What TRIPOD+AI is and who maintains it
TRIPOD+AI, Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis, including artificial intelligence, is an updated reporting guideline for studies that develop or validate a model predicting an individual's diagnosis or prognosis from multiple predictors. It extends the original 2015 TRIPOD statement to explicitly cover machine learning models alongside conventional regression models, and was developed by a large international group of methodologists, clinicians and data scientists (Collins et al., 2024).
The checklist covers the title and abstract, introduction, methods (source of data, participants, outcome, predictors, sample size, missing data, the statistical or machine learning analysis, and how performance was evaluated), results and discussion. A key item throughout is stating clearly whether the study develops a new model, validates an existing model, or does both.
Why it exists
A prediction model is often reported with a single headline number, such as an area under the curve of 0.85, which on its own tells a reader almost nothing about whether the model is ready for use. Readers need to know how the model was built, on what data, whether it was tested on data it had not seen, and how well its predicted probabilities matched what actually happened, not only how well it ranked patients from higher to lower risk.
Machine learning models raise this problem again in a sharper form, because flexible algorithms can fit patterns in the training data that do not generalise, and because reporting practices around data splitting, feature selection and hyperparameter tuning have historically been inconsistent. TRIPOD+AI exists to bring the same discipline that TRIPOD brought to regression models to this wider class of methods.
A further, specific risk with flexible models is data leakage: information from the data used to test a model inadvertently influencing how it was built, for example when feature selection or hyperparameter tuning is carried out on the whole dataset before splitting it into training and test sets. Leakage of this kind inflates apparent performance in a way that can be almost invisible unless the modelling pipeline is reported step by step, which is exactly what TRIPOD+AI's methods items ask for.
Which studies it fits
The guideline also covers model updating, where an existing model is recalibrated or extended for a new population rather than built from scratch; this is treated as its own category because the reporting needs are different again from either a fresh development or a straightforward external validation.
TRIPOD+AI applies to any study that develops, validates or updates a multivariable model to predict an individual's probability of having a condition now (diagnostic prediction) or developing one in the future (prognostic prediction), whether the model is a logistic regression, a Cox model, a random forest, a neural network or another machine learning method.
| Study type | What it reports | TRIPOD+AI expectation |
|---|---|---|
| Development | Building a new model from a dataset | Full detail on predictors, outcome definition, sample size, handling of missing data and model-building steps |
| Internal validation | Testing performance on data from the same source, e.g. by cross-validation or bootstrapping | The validation method used and whether it accounts for the full model-building process, including any feature selection |
| External validation | Testing an existing model on an independent dataset | How the validation population differs from the development population, and performance in that new setting |
| Development and validation combined | Both steps reported in one paper | Clear separation of development and validation results, not a single blended performance figure |
Where TRIPOD+AI does not apply
TRIPOD+AI is for models that predict an individual outcome from multiple predictors, and it should not be used for a study that simply tests whether one exposure is associated with one outcome; that is an observational association study and belongs under STROBE instead (von Elm et al., 2007). It also does not apply to a study evaluating a single index test's accuracy against a reference standard with no combination of predictors, which belongs under STARD (Bossuyt et al., 2015).
A trial that tests whether using a prediction model in clinical practice changes patient outcomes, as opposed to a study that reports the model's statistical performance, is an intervention study and should be reported with CONSORT and its relevant extension, since the question being asked (does using the model help patients) is different from the question TRIPOD+AI reporting answers (how accurate is the model).
It is worth separating these two questions explicitly in a doctoral thesis, because they are often run together in a single chapter. Showing that a model is accurate is a necessary step before testing whether using it changes care, but accuracy alone does not establish that acting on the model's output improves outcomes; that claim needs its own, separate study design.
Why it matters
Prediction models that look promising on paper very often perform worse when tested on new data, and many published models are never properly externally validated before being proposed for use. Reporting that is complete enough to show exactly how a model was built and evaluated is what lets other researchers attempt an external validation at all; a model whose predictors, coding and thresholds are not fully specified cannot be reproduced by anyone else. The same indirectness concerns that apply to diagnostic accuracy studies apply here too: a model built on one population may not transfer to another, which frameworks such as GRADE take into account when rating how much confidence to place in model-based evidence (Guyatt et al., 2008).
How it is applied in practice
Journals publishing clinical prediction model research increasingly require a completed TRIPOD or TRIPOD+AI checklist at submission, and some now explicitly ask authors to state whether their model has been externally validated and, if not, to justify why it is being reported before that step. Sharing code and, where governance allows, data following the FAIR principles is increasingly expected alongside the written report, since a prediction model is, in practice, a piece of software as well as a set of coefficients, and independent testing usually needs access to it (Wilkinson et al., 2016).
For a doctoral researcher, the most common mistake is reporting a model's internal performance (how it does on the data used to build it) as if it were external validation; TRIPOD+AI's insistence on stating the study type up front is a direct response to this confusion.
Where a doctoral project plans to develop a model and validate it on a separate dataset, deciding and recording the analysis plan, including which predictors will be tried and how performance will be measured, before looking at the validation data serves the same purpose that a registered protocol serves for a trial: it keeps a clear line between what was planned and what was found after the fact (Nosek et al., 2018).
An illustrative example
The passage below is a hypothetical illustration, not a quotation from a published study.
| Element | Without TRIPOD+AI | With TRIPOD+AI |
|---|---|---|
| Study type | "We built and tested a model to predict readmission." | "This is a model development study with internal validation by 10-fold cross-validation, repeated 100 times; no external validation was performed." |
| Predictors | "Clinical and demographic variables were used." | "Twelve predictors were included, selected before model fitting based on prior literature; each is listed with its definition, timing of measurement and how missing values were handled (multiple imputation, 20 datasets)." |
| Model | "A machine learning model was trained." | "A gradient-boosted tree model (XGBoost, version 2.0) was trained with hyperparameters tuned by nested cross-validation; the final model and code are available in a public repository." |
| Performance | "The model achieved an AUC of 0.85." | "Discrimination (C-statistic 0.85, 95% CI 0.81–0.89) and calibration (calibration plot and slope, 0.94) are reported together, with performance also reported separately for the internal validation folds." |
| Applicability | Not discussed | "The development cohort was drawn from a single urban tertiary hospital; performance in other settings, such as rural or paediatric populations, has not been tested." |
What to check before you submit
- Does the title or abstract state plainly whether this is a development, validation or combined study?
- Are all predictors defined, with their timing of measurement and how missing values were handled?
- Is the model, including its final form or code, reported in enough detail for another researcher to apply it?
- Are both discrimination and calibration reported, not discrimination alone?
- Is the development or validation population described well enough to judge applicability elsewhere?
- If internal validation was used, does it account for the full model-building process, including any feature selection or tuning?
- Official checklist and source: TRIPOD Statement (tripod-statement.org) .
Key takeaways
- TRIPOD+AI reports how a multivariable prediction model, regression-based or machine learning, was developed and validated (Collins et al., 2024).
- It exists because headline statistics such as a single AUC do not show how well a model will perform on new data or new populations.
- It fits prediction models with multiple predictors; single-test accuracy studies use STARD (Bossuyt et al., 2015) and single-exposure association studies use STROBE (von Elm et al., 2007).
- Development and validation performance should always be reported and labelled separately, never blended.
- Sharing the model's code and data, where governance allows, follows FAIR principles and is often what makes external validation possible (Wilkinson et al., 2016).
- Fixing the analysis plan before looking at validation data protects against the same hindsight problem preregistration addresses in other designs (Nosek et al., 2018).
