A Fuzzy-Enhanced Broad Learning Framework with Minority-Class Ensemble for ECG Arrhythmia Classification

Existing Broad Learning System (BLS)-based Electrocardiogram (ECG) arrhythmia classification methods still face limitations in multi-source feature representation, sample reliability modeling, and minority-class recognition. To address these issues, this study proposes a combinational improvement framework built upon BLS, integrating uncertainty aware gated fuzzy feature enhancement and a minority-class ensemble strategy for ECG arrhythmia classification. First, heartbeat waveform, QRS complex structural information, and RR rhythm features are integrated to construct a hybrid ECG representation that jointly captures morphological and rhythm characteristics. The BLS is then employed for high-dimensional nonlinear feature mapping, upon which an uncertainty-aware gated fuzzy enhancement mechanism evaluates sample reliability through robust class prototype estimation and fuzzy membership calculation. A confidence-based gating strategy is further introduced to achieve adaptive feature enhancement, strengthening discriminative information from reliable samples while reducing noise interference caused by ambiguous boundary samples. Finally, a minority-class-aware ensemble classification strategy is developed, which employs eXtreme Gradient Boosting (XGBoost) as the primary classifier. One-versus-Rest(OvR) classifiers, a complementary Logistic Regression (LR) model, and class-wise probability calibration are incorporated to refine classification decisions and alleviate majority class bias caused by class imbalance. Experimental results demonstrate that the proposed method achieves overall accuracies of 99.04% and 92.58% (Macro-F1 of 93.39% and 61.60%) under the intra-patient and inter-patient evaluation protocols, respectively. The framework trades a moderate training-time and memory footprint (an ensemble of tree models over ANOVA-selected BLS features) for these minority-class gains, while inference remains a CPU-friendly single forward pass through the frozen pipeline. It improves minority-class recognition relative to the compared methods under severe class imbalance, while the recognition of fusion beats under the inter-patient protocol remains challenging(five-seed mean F1 of 6.72%±2.69%)and is explicitly discussed as a limitation.