Pre-Market Guidance for Machine Learning-Enabled Medical Devices – Health Canada Recommendations

Purpose and Scope

This guidance assists manufacturers in preparing pre-market submissions for machine learning-enabled medical devices (MLMD), including software as a medical device (SaMD) and hardware incorporating ML components. It focuses on unique ML considerations such as data dependency, model generalizability, and performance drift, while aligning with Canada’s Medical Devices Regulations and international frameworks (IMDRF, FDA, EU). The document applies to Class II, III, and IV MLMD intended for diagnostic, prognostic, or therapeutic purposes. Source: Pre-Market Guidance for Machine Learning-Enabled Medical Devices - Canada.ca

Key Principles for MLMD Regulation

• Risk-Based Approach: Evidence and oversight proportional to device risk class and ML impact on clinical decisions
• Lifecycle Management: Emphasis on continuous monitoring and management of changes post-market
• Transparency and Explainability: Manufacturers must provide sufficient information for regulators, healthcare providers, and users to understand model behavior and limitations
• Data Quality and Representativeness: Training, validation, and test datasets must be diverse, well-characterized, and representative of the intended population

Recommended Elements in Submissions

Device Description and Intended Use

• Clearly define the ML component’s role (e.g., diagnostic aid, risk stratification)
• Specify input data types, output format, and clinical decision impact
• Describe intended population, use environment, and limitations

Data Management

• Detail source, collection methods, annotation process, and inclusion/exclusion criteria
• Provide demographic breakdowns and evidence of diversity
• Describe data splitting (training/validation/test) and independence

Model Development and Validation

• Explain architecture, training methodology, hyperparameters, and optimization process
• Present performance metrics appropriate to the task (e.g., sensitivity/specificity, AUC, calibration)
• Include independent test set results and subgroup analyses
• Demonstrate robustness to variations in input data

Risk Management

• Integrate ML-specific risks (overfitting, bias, drift, adversarial attacks) into ISO 14971 process
• Identify mitigations such as locked vs. adaptive models, uncertainty estimation, and fail-safe mechanisms

Transparency and Interpretability

• Provide model cards or fact sheets summarizing key characteristics
• Include explainability methods where feasible (e.g., feature importance, saliency maps)
• Describe how users will be informed of model confidence and limitations

Change Management and Post-Market Plans

• Define predetermined change control plans (PCCP) for anticipated modifications
• Outline post-market surveillance strategy, including performance monitoring and retraining triggers
• Commit to reporting significant changes requiring new submissions

Alignment with International Standards

The guidance references IMDRF documents on SaMD and AI/ML, FDA’s Good Machine Learning Practice principles, and IEC 62304/82304 for software lifecycle processes. Manufacturers are encouraged to leverage these frameworks to support global market access.

This document supports safe innovation in MLMD by providing clear expectations for evidence and transparency. Detailed examples, acceptable performance metrics, PCCP templates, and submission checklists are included in the official Health Canada pre-market guidance for machine learning-enabled medical devices. Source: Pre-Market Guidance for Machine Learning-Enabled Medical Devices - Canada.ca