Close
Traditional medical-device development often assumes that the version evaluated during clinical development will remain relatively stable through authorization and commercial use. AI-enabled medical devices can introduce a different challenge. Algorithms may be updated, datasets may evolve, intended-use environments may change, and performance may vary as devices are deployed across new sites, patient populations, users, or clinical workflows. For sponsors, that means AI-enabled medical device clinical development should be planned as a lifecycle evidence strategy—not simply a one-time validation exercise.
FDA’s final 2025 guidance on Predetermined Change Control Plans (PCCPs) for AI-enabled device software functions provides a framework for certain planned modifications to be considered within a marketing submission. FDA states that this approach is intended to support iterative improvement while maintaining reasonable assurance of device safety and effectiveness.
At the same time, FDA’s broader guidance on lifecycle management for AI-enabled device software functions remains draft and not for implementation as of 2026. That draft emphasizes total product lifecycle risk management, documentation, transparency, performance monitoring, and bias considerations.
An AI-enabled device may perform well in its development dataset and still encounter challenges after deployment.
Performance may be affected by:
This means clinical validation should ask more than:
“Does the model perform?”
It should also ask:
“Under what conditions does the model perform reliably, and how will that performance remain controlled over time?”
FDA’s final PCCP guidance applies to AI-enabled devices reviewed through the 510(k), De Novo, and PMA pathways.
A PCCP allows a manufacturer to prospectively describe certain planned modifications and the methods that will be used to develop, validate, and implement those changes.
FDA identifies three interconnected components:
This identifies the specific planned modifications and describes the expected characteristics and performance of those modifications.
This describes the methods used to develop, verify, validate, and implement the planned changes, including predefined acceptance criteria.
This evaluates the benefits and risks associated with the planned modifications and explains how the proposed verification and validation activities continue to support safety and effectiveness.
That structure has major implications for clinical development.
A sponsor cannot build a credible PCCP at the end of development if the evidence-generation process was never designed to support future change.
For AI-enabled devices, sponsors should consider both:
Current Version Performance & Future Controlled Modifications
The clinical evidence strategy should establish a baseline against which future performance can be evaluated.
That means defining:
If those foundations are unclear, evaluating the impact of future changes becomes much more difficult.
An AI model should not drive the intended use.The intended use should drive the evidence strategy.
Sponsors should first determine:
Only then should the sponsor determine whether development and validation datasets adequately represent the intended-use population and environment.
FDA’s PCCP guidance specifically notes the importance of considering intended-use populations and environments as modifications are planned.
High performance in retrospective development data does not automatically establish reliable real-world clinical performance.
Clinical evidence planning may need to evaluate performance across:
The goal is not simply to maximize a performance metric. The goal is to establish whether the device performs reliably under conditions that reflect its intended use.
For AI-enabled IVDs, where applicable, evidence planning may also need to align with relevant IVD standards such as ISO 20916 for clinical performance studies. For other medical devices, applicable clinical-investigation activities may align with ISO 14155.
FDA’s PCCP approach is explicitly risk-based.
The Impact Assessment should evaluate benefits and risks associated with planned changes, including risks such as unintended bias. FDA also references ISO 14971 terminology within the PCCP guidance.
Sponsors should therefore connect AI clinical development with the device risk-management process.
Potential risks may include:
Risk controls should then connect to verification, validation, clinical evidence, labeling, training, software controls, or post-market monitoring as appropriate.
An AI device cannot be managed predictably if sponsors do not define what acceptable performance looks like. Relevant metrics will depend on the intended use and device type, but may include:
The clinical and statistical rationale for these metrics should be established early. The same principle applies to future modifications under a PCCP.
FDA expects the Modification Protocol to describe verification and validation activities and predefined acceptance criteria supporting planned modifications.
Performance monitoring should not begin as an improvised activity after launch.
FDA’s draft lifecycle guidance specifically discusses post-market performance monitoring as part of lifecycle risk management for AI-enabled devices. The guidance remains draft, so its recommendations are not yet for implementation, but it provides useful insight into FDA’s current thinking.
Sponsors should consider in advance:
This is where clinical development, post-market evidence, quality systems, and software lifecycle management intersect.
AI systems can perform differently across populations if development or validation data do not adequately represent intended users. FDA’s draft lifecycle guidance highlights bias and transparency considerations for AI-enabled devices.
Sponsors should therefore consider:
These considerations should not be treated as a post-market discovery problem. They should inform dataset strategy and clinical validation.
AI-enabled devices remain medical devices.
Changes to the algorithm or device software should therefore be controlled within the manufacturer’s quality management system. Depending on the device and development model, relevant frameworks may include:
A PCCP does not eliminate change control. It creates a prospectively reviewed framework for certain defined modifications.
Clinical evidence for an AI-enabled device should be traceable.
Sponsors should be able to connect:
Data Source → Patient/Specimen → Device Version → Model Version → Output → Reference Standard → Endpoint → Analysis
If that traceability breaks, it becomes harder to understand whether a performance change comes from:
Strong traceability therefore supports both clinical evidence quality and future change assessment.
A PCCP is included in a marketing submission, but the decisions behind it should begin much earlier.
If the sponsor intends for the product to evolve, clinical development should ask:
FDA’s final guidance makes clear that the Description of Modifications, Modification Protocol, and Impact Assessment are interrelated elements of the PCCP.
That makes PCCP planning a development strategy issue, not simply a regulatory-writing task.
Dynamill Research supports AI-enabled diagnostic and medical-device sponsors with development planning designed around clinical evidence, regulatory readiness, and lifecycle performance.
Support may include: