Written and maintained by CASRAI Editorial Board
Last updated
TRIPOD+AI (formally the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence statement, published by Collins, Moons, Dhiman, et al. in The BMJ on April 16, 2024; 385:e078378, PMID 38626948) is the international consensus checklist governing how researchers report studies that develop, externally validate, or update diagnostic and prognostic prediction models using either regression or machine learning methods.
Purpose and Scope: Why TRIPOD+AI Replaced the 2015 TRIPOD Statement
What is TRIPOD+AI? In 2015, the original TRIPOD statement (Collins et al., BMJ 2015;350:g7594) established a 22-item checklist focused primarily on regression-based clinical prediction rules. As machine learning (random forests, gradient boosting, support vector machines) and deep neural networks proliferated across Electronic Health Record (EHR), genomics, and clinical imaging datasets, systematic reviews documented widespread reporting defects: omitted hyperparameter tuning procedures, silent data leakage between training and test partitions, lack of calibration reporting, and unexamined performance disparities across race, ethnicity, sex, and socioeconomic subgroups.
Rather than maintaining two separate guidelines (one for regression and one for machine learning), the international Delphi consensus panel retired the 2015 checklist and issued TRIPOD+AI as a single harmonized standard covering all multivariable prediction model studies regardless of modeling technique. Within the broader EQUATOR Network reporting guidelines ecosystem and the clinical AI reporting guidelines (CONSORT-AI, TRIPOD+AI, DECIDE-AI) lifecycle, TRIPOD+AI governs preclinical and retrospective/prospective model development and validation (whereas SPIRIT-AI and CONSORT-AI govern interventional clinical trials evaluating an AI system deployed in care).
The 27-Item TRIPOD+AI Checklist (Grouped by Manuscript Section)
The 2024 TRIPOD checklist comprises 27 main numbered items subdivided into 52 subitems, accompanied by a separate 13-item TRIPOD+AI for Abstracts checklist (embedded under Item 2). Each item is designated as applicable to model Development (D), External Validation (V), or both (D;V):
| Manuscript Section & Item Numbers | Subitem Count & Applicability | Core Reporting Requirement in TRIPOD+AI (BMJ 2024;385:e078378) | Key AI/ML & Methodological Specifics |
|---|---|---|---|
| Title (Item 1) & Abstract (Item 2) | Item 1 (D;V), Item 2 (13-item Abstract checklist, D;V) | Identify the study as developing, validating, or updating a multivariable prediction model, specify the target population and outcome to be predicted, and provide a structured 13-element abstract | Title or abstract must clearly indicate whether statistical regression or machine learning methods were used and report both discrimination and calibration summary metrics with confidence intervals |
| Introduction: Background (Item 3) & Objectives (Item 4) | Items 3a–3c (D;V), Item 4 (D;V) | Explain the clinical context, intended point in the care pathway, intended users (clinicians, patients, or health administrators), existing models, and explicit study objectives | State how the model output is intended to inform clinical decision-making and why a new or updated model is needed over existing rules |
| Methods: Data Sources, Participants & Outcome (Items 5–7) | Items 5a–5b, 6a–6c, 7a–7c (D;V) | Describe study design/data sources (cohort, RCT, registry, EHR), eligibility criteria, care setting, recruitment dates, treatments received, and exact outcome definition/measurement | Report whether outcome assessment was blinded to predictor information (Item 7b) and specify the prediction horizon/time point for prognostic models |
| Methods: Predictors, Data Preparation & Sample Size (Items 8–11) | Items 8a–8c, 9a–9c, 10, 11 (D;V) | Define all candidate predictors (features), timing of measurement, data preprocessing/harmonization, sample size justification, and missing data handling | Detail feature engineering, intensity normalization, text tokenization, class-imbalance adjustments (SMOTE/weighting), and missing-data imputation methods (distinguishing training-time vs. deployment-time missingness) |
| Methods: Analytical Methods & Fairness (Items 12–17) | Items 12a–12g (D), 13a–13b (D), 14 (D;V), 15 (D;V), 16 (D;V), 17 (V) | Specify model architecture, hyperparameter tuning, internal validation (bootstrapping/cross-validation), performance metrics, subgroup/fairness evaluation, and model updating procedures | Requires reporting measures of calibration (calibration slope, calibration-in-the-large, and smoothed calibration curves), discrimination, clinical utility (decision curve analysis net benefit), and subgroup fairness across key sociodemographic groups (Items 12e, 16) |
| Open Science (Item 18) & Public Involvement (Item 19) | Items 18a–18f (D;V), Item 19 (D;V) | Declare funding, conflicts of interest, study registration/protocol availability, data availability, and analytical code/model weight sharing, plus Patient and Public Involvement (PPI) | Item 18e requires stating how complete analysis code, preprocessing scripts, trained model weights/containers, or regression equations can be accessed for independent reproduction |
| Results: Participants, Model Development & Performance (Items 20–24) | Items 20a–20c, 21 (D;V), 22a–22b (D), 23a–23b (D;V), 24 (V) | Report participant flow diagram, baseline characteristics (including sociodemographics and missingness per predictor), final model specification, and overall/subgroup performance with 95% CIs | For regression, provide all coefficients and the baseline intercept/hazard; for black-box ML models, provide executable model artifacts or APIs and report disaggregated performance across demographic and clinical subgroups (Item 23b) |
| Discussion: Interpretation, Limitations & Usability (Items 25–27) | Item 25, Item 26, Items 27a–27c (D;V) | Interpret findings in context of existing evidence, discuss biases/dataset shift/generalizability limitations, and specify clinical usability and deployment requirements | Item 27 requires detailing how poor-quality or unavailable input data should be handled in live clinical use (27a), what user training or expertise is required (27b), and next steps toward clinical pathway integration (27c) |
How Requirements Vary by Study Design (Development vs. Validation vs. Updating)
The TRIPOD+AI consortium provides tailored checklist variants depending on which of the three primary prediction modeling study designs a manuscript reports:
- Model Development Studies (Column D — Items 1–16, 18–23, 25–27): Any study that trains a new multivariable model (or retrains weights from scratch) must never rely solely on an apparent (in-sample) performance check or a single random train/test split in small samples without accounting for optimism. Item 12f mandates reporting how internal validation (such as k-fold cross-validation, nested cross-validation for hyperparameter tuning, or bootstrap optimism correction) was performed, including whether all preprocessing, imputation, and feature selection steps were repeated inside each resampling fold to prevent data leakage.
- External Validation Studies (Column V — Items 1–11, 14–21, 23–27): External validation evaluates an existing, locked prediction model on participant data independent of the development cohort (temporally, geographically, or across different hospital EHRs). Item 21 requires a direct side-by-side comparison of predictor distributions, outcome incidence, and case-mix differences between the original development population and the external validation cohort.
- Model Updating and Recalibration Studies (Item 17 & Item 24): When an existing clinical prediction model exhibits miscalibration in a new health system (for example, overpredicting mortality due to improved standard of care), researchers should update the model (via intercept recalibration, logistic recalibration of the slope, or predictor re-estimation) rather than building an entirely new score from scratch. Items 17 and 24 require reporting both the pre-update performance and the post-update performance. For how diagnostic prediction models compare against single diagnostic test accuracy studies, see STARD, STARD-AI and TRIPOD+AI compared.
Governance, Code Archiving, and the TRIPOD Guideline Family
When preparing a clinical prediction manuscript for submission to journals endorsing EQUATOR standards (including The BMJ, JAMA network journals, The Lancet Digital Health, and Nature Medicine), study teams should maintain three persistent artifacts alongside the completed TRIPOD+AI checklist:
- Study Protocol Registration: Item 18c asks where the study protocol can be accessed or whether the prediction study was prospectively registered (for example, on OSF, ClinicalTrials.gov, or PROSPERO for systematic reviews of prediction models).
- Executable Model and Preprocessing Code Repository: Under Item 18e, manuscripts must provide a persistent identifier (such as a Zenodo DOI or GitHub release commit) containing the exact software environment, package versions, feature-scaling parameters, and trained model weights or regression equation.
- Extension Selection Across the TRIPOD Family: Verify whether a specialized sibling extension applies alongside TRIPOD+AI:
- TRIPOD-Cluster (Debray et al., BMJ 2023;380:e071018): For prediction models developed or validated using clustered or multi-center datasets (accounting for between-hospital heterogeneity).
- TRIPOD-SRMA (Snell et al., BMJ 2023;381:e073538): For systematic reviews and meta-analyses of prediction model studies.
- TRIPOD-P (Dhiman et al., 2023): For protocols of clinical prediction model studies.
- TRIPOD-LLM (Gallifant et al., Nature Medicine 2025): Specialized extension for reporting healthcare studies using Large Language Models.
Standards and Related Methodological Appraisal Tools
TRIPOD+AI is a reporting standard (defining what must be transparently disclosed in a paper), which pairs directly with risk-of-bias appraisal and regulatory standards:
- PROBAST and PROBAST+AI (Prediction model Risk Of Bias Assessment Tool): Used by systematic reviewers and guideline panels to grade risk of bias and applicability across four domains: Participants, Predictors, Outcome, and Analysis.
- FDA Good Machine Learning Practice (GMLP) Guiding Principles: Jointly issued by the FDA, Health Canada, and UK MHRA, requiring representative clinical study participants, strict separation of training and test datasets, and monitoring for dataset drift—principles directly operationalized in TRIPOD+AI Items 6, 12, 16, and 27.
Frequently Asked Questions
What is TRIPOD+AI?
TRIPOD+AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence, published by Collins et al. in The BMJ in April 2024; 385:e078378) is the international EQUATOR Network reporting guideline for studies that develop, externally validate, or update clinical prediction models—whether built with statistical regression (such as logistic or Cox regression) or machine learning/deep learning algorithms. It supersedes the 2015 TRIPOD statement with a unified 27-item (52-subitem) main checklist and a 13-item abstract checklist.
Sources
- Collins GS, Moons KGM, Dhiman P, et al. — “TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods,” BMJ, 2024;385:e078378, PMID: 38626948 (https://pubmed.ncbi.nlm.nih.gov/38626948/).
- Collins GS, Reitsma JB, Altman DG, Moons KGM — “Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD Statement,” BMJ, 2015;350:g7594, PMID: 25569120 (https://pubmed.ncbi.nlm.nih.gov/25569120/).
- TRIPOD Statement Executive Group & EQUATOR Network — TRIPOD+AI Official Checklists and Guidance Repository (https://www.tripod-statement.org/).
