FDA Clarifies Predetermined Change Control Plans for AI-Driven Clinical Decision Support Tools

FDA Updates Details on Predetermined Change Control Plans for AI/ML Software as Medical Device

The U.S. Food and Drug Administration released further information on April 3, 2024, about its framework for artificial intelligence and machine learning software as a medical device. The agency placed new emphasis on predetermined change control plans that let manufacturers manage certain modifications without fresh premarket submissions. This fits into a total product lifecycle strategy that tracks how such software changes after initial review. The material also aligns with international positions on ethical AI use in health settings.

What this means

The update spells out documentation expectations for PCCPs within marketing applications. It supplies examples of modification protocols that fall within acceptable bounds. Manufacturers gain a more defined pathway for iterative updates compared with the 2019 discussion paper and 2021 Action Plan. The material also ties FDA positions to broader international thinking on ethics, bias, and ongoing oversight.

Key takeaways

• Predetermined Change Control Plans serve as the primary regulatory focus for managing post-approval modifications in AI/ML-enabled device software functions without new premarket submissions. [1]
• Good Machine Learning Practice calls for multi-disciplinary expertise, risk management, independent review of data quality, model robustness, and transparency. [2]
• Ten guiding principles developed with international partners in 2021 address data quality assurance, model robustness, and transparency for safe AI/ML medical device deployment. [2]
• FDA guidance aligns with global efforts that highlight ethical use, bias mitigation across diverse populations, and lifecycle monitoring. [3]
• Manufacturers must maintain robust post-market surveillance and change management for adaptive AI models. [1]

Predetermined Change Control Plans

The FDA identified PCCPs as central to its current thinking on AI/ML software. These plans describe anticipated changes and the methods manufacturers will use to control them. Submissions now require specific documentation that shows how modifications will preserve safety and effectiveness.

Good Machine Learning Practice

Ten guiding principles released on October 27, 2021 outline expectations for AI/ML device development. They were created with international regulatory partners. Core elements include rigorous data quality assurance, attention to model robustness, and steps that promote transparency for users and regulators. [2]

International Context and Bias Considerations

A 2021 World Health Organization document sets ethical and governance principles for artificial intelligence in health. FDA guidance aligns with its emphasis on bias identification and mitigation, representation of diverse populations in training data, and continued performance monitoring. [3]

Post-Market Obligations

The framework calls for real-world performance monitoring once devices reach the market. Manufacturers collect and review data as models change over time. This total product lifecycle view replaces one-time premarket review for technologies that adapt.

Limitations

Guidance documents represent current thinking but are not legally binding regulations. Rapid evolution of foundational AI models may outpace even updated lifecycle frameworks. Limited long-term real-world evidence cited for certain bias-mitigation techniques in diverse clinical settings. International harmonization is ongoing, and specific requirements still differ between FDA, EMA, and other regulators.