Deep learning driven electrocardiogram classification with optimized convolutional neural network for accurate arrhythmia detection and explainable clinical decision support

Abstract

Since cardiovascular diseases are the top cause globally, there is a critical need to develop dependable automated electrocardiogram (ECG) classification systems that can detect arrhythmias early. Existing deep learning models for ECG classification demonstrate potential for enhanced accuracy but encounter obstacles related to class imbalance, limited generalization across different datasets, computational inef, and interpretability issues. The paper presents an enhanced CNN-based ECG classification framework which merges multi-scale feature extraction with sophisticated preprocessing methods and explainability tools to resolve prevalent model limitations. The model combines wavelet-based noise reduction with hybrid approaches to class balance through synthetic data creation and domain adaptation, which results in improved detection of uncommon arrhythmias. The use of Grad-CAM and SHAP-based visualizations enhances clinical interpretability by providing clear insights into model predictions. Real-time execution optimization with minimal computational overhead enables this model to function appropriately for wearable cardiac monitoring applications. The experimental findings show that the proposed CNN model reached 98.86% test accuracy, which surpasses the performance of current deep learning models. The model maintains computational efficiency as it achieves better generalization and interpretability. This research advances AI-driven cardiac diagnostics by connect-ing AI-based ECG classification with real-world clinical applications to enable reliable and interpretable arrhythmia detection systems that can be deployed in practice.