Health Canada's New Pre-Market Guidance for Machine Learning-Enabled Medical Devices: PCCP Explained

Executive Summary

On April 1, 2026, Health Canada published its final Pre-market Guidance for Machine Learning-Enabled Medical Devices (MLMD). The guidance applies to Class II, III and IV devices that use machine learning, in whole or in part, to achieve a medical purpose. Its most important feature is the Predetermined Change Control Plan (PCCP), a mechanism that lets manufacturers obtain pre-authorization for planned model changes so they can update their algorithms after licensing without filing a licence amendment each time.

The MLMD guidance did not arrive alone. It is one of five new and revised guidance documents Health Canada released between March and April 2026, covering terms and conditions, significant change interpretation, clinical evidence requirements and post-market summary reports. Together, they move Canada from point-in-time device licensing toward true lifecycle oversight.

If you manufacture, import or plan to license an AI or machine learning medical device in Canada, this article walks you through what the guidance expects, how to build a PCCP that Health Canada will accept, and where most submissions go wrong.

What Is a Machine Learning-Enabled Medical Device (MLMD)?

A machine learning-enabled medical device is a medical device that uses machine learning, in part or in whole, to achieve its intended medical purpose. Machine learning is the subset of artificial intelligence where algorithms learn patterns from data to build models, rather than following rules that were explicitly programmed.

MLMDs fall under the Food and Drugs Act and the Medical Devices Regulations. An MLMD can be standalone software, often called software as a medical device (SaMD), or software embedded in a physical device. It can be an in vitro diagnostic device or a non-IVDD, and its risk classification can range from Class I to Class IV.

Health Canada has adopted the terminology of the International Medical Device Regulators Forum (IMDRF), so manufacturers should align their submissions with the IMDRF key terms and definitions for MLMD. Three terms matter most:

• ML training algorithm: the software procedure that establishes the parameters of a model by analyzing data
• ML model: the mathematical construct that generates a prediction or inference from new input data
• ML system: the ML-enabled software that meets the definition of a medical device, including the models and associated training algorithms

Why April 2026 Matters: A Coordinated Guidance Package

The MLMD guidance is best read as one piece of a larger reform. Health Canada first consulted on a draft version in 2023. The final guidance, dated April 1, 2026, landed within weeks of four companion documents, and regulatory amendments that took effect on January 1, 2026 now allow the Minister of Health to impose or amend terms and conditions on a medical device licence at any point in the device lifecycle.

Here is the package at a glance.

The practical effect is simple to state and demanding to implement. A medical device licence in Canada is no longer a one-time gate. It is the start of an ongoing relationship with the regulator, and software-driven devices sit squarely in the spotlight.

What Is a Predetermined Change Control Plan (PCCP)?

A PCCP is documentation, submitted as part of the device design in a new licence application or an amendment, that describes the changes a manufacturer plans to make to an MLMD after authorization, the limits within which the device will operate, and how each change will be implemented, verified and assessed.

Why does it matter? Machine learning models degrade. Patient populations shift, clinical practice evolves, input data drifts, and a model that performed well at licensing can quietly lose accuracy in the field. Before the PCCP existed, every significant change to address that drift required a licence amendment application, which meant the regulatory process often lagged behind the science.

With an authorized PCCP, changes that stay within the approved plan do not require a licence amendment for a significant change. The manufacturer implements the change, documents it within the quality management system, and carries on. Changes outside the plan, including changes to the PCCP itself, must still be assessed against the significant change guidance and may require an amendment before implementation.

One caution we give every client: a PCCP is not a blank cheque. Changes listed in a PCCP must keep the device within its authorized intended use. Any change to the medical conditions, purposes or uses of the device always requires a licence amendment first.

The Three Components of a PCCP

Health Canada expects a PCCP to be a standalone section of the submission, typically within the device description or software section, with three components.

A few details from the guidance deserve emphasis because they routinely separate acceptable plans from rejected ones.

Traceability is expected. Each change in the change description should trace clearly to the relevant procedures in the change protocol, for example through a traceability table.

Triggers must be defined. Health Canada wants to know what initiates a change: a performance threshold, a scheduled retraining interval, accumulated user feedback. "We will retrain when needed" is not a trigger.

Roll-back is part of the plan. The change protocol should describe corrective actions, including roll-back plans, backup and recovery procedures and customer communications if an update underperforms.

Flag it in the cover letter. Manufacturers should state in the cover letter that the device uses machine learning, and separately that the application includes a PCCP. Leaving those statements out can delay screening.

Health Canada also points manufacturers to the joint Predetermined Change Control Plans for MLMD: Guiding Principles, developed with the US FDA and the UK MHRA, which call for PCCPs that are risk-based, evidence-supported and transparent across the total product lifecycle.

What Belongs in a PCCP and What Does Not

The guidance gives a useful test. Appropriate PCCP changes are those where pre-authorization addresses a known risk while upholding the benefits to the patient. The canonical example is maintaining or improving performance to counter model degradation caused by changes in the environment, the input data or the relationship between input variables and the target variable.

Lifecycle Evidence: What Health Canada Expects in an MLMD Application

The guidance frames its expectations around nine lifecycle components: good machine learning practice, design, risk management, data selection and management, development and training, testing and evaluation, clinical validation, transparency, and post-market monitoring. Four areas deserve particular attention.

Data Quality and Canadian Representativeness

Health Canada expects the data used to develop and evaluate an MLMD to be justified as adequately representative of the Canadian population and clinical practice. Submissions should describe training, tuning and test datasets, including sample sizes, demographics, prevalence comparisons against the intended population, collection methods and devices, and inclusion and exclusion criteria with justification for any removed data.

Manufacturers should also explain how data integrity was maintained during curation, and describe any augmentation or imputation practices such as geometric transformations or synthetic data, with justification.

Bias Management and SGBA Plus

Bias is treated as a risk to be analyzed and controlled, not a checkbox. The risk analysis should address systematic differences in how the device treats certain groups, and the guidance expects Sex and Gender-Based Analysis Plus (SGBA Plus) to be applied across the lifecycle. In practice, that means considering sex, gender, racial and ethnic minorities, pediatric and elderly populations and pregnant people, and collecting disaggregated subgroup data in training data, test data and clinical studies where appropriate.

Testing should produce evidence of performance for relevant subgroups, including at intersections of identity factors and clinical features. The risk analysis should also cover erroneous outputs, overfitting, underfitting, performance degradation, automation bias, alarm fatigue and the risks associated with operating a PCCP. ISO 14971 remains the reference standard for medical device risk management.

Clinical Validation

For Class III and IV MLMDs, clinical evidence including clinical validation studies must be provided in the application. For Class II, it must be available on request. The study population must be independent of the data used to develop, train and tune the model, a point Health Canada underlines and reviewers check. Real-world evidence and post-market clinical experience are explicitly recognized as acceptable evidence types, consistent with the new clinical evidence guidance.

Transparency and Labelling

Transparency in this guidance means giving patients, users, health care providers and regulators clear information about how the device works, what its outputs mean and where its limits are. Labelling should state that the device includes machine learning, explain how to interpret outputs including aids such as confidence scores or saliency maps, and describe how to perform calibrations, local validation and ongoing performance monitoring. Remember that websites, brochures and marketing material with ML claims are all considered labelling under sections 21 to 23 of the Regulations. The joint transparency guiding principles are the reference point.

How Canada's PCCP Compares with the FDA Approach

Manufacturers selling in both markets will recognize the architecture. The FDA finalized its guidance on predetermined change control plans for AI-enabled device software functions in December 2024, and both regulators worked from the same jointly published guiding principles. Both frameworks use a three-part structure: a description of the planned modifications, a protocol governing how they are made, and an impact assessment.

The differences sit in the details. Canada ties the PCCP directly to the significant change framework under the Medical Devices Regulations, treats the PCCP as part of the device design, and reinforces it with the new power to impose terms and conditions on a licence at any time. A PCCP written for an FDA submission is a strong starting point for Canada, but it should be adapted, not photocopied. Canadian population representativeness and SGBA Plus expectations in particular have no exact FDA equivalent.

Preparing an MLMD Application with a PCCP: A Step-by-Step Approach

1. Confirm classification. Apply the classification rules in Schedule 1 of the Regulations and the SaMD classification guidance, and include your justification in the application.
2. Decide whether you need a PCCP. If you expect to retrain or update the model after licensing, you almost certainly do.
3. Define the performance envelope. Set baseline performance, thresholds and limits the device will stay within over time.
4. Write the change description, protocol and impact assessment. Make every change traceable to a procedure and every procedure traceable to a risk.
5. Assemble lifecycle evidence. Data descriptions, training methods, testing results with subgroup analysis, clinical validation and labelling.
6. Consider a pre-submission meeting. Health Canada invites manufacturers to use the pre-submission process to discuss a proposed PCCP before filing. For a novel device or an ambitious plan, take that invitation.
7. Flag ML and the PCCP in your cover letter. Then file, respond to questions and build your post-market monitoring before launch, not after.

MLMD Compliance Checklist

• Device classification justified against Schedule 1 rules
• Cover letter states the device uses ML and includes a PCCP
• ML methods, training algorithms and architecture described
• Training, tuning and test datasets characterized and justified as representative of the Canadian population
• Bias controls documented, SGBA Plus applied, subgroup performance evidence included
• Risk analysis covers erroneous outputs, drift, automation bias, alarm fatigue and PCCP risks per ISO 14971
• Test data independent of training data; reference standard justified
• Clinical validation evidence in the application (Class III and IV) or on file (Class II)
• PCCP includes change description, change protocol and impact assessment with traceability
• Roll-back and corrective action procedures defined
• Labelling discloses ML, explains outputs and supports ongoing performance monitoring
• Post-market monitoring plan addresses degradation, data drift and input changes
• QMS procedures ready to document PCCP-driven changes

Common Mistakes We See in MLMD Submissions

Treating the PCCP as a wish list. Plans that try to pre-authorize vague or open-ended changes get questioned. Specific changes, defined triggers, bounded performance.

Testing on training data. Reviewers look for confirmation that test and clinical study populations are independent of development data. Contaminated splits undermine the entire evidence package.

US data, no Canadian justification. Datasets built entirely on foreign populations need a documented justification of representativeness for Canadian clinical practice, or supplementary evidence.

Ignoring subgroup performance. Overall accuracy can hide poor performance in subgroups defined by skin pigmentation, sex, age or comorbidities. Health Canada asks for the disaggregated picture.

Forgetting that marketing is labelling. A website claim about your algorithm's accuracy is a labelling claim subject to review.

No QMS pathway for PCCP changes. Changes under a PCCP must be documented in your quality management system. If your QMS has no procedure for executing the change protocol, the plan is not operational.

Frequently Asked Questions

Does the MLMD guidance apply to my software if it only uses a fixed, locked algorithm?

If your device uses machine learning in any part of achieving its medical purpose, the guidance applies, whether the model is locked or adaptive. A locked model simply means your PCCP, if you include one, may be narrower.

Is a PCCP mandatory for an AI medical device licence in Canada?

No. A PCCP is optional. Without one, however, every post-market change that qualifies as a significant change requires a licence amendment application before implementation, which is slow for devices that need regular retraining.

Can I add a PCCP to a device that is already licensed?

Yes. A PCCP may be submitted with a licence amendment application for an existing device, not only with new applications.

Do changes made under an authorized PCCP need to be reported to Health Canada?

They do not require a licence amendment application for a significant change, but they must be documented in your quality management system, and they remain subject to other amendment requirements, such as a change in device identifier, and to post-market oversight.

What happens if I make a change outside my PCCP?

The change must be assessed under the significant change guidance. If it is significant, you need an authorized amendment before implementation. Changes to the PCCP itself also fall outside the plan.

How does the January 2026 terms and conditions power affect MLMDs?

The Minister can now impose or amend terms and conditions on any Class II to IV licence at any time, for example requiring ongoing performance reporting or evidence generation in under-represented populations. Health Canada has signalled that MLMDs and PCCPs are a natural focus for this tool.

Does Health Canada accept real-world evidence for MLMD clinical validation?

Yes. Real-world evidence and post-market clinical experience are listed among acceptable forms of clinical evidence, provided the evidence is relevant, reliable and supported by a justification of its sufficiency.

How MFLRC Can Help

MFLRC has supported regulated companies through Canadian licensing for more than 20 years, and our medical device regulatory practice covers exactly the ground this guidance demands.

• Medical Device Licensing: classification strategy, MDL applications and amendments, PCCP development and pre-submission support
• Quality Assurance Services: QMS structure and SOPs so PCCP changes are documented and defensible
• ISO 13485 and MDSAP Readiness: gap analysis and inspection preparation for device quality systems
• Validation Services: computerized system and software validation planning that stands up to review
• Audit Services: internal audits, mock inspections and CAPA review before Health Canada comes knocking

Whether you are preparing a first Canadian application for an AI-enabled device or retrofitting a PCCP onto a licensed product, we build the documentation with you, not for the shelf.

Need help with your MLMD licensing strategy or PCCP? Book a consultation with MFLRC for practical guidance tailored to your device.

Conclusion

Health Canada's April 2026 MLMD guidance is good news for serious manufacturers. It replaces uncertainty with a clear, internationally aligned pathway, and the PCCP finally gives machine learning devices a regulatory mechanism that matches how the technology actually behaves. The price of that flexibility is discipline: representative data, honest bias analysis, traceable change control and a quality system that can execute the plan. Companies that invest in those foundations now will move faster than competitors who treat the guidance as paperwork.

Sources and References

• Pre-market guidance for machine learning-enabled medical devices, Health Canada, April 1, 2026
• Predetermined change control plans for MLMD: Guiding principles, Health Canada, FDA and MHRA
• Good machine learning practice for medical device development: Guiding principles
• Transparency for machine learning-enabled medical devices: Guiding principles
• Guidance on clinical evidence requirements for medical devices, Health Canada
• Guidance for the interpretation of significant change of a medical device, Health Canada
• Guidance on terms and conditions for Class II to IV medical devices, Health Canada
• Software as a medical device (SaMD): Definition and classification, Health Canada
• Medical Devices Regulations, SOR/98-282
• Food and Drugs Act, RSC 1985, c F-27
• IMDRF: Machine learning-enabled medical devices, key terms and definitions
• ISO 14971, Medical devices, Application of risk management to medical devices