This draft guidance, "Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations", provides a comprehensive overview of the FDA’s current thinking on strategies to address transparency and bias throughout the Total Product Life Cycle (TPLC) of AI-DSF in a medical device (referred as AI-enabled devices).
The guidance provides detailed instructions to sponsors and regulatory personnel on what to submit as part of the marketing submission and how to manage an AI-enabled device (i.e., SaMD or SiMD with machine learning, deep learning, neural networks, and other forms of AI) throughout its TPLC.
Major Takeaways:
Here are some key highlights of the FDA recommendations for manufacturers to consider during the pre- and post-market phases of AI-enabled devices:
Marketing Submissions Requirements:
- Comprehensive descriptions of the device’s purpose and function.
- Include in-detail information on Data management, Algorithm training, and User interface.
- Evidence of safety and efficacy through rigorous testing.
- Algorithm Change Protocols (ACP) for adaptive AI systems to manage post-market updates.
- FDA may request “proactive performance monitoring” for the AI-enabled device in a De Novo (as a special control) and in the case of PMA.
- Provide examples of “510(k) summary model card” for the AI-enabled device, including clinical study results, algorithm training description, etc.
Total Product Life Cycle (TPLC) Framework:
- Design and Development: Address biases, ensure transparency, and mitigate risks during design.
- Implementation: Document AI algorithms, intended use, and limitations.
- Post-Market Management: Continuously monitor and update devices to maintain safety and effectiveness.
Risk Management:
- Since AI-enabled devices may present unique or different risks compared to traditional device software functions, the FDA advises sponsors to integrate the considerations detailed in the FDA-recognized voluntary consensus standard, "AAMI CR34971 Guidance on the Application of ISO 14971 to Artificial Intelligence and Machine Learning,"
Cybersecurity:
- Guidance elaborates on the examples of AI risks that can be impacted by cybersecurity threats, such as data poisoning, model stealing, model evasion, data leakage, performance drift, etc., and how to control them.
Overall, the developers of AI-enabled devices should align with these guidance recommendations to ensure a smooth FDA submission and product life cycle process. We understand that incorporating FDA regulations and guidance into the development process and generating the necessary documentation for marketing authorization can be overwhelming. This is where the RQMIS Quality, Clinical, Cyber, and Regulatory team can assist you in navigating each step, starting from developing a Software Quality System, helping in software verification and validation, cybersecurity testing, and acquiring FDA authorization (510k, De Novo, or PMA).