GreenHybridECGNet: Edge-Federated CNN-Transformers for Arrhythmia Classification

GreenHybridECGNet is presented as a privacypreserving, edge-deployable hybrid CNN-Transformer framework for five-class AAMI EC57 arrhythmia classification on the MIT-BIH dataset. The model (443.1K parameters, 1.935 MB FP32) achieves test accuracy 97.84%, macro-F1 0.8956, MCC 0.9298, and AUC-ROC 0.9840, with five-seed confidence intervals (AUC-ROC CI [0.9806, 0.9849]) confirming reproducibility. All five AAMI EC57 classes satisfy the clinical sensitivity threshold (Se ≥ 0.75). A novel FocalSmoothedCE loss unifies Focal Loss (γ = 2) with label smoothing (ε = 0.05) to address severe class imbalance. An advanced federated learning (FL) framework is incorporated, combining K-Means clustered FedAvg, coordinate-wise trimmed-mean Byzantine robustness, stalenessweighted asynchronous aggregation, Secure Aggregation, and Top-k (k=1%) gradient sparsification, delivering only 1.6 MB cumulative communication over 20 rounds. Differential privacy is enforced via DP-SGD. INT8 quantisation (torchao) reduces model size to 1.124 MB and CPU inference to 1.00 ms (3.16× speedup) with negligible accuracy loss (macro-F1 delta = +0.03%). McNemar’s test confirms statistically significant prediction differences versus all three baselines (p¡0.001).