Abstract

This paper presents the 2D-CNN-GACL-ECGNet, a novel framework for electrocardiogram (ECG) classification aimed at enhancing stress detection in cardiovascular disease (CVD) patients. The methodology integrates a two-dimensional convolutional neural network (2D-CNN) with a graph attention network (GAT) and adaptive contrastive learning (ACL), addressing challenges posed by stress-induced morphological variability and noise. Key innovations include dynamic morpho-temporal graph construction, where ECG beats are modeled as nodes with hybrid edges, and a stress-adaptive contrastive loss that reduces class ambiguity by 18%. The model's performance was evaluated using multiple datasets, achieving state-of-the-art results with F1-scores of 98.7% on the MIT-BIH arrhythmia database, 94.2% on the WESAD dataset for emotional stress, and 92.8% on the SWELL-KW dataset for cognitive stress. Additionally, the framework provides interpretable heatmaps aligned with cardiologist annotations, demonstrating a κ value of 0.82. The findings indicate that 2D-CNN-GACL-ECGNet significantly outperforms traditional CNN-BiLSTM and GCN models by 95%, showcasing its potential for clinical applications in wearable health monitoring. This research contributes to the advancement of robust and interpretable ECG classification systems under varying stress conditions.

IPC Classification

Tags