AI/ML Medical Device FDA Guidance: Predetermined Change Control Plans

AI/ML Medical Device FDA Guidance: Predetermined Change Control Plans

Key Takeaways

FDA's AI/ML Regulatory Evolution

FDA's approach to AI/ML medical devices has evolved through a series of discussion papers, action plans, and guidance documents:

Timeline of Key FDA Actions

Locked vs Adaptive AI/ML Algorithms

A fundamental distinction in AI/ML device regulation is between locked and adaptive algorithms.

Locked Algorithms

A locked algorithm is one whose function does not change after deployment. The algorithm produces the same output for a given input every time, and the manufacturer does not update the model in response to new data during clinical use. The algorithm is fixed at the time of premarket authorization.

Locked algorithms include:

• Rule-based systems with fixed decision logic
• Machine learning models trained on a fixed dataset and deployed without retraining
• Neural networks with frozen weights

Locked algorithms are regulated through traditional premarket review. Any change to the algorithm (retraining, weight updates, rule modifications) requires either a new premarket submission or coverage under an authorized PCCP.

Adaptive Algorithms

An adaptive algorithm changes its behavior over time based on new data or learning. The algorithm's output for a given input may change as it processes additional data. Adaptive algorithms include:

• Continuously learning algorithms that update model parameters based on new patient data
• Algorithms that periodically retrain on expanded datasets
• Reinforcement learning systems that modify behavior based on outcome feedback

FDA's 2019 discussion paper distinguished between:

• Algorithms that adapt based on data from the local institution: Higher risk because changes are site-specific and may not be validated across diverse populations
• Algorithms that adapt based on centralized data aggregation: Modifications are developed centrally by the manufacturer and deployed as updates

Adaptive algorithms present regulatory challenges because the device that was authorized at the time of premarket review may behave differently in clinical use. FDA's PCCP framework is designed to address this by requiring manufacturers to pre-specify the types of adaptations, validation methodology, and performance bounds.

Current Regulatory Status

FDA's current public guidance is written around premarket-reviewed software functions and manufacturer-controlled post-authorization modifications. Sponsors developing software that could change after authorization should expect FDA to focus on clearly bounded modifications, validation methods, and postmarket monitoring rather than open-ended field learning.

Predetermined Change Control Plans in Detail

FDA's August 2025 final guidance, "Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions," establishes the current framework for PCCPs.

PCCP Components

A PCCP must include three elements:

1. Description of Modifications (Modification Protocol)

The PCCP must describe:

• What will change: Specific aspects of the AI/ML algorithm that may be modified (e.g., model architecture, training data, feature engineering, decision thresholds)
• Why it will change: The rationale for the planned modification (e.g., performance improvement, expanded clinical scope, new device platform)
• How it will change: The methodology for implementing the modification (e.g., retraining protocol, data selection criteria, validation procedures)
• Bounds of the change: Quantitative or qualitative limits on the extent of modification

The modification description must be specific enough that FDA can evaluate the risk profile of the proposed changes. Vague descriptions such as "the algorithm may be updated to improve performance" are insufficient.

2. Impact Assessment

The PCCP must include an assessment of how the planned modifications will affect:

• Safety: Analysis of potential risks introduced by the modification, including new failure modes
• Effectiveness: Analysis of how the modification may affect device performance, including the possibility of performance degradation
• Benefits: Expected improvements in device performance or clinical utility

The impact assessment should reference the device's risk analysis (per ISO 14971) and demonstrate that the manufacturer has considered the full range of potential impacts.

3. Description of Modified Device

The PCCP must describe:

• Performance specifications: Quantitative performance metrics that the modified device must meet (e.g., sensitivity, specificity, AUC, positive predictive value)
• Acceptance criteria: Criteria that determine whether a modification passes validation and can be deployed
• Labeling updates: Any changes to labeling that will result from the modification

PCCP Evaluation Criteria

FDA evaluates PCCPs against the following criteria:

PCCP Examples

Example 1: Retraining with Expanded Dataset

A manufacturer of an AI-based chest X-ray CADe system might include a PCCP that allows:

• Retraining the model on additional chest X-ray datasets meeting pre-specified diversity and quality criteria
• Maintaining pre-specified clinical performance criteria on a standardized test set
• Validating performance on a held-out dataset with pre-specified demographic subgroups
• Deploying the updated model only if all acceptance criteria are met

Example 2: Decision Threshold Adjustment

A manufacturer of an AI-based diabetic retinopathy screening device might include a PCCP that allows:

• Adjusting the binary classification threshold within a pre-specified range (e.g., 0.40-0.60)
• Validating each threshold change on a standardized test set with known clinical outcomes
• Maintaining pre-specified clinical performance criteria at any selected threshold

Example 3: Platform Expansion

A manufacturer might include a PCCP that allows:

• Deploying the same algorithm on additional hardware platforms (e.g., new tablet models, new operating system versions)
• Validating that algorithm performance is equivalent across platforms
• Documenting any platform-specific adjustments to image preprocessing

What a PCCP Cannot Cover

FDA's PCCP guidance clarifies that certain changes fall outside the scope of a PCCP:

• Changes to the intended use or indications for use
• Changes to the target population (e.g., expanding from adult to pediatric use)
• Changes to the clinical domain (e.g., expanding from chest X-ray to brain MRI)
• Changes that would require a different classification or regulatory pathway
• Open-ended modifications without specific bounds

These changes require a new premarket submission (510(k), De Novo, or PMA supplement).

Total Product Lifecycle Approach

FDA's TPLC approach for AI/ML devices envisions a continuous cycle of oversight rather than a one-time premarket review.

TPLC Framework

Key TPLC Elements

1. Premarket Authorization:

• Traditional premarket review (510(k), De Novo, PMA) with PCCP included
• Clinical evidence supporting initial algorithm performance
• Software documentation per FDA guidance

2. Post-Market Monitoring:

• Real-world performance monitoring comparing actual performance to premarket testing
• Adverse event reporting under 21 CFR Part 803
• Any additional periodic reporting should follow the reporting structure FDA describes in the applicable authorization, guidance, or future program-specific communications

3. Modification Management:

• Modifications within PCCP scope: implement per PCCP methodology, document, report to FDA
• Modifications outside PCCP scope: new premarket submission required
• PCCP updates: manufacturer may submit updated PCCPs through supplemental submissions

4. Transparency:

• Labeling that describes the AI/ML nature of the device
• Disclosure of algorithm modifications to users
• Public listing of AI/ML-enabled devices on FDA's AI/ML database

Good Machine Learning Practice (GMLP)

In October 2021, FDA, Health Canada, and the UK's MHRA jointly published "Good Machine Learning Practice for Medical Device Development: Guiding Principles." These ten principles provide a framework for responsible AI/ML device development:

The Ten GMLP Principles

GMLP and Regulatory Submissions

While the GMLP principles are not legally binding requirements, FDA expects AI/ML device submissions to address these principles. Specifically:

• Principle 3 (Representative data): FDA routinely requests demographic breakdowns of training and testing data and evaluates performance across subgroups
• Principle 4 (Independent datasets): FDA expects clear documentation of dataset partitioning and absence of data leakage
• Principle 7 (Human-AI team performance): For CADe/CADx devices, FDA typically requires reader studies evaluating clinician performance with and without the AI
• Principle 10 (Post-deployment monitoring): FDA may impose post-market surveillance requirements, particularly through Special Controls in De Novo orders

Transparency and Labeling

FDA's Transparency Expectations

FDA expects AI/ML device labeling to include:

FDA's AI/ML-Enabled Device Database

FDA maintains a publicly accessible list of AI/ML-enabled devices that have received marketing authorization. Sponsors should use that official FDA list for current product counts, specialty distribution, and authorization status rather than relying on secondary summaries.

The concentration in radiology reflects both the maturity of AI/ML applications in medical imaging and the availability of large, labeled imaging datasets for algorithm training.

Real-World Performance Monitoring

Why Post-Market Monitoring Matters for AI/ML

AI/ML algorithms can experience performance degradation in real-world deployment due to:

• Dataset shift: Differences between the training data distribution and the real-world data distribution
• Covariate shift: Changes in patient demographics, disease prevalence, or clinical practice patterns
• Concept drift: Changes in the relationship between input features and the target outcome
• Acquisition shift: Changes in data acquisition equipment, protocols, or settings (particularly relevant for medical imaging)

FDA's Evolving Monitoring Framework

FDA's October 2023 draft guidance "Lifecycle Management Considerations for AI/ML-Enabled Device Software Functions" proposed a framework for ongoing performance monitoring, including:

Performance Metrics:

• Tracking key performance metrics (sensitivity, specificity, PPV, NPV) against pre-specified thresholds
• Monitoring for algorithmic bias through subgroup performance analysis
• Detecting performance degradation through statistical process control methods

Reporting Requirements:

• Adverse event reporting per 21 CFR Part 803 (mandatory)
• Periodic performance reports (proposed; format and frequency under development)
• Notification to FDA of performance degradation below pre-specified thresholds

Corrective Actions:

• Retraining or updating the algorithm within PCCP scope
• Issuing field safety corrective actions if performance falls below safety thresholds
• Recalling or withdrawing the device if safety issues cannot be resolved

Bias and Fairness

FDA has identified algorithmic bias as a significant concern for AI/ML medical devices. Sources of bias include:

FDA expects premarket submissions to address bias by:

• Describing the demographic composition of training and testing datasets
• Reporting performance metrics stratified by relevant demographic subgroups (age, sex, race/ethnicity, disease severity)
• Identifying populations where the device has not been evaluated
• Describing the manufacturer's approach to bias detection and mitigation

Practical Implications for AI/ML Device Development

Pre-Submission Strategy

AI/ML device manufacturers should strongly consider a Pre-Submission (Q-Sub) meeting with FDA to discuss:

1. Whether the device meets the SaMD definition
2. Classification and regulatory pathway
3. Clinical evidence expectations, including study design and endpoints
4. PCCP scope and content
5. Algorithm documentation requirements
6. Cybersecurity considerations
7. Post-market monitoring expectations

Documentation Best Practices

Common FDA Review Issues

Based on publicly available FDA review summaries and decision documents, common issues in AI/ML device submissions include:

1. Insufficient subgroup analysis: Performance not broken down by relevant demographics
2. Training data not representative: Training data does not reflect the intended patient population
3. Unclear algorithm description: Insufficient detail about the AI/ML methodology
4. Missing standalone performance data: Only reader study data provided without standalone algorithm performance
5. Inadequate ground truth methodology: Reference standard for labeling not clearly defined or clinically appropriate
6. Cybersecurity gaps: Missing SBOM, inadequate vulnerability management plan
7. PCCP too broad: Planned modifications described without sufficient specificity or bounded acceptance criteria

Key Regulatory References

References