Key result
FedCardioGuard outperforms standard federated learning models in ECG arrhythmia classification with ~0.04-point higher macro-F1.
Why the study?
Personalized federated ECG classification must address patient-level non-IID data, severe class imbalance, unreliable confidence, and edge constraints without centralizing raw signals.
Does the FedCardioGuard personalized federated learning framework improve macro-F1 arrhythmia classification performance and calibration compared to standard federated learning baselines in patient-clients from the MIT-BIH DS1 dataset?
Does the FedCardioGuard personalized federated learning framework improve macro-F1 arrhythmia classification performance and calibration compared to standard federated learning baselines in patient-clients from the MIT-BIH DS1 dataset?
FedCardioGuard significantly improves macro-F1, worst-client F1, and confidence calibration in patient-level federated ECG arrhythmia classification compared to existing federated learning baselines, though rare supraventricular class recognition remains limited by cold-start challenges.
Should not yet change ECG workflows; extends personalized federated learning methods with better calibration on benchmark data.
Personalized federated electrocardiogram classification must address patient-level non-IID data, severe class imbalance, unreliable confidence, and edge constraints without centralizing raw signals. We propose FedCardioGuard, a quality- and uncertainty-aware framework combining a lightweight multi-scale encoder, federated aggregation, local-global prototype personalization, patient-adaptive fusion, and calibrated selective prediction. In a locked development study on 22 patient-clients from MIT-BIH DS1, ten independent seeds, and 15 federated rounds, FedCardioGuard achieved a macro-F1 of 0.556 ± 0.049, exceeding FedProto by 0.042 and Ditto by 0.067. The overall difference was significant (Friedman p = 2.23 × 10⁻⁵), and all paired comparisons remained significant after Holm correction. Calibration reduced ECE from 0.245 to 0.089 without changing hard predictions, while support-aware worst-client F1 improved by 49.7% over FedProto. However, S-class recall remained 0.082, identifying rare-class recognition as the principal unresolved limitation. A subsequent stability-aware configuration reached a macro-F1 of 0.569 ± 0.050 but did not satisfy the frozen rare-class and calibration gates. Controlled architecture, cold-start, representation, metric-head, and patient-disjoint audits localized the remaining failure to rare-class cold start and patient-specific score transfer. These DS1 results support the personalization and calibration design but are intentionally treated as development evidence; DS2 remained unopened throughout the complete study.
No takes yet. Share an insight, caveat, or question.
Saberi et al. (2026) studied Cardiac arrhythmia (n=22). FedCardioGuard (quality- and uncertainty-aware personalized federated learning framework) vs. FedProto and Ditto (personalized federated learning baselines) was evaluated on Macro-F1 score for four-class ECG arrhythmia classification (MD 0.042 vs FedProto; MD 0.067 vs Ditto, p=2.23e-5). FedCardioGuard achieved a macro-F1 of 0.556 in ECG arrhythmia classification, outperforming FedProto by 0.042 and Ditto by 0.067 (Friedman p=2.23e-5) while reducing calibration error to 0.089.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: