FedIoMT-RPM v3.0: Lightweight adaptive split federated learning for patient-independent ECG arrhythmia classification

Abstract

FedIoMT-RPM v3.0: Lightweight adaptive split federated learning for patient-independent ECG arrhythmia classification This release contains the complete code and final results for the revised manuscript submitted to Scientific Reports. It replaces v2.0. Main changes from v2.0 Patient-independent evaluation. All methods now use the standard inter-patient DS1/DS2 protocol of the MIT-BIH Arrhythmia Database with the AAMI classes N, S, V and F. Each of the five federated clients holds whole patient records, and the validation records 118, 124, 215 and 220 are patient-disjoint. Bounded Adaptive Weighted Aggregation. Every client keeps at least half of its sample-size weight, and all clients take part in every round. Retrained methods. FedIoMT-RPM and seven baselines were retrained with three seeds each: Centralized CNN, FedAvg, FedProx, SCAFFOLD, LW-FedAvg and SplitFed. New analyses: paired statistical tests and a patient-level bootstrap; an ablation study; client and patient fairness; a check of the aggregation bias; memory and byte-level communication for all methods; reconstruction and membership-inference attacks; zero-shot validation on INCART and PTB-XL; patient-wise five-fold cross-validation. Main results on DS2, 21 unseen patients, mean ± SD over three seeds Accuracy: 94.90 ± 0.27% MCC: 0.751 ± 0.010 Macro-F1: 58.66 ± 2.61% Client-side model: 7,264 parameters, 28.4 KiB Communication: 32% less than SplitFed