TL;DR

• FDA PCCP guidance allows certain AI algorithm changes to be authorized in advance
• PCCPs apply only to modifications that would otherwise require a new submission
• A PCCP must include a description of modifications, a modification protocol, and an impact assessment
• Changes outside the authorized PCCP still require FDA review
• PCCPs balance post-market AI evolution with regulatory oversight

In response to the rapid evolution of AI-enabled medical devices, the FDA issued Predetermined Change Control Plan (PCCP) guidance document in August 2025 describing how manufacturers may plan for algorithm changes over time while maintaining patient safety.

This article outlines the core elements of the PCCP framework, the types of modifications it applies to, and how the FDA expects manufacturers to use it in practice.

Why PCCPs Matter

AI-enabled algorithms rarely remain static after deployment. As New data becomes available and clinical environments change, manufacturers often seek to update models to improve performance. Historically, many of these updates required new marketing submissions, slowing deployment and increasing regulatory burden. The PCCP framework gives manufacturers a way to describe certain planned, higher-risk modifications in advance. Once the FDA reviews and authorizes a PCCP as part of a device’s marketing submission, the manufacturer may implement those planned modifications without submitting a new 510(k), De Novo request, or PMA supplement, provided the changes remain within the authorized plan.

The PCCP approach did not emerge in isolation. It builds on earlier FDA policy efforts, beginning with the agency’s 2019 discussion paper on AI and machine learning-based software as a medical device and later formalized in statute through section 515C of the FD&C Act. The 2025 guidance clarifies how this concept is applied in practice by describing the structure, content, and FDA review expectations for PCCPs.

Scope of the FDA PCCP Guidance

The guidance applies to any AI-enabled device software function that a manufacturer intends to modify over time. This includes:

• AI models updated manually through human-supervised processes
• AI models updated automatically, also known as continuous learning
• Devices that involve both types of updates

The recommendations focus on modifications that would otherwise require a new marketing submission because they could significantly affect safety or effectiveness. Modifications that would not require a new submission by regulation are not the target of PCCPs.

The Three Required Components of a PCCP

The FDA describes a PCCP as a structured document consisting of three distinct but related components.

Description of Modifications

This section explains exactly what changes the manufacturer plans to make to the AI-enabled device software function. The description must be detailed, specific, and limited in scope so the FDA can evaluate whether the changes can be safely implemented later.

Examples of modifications that may be appropriate include:

• Improvements to performance metrics such as sensitivity, specificity, or false-positive rate
• Expansion of compatible data inputs, such as additional imaging systems
• Modifications addressing specific subpopulations within the original intended use

Manufacturers must also explain whether modifications will be implemented globally across all deployed devices or locally at specific sites, and whether updates will occur automatically or manually. Modifications described at a high level, without clear boundaries or implementation detail, are unlikely to be authorized as part of a PCCP.

Modification Protocol

This is the heart of the PCCP. The protocol outlines how the planned modifications will be developed, validated, and implemented in a way that allows the FDA to assess ongoing safety and effectiveness without reviewing each individual change.

The protocol must include predefined acceptance criteria and address four methodological components:

• Data management practices, including how new training, tuning, and test data will be collected, curated, annotated, sequestered, and evaluated for bias
• Re-training practices, describing what parts of the AI model can be updated, how re-training will be triggered, and how architecture or parameter changes will be justified
• Performance evaluation, outlining study designs, metrics, statistical tests, and acceptance thresholds used to determine whether a modification is safe to deploy
• Update procedures, explaining how updates will be deployed, how labeling will be updated, what user communication will occur, and how real-world performance and safety will be monitored after implementation

These elements ensure that each pre-specified modification follows a rigorous, traceable process even when additional FDA review is not required.

Impact Assessment

The PCCP must include an analysis of the benefits and risks of implementing each planned modification and of the combined effect of all modifications. This includes:

• Comparing the modified model to the original, unmodified version
• Assessing risks such as unintended bias or performance degradation
• Describing how the Modification Protocol mitigates those risks
• Explaining interactions between multiple planned modifications

The FDA uses this section to determine whether the PCCP, taken as a whole, supports safe and predictable device performance. Weak or incomplete impact assessments make it difficult for the FDA to conclude that future modifications can be implemented safely without additional review.

How the FDA Expects PCCPs to Be Used

Once authorized, a PCCP functions as a set of binding conditions that govern when and how post-market algorithm changes may be implemented.

• The manufacturer may implement modifications only if they are specified in the Description of Modifications and carried out according to the Modification Protocol
• Device labeling must be updated to inform users of the presence of a PCCP and any changes implemented under it
• Deviating from the authorized PCCP generally requires a new marketing submission

If a modification is not described in the PCCP or if the manufacturer cannot follow the protocol, the change is presumed to require a new submission.

Real-World PCCP Example: Patient Monitoring Software

The FDA guidance describes an AI-enabled patient monitoring system used in high-acuity settings, such as intensive care units, that analyzes physiological signals to detect early signs of instability and trigger alarms.

The same example also demonstrates the limits of a PCCP. In this scenario, the authorized PCCP allows the manufacturer to retrain the AI model using additional real-world data to reduce the false alarm rate while maintaining sensitivity within predefined non-inferiority margins. The PCCP clearly specifies the performance targets, retraining methodology, and validation criteria that must be met before deployment. Because the modification is pre-specified and implemented according to the authorized Modification Protocol, the manufacturer may deploy the updated model without submitting a new marketing application, while still documenting the change within its quality system and updating labeling as required.

During retraining, the manufacturer identifies that the updated model can predict physiologic instability earlier than the original version, enabling a new clinical claim. Because this new predictive capability was not included in the Description of Modifications or supported by the authorized validation methods, it falls outside the scope of the PCCP. Such a change could significantly affect device performance and clinical use and would require a new marketing submission before implementation.

Related Resources