Explainable Graph Convolutional Network Framework for Robust ECG Arrhythmia Classification and Patient-Level Risk Stratification

Abstract

Accurate and interpretable detection of arrhythmias from electrocardiogram (ECG) signals plays a critical role in the early cardiac risk assessment and patient management. This paper presents a novel, explainable framework that leverages a dynamic Graph Convolutional Network (GCN) to model ECG beat sequences as graphs, where beats act as nodes and temporal RR-interval relationships forms edges. Our approach integrates comprehensive preprocessing steps, including feature selection, Synthetic Minority Oversampling Technique (SMOTE) balancing, and data augmentation, to mitigate severe class imbalance. The model training pipeline followed an 80/20 random beat-level split on the MIT-BIH Arrhythmia dataset with external compatibility checks on the INCART database to assess domain generalizability. Detailed analyses encompass saliency mapping for beat-level interpretability, robustness evaluations under realistic noisy conditions, and personalized risk estimation by mapping the arrhythmia burden to clinically relevant risk zones. Comparative evaluation against baseline models, such as multilayer perceptron and CNN- BiLSTM hybrids, demonstrated consistent macro F1-scores and reliable patient-level stratification. This integrated pipeline not only advances robust arrhythmia classification but also facilitates actionable clinical decision support via transparent risk stratification. All code and analyses are reproducible with fixed random seeds and open-source implementations, underscoring the potential of the framework for real-world deployment in ambulatory cardiac monitoring.