Good Machine Learning Practice Principles for Medical Device Development – Health Canada Joint Guidance

Introduction and Purpose

In October 2021, Health Canada, the U.S. Food and Drug Administration (FDA), and the United Kingdom’s Medicines and Healthcare products Regulatory Agency (MHRA) jointly published a discussion paper titled “Good Machine Learning Practice for Medical Device Development: Guiding Principles.” This document identifies 10 guiding principles intended to promote safe and effective development of machine learning-enabled medical devices (MLMD), foster best practices, and support regulatory convergence across jurisdictions. The principles apply to ML models used as components of medical devices, including software as a medical device (SaMD) and hardware-embedded ML. Source: Good Machine Learning Practice for Medical Device Development - Canada.ca

The 10 Good Machine Learning Practice Principles

1. Multi-Disciplinary Expertise

Leverage expertise from software engineering, data science, clinical/medical domains, and regulatory science throughout the total product lifecycle.

2. Data Independence

Implement processes to manage independence between training, tuning, and test datasets to ensure unbiased performance assessment and reduce overfitting.

3. Data Relevance and Representativeness

Ensure datasets are relevant, representative of the intended use population, and sufficiently diverse to minimize bias and support generalizability.

4. Device Performance Assessment

Establish clinically meaningful performance goals and use appropriate metrics tailored to the intended use and clinical context.

5. Appropriate Reference Datasets

Use high-quality, well-characterized reference datasets (ground truth) for training and validation, with clear annotation protocols.

6. Sound Human Factors and Usability Engineering

Incorporate human-centered design principles to ensure safe and effective interaction between users (healthcare professionals and patients) and the ML-enabled device.

7. Comprehensive Risk Management

Apply risk management throughout the lifecycle, addressing ML-specific risks such as data drift, model degradation, bias, and adversarial attacks.

8. Transparency and Explainability

Provide clear, comprehensive documentation of model architecture, training process, inputs/outputs, limitations, and performance to enable informed use.

9. Lifecycle Management Plan

Develop a predetermined change control plan (PCCP) to manage anticipated modifications, including retraining, while maintaining safety and effectiveness.

10. Post-Market Monitoring

Establish robust post-market surveillance to detect performance changes, emerging risks, or new use cases, with mechanisms for timely updates and communication.

Application and Alignment

These principles are non-binding but reflect current thinking on best practices for MLMD. They align with existing standards (ISO 14971, IEC 62304) and complement agency-specific guidance. Manufacturers are encouraged to integrate GMLP into quality management systems and pre-market submissions.

The joint document supports innovation while ensuring patient safety in ML-enabled medical technologies. Full descriptions, rationale, and examples for each principle are detailed in the official Good Machine Learning Practice guiding principles published by Health Canada, FDA, and MHRA. Source: Good Machine Learning Practice for Medical Device Development - Canada.ca