The landscape of medical device regulation is undergoing a fundamental transformation as the U.S. Food and Drug Administration (FDA) prepares to address the unique challenges posed by generative artificial intelligence (GenAI). Speaking at a high-level panel during The MedTech Conference, hosted by AdvaMed in Boston, Rick Abramson, a Harvard-trained radiologist and the Director of the FDA’s Digital Health Center of Excellence, outlined a visionary approach to oversight. Abramson suggested that the agency may move away from traditional software validation methods in favor of a model that mirrors the competency evaluations used for human physicians. This shift acknowledges that GenAI, characterized by its "unbounded" nature, behaves more like a clinical peer than a static piece of medical equipment.
The implications of this shift are profound for the medical technology industry. For over two decades, the FDA has governed Software as a Medical Device (SaMD) through a deterministic paradigm: a developer provides specific inputs (X), and the software produces a predictable, repeatable output (Y). However, the rise of large language models (LLMs) and agentic AI—tools capable of generating limitless text and autonomous reasoning—has rendered this traditional "black box" testing insufficient. Abramson’s remarks signal that the agency is moving toward a more dynamic, judgment-based framework to ensure patient safety in an era of probabilistic healthcare technology.
The Evolution of Regulatory Thinking: From Static Code to Clinical Judgment
To understand the magnitude of the proposed changes, it is necessary to examine the history of the FDA’s oversight of digital health. For years, the agency has relied on the "Pre-market Notification" (510(k)) or "De Novo" pathways to clear software. These pathways were designed for tools with fixed algorithms. In the radiology world, for instance, an AI tool might be programmed to identify a pneumothorax (a collapsed lung) on a chest X-ray. The evaluation is binary: the software either detects the condition or it does not. The evidentiary proof is straightforward because the scope of the task is narrowly defined.
Generative AI, however, introduces what Abramson calls an "unbounded evaluation space." Unlike traditional AI, which might classify an image, GenAI can synthesize patient histories, suggest differential diagnoses, and interact with clinicians in natural language. The inputs are as limitless as the potential outputs. Consequently, the FDA lacks the existing methodology to provide exhaustive evidentiary proof for every possible scenario a GenAI tool might encounter.
Abramson’s analogy to physician training offers a solution to this regulatory bottleneck. Just as a human radiologist is not tested on every single X-ray they will see in a 30-year career, the FDA may begin to evaluate GenAI based on its "knowledge, clinical judgment, and communication ability." This approach would involve assessing whether the AI can adhere to safety boundaries and demonstrate a consistent level of competency, rather than attempting to predict every individual response the model might generate.
Chronology of Recent FDA Actions on Artificial Intelligence
The FDA’s current trajectory is the result of several years of accelerating internal policy development. The timeline below highlights the key milestones leading to the current debate over GenAI regulation:
- April 2019: The FDA published its first discussion paper on a "Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device." This introduced the concept of a "Predetermined Change Control Plan" (PCCP).
- January 2021: The agency released the "Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device Action Plan," emphasizing the need for harmonized standards and transparency.
- October 2023: President Biden issued the "Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence," which specifically tasked the Department of Health and Human Services (HHS) with creating a framework for AI in healthcare.
- August 2024: The FDA released a detailed discussion paper titled "Multi-Purpose General Purpose AI in Medical Devices," seeking industry feedback on how to handle the "unbounded" nature of GenAI.
- October 2024: The public comment period for the August discussion paper closed, garnering over 300 detailed responses from medical societies, technology developers, and patient advocacy groups.
During the panel in Boston, Abramson emphasized that the August paper was not merely a "thought exercise" or an "academic paper." He confirmed that the agency is moving directly toward drafting formal policy, driven by high-level mandates from the White House and the Secretary of HHS.
Supporting Data: The Rising Tide of AI in Healthcare
The push for new regulation is fueled by the rapid integration of AI into clinical settings. As of mid-2024, the FDA has cleared or authorized approximately 950 AI- and machine-learning-enabled medical devices. The vast majority of these (over 75%) are in the field of radiology, followed by cardiovascular and neurological applications.
However, the majority of these cleared devices utilize "locked" algorithms—those that do not change their logic after being deployed. The industry is now pivoting toward "adaptive" or "generative" models that can learn from new data or generate novel content. This transition represents a significant leap in complexity. Market analysts suggest that the healthcare AI market is expected to grow from approximately $20 billion in 2023 to over $180 billion by 2030, with a significant portion of that growth attributed to GenAI-driven diagnostics and administrative automation.
The FDA’s challenge is to foster this innovation without compromising the "gold standard" of safety and efficacy. The 300+ comments received by the agency reflect a diverse range of concerns, from the technical difficulties of monitoring "hallucinations" (where AI generates false information) to the legal questions of liability when an AI makes a clinical recommendation.
Industry and Stakeholder Responses: The Question of Accountability
The shift toward a physician-like evaluation model has met with both intrigue and caution from the medical community. One of the most significant responses came from the American Medical Association (AMA). In a letter addressed to Abramson, AMA CEO John Whyte highlighted the practical difficulties of the FDA’s proposed path.
While Whyte acknowledged the necessity of modifying traditional testing for GenAI, he warned against the diffusion of accountability. The AMA’s position is that the probabilistic nature of GenAI—meaning it provides the most likely answer rather than a definitive one—should not excuse developers from rigorous oversight. Whyte argued that neither the complexity of the models nor the impossibility of testing every scenario should shift the burden of responsibility away from the developers and onto the clinicians or patients.
The AMA has recommended that the FDA develop a framework that includes:
- Transparency: Clear documentation of the data used to train the models.
- Post-market Surveillance: Continuous monitoring of the AI’s performance in real-world clinical settings to detect "drift" or declining accuracy.
- Algorithmic Bias Mitigation: Ensuring that GenAI tools perform equitably across different patient demographics.
These recommendations align with Abramson’s suggestion that AI should be evaluated on its "ability to adhere to certain safety boundaries." However, the industry remains concerned about how these "competency tests" will be standardized.
Future Outlook: Draft Policy on the Horizon
Stakeholders in the MedTech sector are eager for a timeline on when these new rules will take effect. Geeta Pamidimukkala, AdvaMed’s Vice President of Technology and Regulatory Affairs, pressed Abramson on the next steps during the panel discussion.
Abramson was careful not to provide a specific date but offered a general window, stating that the agency intends to be "very thoughtful" in synthesizing the 300+ comments. He jokingly noted that the FDA would not be using an AI engine to summarize the feedback, drawing laughter from the audience, but underscored the seriousness of the task. He indicated that draft policy guidance should be expected within "months," rather than weeks or years.
This timeline suggests that the FDA is working at an accelerated pace to keep up with the speed of technological development. Once the draft guidance is issued, there will be another period for public comment, allowing stakeholders to weigh in on the specific mechanics of the "physician competency" model.
Broader Impact and Implications for Global Healthcare
The FDA’s approach to GenAI regulation will likely set a global precedent. As the first major regulatory body to propose treating AI as a "clinical peer," the FDA is navigating uncharted territory that could influence the European Union’s AI Act and other international frameworks.
The move toward competency-based evaluation also raises questions about the future of medical education and the division of labor in hospitals. If an AI is evaluated like a physician, will it eventually be granted similar levels of autonomy? For now, the FDA maintains that AI should serve as a "decision support" tool, keeping a human clinician in the loop. However, as GenAI becomes more sophisticated, the line between "tool" and "practitioner" may continue to blur.
Ultimately, the goal of the Digital Health Center of Excellence is to create a regulatory environment where the benefits of GenAI—such as reduced clinician burnout, faster diagnostics, and personalized treatment plans—can be realized safely. By moving toward a model of "clinical judgment" for software, the FDA is attempting to build a bridge between the rigid requirements of engineering and the fluid, complex reality of modern medicine. The coming months will be critical as the agency translates these philosophical concepts into the concrete policies that will define the next generation of healthcare technology.
