Background: Electrocardiogram (ECG)-based arrhythmia detection is important for the early diagnosis and monitoring of cardiovascular diseases. However, accurate ECG classification remains difficult because of class imbalance, variations in ECG morphology, and noise in real-world recordings. Traditional data augmentation methods may distort important clinical characteristics of ECG signals, while conventional classification models often struggle to recognize minority arrhythmia classes. Methodology: This thesis presents a unified framework that combines a Transfer-Guided Multi-Stage Generative Adversarial Network (TG-MS-GAN), an EfficientNetV2-S classifier, and an adaptive learning strategy. The proposed GAN generates realistic ECG beats while preserving the morphology of the P-wave, QRS complex, and ST segment. The synthetic data are then used to train the EfficientNetV2-S classifier with an adaptive hybrid loss function that improves learning from imbalanced datasets and enhances the recognition of minority classes. Results: The proposed framework was evaluated on four publicly available ECG datasets: MIT-BIH, Chapman–Shaoxing, PTB-XL, and Fetal ECG. It achieved classification accuracies of 99.1%, 97.6%, 91.4%, and 99.0%, respectively. The corresponding Macro-F1 scores were 0.987, 0.962, 0.839, and 0.989, while the AUPRC values reached 0.989, 0.891, 0.846, and 0.991. These results demonstrate significant improvements in the recognition of minority arrhythmia classes. Conclusion: The proposed framework effectively addresses the class imbalance problem while preserving clinically important ECG morphology. By combining transfer-guided data augmentation, efficient feature extraction, and adaptive learning, it achieves high classification accuracy, strong generalization across different datasets, and low computational complexity, making it suitable for real-time wearable and remote cardiac monitoring applications.
