Guiding Principles for Good Machine Learning Practice

The U.S. Food and Drug Administration (FDA), Health Canada, and the United Kingdom’s Medicines and Healthcare Products Regulatory Agency (MHRA) have jointly outlined ten guiding principles to support the development of Good Machine Learning Practice (GMLP). These principles aim to advance the creation of medical devices that incorporate artificial intelligence and machine learning (AI/ML) in a manner that ensures safety, effectiveness, and high quality.

Guiding Principles

1. Multi-Disciplinary Expertise Is Applied Across the Total Product Life Cycle – Effective integration of machine learning within clinical workflows requires a thorough understanding of intended use, clinical context, patient benefits, and potential risks. This ensures that AI/ML-enabled medical devices perform safely and effectively.

2. Strong Software Engineering and Security Practices Are Followed – Device development must incorporate robust software engineering, strong data quality controls, comprehensive data management, and cybersecurity protections.

3. Clinical Study Participants and Data Sets Reflect the Intended Population – Data collection processes should ensure that the characteristics of the target patient population and clinical use scenario are adequately represented in training, testing, and clinical datasets. This supports generalizable performance, reduces bias, and identifies conditions where the model may not perform as expected.

4. Training Data Sets Remain Separate from Test Data Sets – Training and test datasets must be independently selected and maintained to ensure objective evaluation of model performance.

5. Reference Datasets Use the Best Available Methodologies – High-quality reference datasets should be developed using accepted, validated methods. These datasets must contain clinically meaningful, well‑characterized information, along with clear documentation of any limitations.

6. Model Design Aligns with Available Data and Intended Use – The model should be designed to mitigate known risks such as overfitting, loss of performance, and security vulnerabilities while supporting its intended clinical purpose.

7. Emphasis Is Placed on Human-AI Team Performance – For systems involving human interaction, human factors engineering and interpretability of outputs are essential. Overall performance should be evaluated based on the combined effectiveness of the human-AI team rather than the model alone.

8. Testing Shows Device Performance Under Realistic Clinical Conditions – Testing should evaluate how the device performs in conditions that closely match clinical use, including consideration of patient subgroups, environments, input variability, and possible confounding influences.

9. Users Receive Clear and Essential Information – Users should be informed about device indications, intended use, data sources used for model training and testing, known limitations, workflow integration considerations, and any updates resulting from ongoing real‑world monitoring. Mechanisms for users to report concerns should also be provided.

10. Deployed Models Are Monitored and Re-Training Risks Are Managed – Once deployed, models must be continuously monitored under real-world conditions to ensure that safety and performance are maintained or improved. Risks associated with model updates or retraining must be proactively managed.