FedLite-Med: Personalized Federated ECG Classification Under MCU Memory Budgets

Time-series models that generalize across patients cannot train where they are needed most: on the wearable device itself. Uploading raw physiological data raises privacy concerns, and per-patient differences in waveform morphology limit the accuracy of a single global model. Federated learning (FL) keeps data on-device, but on-device learning on microcontroller-class hardware faces two coupled challenges: per-patient data is strongly non-IID, and full backpropagation through a convolutional network can exceed typical MCU SRAM budgets. On MIT-BIH ECG with per-patient client splits, we show that FedAvg's global model collapses to a macro F1 of 0.36 versus 0.74 for purely local training. We then present FedLite-Med, which pretrains a compact 1D-CNN backbone on public ECG data, freezes most of it, and federates sparse updates of only the final convolutional block, while each patient keeps a private classifier head on-device. FedLite-Med thereby matches the strongest personalized FL baselines within seed noise, at 146 KB peak training memory, inside the 192 KB SRAM budget that full-backpropagation baselines exceed (204-252 KB). Ablations show that prototype-based transfer is ineffective or harmful under per-patient domain shift and that even a bias-only update of a few hundred parameters recovers most of the achievable accuracy, characterizing feasible operating points for privacy-preserving time-series analytics on wearable hardware.