Rockville, MD — On January 7, 2025 — The Food and Drug Administration (FDA) published a landmark draft guidance titled “Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations“. This document provides a comprehensive roadmap for manufacturers to navigate the regulatory requirements for AI-enabled medical devices across their entire lifespan, marking a significant step in the agency’s oversight of rapidly evolving digital health technologies.
This guidance is a critical update for developers of AI-enabled device software functions (AI-DSFs), emphasizing safety, transparency, and the proactive mitigation of bias throughout the Total Product Life Cycle (TPLC).
THE TOTAL PRODUCT LIFE CYCLE (TPLC) APPROACH
The FDA is shifting from traditional static oversight to a Total Product Life Cycle (TPLC) model. This approach requires manufacturers to consider the safety and effectiveness of their AI from initial design through development, validation, and post-market performance monitoring.
- Continuous Oversight: Manufacturers are expected to identify possible hazards in both normal and fault conditions throughout the device’s life.
- Quality System Integration: The guidance explains how documentation related to the Quality System (QS) Regulation such as design controls, validation procedures, and corrective actions can be leveraged pre-market to demonstrate how risks are addressed.
ELIMINATING THE "BLACK BOX": TRANSPARENCY AND LABELING
To foster trust and ensure safe use, the FDA demands that AI functionality be clear and accessible to clinicians, patients, and caregivers.
- The “Model Card”: While not a strict requirement, the FDA strongly encourages using a “Model Card” (Appendix E) to consistently summarize key aspects like model architecture, training data, and known limitations.
- User-Centered Design: Transparency involves ensuring information is contextually relevant and functionally comprehensible.
- Detailed Labeling: Labeling must include an explanation of how AI achieves the intended use, the meaning of model outputs, and the intended degree of automation.
COMBATTING AI BIAS AND DATA DRIFT
The guidance places a heavy emphasis on ensuring that AI devices work equitably across all demographic groups to prevent systematic incorrect results.
- Representativeness: Manufacturers must ensure validation data sufficiently represents the intended use population, including subgroups defined by race, ethnicity, sex, and age.
- Managing Data Drift: AI devices can be sensitive to “data drift,” where performance degrades because real-world input data differs from the data used during development.
- Subgroup Analysis: Performance must be evaluated and reported across important subgroups to identify if a device performs worse in specific populations.
KEY MARKETING SUBMISSION REQUIREMENTS
Manufacturers must now include specific AI-related documentation in their marketing submissions, such as 510(k), PMA, or De Novo requests:
Submission Section | Recommended Documentation |
Device Description | A statement that AI is used, descriptions of inputs/outputs, and intended workflow. |
User Interface | Graphical representations or videos showing how the user interacts with the AI output. |
Data Management | Detailed info on data collection, cleaning, annotation, and how test data was sequestered. |
Model Description | Detailed technical characteristics of the architecture, features, and training paradigms. |
Validation | Objective evidence of performance on independent datasets, including “Human-AI Team” performance. |
Cybersecurity | Strategies to prevent “data poisoning,” model theft, and model evasion attacks. |
FLEXIBILITY THROUGH PCCPS
One of the most innovative aspects of the guidance is the reinforcement of Predetermined Change Control Plans (PCCPs). A PCCP allows manufacturers to pre-specify intended modifications such as retraining an algorithm to improve performance and implement them without needing a new marketing submission, provided the changes are consistent with the authorized plan.
WHAT SHOULD MANUFACTURERS DO NOW?
- Prioritize Transparency: Integrate transparency considerations starting at the design phase to avoid costly late-stage changes.
- Implement Performance Monitoring: Develop robust plans to proactively capture device performance after deployment to manage real-world risks.
- Engage Early: The FDA highly encourages sponsors to use the Q-Submission Program to discuss novel AI technologies, the use of real-world data, or emerging validation methods.
 - Submit Comments: This guidance is currently in draft form. The FDA is inviting comments and suggestions from the industry and public within 90 days of its publication.