Published
The TRIPOD+AI checklist is the reporting standard that journals now expect when a manuscript develops or evaluates a clinical prediction model, whether the model uses logistic regression or a random forest. Published in the BMJ in April 2024 by Collins and colleagues, it holds 27 numbered items, replaces the 2015 TRIPOD checklist, and asks for details that older papers never gave, such as fairness, code access and patient and public involvement.1
Most guides stop at listing the items. This one goes further. After the checklist itself, we score a real open-access machine learning paper item by item, with the section of the paper beside every verdict, so you can see what a completed checklist looks like and where authors usually lose marks.
In Simple Words
TRIPOD+AI is a list of 27 things your paper must tell the reader if you built or tested a tool that predicts a patient outcome. It tells you what to report, not how to build the model. Fill it in with a page or line number for every item, submit it with your manuscript, and a reviewer can check your paper in minutes instead of guessing what is missing.
What the TRIPOD+AI Checklist Is and What Changed From TRIPOD 2015
TRIPOD stands for Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis. The 2015 version listed 37 entries once sub-items were counted, and it was written when most prediction models were regression equations.1 Since then, machine learning has become common in prediction research, and the 2024 update was built to cover both traditions in one document. The authors used systematic reviews of the literature, a two-round Delphi survey and an online consensus meeting. The first Delphi round was sent to 292 people and 170 completed it. The second round reached 200 completers from 27 countries.1
Three points matter for your submission. First, the guideline states plainly that the new checklist supersedes the 2015 version, which should no longer be used.1 If your target journal still links the old checklist, follow the 2024 one and say so in the cover letter. Second, TRIPOD+AI is a reporting guideline and not a quality appraisal tool. It checks whether you described what you did, not whether what you did was sound. For appraisal, the same research community published PROBAST+AI in 2025, an updated tool for risk of bias and applicability in prediction model studies.3 Third, the guideline comes with a 13-item checklist for abstracts, so the abstract is held to its own standard.
If you are still deciding between a diagnostic accuracy guideline and a prediction model guideline, read our comparison of STARD and TRIPOD first. It explains which reporting guideline fits which study question, and this article assumes you have already settled on TRIPOD+AI. Our decision tree for choosing a reporting guideline covers the wider set of options.
The 27 Items of the TRIPOD+AI Checklist at a Glance
The table below condenses the full checklist into its 27 numbered items. We paraphrased the wording to keep it short, so use the official checklist from the TRIPOD website for the exact text you will submit. The last column shows who owes the item. D means a model development study, E means a model evaluation (validation) study, and D;E means both.
Table 1. The 27 TRIPOD+AI checklist items, condensed. Source structure from Collins et al., BMJ 2024.
| Item | Topic | Applies to | What to report (paraphrased) |
|---|---|---|---|
| 1 | Title | D;E | Name the study as developing or evaluating a prediction model, and give the target population and the outcome. |
| 2 | Abstract | D;E | Follow the separate 13-item TRIPOD+AI for Abstracts list. |
| 3a to 3c | Background | D;E | Healthcare context and rationale, existing models, target population and intended use in the care pathway, intended users, and known health inequalities between sociodemographic groups. |
| 4 | Objectives | D;E | State whether the study develops a model, validates one, or both. |
| 5a, 5b | Data | D;E | Data sources for development and evaluation data separately, why they were chosen, how representative they are, and the start and end dates of participant accrual. |
| 6a to 6c | Participants | D;E | Setting, number and location of centres, eligibility criteria, and any treatments received and how they were handled. |
| 7 | Data preparation | D;E | Pre-processing and quality checks, including whether they were similar across sociodemographic groups. |
| 8a to 8c | Outcome | D;E | Exact outcome definition and time horizon, how and when it was assessed, who assessed it if judgement was needed, and any blinding. |
| 9a to 9c | Predictors | D (9a), D;E (9b, 9c) | How the initial predictors were chosen, precise definitions and timing of each predictor, and who measured those needing judgement. |
| 10 | Sample size | D;E | How the study size was reached, with the calculation, separately for development and evaluation. |
| 11 | Missing data | D;E | How missing data were handled and why any data were omitted. |
| 12a to 12g | Analytical methods | D, E or both by sub-item | Data partitioning, handling of predictors, model type and tuning, heterogeneity across clusters, performance measures, model updating, and how predictions were calculated in evaluation. |
| 13 | Class imbalance | D;E | If imbalance methods were used, why and how, and how predictions were recalibrated afterwards. |
| 14 | Fairness | D;E | Any approach used to address model fairness, and the rationale. |
| 15 | Model output | D | The form of the output (probability or class) and how any threshold was set. |
| 16 | Training versus evaluation | D;E | Differences between development and evaluation data in setting, eligibility, outcome and predictors. |
| 17 | Ethical approval | D;E | The approving board and the consent arrangement or waiver. |
| 18a to 18f | Open science | D;E | Funding and funder role, conflicts of interest, protocol access, registration, data availability, and code availability. |
| 19 | Patient and public involvement | D;E | What involvement took place, or a statement that there was none. |
| 20a to 20c | Participants (results) | D;E (20c is E only) | Participant flow with outcome counts, characteristics by data source, and for evaluation a comparison with the development data. |
| 21 | Model development | D;E | Number of participants and outcome events in each analysis. |
| 22 | Model specification | D | Full model details so that a third party can make predictions and evaluate the model, with any access restrictions. |
| 23a, 23b | Model performance | D;E | Performance estimates with confidence intervals, including key subgroups, and heterogeneity across clusters if examined. |
| 24 | Model updating | E | Results of any updating, such as recalibration, and the performance afterwards. |
| 25 | Interpretation | D;E | Overall interpretation of the main results, including fairness. |
| 26 | Limitations | D;E | Limitations such as a non-representative sample, sample size, overfitting and missing data, and their effect on bias and generalisability. |
| 27a to 27c | Usability in current care | D (27a, 27b), D;E (27c) | Handling of poor or unavailable inputs, the user interaction and expertise needed, and next steps for research. |
Development, Evaluation or Both? Which Items You Owe
Counting every lettered entry, the checklist has 52 rows. Forty apply to both development and evaluation studies, eight apply to development only, and four apply to evaluation only. The development-only rows are 9a, 12a, 12b, 12c, 15, 22, 27a and 27b. The evaluation-only rows are 12f, 12g, 20c and 24.
That split changes what you submit. A team that builds a model and tests it on a held-out sample and on an outside dataset owes the whole list. A team that only tests someone else’s published model owes the evaluation rows, and must describe how it calculated that model’s predictions (12g) and how its data differ from the original development data (20c, 16). Skipping 12g leaves readers unable to tell whether the original equation or a refitted version was tested.
The same logic helps when you plan the analysis. If you expect a reviewer to ask for a decision curve, build it in from the start and report it under 12e. Our short explainer on decision curve analysis shows what net benefit adds beyond discrimination, and our guide to ROC curves and predictive values covers the discrimination side.
Three Groups of New Items That Catch Authors Out
The TRIPOD+AI authors state that the 2015 version did not cover several developments, naming fairness, open science practices and patient and public involvement.1 Those three areas map to items 14, 18 and 19, and the 2015 checklist gave authors no prompt for any of them, so gaps are easy to leave.
Fairness (items 3c, 7, 8a, 14, 23a and 25)
Fairness is threaded through six items, not one. You are asked to describe known health inequalities (3c), to say whether data preparation and outcome assessment were consistent across sociodemographic groups (7, 8a), to describe any fairness approach (14), to report performance for key subgroups (23a), and to discuss fairness in the interpretation (25). If your dataset has no race, sex or age breakdown, say so, and say what that means for the groups you can and cannot speak about.
Open science (items 18a to 18f)
Item 18 bundles six disclosures that used to be scattered. A protocol statement, a registration statement, a data availability statement and a code availability statement each need an answer, even if the answer is that no protocol was prepared or the study was not registered. The checklist accepts a negative answer. It does not accept silence.
Patient and public involvement (item 19)
Item 19 asks what involvement took place during design, conduct, reporting or dissemination, or a statement that there was none. The guideline article records that this item reached 69% agreement in the second Delphi round, just under the 70% threshold, and the steering group kept it for the consensus meeting.1 Treat it as a real requirement. One sentence is enough when involvement did not happen.
Worked Example Applying the TRIPOD+AI Checklist to a Real Open-Access Model
To show the checklist in use, we scored one published paper. Lei and colleagues developed a stacking ensemble to predict sepsis-associated liver injury (SALI) in intensive care patients, using MIMIC-IV for training and internal validation and eICU-CRD for external validation. It was published in the Journal of Medical Internet Research in 2025 under a CC BY 4.0 licence, which is why we can quote it, and it states that it followed TRIPOD+AI and supplies the completed checklist.2
We chose it for a practical reason. The authors did many things right, so the exercise shows what a strong submission looks like, and the few queries that remain are the kind a careful reviewer would raise. We read the full text. Where a verdict depends on judgement, the row says Judgement call. Quotes are short and verbatim from the open-access full text.
How to read the verdicts
Reported means the item is answered in the text. Partly reported means some elements are present and others are missing. Not found means we could not locate it in the full text and appendices we could read. Judgement call means a reasonable reviewer could score the row either way, and the note explains why. This is an Editverse reading of the text and not an official TRIPOD+AI assessment.
Table 2. TRIPOD+AI items scored against Lei et al., J Med Internet Res 2025 (CC BY 4.0). Editverse reading of the full text.
| Item | Where to look in the paper | What the paper reports | Verdict |
|---|---|---|---|
| 1 | Title | Names development and validation, the population (patients with sepsis) and the outcome (SALI). The title ends with “Development and Validation Study Based on a Multicenter Cohort Study”. | Reported |
| 2 | Abstract | States data sources, outcome definition, models, metrics and AUCs. We did not score it against the 13 abstract items. | Judgement call |
| 3a | Introduction, paragraphs 1 to 3 | Gives the clinical context and the case for explainable machine learning. Earlier SALI prediction models are not cited in the Introduction. | Partly reported |
| 3b | Introduction, final paragraph | Intended users are named only as “health care professionals”. Position in the care pathway is not described. | Judgement call |
| 3c | Whole text | No statement on health inequalities between sociodemographic groups. | Not found |
| 4 | Introduction, final paragraph, and Abstract | The aim is to develop an explainable stacking model to predict SALI. Validation is described in Methods. | Reported |
| 5a | Methods, Data Sources | MIMIC-IV for development and internal validation, eICU-CRD for external validation, each described separately. | Reported |
| 5b | Methods, Data Sources | Database versions and the data years (2008 to 2019 for MIMIC-IV, 2014 to 2015 for eICU-CRD) are given. Exact start and end dates of accrual are not. | Partly reported |
| 6a | Methods, Data Sources | MIMIC-IV comes from one US hospital. eICU-CRD is described as multicenter across the US. The number of eICU centres is not given. | Partly reported |
| 6b | Methods, Participants | Five inclusion criteria and three exclusion criteria are listed. | Reported |
| 6c | Methods, Predictors | Renal replacement, ventilation and vasopressors enter as predictors. How later treatment was handled relative to the prediction time is not stated. | Judgement call |
| 7 | Methods, Data Preparation | Cleaning, imputation, normalisation and dummy coding are described. No check across sociodemographic groups is reported. | Partly reported |
| 8a | Methods, Outcome | SALI is defined by INR above 1.5 and total bilirubin above 2 within a stated window. The rationale is given as “Given the absence of a gold standard for diagnosing SALI, we defined it based on criteria from previous studies”. | Reported |
| 8b, 8c | Methods, Outcome | The outcome comes from laboratory thresholds, so subjective assessors and blinding may not apply. The text does not say so. | Judgement call |
| 9a | Methods, Outcome, and Results, Participants | An initial pool was reduced to 42 variables by domain knowledge and a data-driven step, and highly correlated continuous variables were dropped. | Reported |
| 9b | Methods, Predictors and Outcome | Laboratory predictors cover “Laboratory data within the first 72 hours of ICU admission”, while the outcome window runs from 24 hours to 1 week. Total bilirubin and INR appear both as predictors and in the outcome definition. This deserves a clarifying sentence. | Judgement call |
| 10 | Methods, Sample Size | A calculation is given, built on the rule that “model development must adhere to the principle of at least 10 events per variable”. See the next section for why this is a dated method. | Reported |
| 11 | Methods, Missing Data | Variables over 30% missing were excluded, and multiple imputation with five imputed datasets and Rubin rules is described. | Reported |
| 12a, 12b | Methods, Data Preparation | A 70/30 split of MIMIC-IV with eICU-CRD held out for external validation, and standardisation plus dummy coding. | Reported |
| 12c | Methods, Analytical Methods, and Results | Nine algorithms and a stacking ensemble are described. Methods say the meta-model was “typically logistic regression” and later that Lasso “was selected as the meta-model”. Tuning is described as grid search in Methods and as random grid search plus Bayesian optimisation in Results. | Judgement call |
| 12d | Whole text | eICU-CRD pools many hospitals, but no handling of between-hospital heterogeneity is described. | Not found |
| 12e | Methods, Model Performance Evaluation | ROC-AUC, PR-AUC, Brier score, calibration plots and decision curve analysis are specified, each with its purpose. | Reported |
| 12f and 24 | Results, Model Updating, and Table 2 | Table 2 reports observed and expected events before and after calibration. Whether this counts as model updating in the TRIPOD+AI sense is open. | Judgement call |
| 12g | Data Availability | Code is on GitHub, which is where predictions can be reproduced. The text does not say how the stacked model’s predictions were calculated for evaluation. | Judgement call |
| 13 | Methods, Discrimination | The authors say PR-AUC “is particularly valuable when the classes are imbalanced, as is the case with SALI prediction”. No resampling statement either way. | Judgement call |
| 14 | Whole text | No fairness approach is described. | Not found |
| 15 | Methods, Clinical Utility | Decision curves cover a range of thresholds. How a single operating threshold was chosen is not stated. | Partly reported |
| 16 | Results and Strengths and Limitations | Event rates differ between development and external data, and the authors attribute the lower external AUC to interhospital variability. | Partly reported |
| 17 | Methods, Ethical Considerations | Cites the MIT institutional review board protocol covering MIMIC-IV and explains why consent was waived. | Reported |
| 18a | Acknowledgments | Three grants are named. The funders’ role is not described. | Partly reported |
| 18b | End matter | States “Conflicts of Interest: None declared.” | Reported |
| 18c, 18d | Whole text | No protocol or registration statement. | Not found |
| 18e, 18f | Data Availability | States “All intermediate data can be obtained from the authors upon reasonable request, and the related code is available on GitHub”. | Reported |
| 19 | Whole text | No statement on patient and public involvement. | Not found |
| 20a, 20b, 20c | Figure 1, Sample Size, and Table 1 | A flow diagram with exclusion counts, a baseline table covering training, internal and external sets, and event counts for each. | Reported |
| 21 | Methods, Sample Size, and Results, Participants | Event counts and percentages differ between sections (see below), and the external set is given as 4236 patients in Methods and Table 1 but 4358 in one Results sentence. | Judgement call |
| 22 | Methods, Analytical Methods, and GitHub | Stacking structure is described. The fitted meta-model coefficients are not given in the text, so a third party depends on the repository. | Partly reported |
| 23a | Results | AUCs are given as point estimates such as 0.8401 and 0.7077. We found no confidence intervals. | Not found |
| 23b | Whole text | No performance by hospital cluster. | Not found |
| 25 | Discussion, Principal Findings | Main results are interpreted against clinical literature. Fairness is not discussed. | Partly reported |
| 26 | Discussion, Strengths and Limitations | Names retrospective design, lower external performance, a laboratory-based outcome definition, US-only data and the absence of dynamic variables, including that “the definition of SALI based on laboratory thresholds rather than histological confirmation may have misclassified subclinical liver injury”. | Reported |
| 27a, 27b | Whole text | No guidance on poor or missing inputs at deployment, or on the user expertise needed. | Not found |
| 27c | Discussion, Strengths and Limitations | Future prospective work and real-time data streams are proposed. | Reported |
Across the 43 grouped rows, we scored 16 as Reported, 9 as Partly reported, 10 as Judgement call and 8 as Not found. The Not found rows cluster in exactly the areas the 2024 update added or sharpened, which are fairness, protocol and registration, public involvement, subgroup performance and deployment guidance. The paper’s modelling description is thorough, and the gaps sit in the newer items.
What the Annotated Example Teaches
A rule of thumb is not a sample size calculation
The example justifies its sample with a ten events per variable rule. Using its own numbers, 428 training events across 42 predictors gives 10.2 events per predictor, so it clears the rule by a hair. TRIPOD+AI item 10 asks you to explain how the study size was arrived at, and the guideline’s background cites newer sample size methods for developing and evaluating models.1 One of them, by Riley and colleagues in the BMJ, replaces the single rule with criteria that depend on the number of candidate parameters, the outcome proportion and the expected model fit.4 If your model uses flexible machine learning, the parameter count is larger than the predictor count, which makes the ten events rule less reassuring. Report the calculation you used and say why it fits your model type.
Numbers must agree across sections
Item 21 asks for participants and events in each analysis. In the example, the Methods paragraph on sample size lists the training events as 428 (10.10%) and the internal validation events as 165 (3.90%). The Results section gives the same counts as 6.92% and 6.22%, and the arithmetic supports the Results figures, since 428 of 6183 is 6.92% and 165 of 2651 is 6.22%. The external set is 4236 patients in Methods and Table 1, and 648 of 4236 is 15.30%, yet one Results sentence says 4358. None of this changes the findings. It is the kind of discrepancy a reviewer can spot quickly, and an internal consistency pass before submission removes it.
Say what a method is for, and say it once
Item 12c asks for the type of model, all model-building steps and the tuning. When the Methods and the Results describe tuning differently, or name two different meta-models, the reader cannot tell which description is true. Write the model-building paragraph once, in Methods, and let Results refer back to it.
Timing of predictors is not a detail
Item 9b asks how and when each predictor was measured. A prediction model is only useful if its inputs are available before the outcome occurs. When predictor windows and the outcome window overlap, or when a predictor is part of the outcome definition, the reader needs an explicit sentence explaining the prediction time point. This is a clarity query and not an accusation, and it is cheap to fix.
Confidence intervals are expected, not optional
Item 23a asks for performance estimates with confidence intervals. A single AUC from a validation set of a few thousand patients carries real uncertainty, and the external AUC in the example is the number a reader most wants to trust. Bootstrap intervals are inexpensive to add. For a refresher on how regression-based models are built and reported, see our overview of regression models in medical research.
How to Complete the TRIPOD+AI Checklist Before You Submit
The guideline recommends that authors submit a completed checklist that points to the page or line where each item can be found.1 The routine below works for a finished manuscript.
- Download the fillable checklist from the TRIPOD website and keep the version date in your file name.
- Decide D, E or both for your study, then strike the rows that do not apply, with a reason where it is not obvious.
- Fill every row with a location, using page and line numbers from the submission file, not section names alone.
- Write the negative answers for items 18c, 18d, 19 and 3c rather than leaving blanks. A statement that no protocol was prepared satisfies the item.
- Run a consistency pass comparing every count, percentage and model description across abstract, Methods, Results, tables and figures.
- Add confidence intervals to every headline performance estimate and to the key subgroup estimates.
- Check the abstract against the 13-item TRIPOD+AI for Abstracts list.
- Check the journal’s instructions, since some ask for the checklist as a supplement and others as part of the submission form.
Prediction models built on routine data also need a sound data preparation plan. Our walkthrough of preparing patient readmission data for predictive modelling shows the steps that feed items 7 and 11, and our guide to when and how to use machine learning in research helps you decide whether a flexible model is justified at all. For the broader context on clinical applications, see AI in healthcare implementation.
Prospective projects should write the protocol first, since item 18c asks where it can be found. Our study protocol development guide gives a structure you can adapt, and students can map their thesis chapters to the right reporting guideline with the medical dissertation reporting guidelines guide.
After submission, reviewers may appraise your model with PROBAST+AI, which assesses the quality, risk of bias and applicability of prediction models.3 Reading that tool before you write is a cheap way to see your paper through a critical eye. You can find the guideline family and the checklists on the TRIPOD website and the EQUATOR Network TRIPOD page.
Free Research Tools from Editverse
- GraphPro: build publication-ready charts, 25 types, with built-in t-test, ANOVA, correlation, and chi-square p-values. Free, no signup.
- StatPro: free statistical calculators and a test-selector tool covering group comparisons, categorical outcomes, time-to-event analysis, and diagnostic accuracy.
- Flowchart Pro: generate PRISMA 2020, PRISMA-ScR, CONSORT, STROBE, STARD, MOOSE, and ARRIVE flow diagrams directly in your browser.
Professional Support for TRIPOD+AI Reporting
Writing to a 27-item checklist is easier when the analysis and the manuscript were built with it in mind. Two Editverse services map directly onto the rows that cost authors the most time. The first covers model building and interpretation with documented code for projects built on open data, which feeds items 12c, 12g and 18f. The second covers substantive editing and language polishing of the manuscript itself. Editverse editing and writing support aligns with COPE, ICMJE and GPP 2022, and with the ethics requirements of your institution and country.
Machine learning analysis
$1,700
Extended machine learning analysis on public data with multi-source construction, expanded algorithm comparison, SHAP and R/Python code. 4 weeks.
Advanced editing
$550
Substantive editing, scientific language polishing, APA, AMA or Vancouver style editing and reference list verification. 7-10 business days, one free revision round.
If your project needs the statistics rather than the machine learning, our statistical analysis service covers test selection, formatted output and a results narrative.
Conclusion
The TRIPOD+AI checklist asks for 27 things, spread over 52 rows, and the rows that now matter most are the ones the 2015 version never asked about. Our worked example shows the pattern clearly. A paper can describe its modelling in detail and still leave out confidence intervals, a protocol statement, fairness and deployment guidance. Start from the checklist when you plan the analysis, write a location against every row when you draft, and run a consistency pass before you submit. A reviewer who can find every item in minutes is a reviewer who spends the review on your science.
Frequently Asked Questions
Does TRIPOD+AI replace TRIPOD 2015?
Yes. The TRIPOD+AI authors state that the new checklist supersedes the 2015 checklist, which should no longer be used. Follow the 2024 version for any new submission.
How many items does the TRIPOD+AI checklist have?
The checklist has 27 numbered items. Counting the lettered sub-items, it has 52 rows. Forty apply to both development and evaluation studies, eight to development only and four to evaluation only. A separate list of 13 items covers abstracts.
Do I need TRIPOD+AI if I used logistic regression and not machine learning?
Yes. TRIPOD+AI covers prediction model studies that use regression or machine learning methods, so a logistic regression risk model falls within it.
Is TRIPOD+AI the same as PROBAST+AI?
No. TRIPOD+AI is a reporting guideline that checks whether a paper describes what was done. PROBAST+AI is an appraisal tool that assesses risk of bias and applicability. Reviewers and systematic reviewers use both, for different purposes.
Should I submit the completed checklist with my manuscript?
The guideline recommends submitting a completed checklist that gives the page or line where each item can be found. Many journals ask for it as a supplement, so check the author instructions.
What if an item does not apply to my study?
Mark it as not applicable and give a short reason, for example that the outcome comes from a laboratory threshold and has no subjective assessors. A stated reason is better than a blank row.
Where do I find the official TRIPOD+AI checklist?
The official fillable checklist is on the TRIPOD website, and the full guideline article is open access in the BMJ with a CC BY licence.
References
- Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. doi:10.1136/bmj-2023-078378. PMID: 38626948. PubMed, PMC.
- Lei J, Zhai J, Zhang Y, Qi J, Sun C. Supervised machine learning models for predicting sepsis-associated liver injury in patients with sepsis: development and validation study based on a multicenter cohort study. J Med Internet Res. 2025;27:e66733. doi:10.2196/66733. PMID: 40418571. PMCID: PMC12149780. CC BY 4.0. PubMed, PMC.
- Moons KGM, Damen JAA, Kaul T, Hooft L, Andaur Navarro C, Dhiman P, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025;388:e082505. doi:10.1136/bmj-2024-082505. PMID: 40127903. PMC.
- Riley RD, Ensor J, Snell KIE, Harrell FE Jr, Martin GP, Reitsma JB, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441. doi:10.1136/bmj.m441. PMID: 32188600. PubMed.