Health Canada’s Machine Learning Device Guidance: What Changed in 2026
Health Canada's 2026 MLMD update strengthened lifecycle oversight rather than creating a new AI-device framework. This article explains what changed around licence conditions, PCCPs, significant changes, model monitoring and regulatory submissions - and what Canadian medical-AI founders should account for before development is locked.
Health Canada’s 2026 update did not replace its machine learning medical device framework. It made the lifecycle-regulation model more operational.
Health Canada finalised its Pre-market guidance for machine learning-enabled medical devices in February 2025. A revised version was published on 1 April 2026. Concepts including good machine learning practice, representative data, clinical validation, transparency, post-market monitoring and Predetermined Change Control Plans were therefore already part of the framework before 2026.
The more important 2026 development is what happens after licensing.
Amendments to the Medical Devices Regulations took effect on 1 January 2026, strengthening Health Canada’s ability to impose or amend terms and conditions on Class II, III and IV medical-device licences over the device lifecycle. The April 2026 MLMD guidance now operates within that wider lifecycle framework.
For founders, the practical message is:
Design the regulatory strategy around the model’s lifecycle, not only around the version submitted at launch.
First: what counts as a machine learning-enabled medical device?
Health Canada describes a machine learning-enabled medical device, or MLMD, as a medical device that uses machine learning, in whole or in part, to achieve its intended medical purpose.
That can include standalone software as well as software contained within a physical medical device.
An MLMD can fall within Class I, II, III or IV, depending on its intended purpose and the applicable classification rules. Health Canada’s current pre-market guidance is primarily directed at new and amendment applications for Class II, III and IV MLMDs.
This distinction matters because:
AI does not automatically mean medical device.
Similarly:
Machine learning does not create a separate Canadian device class.
The intended medical purpose and the Medical Devices Regulations determine whether the product falls within the medical-device framework.
So what actually changed in 2026?
The publication date needs careful interpretation.
Health Canada’s current MLMD guidance is dated 1 April 2026, but the Government of Canada catalogue confirms that final guidance had already been published in February 2025.
Therefore, it would be inaccurate to present every concept in the April 2026 document as a new 2026 requirement.
Instead, three core regulatory developments matter most.
1. Licence terms and conditions became a stronger lifecycle tool
This is the clearest substantive regulatory development surrounding the 2026 MLMD guidance.
Health Canada’s current terms-and-conditions guidance reflects expanded authority under section 36 of the Medical Devices Regulations for Class II, III and IV devices. The regulator can use terms and conditions as part of ongoing oversight rather than treating licensing as the end of regulatory scrutiny.
Those controls can support objectives such as:
- maintaining device safety and effectiveness;
- managing benefits and risks;
- identifying changes in those benefits and risks;
- addressing uncertainty throughout the lifecycle.
Why this matters for machine learning
ML systems create a challenge that conventional static software may encounter less frequently:
Performance can change even when intended use remains the same.
For example:
- data distributions can shift;
- clinical environments can change;
- input sources can change;
- model performance can degrade.
Health Canada’s MLMD guidance specifically discusses post-market performance monitoring and the need to consider degradation and changes in input distributions.
Consequently, founders should increasingly think of regulatory status as:
licence + continued evidence + monitoring + controlled change
rather than:
licence = regulatory work finished.
2. The PCCP now sits inside a more mature change-management framework
A Predetermined Change Control Plan, or PCCP, is one of the most important concepts for an ML product expected to evolve after launch.
However, PCCPs are not new in 2026.
Health Canada’s 2025 final guidance already set expectations around PCCPs, and the Government of Canada has confirmed that those provisions were part of the final February 2025 MLMD framework.
What has changed is the wider regulatory environment around them.
What a PCCP actually does
A PCCP describes specified future modifications to an MLMD and explains how those changes will be implemented, assessed and controlled.
Health Canada’s framework addresses three central components:
1. Change description
What modifications are planned.
2. Change protocol
How those modifications will be developed, verified, validated, implemented and controlled.
3. Impact assessment
How the modifications could affect the device’s benefits, risks, safety and effectiveness.
When Health Canada authorises a PCCP, a subsequent significant change implemented according to that authorised plan does not require a licence amendment solely because it constitutes a significant change.
However, manufacturers still need to document those changes within their quality system, and other licence-amendment or post-market requirements may continue to apply.
That can materially change the regulatory strategy for a machine-learning product expected to evolve after release.
But a PCCP is not permission for unlimited model updates
An authorised PCCP covers defined and bounded modifications.
It should not be interpreted as general permission for a model to change in any way the development team chooses.
For example:
Retraining a model to maintain previously established performance
may potentially fit within an appropriately authorised PCCP.
But:
Changing a monitoring application into a diagnostic application
changes a much more fundamental part of the product’s medical purpose.
A founder should therefore identify foreseeable model changes before submission rather than attempting to define the change strategy only after commercial deployment.
3. Significant-change guidance now connects ML updates more clearly with PCCPs
Health Canada published revised guidance on interpreting significant changes on 31 March 2026.
The guidance now specifically discusses MLMDs and explains that a PCCP can help manage planned significant modifications associated with known risks, including model-performance degradation caused by model drift.
This matters because development teams routinely make changes that may appear technically ordinary:
- retraining a model;
- changing datasets;
- adjusting thresholds;
- updating inputs;
- modifying architecture;
- changing how outputs enter a workflow.
The regulatory question is different:
Does the modification affect the safety or effectiveness of the licensed device in a way that makes it a significant change?
If the modification sits inside an authorised PCCP, the authorised pathway may apply.
If it sits outside that PCCP, the manufacturer needs to assess the Medical Devices Regulations and Health Canada’s significant-change guidance before implementing it.
What did not suddenly become new in 2026
Several concepts appear prominently in the current guidance but already formed part of Health Canada’s 2025 MLMD framework.
These include:
- good machine learning practice;
- representative datasets;
- Sex and Gender-Based Analysis Plus;
- intended use and clinical workflow;
- model architecture and training information;
- PCCPs;
- risk management;
- training and testing data;
- clinical validation;
- transparency;
- labelling;
- post-market performance monitoring.
Therefore, the 2026 story is not that Canada suddenly started regulating these areas.
A more accurate interpretation is:
Health Canada’s existing MLMD framework now operates within a more developed total-product-lifecycle regulatory environment.
The lifecycle approach is the real story
Health Canada’s current MLMD guidance addresses the product lifecycle through areas including:
- good machine learning practice;
- design;
- risk management;
- data selection and management;
- development and training;
- testing and evaluation;
- clinical validation;
- transparency;
- post-market monitoring.
These are not isolated documentation exercises.
The product architecture affects whether the organisation can later establish:
- which dataset generated a particular model;
- whether subgroup performance was assessed;
- what changed between model versions;
- whether drift can be detected;
- whether validation can be reproduced;
- whether an authorised PCCP can actually be implemented.
For that reason:
Regulatory design begins before regulatory submission.
Data representativeness remains a major expectation
Health Canada’s guidance says data used to develop an MLMD or demonstrate its safety and effectiveness should appropriately represent the intended population and relevant clinical practice.
The guidance highlights considerations including:
- skin pigmentation;
- biological differences between sexes;
- sex and gender;
- racial and ethnic groups;
- older adults;
- paediatric populations;
- pregnancy;
- relevant subgroup performance.
For an AI founder, that means:
“We trained on a large dataset” is not enough.
The more useful question is:
Does the dataset represent the people and environments in which the device claims to work?
Clinical workflow and degree of autonomy need to be explicit
Health Canada’s framework also asks manufacturers to describe how the ML function sits inside its intended clinical environment.
That includes considerations such as:
- intended user;
- intended workflow;
- how outputs are used;
- degree of clinical autonomy.
Two products can therefore use technically similar models while creating different regulatory risks.
For example:
AI output → clinician reviews → clinician decides
is not equivalent to:
AI output → clinical action occurs automatically.
The intended use and the workflow cannot be separated.
Explain the model at the level relevant to risk
Health Canada asks manufacturers to describe relevant machine-learning methods, model architecture and training approaches.
Where appropriate, the guidance also discusses communicating information about factors that influence outputs and how device behaviour can be interpreted.
This should not be interpreted as a blanket rule requiring every machine-learning model to be completely explainable.
Instead, transparency should support safe and effective use by relevant stakeholders, including:
- users;
- healthcare professionals;
- patients where applicable;
- regulators.
Monitoring after deployment needs to be designed before launch
Health Canada’s MLMD guidance addresses surveillance and performance monitoring after deployment.
Manufacturers should consider potential effects from:
- model-performance degradation;
- changes in input distributions;
- changes to input sources;
- changes to compatible systems handling the output.
This has an important product implication.
Monitoring should not simply be treated as a dashboard feature added after release.
It may be part of the regulatory control system around the model.
If the device also relies on a PCCP, monitoring and modification triggers become even more closely connected.
A model update should have a regulatory state
A practical development team should be able to answer:
What model version is deployed?
Which data produced it?
What changed from the previous version?
Why was the change made?
What evidence shows performance remains acceptable?
Was the change covered by the authorised PCCP?
If not, was it assessed as a potential significant change?
What monitoring follows deployment?
Health Canada does not present this as one formal checklist.
Rather, it is the practical implication of combining its expectations around model lifecycle management, PCCPs, significant changes, quality systems and post-market monitoring.
What this means for a founder building a medical AI product in Canada
1. Determine whether the intended product is actually a medical device
Start with the intended medical purpose.
Do not begin with:
“It uses AI, so which AI regulation applies?”
Begin with:
“What medical function does this software perform?”
2. Establish the likely device classification
An MLMD may fall within Class I, II, III or IV depending on its intended purpose and the Medical Devices Regulations.
For Class II, III and IV applications, Health Canada’s MLMD guidance provides the relevant pre-market pathway.
3. Decide whether the model is expected to change after approval
Ask this before submission.
Possible future changes could include:
- retraining;
- new datasets;
- updated thresholds;
- additional inputs;
- local adaptation;
- changes intended to maintain performance.
If predictable changes could constitute significant changes, a PCCP may become strategically important.
4. Build evidence collection into development
Do not reach submission and then discover that the team cannot reconstruct:
- training datasets;
- test populations;
- model versions;
- subgroup results;
- validation evidence;
- modification history.
Health Canada’s regulatory model increasingly assumes that lifecycle evidence can be produced and traced.
5. Define post-market monitoring before launch
Consider:
- model drift;
- input-distribution changes;
- differences between sites;
- subgroup performance;
- emerging failure modes;
- changes in workflow.
These issues closely reflect Health Canada’s current post-market MLMD guidance.
6. Treat regulatory status as part of release management
A conventional software team might classify releases as:
major / minor / patch
A medical-AI team also needs to understand whether a change is:
within the authorised device state / within the PCCP / potentially significant / potentially subject to an amendment
Those classifications answer a different question from semantic software versioning.
One separate 2026 operational change founders should know about
Another April 2026 change affects how manufacturers submit material to Health Canada, but it should not be confused with the substantive MLMD guidance.
From 1 April 2026, the Regulatory Enrolment Process and Common Electronic Submission Gateway became part of the required submission workflow for applicable medical-device regulatory transactions.
The REP collects information relating to the manufacturer, dossier, device and regulatory activity.
This is principally an administrative submission change.
It does not rewrite the substantive evidence expectations for machine-learning medical devices.
However, founders planning Canadian submissions need to account for it operationally.
A newer 2026 submission expectation
Health Canada issued another relevant notice on 16 July 2026 concerning Class III and IV medical-device licence applications.
The regulator said that manufacturers should submit only information needed to support the application and should provide clear, concise summaries that accurately reflect the supporting evidence. Health Canada noted that unnecessarily large or duplicative submissions can slow review.
The notice contains a particularly relevant point for AI teams:
If an applicant uses an AI tool to generate summaries or explanatory material, the applicant remains responsible for ensuring that the resulting content is accurate and reflects the supporting data.
This does not create a new MLMD classification rule.
Instead, it reinforces a broader principle that founders should already recognise:
using AI to prepare regulatory material does not transfer regulatory responsibility to the AI system.
What the 2026 framework does not mean
The current Health Canada position does not mean:
- every healthcare AI product is a medical device;
- every ML model requires a Canadian medical-device licence;
- every model update can be pre-authorised through a PCCP;
- a PCCP allows changes to intended medical purpose;
- licensing removes post-market obligations;
- every MLMD requires an identical evidence package;
- Health Canada prescribes one mandatory ML architecture;
- an authorised PCCP allows unrestricted self-modification.
Health Canada’s approach remains risk-based and dependent on the device’s intended use, risk and supporting evidence.
The most important change is architectural, not documentary
A founder could read the guidance and conclude that the answer is simply to produce more regulatory documentation.
That misses the larger implication.
A credible MLMD programme needs the underlying product and development process to support:
traceability → evidence → monitoring → controlled modification → reassessment
If those capabilities are absent from the technical architecture and product process, writing a PCCP later will not solve the underlying problem.
Similarly, if the organisation cannot identify which data generated a specific model version, investigating post-market performance becomes much harder.
The regulatory strategy and the technical architecture therefore need to develop together.
Where this leaves Canadian medical AI founders
Health Canada’s 2026 position should not be presented as an entirely new AI-device rulebook.
The foundational MLMD framework — including GMLP, PCCPs, representative data, clinical validation, transparency and post-market monitoring — was already established in the February 2025 final guidance.
What changed materially in 2026 is the surrounding lifecycle-regulation environment.
Health Canada’s expanded use of licence terms and conditions, revised significant-change guidance, April submission-process changes and later July expectations for concise and accurate Class III/IV submissions all reinforce the same direction:
Do not design a medical AI product only for the first licence submission. Design it so that evidence, model changes and performance remain controlled throughout the product lifecycle.
That is the more important 2026 change.
Building a healthcare AI product for Canada?
Before the architecture is locked, define the intended use, regulatory pathway, evidence requirements and how future model changes will be handled. KastHunt’s Discovery Workshop helps turn those questions into a practical product and delivery plan.
Frequently asked questions
What is Health Canada’s current guidance for machine learning-enabled medical devices?
Health Canada’s current Pre-market guidance for machine learning-enabled medical devices is dated 1 April 2026. It covers areas including design, GMLP, PCCPs, risk management, data, model development, testing, clinical validation, transparency and post-market monitoring for relevant MLMD licence applications.
Was Health Canada’s MLMD framework first introduced in 2026?
No. Health Canada finalised the MLMD guidance in February 2025. The current version was published in April 2026 within an updated regulatory environment.
What was the major regulatory development in 2026?
One of the most important changes was the stronger lifecycle role of terms and conditions for Class II, III and IV medical-device licences, alongside revised significant-change guidance and updated submission processes.
What is a PCCP?
A Predetermined Change Control Plan defines specified future changes to an MLMD, how they will be implemented and validated, and how their potential impact on safety and effectiveness will be assessed.
Can a manufacturer update an ML model without filing a licence amendment?
Potentially. When a significant modification is made according to an authorised PCCP, Health Canada says the manufacturer does not need to submit a licence amendment solely because that modification constitutes a significant change. Other regulatory requirements may still apply.
Can a PCCP change the intended medical use of the device?
A PCCP is intended for specified, controlled modifications within the authorised regulatory framework. It should not be treated as permission to make an unrestricted change to the product’s medical purpose.
Does Health Canada expect representative training and testing data?
Yes. Health Canada’s guidance discusses representative data and consideration of relevant population characteristics and subgroups when demonstrating device performance and safety.
Does the guidance apply to generative AI?
The MLMD framework is based on medical purpose and the use of machine learning rather than one specific model architecture. Whether a generative-AI system falls within the medical-device framework therefore depends on what the product is intended to do.
Does Health Canada expect post-market model monitoring?
Health Canada’s guidance discusses performance surveillance and risks including model degradation, changes in input distributions and changes to input sources after deployment.
Did Health Canada change the submission process in 2026?
Yes. The Regulatory Enrolment Process and Common Electronic Submission Gateway became relevant to applicable medical-device regulatory transactions from April 2026. Health Canada also issued a July 2026 notice reminding Class III and IV applicants to submit concise, relevant and accurate supporting material.
Can AI be used to prepare a medical-device licence submission?
Health Canada’s July 2026 notice does not prohibit using AI to prepare summaries or explanatory material. However, it states that the applicant remains responsible for ensuring AI-generated material is accurate and reflects the underlying evidence.
About the author
Kamna Singh Thakur is CEO and Founder of KastHunt, an AI-native healthcare technology firm serving healthcare organisations and healthtech teams across multiple international markets.
She has spent nine years in healthcare technology across business development, account management and digital transformation.
She writes about clinical AI governance, healthcare product delivery and what survives deployment in real healthcare environments.
Questions or corrections: info@kasthunt.com
Related reading
- FHIR Is a Standard. EHR Implementations Are Not
- What Healthcare Architecture and Database Diagrams Need to Show
Sources
- Health Canada — Pre-market guidance for machine learning-enabled medical devices, current April 2026 version — MLMD scope, PCCPs, lifecycle evidence and post-market monitoring.
- Government of Canada Publications — Pre-market guidance for machine learning-enabled medical devices, February 2025 — confirms that the final MLMD guidance predates the 2026 revision.
- Health Canada — Guidance on terms and conditions for Class II to IV medical devices — current lifecycle T&C framework.
- Health Canada — Guidance on interpreting significant change of a medical device, 31 March 2026 — includes MLMD and PCCP treatment.
- Health Canada — Regulatory Enrolment Process for medical devices, 1 April 2026 — REP and submission workflow.
- Health Canada — Expectations for information submitted with Class III and IV medical-device licence applications, 16 July 2026 — concise submissions and responsibility for AI-generated summaries.
