Model card

A model card is a short document that accompanies a trained machine-learning model and reports what it is for, how it was built, and how it performs, including across the demographic or clinical groups relevant to its use 1. Mitchell and colleagues proposed the format in 2019. It has since become the common container for AI transparency in healthcare.

Why do model cards matter in healthcare?

A clinician or buyer usually cannot inspect a model's code or training data. The card is the substitute. The original proposal has nine sections: model details, intended use, factors, metrics, evaluation data, training data, quantitative analyses, ethical considerations, and caveats and recommendations 1. Two ideas do most of the safety work. The intended-use section names uses the model was never meant for. And disaggregated evaluation reports performance for each relevant group, so a weak subgroup cannot hide inside a strong average.

For bedside use, Duke researchers proposed a one-page "Model Facts" label, so that front-line clinicians "actually know how, when, how not, and when not to incorporate model output into clinical decisions" 2.

How do regulators use the format?

Two US tracks converge on it. The FDA's January 2025 draft guidance on AI-enabled device software recommends a model card in device labeling and includes an example model card and an example 510(k) summary with one 3. The ONC's HTI-1 rule requires certified health IT to support 31 source attributes for predictive decision support interventions, which cover much of what a model card holds 4.

Both are in flux. The FDA guidance is still a draft. A December 2025 HTI-5 proposed rule would remove the HTI-1 source-attribute requirement 5, and it had not been finalized as of September 2026. The full picture is in transparency and labeling requirements.

What do real disclosures look like?

Thin. A review of 1,012 FDA summaries for AI/ML-enabled devices scored transparency across 17 categories. The average score was 3.3. Nearly half of devices reported no clinical study, and over half reported no performance metric 6. Among machine-learning-enabled devices authorized in 2024, 15.5% of summaries provided demographic data 7.

Common misunderstandings

A model card proves the model works. It documents claims. It does not verify them. Performance at your site still needs external validation.

One card covers every version. A card describes one model at one point in time. After retraining, or as populations shift, it goes stale; see model drift.

A complete card means a fair model. Disaggregated numbers show where performance differs. Deciding whether a gap is acceptable is a clinical and governance judgment; see subgroup performance and bias audits.

What to ask a vendor

• Can we see the model card before the demo?
• Which model version does it describe, and when was it last updated?
• Does it report performance by age, sex, race and ethnicity, and site, with the prevalence in each evaluation set?
• What uses does it list as out of scope?
• Was the evaluation data external to the training data?
• For a regulated device, does the card match the 510(k) or De Novo decision summary?

Related terms

• Transparency and labeling requirements: the rules that ask for this information.
• External validation: the test a card's numbers should survive.
• The hospital AI governance committee playbook: where a card gets read at intake.