§01
Pre-market expectations for ML-enabled devices
The guidance walks manufacturers through risk-based scoping, data quality, model development, validation, transparency, and post-market monitoring · with explicit hooks for adaptive vs. locked models. The framing is GMLP-first, with regulatory specifics layered on top.
- Clear articulation of intended use, deployment environment, and user.
- Documented data lineage, quality, and bias mitigation.
- Validation against representative populations with subgroup analysis.
- Defined performance monitoring with thresholds and escalation paths.
§02
Predetermined changes
Like the FDA's PCCP, Health Canada expects manufacturers to declare in advance which model changes are anticipated and how they will be controlled. Anything outside that envelope triggers a licence amendment.
§03
Cybersecurity for medical devices
Health Canada's cybersecurity guidance aligns with IMDRF principles and FDA expectations: threat modelling, SBOM, secure update, vulnerability disclosure, and lifecycle management. For AI SaMD, the model layer is increasingly part of the conversation.
§04
Real-world performance
The guidance signals a clear expectation that real-world performance · not just pre-market validation · must be monitored, characterised, and reported. This is where adaptive AI either earns trust or attracts scrutiny.