What does a predetermined change control plan mean for retraining a deep learning model in a medical device?

Why

A deep learning device would otherwise face a dilemma: retraining on new data may improve it, but a significant change to a cleared device normally needs a new marketing submission. FDA's final guidance on PCCPs for AI-enabled device software functions, first issued in December 2024 and reissued in August 2025, sets out an alternative. The plan has three parts.

A description of modifications lists exactly which changes are planned, for example retraining with data from additional sites or adjusting an operating threshold. A modification protocol describes how each change will be developed, validated and implemented, covering data management practices, retraining practices, performance evaluation and update procedures. An impact assessment weighs the benefits and risks of the changes.

For an engineer, the practical meaning is that retraining becomes a controlled, documented procedure with fixed test datasets, prespecified metrics and thresholds, and version control for data and models, rather than an ad hoc improvement. A plan can even cover changes that software applies automatically, but only those described and validated in it. Changes included in a plan must keep the device within its intended use, and a change outside the authorized plan still needs a new submission.

Other regulators handle this differently, so a device marketed in the EU follows its own rules for significant changes.