Written and maintained by CASRAI Editorial Board
Last updated
MINIMAR (MINimum Information for Medical AI Reporting) is a reporting checklist for clinical artificial-intelligence and machine-learning models, proposed by Tina Hernandez-Boussard, Selen Bozkurt, John P. A. Ioannidis and Nigam H. Shah in the Journal of the American Medical Informatics Association (JAMIA) in December 2020. It sets out the minimum information a paper, model card or technical report needs to include so a reader can judge what a clinical AI model was trained to predict, who it was trained and validated on, and how well its performance is likely to generalize to a different population — the three questions the authors identified as most often left unanswered in published health-AI research.
MINIMAR is broader in scope than the disease- or task-specific AI reporting guidelines that came after it. Where STARD-AI covers diagnostic-accuracy studies specifically and TRIPOD+AI covers multivariable prediction-model development and validation, MINIMAR applies to any clinical AI/ML model — diagnostic, predictive, triage, risk-scoring or clinical-decision-support — regardless of the specific study design built around it.
The Four MINIMAR Reporting Domains
MINIMAR organizes its minimum-information requirements into four domains. The original paper does not present them as a numbered checklist of discrete items the way STARD or TRIPOD+AI do; instead it groups the information a reader needs into these four areas:
| Domain | What it requires reporting on |
|---|---|
| 1. Study population and setting | Where the data came from — population source, care setting (academic medical center, community health system, etc.), the underlying data source (e.g. EHR-derived), and the cohort’s inclusion/exclusion criteria. |
| 2. Patient demographic characteristics | Age, sex, race, ethnicity and socioeconomic measures for the training population, reported so a reader can compare that population against the one the model will actually be deployed on. |
| 3. Model architecture | The model’s output (e.g. a risk score), its intended user (clinician, care team, hospital administration), how data were split for training/testing, the source of ground-truth/gold-standard labels, the prediction task (classification vs. prediction), the algorithm type, the feature list with any transformations applied, and how missing data were handled. |
| 4. Model evaluation | How the model was optimized and tuned, the internal validation approach used, whether and how it was externally validated, and whether code and data are shared to support reproducibility. |
Domains 1 and 2 exist to answer a single underlying question: does the population the model was built on look like the population it will be used on? Domains 3 and 4 exist to make the model itself, and the evidence for its performance, auditable rather than a black box. A reader who has all four domains in front of them can independently judge whether a model’s reported accuracy is likely to hold up in their own setting — the specific failure mode MINIMAR was written to prevent.
Why MINIMAR Exists
Hernandez-Boussard and coauthors wrote MINIMAR in response to a documented pattern in early clinical-AI publications: papers reporting strong model performance while omitting basic facts about who the training population was, leaving readers unable to tell whether a model validated on one hospital’s patients would generalize to a different demographic mix elsewhere. That gap matters more in health AI than in most other applied-ML domains, because a model trained on an unrepresentative population can silently underperform — or actively harm — patients in groups it was not adequately trained or validated on. MINIMAR frames complete reporting as a prerequisite for catching that failure mode before deployment, not an optional transparency nicety.
The paper is explicit that MINIMAR is a proposal meant to start a broader community conversation about AI reporting standards in health care, not a finished, consensus-ratified checklist. That distinction matters when comparing it to what came afterward.
MINIMAR vs. STARD-AI, TRIPOD+AI, CONSORT-AI and DECIDE-AI
MINIMAR (2020) predates all four of the major AI-specific extensions to established clinical reporting guidelines, and it occupies a different niche than any of them:
- MINIMAR — general-purpose minimum-information checklist for any clinical AI/ML model. Proposed by a single author group, not developed through a formal multistakeholder Delphi consensus process.
- STARD-AI — AI-specific extension of STARD 2015, scoped to diagnostic-accuracy studies where the index test is an AI/ML system. Developed via a 240+ expert international Delphi consensus, published in Nature Medicine.
- TRIPOD+AI — 2024 update to TRIPOD, scoped to multivariable prediction-model development and validation studies, ML-based or regression-based.
- CONSORT-AI — extension of CONSORT, scoped to randomized controlled trials of an AI intervention.
- DECIDE-AI — scoped to early-stage, small-sample clinical evaluation of an AI decision-support system, between development and a full trial.
In practice, the four newer guidelines have absorbed most of the study-design-specific reporting requirements in their respective niches, and journals increasingly ask authors to complete whichever one matches their study design. MINIMAR’s four domains still function well as a general baseline — particularly for AI systems and technical reports (model cards, internal validation summaries, vendor documentation) that do not fit neatly into a trial, prediction-model-validation, or diagnostic-accuracy study design, where none of the newer, more specific checklists apply.
Where MINIMAR’s Domains Overlap with CASRAI’s Own Vocabulary
MINIMAR’s four domains map directly onto concepts CASRAI’s Dictionary already defines independently, which is worth knowing if you’re documenting a model rather than just checklisting a manuscript:
- Domain 1 (study population and setting) and its “data source” item are the same concern as data provenance and, for training datasets specifically, training data provenance — and, where that training data itself needs a durable, citable identity, persistent identifiers for AI models and training datasets.
- Domains 1 and 2 together are the underlying substance of what CASRAI’s Dictionary defines as AI fairness and the closely related AI bias — a model can only be fairness-audited against the demographic breakdown of the population it was actually trained and validated on.
- Domains 3 and 4 (model architecture and evaluation) are what a model card is designed to capture in a standardized, machine-readable format, and what a model bias audit formally tests against.
- The “generalize to a different population” question MINIMAR’s domains 1–2 exist to let a reader answer is the same question this site’s guide to generalizability in research and its internal-vs-external-validity comparison cover for research design more broadly — MINIMAR is that same concern applied specifically to a deployed clinical AI model rather than a study sample.
Using MINIMAR When Preparing a Manuscript, Model Card or Technical Report
MINIMAR was not built as a fillable checklist with numbered items the way STARD or TRIPOD+AI are, so there is no official downloadable form to attach as a submission file. In practice, authors and documentation teams use it as a coverage check: read the four domains above, and confirm the manuscript, model card or internal report states something concrete under each one. If a domain has nothing written against it — most commonly domain 2 (demographic characteristics of the training population) or the external-validation half of domain 4 — that is the specific gap MINIMAR is designed to surface before the AI system reaches a reader who needs to judge whether it applies to their own patients.
Because MINIMAR has not gone through the same multistakeholder consensus process as STARD-AI or TRIPOD+AI, individual journals rarely mandate it by name the way they now mandate those two for their respective study designs. It is, however, listed in the EQUATOR Network’s reporting guidelines library, and its four domains are frequently cited as a baseline reference point in subsequent AI-reporting-guideline development work — including, directly, in the introductions to both the STARD-AI and TRIPOD+AI consensus papers.
Frequently Asked Questions
What does MINIMAR stand for?
MINimum Information for Medical AI Reporting.
Who wrote MINIMAR, and where was it published?
Tina Hernandez-Boussard, Selen Bozkurt, John P. A. Ioannidis and Nigam H. Shah, published in the Journal of the American Medical Informatics Association, volume 27, issue 12, pages 2011–2015, December 2020 (DOI: 10.1093/jamia/ocaa088).
Is MINIMAR the same thing as TRIPOD+AI or CONSORT-AI?
No. MINIMAR is a general-purpose minimum-information checklist for any clinical AI/ML model, published in 2020 by a single author group. TRIPOD+AI, CONSORT-AI, DECIDE-AI and STARD-AI are later, study-design-specific extensions of established reporting guidelines, each developed through a formal multistakeholder consensus process and each scoped to a particular kind of study (prediction-model validation, randomized trial, early-stage evaluation, and diagnostic-accuracy study respectively).
Is MINIMAR a mandatory reporting standard?
No journal is known to mandate MINIMAR by name as a required submission checklist, unlike STARD-AI and TRIPOD+AI, which are increasingly required for their respective study types. MINIMAR functions as a baseline reference and coverage check, and is most useful for AI systems and technical documentation — model cards, internal validation summaries, vendor documentation — that fall outside the specific study designs the newer guidelines cover.
What are the four MINIMAR domains?
Study population and setting; patient demographic characteristics; model architecture; and model evaluation.
Does MINIMAR apply to diagnostic AI, predictive AI, or both?
Both, and more broadly than that. MINIMAR was written to apply to any clinical AI/ML model — diagnostic, predictive, risk-scoring, triage or clinical-decision-support — rather than being scoped to one study design the way STARD-AI (diagnostic accuracy) and TRIPOD+AI (prediction-model development/validation) are.
