FDA published two hypothetical predetermined change control plans on October 20. Both cover machine learning devices. The first is a sepsis alert system that reads EHR data and flags patients at risk within 24 hours. The second is arrhythmia detection software running on ECG telemetry.

The sepsis example allows three modifications: reweighting structured nursing documentation inputs, retraining the final layers on site data, and moving the alert threshold. Each site adaptation is manual and needs manufacturer review before deployment. The arrhythmia example permits periodic full retraining and operating point changes, pushed automatically across sites behind validation checks.

Both run the three PCCP components in order: description of modifications, modification protocol, and impact assessment. FDA cites section 515C of the FD&C Act, 21 CFR 807.81(a)(3) for 510(k) devices and 21 CFR 814.39(a) for PMA supplements. De Novo requests are not addressed.

The arrhythmia case is the one to read twice. Automated retraining with no human in the loop is where review questions concentrate. FDA says both are illustrative and are not complete PCCPs.