FDA’s Evolving Approach to AI/ML-Enabled Medical Devices
Opportunities and Implications
The Food and Drug Administration (FDA) has undergone a significant transformation in its regulatory approach to address the unique challenges posed by artificial intelligence and machine learning (AI/ML) in medical devices. This evolution reflects the agency's recognition that traditional regulatory frameworks are insufficient for software-based devices capable of autonomous, iterative updates. The FDA's journey toward a more adaptive regulatory model exemplifies a broader governmental shift toward operational efficiency, particularly highlighted by recent institutional developments.
Digital Transformation through INFORMED
In 2015, during my tenure at the FDA, I had the privilege of launching the Information Exchange and Data Transformation (INFORMED) initiative within the Oncology Center of Excellence. The initiative marked a pivotal step toward modernizing regulatory processes, operating under special authorities from the Department of Health and Human Services. INFORMED was designed to push the boundaries of data science, AI, and real-world evidence, bringing these powerful tools into regulatory decision-making and reshaping the broader ecosystem to be more adaptive and data-driven.
An important opportunity within INFORMED that remains only partially realized was the move toward fully digitized safety reporting and analysis, expected to save the FDA hundreds of staff hours per month at full implementation. This digital safety review framework was successfully piloted to allow safety data submissions as machine-readable datasets, making it possible to analyze trends and detect safety signals in near real-time. Such a system would reduce the manual workload for FDA reviewers and significantly improve the precision of safety assessments. As government agencies look for ways to modernize, INFORMED’s work in digital transformation can serve as a model for how technology can drive efficiencies, ensuring regulatory processes that are both effective and resource-conscious.
Evolution of Software Regulation: From PRE-CERT to PCCPs
The Pre-Certification Pilot Program
In 2017, the FDA introduced the Software Precertification (PRE-CERT) Pilot Program, representing an initial attempt to create a flexible regulatory model for Software as a Medical Device (SaMD). This program shifted focus from product-specific reviews to evaluating companies' quality management practices, aiming to establish a lifecycle-based approach that could reduce review times and resource demands.
By 2022, the PRE-CERT Pilot had underscored a critical insight: frameworks developed for traditional, hardware-based devices did not fully align with the rapid, iterative needs of digital health technologies. The program highlighted the need for an adaptable model capable of supporting ongoing innovation in digital health. However, participation in the pilot was limited, and implementation remained constrained under existing statutory authorities. These limitations underscored the importance of exploring alternative pathways, leading to the FDA’s shift toward Predetermined Change Control Plans (PCCPs), which allow manufacturers to modify software within established parameters without requiring new marketing submissions for each change.
Transition to PCCPs
In 2019, the FDA released a discussion paper introducing PCCPs, as a new regulatory framework suited for the iterative nature of AI/ML-based devices. PCCPs allow manufacturers to make pre-approved, predefined modifications to software without submitting new applications, thereby supporting ongoing improvements while upholding safety standards. This shift was formalized in the FDA’s AI/ML-Based Software as a Medical Device Action Plan in 2021, which outlined a flexible, multi-pronged regulatory approach that encourages responsible innovation while contributing to a leaner, more adaptable regulatory process.
Building on this framework, the FDA issued draft guidance in 2023 detailing how PCCPs could be included in marketing submissions for AI/ML-enabled device software functions (ML-DSFs). The guidance emphasizes three key elements:
1. Description of Modifications: Provides an outline of intended updates.
2. Modification Protocol: Establishes structured methods for validating and implementing updates.
3. Impact Assessment: Examines the risks and benefits associated with modifications, including any mitigation strategies.
By implementing these structured pathways, the FDA aims to streamline regulatory reviews, with the hope of reducing delays and enhancing the agency’s overall efficiency. For innovators, PCCPs offer a way to potentially bring devices to market more quickly, with greater clarity on regulatory requirements.
Institutional Support for Regulatory Evolution
The recent announcement of the Department of Government Efficiency (DOGE), to be led by Elon Musk and Vivek Ramaswamy, signals strong institutional support for streamlined regulatory processes. DOGE's mandate to reduce redundancies and eliminate inefficiencies aligns with and can potentially accelerate the FDA's modernization efforts and its approach to AI use and regulation. This development suggests increased support for regulatory science initiatives such as INFORMED and adaptive frameworks such as PCCPs.
Technological Integration and Future Directions
Role of Large Language Models
The integration of Large Language Models (LLMs) represents a promising frontier in regulatory efficiency. These AI tools could significantly streamline FDA workflows by automating data extraction, document analysis, and pattern recognition tasks. This automation potential aligns with broader efficiency objectives while maintaining regulatory rigor.
Persistent Challenges and Opportunities
Despite significant progress in regulatory frameworks, the field continues to face complex technical and operational hurdles. The validation of continuous learning AI systems demands standardized methodologies that can reliably assess ongoing performance and safety, particularly through rigorous clinical trials that demonstrate real-world clinical benefit. This clinical validation represents a significant challenge for technology companies, many of which are unfamiliar with the extensive requirements and timelines of clinical research.
The disconnect between traditional software development cycles and the longer-term investment needed for clinical validation often creates tension with investor expectations. Unlike pure software ventures, companies developing AI/ML medical devices must operate more like biotechnology firms, requiring substantial pre-revenue research and development followed by appropriately-designed clinical trials— a model that parallels therapeutic and companion diagnostic development. Real-world data integration presents another significant challenge, requiring robust governance frameworks to ensure data quality and appropriate usage. Furthermore, the industry must develop and implement comprehensive standards for data accuracy and interoperability across different platforms and systems. The generation of convincing clinical evidence requires innovative trial designs that can effectively evaluate both the initial and ongoing performance of AI/ML systems in clinical settings.
Addressing these interconnected challenges requires sustained collaboration between the FDA, industry stakeholders, and data science experts to develop effective solutions while maintaining operational efficiency. This collective expertise is essential for creating practical approaches that balance innovation with safety and appropriate regulatory oversight, while ensuring that clinical validation remains at the forefront of technological advancements. Companies that successfully navigate these challenges and generate robust clinical evidence will secure a distinctive market advantage, as such evidence not only satisfies regulatory requirements but can also address the critical demands of payers for reimbursement decisions, providing the confidence needed for widespread adoption by healthcare providers and health systems.