Rare-Beat Detection in Heterogeneous ECG: Federated Learning Under Severe Class Imbalance and Class-Missing Clients

Abstract

Remote arrhythmia monitoring in Internet of Medical Things (IoMT) systems is challenged by electrocardiogram (ECG) class imbalance, acquisition heterogeneity, and privacy restrictions that limit centralized data sharing. These challenges become more severe in federated learning (FL), where clients may exhibit statistically heterogeneous, non-independent and non-identically distributed (non-IID) data, label skew, and few or no examples of rare arrhythmia classes. This study investigates beat-level arrhythmia classification under heterogeneous and class-missing federated ECG distributions. Beat segments from the MIT–BIH Arrhythmia and St. Petersburg INCART databases were allocated without sample reuse across the original pathology-aware synthetic clients, producing controlled class imbalance and class-missing distributions. Stratified beat-level training, validation, and test subsets were then used to train and evaluate a compact one-dimensional convolutional neural network–bidirectional long short-term memory (1D CNN–BiLSTM) model augmented with a convolutional block attention module (CBAM). To address class imbalance and client heterogeneity, the Synthetic Minority Oversampling Technique (SMOTE) is applied only to local training data and combined with cross-entropy loss, macro-F1-based checkpoint selection, client-specific batch-normalization parameters, and a validation-weighted rare-class-aware aggregation rule. Evaluation is reported using both the native six-class label space <inline-formula> <tex-math notation="LaTeX">$\{N,L,R,V,A,F\}$ </tex-math></inline-formula> and the Association for the Advancement of Medical Instrumentation (AAMI) EC57 grouped categories <inline-formula> <tex-math notation="LaTeX">$\mathcal {N}/\mathcal {S}/\mathcal {V}/\mathcal {F}$ </tex-math></inline-formula>. Across seven non-IID clients and six matched random seeds, the proposed method achieves the highest final-round FL performance among the evaluated strategies, with an accuracy of <inline-formula> <tex-math notation="LaTeX">$0.8590 \pm 0.0053$ </tex-math></inline-formula>, a six-class macro-F1 of <inline-formula> <tex-math notation="LaTeX">$0.8428 \pm 0.0109$ </tex-math></inline-formula>, and an AAMI macro-F1 of <inline-formula> <tex-math notation="LaTeX">$0.8256 \pm 0.0110$ </tex-math></inline-formula>. The largest gains are observed on low-resource minority-bearing clients, particularly the atrial/supraventricular-bearing and fusion-bearing virtual users. Rare-class diagnostics show that, under the evaluated low-support setting, fusion-class performance on the only fusion-positive downstream client is limited primarily by precision rather than recall. Deployment-oriented single-window inference feasibility is evaluated on a Raspberry Pi 5 using frozen TensorFlow Lite artifacts, achieving a 95th-percentile inference latency below 34 ms and a serialized model size below 1 MB. These results indicate that rare-class-aware personalized FL can improve minority-class-sensitive ECG classification under heterogeneous, label-sparse client distributions without sharing raw ECG data. However, persistent <inline-formula> <tex-math notation="LaTeX">$N$ </tex-math></inline-formula>–<inline-formula> <tex-math notation="LaTeX">$V$ </tex-math></inline-formula> ambiguity and limited fusion-beat support remain important targets for future rhythm-context modeling.