FDA REQUIREMENTS FOR AI-ENABLED MEDICAL DEVICES - OVERVIEW OF THE NEW GUIDANCE

The U.S. Food and Drug Administration (FDA) recently issued a draft guidance titled "Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations" (January 7, 2025). This guidance addresses the unique considerations associated with AI-enabled devices. It provides specific recommendations on the content of FDA marketing submissions and lifecycle management considerations consistent with a total product lifecycle (TPLC) approach for AI-enabled devices.

The body of the guidance provides comprehensive recommendations for organizing the content of marketing submissions for devices that include one or more AI-enabled device software functions (AI-DSFs) and addresses the following points specifically.

Manufacturers must provide a detailed device description, highlighting its key features, intended use, and AI integration, including inputs and outputs. It should outline the environments where the device will be used and the intended users' qualifications and training. The description must also cover configuration options, calibration, maintenance procedures, and workflow, including the level of automation and the device’s decision-making role in the clinical or user context.

Manufacturers should supply a comprehensive explanation of the device’s user interface to ensure the workflow is clear and intuitive. This includes outlining how the device operates and the specific information it provides to users at various stages of interaction.

Additionally, labeling requirements should be addressed comprehensively, ensuring that clear, concise instructions are provided to users. Labels should include key details about the model’s capabilities, limitations, and appropriate usage scenarios.

User interface and Labelling information requested by the guidance should be included within the Software Description section of marketing submission.

Appendix B of the guidance further provides information on designing transparent user interfaces for AI-enabled devices, including using model cards to clarify indications for use. While not mandatory, model cards can help organize key information, as illustrated in Appendices E and F, with examples for users, healthcare providers, and regulatory submissions.

The FDA recommends incorporating the principles of the AAMI CR34971 guidance on applying ISO 14971 to AI and machine learning into their risk management documentation. A robust strategy should address hazards under normal and fault conditions, user errors, data integrity, algorithm performance, and bias mitigation across the TPLC. Particular attention should be paid to the management of risks that are related to understanding information that is necessary to use or interpret the device, including risks related to lack of information or unclear information.

The guidance underscores the importance of a comprehensive data management plan that details the sources, diversity, and quality of the data used for model training, tuning, and validation. Manufacturers should explain how data was curated to ensure representativeness across the intended user population, addressing potential gaps or biases. The plan should also describe the methodologies used for data annotation, storage, and security to protect data integrity throughout the device's lifecycle. 

Additionally, manufacturers must demonstrate how the data aligns with the device’s intended use and clinical requirements, providing examples of how data variability has been accounted for during development. Strategies to reduce AI bias, such as oversampling underrepresented populations or employing fairness-focused algorithms, should be detailed.

As part of the Software Description section of the marketing submission, manufacturers must thoroughly explain the AI model’s architecture and the methodologies employed during its development. The submission should also include a justification for the chosen machine learning techniques, such as neural networks or decision trees, and how these methods align with the device’s functional goals. Furthermore, manufacturers should describe the iterative process of model development, including steps taken to optimize performance and mitigate overfitting or underfitting issues.

Performance validation of AI-enabled devices involves confirming safe and effective performance in real-world conditions while consistently meeting specifications. This includes rigorous testing under simulated use cases, subgroup analyses across demographics to ensure equitable outcomes, and reporting metrics like sensitivity and specificity with explanations. Performance validation must also assess the device's robustness to data drift and external changes, ensuring effectiveness throughout its lifecycle. The FDA strongly encourages manufacturers to use the Q-Submission Program to obtain the FDA’s feedback on their proposed AI-enabled device development and validation approaches.

Additionally, manufacturers should ensure that users understand and interact with the device as intended. This process typically includes incorporating human factors evaluation and assessing overall usability.

Including a performance monitoring plan in the submission for an AI-enabled device is essential to address the risk of performance changes or degradation in real-world settings, which could impact patient safety. Manufacturers should proactively monitor and manage performance changes, device inputs, and usage contexts while adhering to comprehensive risk management and quality system regulations (21 CFR Part 820). This includes responsibilities such as design validation, managing design changes, addressing nonconformities, and implementing corrective and preventive actions. 

Additionally, manufacturers must report serious adverse events and malfunctions to the FDA, ensuring continued device safety and effectiveness post-market. The FDA’s guidance encourages manufacturers to include Predetermined Change Control Plans (PCCPs) in their submissions to support innovation and iterative improvements. These plans enable prospective planning for modifications to AI models and streamline regulatory processes for pre-approved changes.

To protect patient safety and data integrity, the FDA advises incorporating robust cybersecurity into device design and lifecycle management. Manufacturers should detail real-time threat monitoring, encryption protocols, user authentication, and fail-safe systems. Submissions must outline procedures for updating security measures, including patch management and user notifications, ensuring safe and reliable operation in interconnected healthcare environments.

Public submission summaries, required for most marketing authorizations, play a key role in fostering trust and understanding by providing detailed descriptions of the device and the regulatory basis for its approval. Including this information in public-facing documents enhances transparency and supports public confidence in AI-enabled medical devices.

The guidance includes several appendices to support stakeholders in implementing its recommendations. These appendices cover topics such as an overview of the recommended information and where to include it in the marketing submission,  transparency design considerations, example Model Cards, performance validation instructions, and usability evaluation recommendations. These resources provide practical tools to help manufacturers navigate the complexities of developing AI-enabled devices.

If you need help with the preparation or revison of your FDA marketing submission, please contact us at: info@quaregia.com. We have 20 years of experience with medical device software in all its facets. For an initial consultation free of charge, you can BOOK A FREE SLOT HERE.

Last updated 2024-01-30