FDA AI-Enabled Device Software Functions: Understanding the January 2025 TPLC Draft Guidance for Lifecycle Management, Bias Assessment, and Marketing Submissions

Why This Guidance Matters

On January 7, 2025, the FDA published its most comprehensive document yet on artificial intelligence in medical devices: the draft guidance "Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations." This is the first FDA guidance to address the full total product life cycle (TPLC) of AI-enabled devices — from initial design through post-market monitoring — with specific, actionable recommendations for what to include in marketing submissions.

The guidance had been anticipated for over a year. It appeared on FDA's "A list" of priority guidances for fiscal year 2024, was transferred to the FY2025 list when that deadline was missed, and was finally published in January 2025. When finalized, it will fundamentally change how manufacturers develop, document, submit, and monitor AI-enabled medical devices.

This guide explains the guidance's key recommendations, how they differ from existing FDA frameworks, and what manufacturers should do now to prepare — even while the guidance remains in draft form.

How This Guidance Relates to Existing FDA AI Frameworks

The FDA has been building its AI regulatory framework incrementally since 2019. Understanding where this guidance fits within the broader landscape is essential.

The TPLC draft guidance is the umbrella document that ties together all prior work. It does not replace the PCCP guidance — instead, it provides the broader context within which PCCPs operate. Where the PCCP guidance addresses how to manage future algorithm changes, the TPLC guidance addresses how to design, develop, validate, document, and monitor the AI system throughout its entire life.

Key Definitions

The guidance introduces important terminology that manufacturers must understand:

AI-Enabled Device: A device that includes one or more AI-enabled device software functions (AI-DSFs). The AI component is part of the device's mechanism of action.

AI-DSF (AI-Enabled Device Software Function): A device software function that implements artificial intelligence. This includes machine learning models, deep learning systems, and other AI techniques embedded in the device.

Model: The algorithm and the data used to train it. FDA considers the model part of the device's mechanism of action, not just its implementation.

Data Management: The practices for collecting, curating, labeling, and managing data used in model training, validation, and testing.

The Total Product Life Cycle Approach

The FDA's TPLC approach requires manufacturers to manage AI device risks from initial concept through decommission. The guidance organizes its recommendations across these lifecycle stages:

1. Design and Development

The guidance recommends that manufacturers integrate AI-specific considerations into design controls from the earliest stages, including:

• Intended use clarity: Explicitly state what the AI component does, how it contributes to the device's overall function, and its role in the clinical workflow
• Model architecture selection: Document why a particular AI approach was chosen and its limitations
• Data strategy: Define the data requirements (quantity, quality, diversity, representativeness) needed to train and validate the model for the intended use population
• Risk management: Address AI-specific risks such as model overfitting, data drift, adversarial inputs, and automation bias

2. Data Management

This is the most detailed section of the guidance and reflects FDA's core concern: data quality determines AI device safety. The guidance states that "reviewers will evaluate the quality, diversity, and quantity of data used to test an AI-DSF to evaluate the safety and effectiveness of the AI-enabled device."

Key recommendations:

Data Collection and Curation

• Document the source, collection methodology, and curation process for all training, validation, and test datasets
• Demonstrate alignment between collected data and the device's intended use and target population
• Maintain sufficient segregation between training and validation datasets to prevent data leakage

Data Representativeness

• Ensure datasets represent the diversity of the intended use population across relevant demographic dimensions (race, ethnicity, sex, age, geography)
• Document and justify any gaps in demographic representation
• Identify potential confounders — for example, if all diseased cases come from one scanner type or clinical site, the model may learn scanner artifacts rather than disease features

Data Quality

• Implement data quality controls throughout the collection and annotation process
• Document data cleaning, preprocessing, and augmentation steps
• Maintain traceability from raw data to processed inputs

3. Bias Mitigation

The guidance treats bias mitigation as a safety requirement, not an aspirational goal. AI bias is defined as "a potential tendency to produce incorrect results in a systematic, but sometimes unforeseeable, way due to limitations in the training data or erroneous assumptions in the machine learning process."

FDA's bias framework requires manufacturers to:

• Identify sources of bias in data collection, labeling, model architecture, and clinical deployment
• Mitigate bias through representative data, diverse development teams, and bias testing protocols
• Evaluate bias by measuring model performance across relevant subgroups (demographics, clinical conditions, device settings)
• Monitor bias post-market through ongoing subgroup performance tracking

The guidance provides specific examples of bias pathways:

4. Transparency

Transparency is presented as both a design principle and a regulatory requirement. The guidance recommends:

For Healthcare Providers

• Clear explanation of what the AI does and does not do
• Description of the clinical evidence supporting the AI function
• Disclosure of known limitations and failure modes
• Information about how the AI output should be used in clinical decision-making
• Description of how the model was validated across patient subgroups

For Patients

• Plain-language description of the AI's role in their care
• Explanation of how the AI output may affect treatment decisions
• Information about data privacy and security practices

Model Cards The guidance encourages the use of model cards — structured summaries that document a model's intended use, performance metrics, limitations, and evaluation results. Model cards provide a standardized format for organizing transparency information.

5. Marketing Submission Content

This section provides specific recommendations for what to include in 510(k), De Novo, PMA, HDE, and BLA submissions for AI-enabled devices.

Device Description

• Declaration that AI is used in the device and its role in the intended use
• Description of the AI model class (e.g., convolutional neural network, transformer)
• Comparison of training, validation, and test datasets
• Statistical confidence levels for predictions, including uncertainty metrics

Software Documentation

• Software description including architecture, inputs, outputs, and decision logic
• Model development and training methodology
• Data management practices
• Performance validation results
• Risk analysis addressing AI-specific hazards

Performance Testing

• Validation study design and statistical methods
• Performance metrics (sensitivity, specificity, PPV, NPV, AUC, etc.)
• Subgroup performance analysis across demographics
• Failure mode analysis and error characterization
• Comparison to clinical standard of care

Post-Market Monitoring Plan

• Performance monitoring strategy
• Data collection and analysis plan
• Thresholds for triggering corrective action
• PCCP (if applicable) for managing future model updates

6. Post-Market Performance Monitoring

The guidance recommends that manufacturers describe in their marketing submissions how they will monitor AI device performance after deployment. This includes:

• Performance drift detection: Monitoring whether the model's accuracy changes over time as the patient population or clinical practice evolves
• Subgroup performance tracking: Ongoing evaluation of performance across demographic groups to detect emerging disparities
• Real-world data collection: Gathering data on how the device performs in routine clinical use, beyond the controlled study setting
• Corrective action protocols: Predefined thresholds and procedures for when monitoring reveals performance degradation

The guidance encourages manufacturers to use PCCPs to manage post-market model updates, linking the monitoring plan to the PCCP modification protocol.

Human Factors Considerations for AI Devices

The guidance includes recommendations for human factors engineering specific to AI-enabled devices, expanded in a February 2026 Emergo by UL analysis:

Automation Bias

Manufacturers must address the risk that users may over-rely on AI recommendations — accepting outputs without appropriate critical evaluation. The guidance recommends:

• Designing user interfaces that present AI outputs as decision support, not directives
• Including confidence levels and uncertainty indicators alongside AI recommendations
• Training materials that explain the AI's limitations and appropriate use

User Interface Design

• Display relevant context for AI outputs (what data was used, how the output was generated)
• Provide clear indication when the AI is operating outside its validated parameters
• Design override mechanisms that are intuitive and accessible

Validation Testing

• Human factors validation should include scenarios where the AI produces incorrect or uncertain outputs
• Test participants should include the range of intended users (specialists, general practitioners, nurses)
• Evaluate whether users correctly identify and respond to AI errors

What This Guidance Does NOT Cover

Understanding the boundaries is important:

• It does not replace existing software guidance. Manufacturers must still comply with IEC 62304, FDA's software guidance, and device-specific requirements.
• It does not establish legally binding requirements. As a draft guidance, it represents FDA's current thinking but is not for implementation until finalized.
• It does not address AI used in drug development. FDA issued separate draft guidance on AI in drug and biological product development.
• It does not provide specific threshold standards for subgroup performance or bias metrics. The guidance encourages monitoring but does not prescribe pass/fail criteria.
• It does not comprehensively address continuously learning systems. The guidance acknowledges adaptive AI but leaves ambiguity for models that update autonomously post-approval.

How This Differs From the PCCP Guidance

The two guidances are complementary. The TPLC guidance describes what your AI device should look like at every lifecycle stage; the PCCP guidance describes how to manage approved changes to it.

Organizational Capability Maturity

MD+DI's analysis of the TPLC guidance proposes a maturity framework for companies building AI device compliance capabilities:

The most impactful organizational change in 2026 is establishing a cross-functional AI Review Board comprising Regulatory, Quality, R&D, Clinical, and IT Security leaders with authority over PCCP authorization, TPLC compliance, bias assessment, and incident escalation.

Practical Steps for Manufacturers

1. Align Development Practices Now

Even though the guidance is in draft, FDA has stated that it reflects the agency's current thinking. Manufacturers who align their development practices with these recommendations will be better positioned when the guidance is finalized — and may receive fewer clarification requests on current submissions.

2. Implement Data Management Frameworks

Invest in data governance infrastructure that enables you to document:

• Data provenance (where data came from, how it was collected)
• Data processing pipeline (cleaning, augmentation, preprocessing)
• Dataset composition (demographics, clinical settings, device types)
• Dataset segregation (training, validation, test splits)
• Bias assessment results

3. Conduct Bias Audits

Before submitting, evaluate your model's performance across relevant subgroups. Document results transparently, including any disparities found and mitigation steps taken.

4. Develop Transparency Materials

Create model cards and user-facing documentation that clearly explain:

• What the AI does and its role in clinical decision-making
• The evidence supporting the AI function
• Known limitations and failure modes
• How to interpret AI outputs appropriately

5. Plan for Post-Market Monitoring

Design performance monitoring systems that can detect drift, track subgroup performance, and trigger corrective action when thresholds are exceeded. Consider integrating this with a PCCP for managing future updates.

6. Engage with FDA Early

The guidance explicitly encourages early engagement — through pre-submissions, Q-submissions, and direct interaction with the Digital Health Center of Excellence. Early feedback on your AI development strategy can prevent costly misalignment later.

What Industry Is Saying

The draft guidance has generated significant industry discussion. Key themes from public commentary:

• Transparency requirements are welcomed but need more specificity on format and content
• Bias mitigation expectations are ambitious and may be challenging for smaller companies with limited access to diverse datasets
• Post-market monitoring recommendations lack specific thresholds and enforcement mechanisms
• Continuously learning systems remain inadequately addressed — the guidance does not clearly define how autonomous model updating would be regulated
• Patient-facing transparency is underdeveloped — the guidance focuses on HCP transparency but provides less detail on how patients should be informed

Timeline and Next Steps

The guidance is not expected to be finalized before late 2026 or 2027. However, as the CenterWatch analysis noted: "To reduce the risk of costly delays, it is important to align to these FDA expectations before submission and before marketing an AI-enabled software device."

Key Takeaways

1. This is the FDA's most detailed AI device document to date. It covers the entire product lifecycle with specific, actionable recommendations.

2. Data management is the centerpiece. FDA will evaluate the quality, diversity, and quantity of your training and testing data as part of its safety and effectiveness assessment.

3. Bias is a safety issue. Not a diversity initiative — a regulatory requirement. Manufacturers must identify, mitigate, evaluate, and monitor bias throughout the device lifecycle.

4. Transparency is required for both HCPs and patients. Model cards, clear labeling, and plain-language documentation are recommended.

5. The guidance complements, not replaces, the PCCP framework. Use the TPLC guidance for overall lifecycle management and the PCCP guidance for managing approved future changes.

6. Start aligning now. Even in draft form, this guidance reflects FDA's current expectations. Companies that wait for finalization risk significant rework on their AI device submissions.