TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis)

What it is

— TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) is a reporting guideline for studies that develop, validate (internally or externally), or update a multivariable prediction model intended to estimate an individual's probability of a current condition (a diagnostic model) or a future event (a prognostic model). The original 2015 statement (Collins, Reitsma, Altman, Moons) is a 22-item checklist with a companion item-by-item Explanation & Elaboration paper (Moons et al. 2015); TRIPOD+AI (2024) is the current, EQUATOR-listed update that harmonises reporting across regression and machine-learning models and strengthens items on data sources, sample size, fairness, and open code/data. TRIPOD-Cluster (2023) is a sibling extension for models developed or validated in clustered/multi-database data (e.g., individual-participant-data meta-analysis or federated multi-site analyses). TRIPOD is maintained within the EQUATOR Network alongside its sister appraisal tool PROBAST (risk-of-bias) — TRIPOD is the reporting checklist; PROBAST is the risk-of-bias instrument. Its purpose is to make a prediction model reproducible and independently assessable: a reader should be able to see exactly which population and data produced the model, which predictors entered it, how it was built and tuned, and how well it discriminated and calibrated in development and in validation.

When to use

— Apply TRIPOD whenever the study's deliverable is a prediction model — a risk score, a prognostic index, a diagnostic probability rule, or an ML classifier producing individual-level risk — regardless of data source (claims, EHR, registry, linked, or prospectively collected). It is the correct checklist for a journal manuscript reporting model development and/or validation, for the model-development component of an HTA/payer submission (e.g., a risk-stratification or enrichment tool), and for FDA/EMA packages where a model supports a device, an enrichment strategy, or a prognostic claim. Decision rules for choosing the right family member: use TRIPOD+AI (2024) for any model built or validated with machine-learning methods, or whenever you want the current, most complete item set (it supersedes the 2015 list for new work); use TRIPOD-Cluster when development or validation uses clustered/multi-database data; use TRIPOD (not STROBE) when the objective is prediction even if the data come from an observational cohort. Conversely, if the objective is an exposure–outcome causal effect, TRIPOD is the wrong guideline — use STROBE/RECORD-PE/HARPER. If the study evaluates a single diagnostic test's accuracy rather than a multivariable model, use STARD.

What it requires

— TRIPOD's items force documentation of the elements that make a model trustworthy and reproducible: a clear statement of the prediction objective (diagnostic vs prognostic; development, validation, or update) and the intended use and target population; source of data and study design (cohort, case-control, registry, routinely collected data) with participant eligibility, setting, and dates; rigorous specification of the outcome (definition, how and when assessed, blinding to predictors) and of all candidate predictors (definitions, timing of measurement, and — critically for real-world data — that predictors are measured at or before the prediction time point, never using future information); sample size / events-per-variable justification; explicit handling of missing data (the imputation model, not a silent complete-case default); the full model-building procedure (predictor selection, functional forms, interactions, penalisation/regularisation or ML hyperparameter tuning, and how overfitting/optimism was addressed); internal validation (bootstrap/cross-validation, optimism correction) and any external/temporal/geographic validation; and performance reporting that includes both discrimination (e.g., C-statistic/AUC) and calibration (calibration plot, calibration-in-the-large and slope) — not discrimination alone — plus, where relevant, clinical utility (decision-curve/net-benefit). TRIPOD+AI adds explicit items on data provenance and representativeness, fairness/subgroup performance, and availability of code and the full model (coefficients or the deployable algorithm). For models built on claims/EHR/registry data these generic items carry teeth: the source must document fitness-for-purpose of the data, the phenotype/algorithm defining outcome and predictor variables (with validation metrics such as PPV/sensitivity), the alignment of the prediction time point so no post-baseline information leaks into predictors, and how attrition and loss to follow-up were handled in a longitudinal prognostic model.

When NOT to use — limitations and common misapplications

— TRIPOD is a reporting checklist; it is not a risk-of-bias instrument, not a quality score, and not a guarantee that a model is valid, fair, or clinically useful. Concrete failure modes a reviewer will flag: (1) Wrong objective — using TRIPOD for an exposure–outcome causal/comparative-effectiveness study; that work belongs to STROBE/RECORD-PE/HARPER, and conversely using STROBE for a prediction-model paper omits the development, validation, and calibration items TRIPOD exists to enforce. (2) Mistaking it for an appraisal tool — TRIPOD says nothing about whether a model is at low risk of bias; that judgement is made with PROBAST, and the two are routinely confused. (3) Stale version — using the 2015 list for a machine-learning model published after 2024 when TRIPOD+AI is the expected standard; or using single-study TRIPOD where TRIPOD-Cluster fits the clustered/multi-database design. (4) Discrimination-only reporting — quoting an AUC with no calibration plot or slope is a classic TRIPOD violation; a model can discriminate well yet be badly miscalibrated and unsafe to deploy. (5) Optimism unacknowledged — reporting apparent (development-set) performance as if it were validated, with no internal validation or optimism correction. (6) Wrong tool entirely — applying TRIPOD to a single diagnostic test's accuracy (use STARD) rather than to a multivariable model. (7) Checklist-as-theater — ticking 22 (or the TRIPOD+AI) items while leaving the predictor definitions, the model equation, the imputation approach, or the calibration evidence vague; completing the checklist does not make the model reproducible, and it certainly does not make a model developed in one database transportable to another.

How it maps to this catalog

— In this repo, TRIPOD's requirements are implemented by the prediction/validation and data-fitness family of concepts (not the causal-inference concepts a comparative-effectiveness study would use): - The core development/validation discipline: prediction-model-validation-recalibration-rwe (internal/external validation, calibration, recalibration) and predictive-and-causal-ml-models-rwe (the ML modelling and tuning that TRIPOD+AI items govern). - Outcome/predictor definition as a measurement-validity problem: diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe, claims-outcome-algorithm-ppv-sensitivity-rwe, ehr-phenotyping-algorithms-rwe, and algorithm-validation implement the phenotype-definition-and-validation items. - Data fitness and the prediction-time-point spine: fit-for-purpose-data-assessment-rwe and time-zero-index-date-alignment-rwe (so predictors are measured before the prediction point and no future information leaks in). - External validation / transportability across databases: generalizability-transportability-external-validity-rwe implements TRIPOD's external-validation and (with TRIPOD+AI) fairness/representativeness items. - Missing data and attrition items: missing-data-pattern-table-rwe, multiple-imputation-longitudinal-rwe, and attrition-and-loss-to-follow-up-rwe. - Sample-size justification and reporting visuals: sample-size-power-precision-rwe and visualizations-pharmacoepidemiology-rwe (calibration plots, ROC/AUC, decision curves). - The structured-question and pre-specification habits: picots-framework-rwe and study-protocol-or-sap-elements.

Applied note (claims/EHR/registry RWE)

A prognostic model built in administrative claims — say, a 12-month hospitalisation-risk score — should report, per TRIPOD: the database and its fitness-for-purpose; continuous-enrollment and observability windows; the prediction time point and proof that every predictor is measured at or before it; the phenotype/algorithm (with PPV) defining the outcome and key predictors; the events-per-variable and missing-data/imputation strategy; the full model (coefficients or deployable algorithm); internal-validation optimism correction; and external or temporal validation in a different database with both discrimination and a calibration plot. Reporting only an in-sample C-statistic, omitting the model equation, or skipping external calibration are precisely the gaps TRIPOD (and TRIPOD+AI) exist to close.