Model Card
Structured documentation of a machine learning model's intended use, performance, and limitations.
Definition
A model card summarizes a model's purpose, training and evaluation data, performance metrics across relevant subgroups, ethical considerations, and known limitations to support transparent, responsible use.What this means in practice
Increasingly expected by regulators (e.g., FDA's transparency principles for ML-enabled devices) and procurement teams. Often combined with 'Indications-for-Use'-style cards for clinical AI.Examples
- A MedTech company developing an AI diagnostic tool for radiology creates a model card detailing the training data's demographic representation and the model's sensitivity and specificity across different patient populations.
- Before deploying an AI-powered insulin pump, the manufacturer publishes a model card outlining the algorithm's performance under various physiological conditions and known limitations, such as interaction with specific medications.
- A hospital procurement team reviews the model card for an AI-driven surgical planning system to assess its suitability for their patient demographics and to understand its validation against relevant clinical endpoints.
- •Failing to update model cards when models are retrained or modified, leading to outdated or inaccurate information, is a common pitfall.
- •Overlooking the inclusion of performance metrics for all relevant subgroups, which can mask biases or disparities, is a mistake to avoid.
- •Presenting overly technical information without clear, concise summaries for diverse audiences can hinder effective communication.
- •Ignoring the iterative nature of AI development and not treating the model card as a living document can lead to compliance issues.
Frequently asked questions
Cross-references
Used by
Related terms
Shared paths + categoryGuiding principles for the development of AI/ML-enabled medical devices.
Medical device that uses artificial intelligence or machine learning to perform its intended use.
Ongoing surveillance of input data and model outputs to detect performance degradation post-deployment.
Methods (SHAP, saliency maps, prototype explanations) that help users understand why a model made a given prediction.
Large pretrained model adaptable to many downstream clinical tasks via fine-tuning or prompting.
Devices that use LLMs or diffusion models to generate text, images, or recommendations within a clinical workflow.
Latest in MedTech
Primary references
3 sources- 1FDA Transparency Principles for ML DevicesVerifiedFDAfda.gov
- 2FDA - Software as a Medical Device (SaMD)VerifiedFDAfda.gov
- 3FDA - AI/ML-Enabled Medical DevicesVerifiedFDAfda.gov
Inline markers like [1] jump to the matching reference above.