PROBAST+AI before submission. Reporting completeness and methodological quality are different questions. RigorMD assesses reporting-guideline adherence and appraises model development, missing-data handling, validation, and whether the conclusions match the reported performance. It runs applicable numerical checks when the manuscript supplies the required inputs. It does not refit the model or recalculate discrimination, calibration, or reclassification from the underlying data. Its report does not replace a complete independent PROBAST+AI assessment or a statistician's input on design.

See the full sample report →, read how the engine works, or review pricing — the pre-submission review is $99. For other designs, see the STROBE, CONSORT, and PRISMA walkthroughs.