Two examples with different release controls
FDA’s Center for Devices and Radiological Health listed a resource on hypothetical predetermined change control plans (PCCPs) in its October 20, 2026 updates. The cases illustrate existing policy; they are neither product authorizations nor a new guidance document. FDA also cautions that they do not contain everything expected in a submitted PCCP.
The hospital-localized sepsis example allows specified input weighting, local fine-tuning and alert-threshold adjustments. The model is locked before validation, with manufacturer review and approval before site deployment. An unchanged global model remains available.
The cardiac example covers automated, periodic retraining of inpatient ECG arrhythmia software using newly annotated data. Thresholds are reassessed, and validation covers individual arrhythmia classes and predefined subgroups. A failed acceptance criterion blocks deployment, alerts the manufacturer and leaves the device unchanged. Successful updates follow a staggered, QMS-controlled rollout, with version information, an audit trail and ongoing performance monitoring.
The distinction matters: automated release is contemplated within a specified process. The example does not describe unrestricted learning during patient monitoring. Its clinical setting also should not be generalized automatically to consumer wearables or other ECG applications.
The authorized plan defines the scope
FDA’s PCCP overview explains the regulatory benefit: an authorized plan can permit specified future changes without a new marketing submission for each change, provided they are included in and implemented according to that plan.
Eligible modifications generally would otherwise require a new submission, remain within the device’s intended use and, for 510(k) devices, preserve substantial equivalence to the predicate. PCCPs can accompany 510(k), De Novo or PMA submissions.
Three components connect the development proposal to its evidence: the Description of Modifications defines what may change; the Modification Protocol explains evaluation and implementation; and the Impact Assessment addresses benefits, risks, dependencies and cumulative effects.
For a product team, the useful question is therefore specific: does the proposed update match both the authorized change description and the authorized method? A better headline performance result alone cannot establish that. A team proposing a different patient population, input specification or clinical function needs a separate scope assessment rather than assuming that an existing PCCP covers it.
Retraining evidence needs its own design
FDA’s August 18, 2025 final AI PCCP guidance supplies the broader submission recommendations. Its data-management discussion addresses representative datasets, reference standards, separation of training, tuning and testing, and safeguards when test data are reused. These considerations remain important when an update pipeline is automated.
For ECG teams, this creates a practical design question: what evidence will show that each future release still supports the intended use? An overall accuracy score may conceal a trade-off between sensitivity and specificity or a weaker result in a clinically important subgroup. The guidance specifically discusses evaluating such trade-offs and high-risk subpopulations.
Threshold adjustment deserves attention alongside retraining. A model produces scores; the operating point determines how those scores become classifications or alerts. Reviewing the model weights while treating the threshold as an incidental setting would leave part of the product decision unexplained. The engineering evidence should connect the two, with data-use rules defined before successive update cycles begin.
These are considerations for a device-specific plan, not numerical acceptance limits announced by the October examples. The appropriate metrics, sample sizes and release criteria still need justification for the product and its intended population.
Quality oversight continues after deployment
The joint FDA, Health Canada and MHRA guiding principles place PCCPs within lifecycle risk management. They emphasize bounded changes, evidence, transparency and mechanisms to stop or reverse changes that fail specified performance criteria. Monitoring and responding to safety concerns remain part of that lifecycle approach.
For manufacturers, a useful reading of the October cases is to examine the entire update path, including failure handling. Who receives a rejected-update alert? Can the installed configuration be identified reliably? What happens when post-deployment performance crosses a defined limit? These questions turn an automation proposal into an assessable operating process.
An automated decision at release does not eliminate the manufacturer’s responsibility for that process. The evidence package needs to explain how acceptable updates reach users and how unacceptable or deteriorating performance is detected and addressed. That is the practical contribution of the examples for teams planning adaptive AI devices.
Sources
- Primary official sourceNovel Medical Devices with Predetermined Change Control Plans: Hypothetical Case Examples (opens in a new tab)U.S. Food and Drug Administration
- Supporting sourceCDRH New - News and Updates (opens in a new tab)U.S. Food and Drug Administration
- Supporting sourcePredetermined Change Control Plans (opens in a new tab)U.S. Food and Drug Administration
- Supporting sourceMarketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions (opens in a new tab)U.S. Food and Drug Administration
- Supporting sourcePredetermined Change Control Plans for Machine Learning-Enabled Medical Devices: Guiding Principles (opens in a new tab)U.S. Food and Drug Administration, Health Canada, and Medicines and Healthcare products Regulatory Agency