Model Cards

A model card must explicitly define the use cases for which the model was designed and validated, as well as out-of-scope applications where performance is untested or likely to fail. This section prevents dangerous off-label deployment.

• Specifies target patient populations, imaging modalities, or data types
• Enumerates known failure modes and demographic exclusions
• Distinguishes between primary and secondary intended uses per FDA guidance

This section reports quantitative performance across relevant evaluation datasets, broken down by meaningful subgroups to surface hidden disparities. It moves beyond aggregate accuracy to reveal where a model underperforms.

• Reports sensitivity, specificity, AUC, and PPV for diagnostic models
• Includes slice-based analysis across age, sex, and race strata
• Documents the evaluation dataset's provenance, size, and labeling methodology

Model cards require a candid assessment of ethical risks, including potential biases and disparate impacts. This section demonstrates that developers have proactively interrogated their model for harm before release.

• Documents fairness metrics such as equal opportunity difference
• Describes mitigation strategies applied during training or post-processing
• Acknowledges limitations in the fairness analysis itself

Transparency about the training data's composition, collection process, and preprocessing allows downstream users to assess representativeness and potential distribution shift. This section provides the provenance necessary for auditability.

• Describes data sources, inclusion criteria, and annotation procedures
• Reports demographic distributions and known gaps in coverage
• Summarizes the model architecture, training objective, and hyperparameters

Every model card must honestly communicate what the model cannot do and under what conditions its predictions become unreliable. This section also addresses the model's lifecycle and update cadence.

• Warns against distribution shift and concept drift over time
• Specifies environmental constraints such as input resolution or format
• Indicates the versioning policy and planned retraining schedule

A game-theoretic approach to explain the output of any machine learning model by computing the contribution of each feature to a prediction. In the context of model cards, SHAP values provide:

• Quantitative feature attribution for documented performance breakdowns
• Patient-level explanations required for clinical decision support transparency
• Fairness evaluation across demographic subgroups

SHAP analysis is often reported in the evaluation section of model cards to demonstrate post-hoc explainability.

The process of estimating and characterizing the different sources of noise and model ignorance that contribute to the total predictive uncertainty of a model's output. Model cards should report:

• Aleatoric uncertainty: Inherent noise in the data distribution
• Epistemic uncertainty: Model ignorance due to limited training data
• Calibration metrics such as Expected Calibration Error (ECE)

Reporting uncertainty bounds is critical for clinical deployment, where overconfident misdiagnosis can have life-threatening consequences.

A mathematical framework that provides a provable guarantee of privacy by adding calibrated noise to computations, limiting the risk of inferring any single individual's data. Model cards intersect with differential privacy when:

• Reporting the privacy budget (ε, δ) used during training
• Documenting the trade-off between privacy guarantees and model accuracy
• Demonstrating compliance with HIPAA and GDPR requirements

This is especially relevant for models trained on sensitive diagnostic datasets.

Quantitative measures that assess how accurately an explanation method reflects the true reasoning process of the underlying machine learning model. When model cards report explainability results, faithfulness metrics validate that:

• Feature attributions genuinely reflect model decision boundaries
• Explanations are not misleading or superficially plausible
• Completeness and sufficiency axioms are satisfied

Without faithfulness verification, reported explanations in model cards may create a false sense of transparency.