Health Canada Machine Learning Medical Device (MLMD) Guidance 2026: PCCP, Bias Management, Data Quality, and Lifecycle Compliance for AI/ML Devices

What This Article Covers

On April 1, 2026, Health Canada published the Pre-market Guidance for Machine Learning-Enabled Medical Devices (MLMD Guidance), establishing specific regulatory expectations for AI/ML-powered medical devices across their entire lifecycle. The guidance introduces the Predetermined Change Control Plan (PCCP) as a mechanism to manage planned model modifications after market authorization, sets detailed requirements for data quality, training, validation, bias management, performance monitoring, and transparency, and creates a framework for lifecycle risk management of adaptive algorithms.

This guidance, combined with Health Canada's mandatory digital submission requirements (effective April 1, 2026), expanded terms and conditions enforcement powers (effective January 1, 2026), and revised medical device license (MDL) application guidance (effective February 2, 2026), represents the most significant update to Canada's medical device regulatory framework for AI-enabled products.

This article covers the MLMD guidance in detail, explains how it interacts with other recent Health Canada regulatory changes, compares Canada's approach to the FDA's AI/ML framework and the EU AI Act, and provides practical guidance for manufacturers preparing submissions under the new requirements.

What This Article Does NOT Cover

This article focuses on Health Canada's MLMD guidance and related 2026 regulatory changes. It does not cover general Health Canada medical device registration (see the Health Canada Device License Guide), SaMD classification basics (see the SaMD Regulatory Guide), or FDA PCCP requirements (see the FDA PCCP Guide). For broader Health Canada reform context, see the Health Canada Regulation Reform Guide.

The 2026 Regulatory Landscape: Four Interlocking Changes

1. MLMD Pre-market Guidance (April 2026)

The centerpiece is the MLMD guidance, which establishes Health Canada's expectations for machine learning-enabled medical devices. It applies to Class II, III, and IV devices where the machine learning system is part of the device's intended function. The guidance does not cover non-ML software components of a medical device — those continue to be governed by existing SaMD guidance.

2. Mandatory Digital Submissions (April 2026)

Effective April 1, 2026, manufacturers must submit all Class II, III, and IV device applications through the Common Electronic Submission Gateway (CESG) using the Regulatory Enrolment Process (REP). Email submissions are no longer accepted. This applies to new MDL applications, amendments, and all other regulatory transactions.

3. Expanded Terms and Conditions Powers (January 2026)

Regulatory amendments effective January 1, 2026, gave the Minister of Health authority to impose or amend terms and conditions on medical device licenses at any point during the device lifecycle. For MLMDs, this means Health Canada can require post-market studies, long-term follow-up, real-world evidence generation, data collection on underrepresented populations, or specific performance monitoring as conditions of maintaining market authorization.

4. Revised MDL Application Guidance (February 2026)

The updated Guidance on Managing Applications for Medical Device Licenses, effective February 2, 2026, introduces structured, lifecycle-driven application management. Key changes include higher expectations for complete dossiers, strictly enforced review timelines, increased need for clinical and technical evidence, and mandatory digital submission formats.

MLMD Guidance: Core Requirements

Scope and Classification

The MLMD guidance applies to manufacturers submitting new or amendment applications for Class II, III, and IV medical devices where the device incorporates machine learning. The guidance covers the ML system specifically — it does not address the non-ML software components, which are governed by existing Health Canada SaMD guidance and medical device software requirements.

A device is considered "ML-enabled" when it uses algorithms that can learn from data to improve performance without explicit programming of the improvement. This includes supervised learning, unsupervised learning, reinforcement learning, and deep learning approaches. Traditional statistical models with fixed coefficients are not considered ML.

Good Machine Learning Practice (GMLP)

The guidance establishes GMLP requirements organized into lifecycle phases:

Design Phase

• Define the clinical problem and intended clinical workflow integration
• Specify the target patient population and clinical use context
• Document design inputs including clinical requirements, performance targets, and safety constraints
• Establish the intended model behavior, including expected outputs and clinical interpretation

Risk Management

• Perform risk analysis specific to the ML system, including risks from data quality issues, model bias, distribution shift, adversarial inputs, and edge cases
• Apply ISO 14971 risk management principles adapted for adaptive algorithms
• Document risk control measures for each identified hazard
• Establish risk acceptability criteria specific to ML performance

Data Selection and Management

This is the most detailed section of the guidance. Health Canada requires manufacturers to document:

• Data provenance: Origin, collection methods, consent, and licensing for all training, validation, and test data
• Data representativeness: Demographic, geographic, and clinical characteristics of the patient population represented in the data, and identification of potential gaps
• Data quality: Completeness, accuracy, consistency, and timeliness of the data
• Data preprocessing: Methods for cleaning, normalization, augmentation, and splitting data into training, validation, and test sets
• Data versioning: Procedures for tracking data changes throughout development and post-market updates
• Data governance: Access controls, privacy protections, and data security measures

Bias Management

Health Canada explicitly requires bias assessment and mitigation. Manufacturers must:

1. Identify potential sources of bias in training data (demographic, geographic, clinical, device-related)
2. Quantify bias through subgroup performance analysis across clinically relevant demographic categories
3. Implement mitigation strategies during data selection, model training, and post-market monitoring
4. Document residual bias and its potential clinical impact
5. Establish ongoing bias monitoring as part of the post-market surveillance plan

Health Canada specifically calls for Sex and Gender Based Analysis Plus (SGBA+), requiring manufacturers to evaluate whether model performance differs across sex, gender, age, ethnicity, and other relevant patient characteristics. Clinical evidence submitted with MLMD applications is increasingly assessed through an SGBA+ lens across the full product lifecycle.

Development and Training

• Document the ML architecture, hyperparameter selection rationale, and training methodology
• Describe the training infrastructure, including computational resources and software frameworks
• Establish reproducibility through version control of code, data, and model parameters
• Implement verification and validation protocols specific to ML model behavior
• Document the relationship between training objectives and clinical performance requirements

Testing and Evaluation

• Performance evaluation must be conducted on data that is independent of training and validation data
• Report performance metrics that are clinically meaningful for the intended use (not just technical metrics)
• Include subgroup analysis to assess performance across clinically relevant patient populations
• Evaluate robustness to distribution shift, noise, and adversarial conditions
• Conduct user interface and usability testing for ML-specific output displays and alerts

Clinical Validation

• Design clinical validation studies appropriate to the device class and risk level
• For Class III and IV devices, clinical evidence requirements may include prospective clinical studies, retrospective studies with independent datasets, or real-world evidence
• Define clinically meaningful endpoints that directly relate to the intended medical purpose
• Assess clinical validation results in the context of known biases and limitations

Transparency

Health Canada requires manufacturers to document:

• What the ML system does and does not do (intended use and limitations)
• How the ML system was developed (methodology, data, training)
• How the ML system performs (metrics, subgroup analysis, known limitations)
• How the ML system may change over time (PCCP or update mechanism)
• How users should interpret and act on ML outputs (clinical workflow integration)

This information must be included in device labeling and instructions for use.

The Predetermined Change Control Plan (PCCP)

What Is a PCCP?

The PCCP is the central innovation in Health Canada's MLMD framework. It is a documented plan that specifies:

• The types of modifications the manufacturer plans to make to the ML system after authorization
• The methodology for implementing and validating those modifications
• The impact assessment protocol for evaluating the effect of modifications on device safety and effectiveness

A PCCP allows certain post-market model changes to be implemented without being classified as significant changes requiring a new MDL amendment — as long as the changes remain within the scope of the authorized plan.

PCCP Structure

Health Canada expects a PCCP to include:

PCCP vs. Significant Change

The critical distinction is between changes within an authorized PCCP (which can proceed without a license amendment) and changes outside the PCCP (which require evaluation under existing significant change guidance). Changes to the PCCP itself also require a license amendment.

Manufacturers should consider using the pre-submission process to discuss a proposed PCCP with Health Canada before submitting the license application.

How Canada Compares to FDA and EU

FDA Approach to AI/ML Devices

The FDA has been developing its AI/ML device framework since 2019 through a series of discussion papers, action plans, and guidances. Key elements include:

• PCCP: The FDA finalized guidance on Predetermined Change Control Plans for AI/ML-enabled devices in August 2025, establishing a similar concept to Health Canada's PCCP. The FDA PCCP allows "locked" and "adaptive" algorithm changes within authorized boundaries.
• Good Machine Learning Practice: FDA has endorsed the IMDRF GMLP guiding principles (published January 2025) developed jointly with Health Canada and MHRA
• SaMD framework: FDA applies its existing device classification framework (510(k), De Novo, PMA) to AI/ML devices based on intended use and risk
• Cybersecurity: FDA has separate premarket cybersecurity guidance requiring SBOM and vulnerability management for connected devices

EU AI Act + MDR

The EU applies a dual regulatory framework to AI-enabled medical devices:

• MDR/IVDR: Device safety and performance requirements, including clinical evaluation and post-market surveillance
• AI Act: Additional requirements for AI systems classified as high-risk, which includes AI used in medical devices (Annex III, Section 6). Requirements include data governance, technical documentation, transparency, human oversight, accuracy, robustness, and cybersecurity
• PCCP equivalent: The EU does not yet have a formal PCCP mechanism, though the MDR's post-market surveillance and vigilance requirements create obligations for monitoring algorithm performance

Comparative Summary

Health Canada's MLMD guidance is notable for its specificity and its integration with the broader regulatory framework. By establishing the PCCP as a formal regulatory tool with explicit requirements, Canada provides a clear pathway for manufacturers that is more prescriptive than the EU's approach and comparable in detail to the FDA's framework.

Practical Guidance for Manufacturers

Preparing an MLMD Submission

Manufacturers preparing a Health Canada submission for an ML-enabled device should:

1. Conduct a gap analysis against the MLMD guidance requirements, identifying any areas where current documentation is insufficient
2. Develop the PCCP early — it should be designed as part of the product development process, not as a regulatory afterthought
3. Invest in data governance — Health Canada's data quality expectations are specific and rigorous. Manufacturers need systems for tracking data provenance, quality, representativeness, and versioning
4. Build bias assessment into the development process — retroactive bias analysis is significantly more difficult than proactive bias management throughout development
5. Prepare for digital submission — the CESG/REP system requires structured, electronic dossier formats. Manufacturers should familiarize themselves with the system before the submission deadline

Leveraging the Pre-Submission Process

Health Canada's pre-submission process allows manufacturers to discuss regulatory strategy, PCCP design, and clinical evidence requirements before submitting a formal application. For MLMDs, a pre-submission is particularly valuable for:

• Discussing the appropriateness of the proposed PCCP scope
• Aligning on clinical validation study design for Class III/IV devices
• Clarifying bias assessment expectations for the specific clinical application
• Understanding how terms and conditions may apply to the authorized device

Planning for Post-Market Obligations

The expanded terms and conditions powers mean that Health Canada can impose post-market requirements at any time during the device lifecycle. Manufacturers should anticipate and plan for:

• Ongoing performance monitoring of the ML system in clinical use
• Bias surveillance across demographic subgroups in the real-world patient population
• Data collection on device performance in underrepresented populations
• Periodic review and potential modification of the PCCP as the model evolves
• Real-world evidence generation to support continued safety and effectiveness

Key Takeaways

1. Health Canada's MLMD Guidance (April 2026) establishes specific, detailed regulatory expectations for AI/ML-enabled medical devices, including PCCP, data quality, bias management, transparency, and lifecycle risk management.

2. The PCCP mechanism allows authorized model modifications without a license amendment, but requires detailed planning, validation protocols, performance monitoring, and rollback mechanisms documented in advance.

3. Health Canada now has expanded powers to impose terms and conditions on device licenses at any point in the lifecycle, enabling post-market study requirements, real-world evidence generation, and population-specific monitoring.

4. Mandatory digital submissions through CESG/REP (effective April 1, 2026) replace email submissions for all Class II–IV device applications, requiring structured electronic dossier formats.

5. Canada's MLMD framework is comparable in specificity to the FDA's AI/ML guidance and more prescriptive than the EU's approach, with the PCCP as a shared innovation between the two North American regulators. Manufacturers pursuing multi-market authorization should align their PCCP design across FDA and Health Canada submissions.