Model Card

Basic identifying information about the model artifact itself.

• Model Name and Version: A unique identifier and semantic version string (e.g., chest-xray-pneumonia-v2.1).
• Model Type: The specific architecture used, such as a DenseNet-121 convolutional neural network.
• Development Organization: The legal entity responsible for training and releasing the model.
• Release Date: The ISO 8601 date when this specific version was finalized.
• Contact Information: A monitored email alias or repository for reporting errors and ethical concerns.

A precise, unambiguous description of what the model is designed to do and, critically, what it is not designed to do.

• Primary Use Case: The specific clinical or operational task, e.g., 'Assist radiologists by flagging suspicious regions in posteroanterior chest radiographs for patients aged 18+.'
• Out-of-Scope Applications: Explicitly forbidden uses. For example, 'This model is not intended for pediatric patients, for use as a standalone diagnostic device, or for detecting non-pulmonary pathologies.'
• Suitable User Profile: The intended operator, such as a board-certified radiologist.

Quantitative evaluation results broken down by demographic and environmental factors to surface hidden biases.

• Overall Metrics: Standard measures like Area Under the Receiver Operating Characteristic (AUROC), sensitivity, and specificity on a held-out test set.
• Disaggregated Evaluation: Performance sliced by sex, age, ethnicity, and equipment manufacturer. A model may achieve 0.95 AUROC overall but only 0.82 on portable X-ray machines.
• Decision Threshold Analysis: The precision-recall trade-off at the specific operating point used in production.

A comprehensive description of the datasets used, establishing lineage and identifying potential sources of bias.

• Source Institutions: The hospitals or public repositories (e.g., NIH ChestX-ray14, CheXpert) that contributed data.
• Demographic Composition: The distribution of age, sex, and race in the training set. A model trained on 90% male patients is a critical finding.
• Labeling Methodology: How ground truth was established—was it a majority vote of three radiologists, or a natural language processing extraction from radiology reports?
• Preprocessing Steps: The exact normalization, windowing, and resizing operations applied.

A candid assessment of the model's risks, biases, and failure modes.

• Known Biases: Documented performance disparities. For instance, 'The model exhibits a 7% lower sensitivity for patients with a BMI > 40 due to increased image noise.'
• Failure Modes: Common scenarios where the model is unreliable, such as the presence of surgical drains or severe scoliosis.
• Fairness Analysis: Results from counterfactual fairness testing or equalized odds assessments.
• Environmental Impact: The estimated carbon footprint of training, measured in kg of CO2 equivalent.

Guidance for safe deployment and ongoing monitoring, transforming the card from a static report into an operational tool.

• Human-in-the-Loop Mandate: An explicit statement that model output is a decision support tool, not a clinical decision.
• Distribution Shift Monitoring: A recommendation to continuously monitor input data drift using metrics like Maximum Mean Discrepancy (MMD).
• Update Cadence: The planned frequency for retraining or fine-tuning the model to prevent concept drift.
• Adversarial Robustness: Known vulnerabilities to specific adversarial patch attacks or image perturbations.