Morphology-preserving transfer-guided GAN with adaptive discriminative loss for robust and imbalanced ECG arrhythmia detection.

Electrocardiogram (ECG) based arrhythmia detection remains challenging due to severe class imbalance, morphological variability, and noise commonly present in wearable monitoring environments. Traditional augmentation techniques often distort clinically meaningful waveform characteristics, while deep learning classifiers typically underperform on minority arrhythmia categories. This study proposes a unified morphology-preserving framework that integrates a Transfer-Guided Multi-Stage Generative Adversarial Network (TG-MS-GAN), an EfficientNetV2-S classifier, and an adaptive discriminative loss to enhance minority-class representation and improve overall diagnostic robustness. The TG-MS-GAN progressively reconstructs ECG beats across multi-resolution stages, synthesizing high-fidelity minority-class waveforms that preserve P-wave integrity, QRS morphology, and ST-segment structure. A transfer-guidance mechanism introduces physiological priors into early-stage generators, stabilizing convergence under extreme sample scarcity. EfficientNetV2-S serves as a lightweight yet expressive backbone capable of capturing fine-grained temporal features with low computational cost, making the method suitable for real-time wearable deployment. An adaptive hybrid loss combining weighted cross-entropy, center loss, and difficulty-based focusing improves inter-class separability and stabilizes optimization under severe imbalance. Experimental results demonstrate consistent performance across different ECG classification tasks, achieving accuracies of 99.1%, 97.6%, 91.4%, and 99.0% on the MIT-BIH, Chapman-Shaoxing, PTB-XL, and Fetal ECG datasets, respectively. The corresponding Macro-F1 scores reach 0.987, 0.962, 0.839, and 0.989, while the AUPRC values attain 0.989, 0.891, 0.846, and 0.991, respectively. These results indicate that the proposed framework effectively addresses class imbalance while maintaining robust classification performance across heterogeneous ECG datasets. The source code is available at https://github.com/phdmajidsepahvand/TG-MS-GAN-ECG.