Key result
FedIoMT-RPM classifies arrhythmias with ~94% accuracy while drastically reducing device model size vs. SplitFed.
Why the study?
Does the FedIoMT-RPM split federated learning framework maintain patient-independent ECG arrhythmia classification performance while reducing device memory and communication overhead compared with federated and split learning baselines?
Does the FedIoMT-RPM split federated learning framework maintain patient-independent ECG arrhythmia classification performance while reducing device memory and communication overhead compared with federated and split learning baselines?
FedIoMT-RPM enables collaborative training of arrhythmia classifiers on resource-constrained wearable ECG devices with a 28.4 KiB on-device footprint and competitive inter-patient accuracy, though transmitted activations remain vulnerable to reconstruction attacks.
Answers are grounded in this paper and cite their sources.
May enable efficient wearable ECG classification without data sharing; extends SplitFed but leaves open privacy safeguards and clinical validation.
Wearable electrocardiogram monitoring in the Internet of Medical Things requires arrhythmia classifiers that run on resource-constrained devices and learn across hospitals without pooling patient data. We propose FedIoMT-RPM, a split federated learning framework in which each device trains only a 28.4 KiB client-side network, 6.8 times smaller than in full-model federated learning, while a bounded adaptive weighted aggregation adjusts each client''s contribution to its smoothed validation performance. Under the patient-independent inter-patient protocol of the MIT-BIH Arrhythmia Database, with five clients holding complete patient records and six seeds per method, its detection performance was statistically comparable to that of the best federated baselines, with an accuracy of 93.97 ± 1.72%, a Matthews correlation coefficient of 0.731 ± 0.039 and a macro-F1 score of 60.07 ± 2.41%. This macro-F1 reflects limited detection of supraventricular beats, with a sensitivity of 64.8%, and near-zero detection of fusion beats, and 1,381 of the 1,781 supraventricular test beats come from a single record. Among the split methods, FedIoMT-RPM averaged 18.9% less communication than SplitFed. Zero-shot INCART and PTB-XL benchmarks gave results comparable to the federated baselines. Reconstruction attacks recovered the ECG from the transmitted activations, which need protection before clinical use; membership inference stayed near chance.
No takes yet. Share an insight, caveat, or question.
Venkatesh et al. (2026) studied ECG arrhythmia classification. FedIoMT-RPM (split federated learning framework with 28.4 KiB client-side network and bounded adaptive weighted aggregation) vs. Federated learning baselines (FedAvg, FedProx, SCAFFOLD, SplitFed, LW-FedAvg) was evaluated on Four-class AAMI ECG heartbeat classification performance (accuracy, Matthews correlation coefficient, and macro-F1 score). FedIoMT-RPM achieved 93.97% accuracy and 60.07% macro-F1 for patient-independent arrhythmia classification with a 6.8-fold smaller device model and 18.9% less communication than SplitFed.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: