A Hybrid Deep Learning Approach for ECG Arrhythmia Detection: GPT, GANs, and Triplet Loss Integration

Abstract

This paper proposes a novel deep-learning method to detect arrhythmias from the ECG data by adopting pre-trained GPT models and other powerful state-of-the-art DL algorithms. Most traditional ECG classification models face challenges in capturing complex temporal dependencies and handling class imbalances. To meet these challenges, our system leverages GPT to capture complex temporal patterns and contextual relationships within ECG signals, enabling us to better under­stand the more intricate depen­dencies in the data. Finally, the proposed system lever­ages data augmentation with Generative Adversarial Networks (GANs) to generate a wide variety ofcomplex samples, which help improve model capability and robustness. It also uses Triplet Loss, which shows it can work better on imbalanced classes and tiny differencesin different cardiac arrhythmias. Compared with other methods, our results exhibit great im­provements in classification performance, particularly for rare arrhythmias. Model Interpretability is based on SHapley Additive exPlanations(SHAP) and Gradient weighted Class Activation Map (Grad-CAM), which interpret the model decisions.