AI Regulation Tracker  /  Regulator guidance

Health Canada guidance asks for a change-control plan on machine-learning medical devices

Health Canada's pre-market guidance, reported effective April 1, 2026, tells manufacturers of machine learning-enabled medical devices to declare the ML, classify the device Class II to IV, and file a plan for future model updates. Manufacturers and the hospitals that buy from them share the exposure.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

Health Canada has set out how it expects machine learning to be handled when a medical device comes up for authorization. Its "Pre-market guidance for machine learning-enabled medical devices," reported to take effect April 1, 2026, asks manufacturers to be explicit about three things: that the device uses machine learning, where it sits in the Class II to IV risk range, and how the underlying model will be allowed to change once it is on the market. The document is guidance issued under the Medical Devices Regulations, not a new statute, but it tells applicants what a satisfactory submission now looks like.

The declaration and the class

The first duty is candor. A manufacturer has to declare that machine learning is part of how the device achieves its medical purpose, in whole or in part. That declaration then drives classification. Health Canada frames these as Class II to IV devices, the classes that carry a licensing obligation, so a machine learning component is not something a manufacturer can quietly fold into a lower-risk filing. Getting the class right at the outset determines the evidence burden for everything that follows, and misjudging it can send a submission back to the start. The higher the class, the more Health Canada will expect a manufacturer to demonstrate about how the model was built and validated before it reaches a patient.

The Predetermined Change Control Plan

The centerpiece is the Predetermined Change Control Plan. Machine learning models are not static. They get retrained on new data, recalibrated, and tuned, and under a conventional device regime each material change can trigger a fresh review. The PCCP is Health Canada's answer to that friction. A manufacturer describes, in advance, the changes it expects to make to the model, the methods it will use to make them, and how it will verify that safety and performance hold. If a later change falls inside the authorized plan, the manufacturer can implement it without returning for a new licence.

That is the reach of the tool, and it is genuinely useful. It is also the limit. A change that sits outside the plan is still a change that needs Health Canada's attention. The PCCP does not hand manufacturers a blank cheque to alter a model however they like; it rewards those who can predict and bound their modifications and offers nothing to those who cannot. Writing a plan that is both specific enough to be credible and wide enough to be useful is the real work.

Duties that do not end at approval

The guidance does not treat authorization as the finish line. It carries continuing expectations on data quality, on managing bias in the model, and on monitoring performance once the device is in clinical use. For a machine learning device, those obligations matter as much as the initial submission, because a model that performed well on its development data can drift or underperform on populations it was not built around. Bias management in particular puts the burden on the manufacturer to show it has looked for and addressed disparate performance, not merely reported aggregate accuracy. Data quality sits underneath all of it: the guidance signals that a manufacturer should be able to account for how its training and validation data were sourced, curated, and kept representative. Performance monitoring then closes the loop, giving the manufacturer a basis to notice when a model in the field has stopped behaving the way the submission promised.

What it means across the border

For a United States reader, the framework should look familiar. It runs closely parallel to the US Food and Drug Administration's own Predetermined Change Control Plan approach for AI and machine learning-enabled devices. A medtech firm operating in both markets is not facing two unrelated regimes so much as two versions of the same idea, and the practical task is to map one onto the other: build the change plan once, then reconcile the Canadian and US expectations rather than authoring each from scratch. Hospitals that buy these devices, on either side of the border, now have a concrete document to ask for. If a vendor cannot produce a coherent change-control plan and a performance-monitoring story, that absence is itself a procurement signal.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.