Does AI retraining require regulatory approval?
Retraining a medical AI model is not automatically a regulated change — but if it shifts intended use, the input/output specification or performance beyond pre-specified bounds, regulators treat it as a modification. The decisive question is whether it stays inside an authorised PCCP.
Run this assessment free →When retraining becomes a significant change
A change is significant when it affects intended use, the performance envelope or safety. Retraining that stays within pre-specified, validated bounds may be a documented change; retraining that moves performance on a key subgroup, or expands intended use, typically triggers a fuller assessment or submission. The act of retraining itself is not the trigger — its effect is.
How a PCCP changes everything
A Predetermined Change Control Plan pre-authorises a bounded set of modifications, so changes inside the envelope don't require re-submission. The FDA's final guidance fixes three components: a description of modifications, a modification protocol, and an impact assessment. A well-built PCCP turns recurring retraining from a regulatory event into routine operation.
Post-market obligations after retraining
Even an in-bounds retraining re-baselines your monitoring: drift thresholds, performance tracking and the risk file all update against the new model. The most common mistake is quantifying change against a moving test set instead of a frozen baseline, which makes a real improvement impossible to separate from sampling noise.