Medical devices have traditionally been validated once, before they reach the market, and that model assumes the product stays the same afterward. AI systems change through updates, retraining, and variations in the data they process, so a certificate issued at launch can describe a product that no longer exists a year later. Continuous certification, modeled on the way physicians are licensed, is one interesting angle on the added complexity AI brings to medical devices, and it also raises questions about responsibility and oversight that the industry has yet to answer.
What have regulators put in place?
Approval in both the US and Europe already takes considerable time, so a full review for every modification to an AI system would be impractical. The FDA’s Predetermined Change Control Plan (PCCP) framework, described in the FDA’s official guidance published in December 2024, addresses this. The manufacturer describes in advance which modifications it expects to make and how they will be developed, validated, and implemented. If the FDA authorizes that plan, modifications made within its specified boundaries can be implemented without seeking separate FDA clearance or approval for each one.
The EU AI Act takes a complementary route by classifying AI-enabled medical devices as high-risk if they require assessment by an independent notified body before they can be placed on the EU market, which adds requirements for risk management, logging, human oversight, and post-market monitoring on top of the Medical Device Regulation (MDR). Under the Digital Omnibus agreement, these obligations are expected to apply to medical devices from August 2028.
The PCCP framework is primarily concerned with how an authorized device may change over time. But a separate challenge arises when AI does more than perform a defined technical function and begins to take on tasks traditionally performed by medical professionals, such as interpreting diagnostic information or recommending therapy. In that case, regulators must consider not only whether the technology performs safely and as intended, but also how its role fits into clinical decision-making and where responsibility lies when its output contributes to a patient outcome.
The physician licensure model as a template
A recent Journal of the American Medical Association paper proposes continuous or periodic verification of AI systems along the lines of physician continuing education, and the licensure model translates to the lifecycle of a clinical AI product as follows:
• Continuous validation against standardized benchmarks replaces the initial examination.
• Real-time monitoring for data shift and demographic drift parallels continuing education and peer review.
• Re-examination before each new version is released corresponds to periodic re-licensure.
• Automatic withdrawal of autonomy when performance drifts serves the function of suspension.
The proposal is theoretical, and it leaves an important design question open. A system that evolves has to be recertified at some point, and nobody has established how large a change must be before it triggers recertification, whether measured in code, data, or observed behavior.
Who holds the license?
At present, physicians carry the liability for clinical decisions. If AI takes on part of that role, accountability has to be redefined, and the manufacturer appears to be the most obvious candidate. Manufacturers can be expected to resist that outcome, and a shared model between vendors, healthcare providers, and certifying bodies is a plausible alternative that no regulator has yet endorsed.
The identity of the certifier is also unresolved. Physicians answer to medical boards, and engineers in other safety-critical fields answer to professional bodies with long track records, while no equivalent institution exists for an AI system that offers a diagnosis. Regulators in Europe and the US are exploring the question while trying to avoid two outcomes at once, which are slowing progress and putting patients at risk.
Explainability as a precondition
Continuous certification only works when the system can be examined. A certifying body is unlikely to accept accountability for a model whose reasoning cannot be inspected, no matter how well it performed on a benchmark. That requirement connects directly to the architecture discussed in our companion article, “Neurosymbolic AI in Healthcare and Life Sciences: Why Generative AI Needs a Verification Layer,” where the symbolic layer supplies the explicit, checkable reasoning that a license would depend on.
Progress is likely to arrive in increments, with individual frameworks maturing while the broader questions of liability and certification remain open. Organizations building clinical AI today can prepare by designing for explainability and continuous monitoring from the start, since both will be prerequisites under any model that emerges





