Graph-Driven Feature Selection for ECG Classification: Leveraging Structural Dependencies Between Leads

The precise identification of ECG patterns serves as an essential requirement for developing automated cardiac diagnosis systems. While Convolutional Neural Network (CNN) models are widely used, they often overlook inter-lead relationships. This research introduced a graph-based feature selection approach that depends on ECG-lead structural relationships to enhance diagnostic precision. Rather than treating all the leads equally, an adjacency matrix is formed based on lead correlations. The approach captures inter-lead relationships, and features are extracted from the upper triangle of the matrix to prioritize significant lead interactions. The methodology selects important lead interactions so that only the most relevant features influence the classification. Two Convolutional Neural Network (CNN) models are proposed: a binary classifier distinguishing normal from abnormal ECGs, achieving 97.61% accuracy, and a multi-class classifier identifying eight specific cardiac conditions, with 91.78% accuracy. These results demonstrate the effectiveness of the proposed approach in improving ECG classification by prioritizing meaningful lead interactions.