Robust ECG signal classification using spiking neural networks with axonal delays

Abstract

Cardiovascular diseases (CVD) continue to be the primary cause of mortality, with myocardial infarctions and strokes being the main contributors; according to the World Health Organization (WHO), timely diagnosis and intervention are crucial to reducing mortality rates. Electrocardiography (ECG) functions as a fundamental diagnostic instrument for detecting CVD conditions such as arrhythmias. However, the complex and noisy nature of ECG signals has become a significant challenge for accurate classification. This paper proposes the use of Spiking Neural Networks with axonal delays (D-SNNs) for ECG signal classification. Unlike traditional artificial neural networks, SNNs emulate the biological behavior of neurons by processing information through discrete spikes over time, making them well-suited for capturing temporal dependencies in sequential data. A key component of the proposed methodology is integrating Leaky Integrate-and-Fire (LIF) neurons with axonal delays, which enhance the network’s ability to learn and process temporal patterns in ECG signals. These features provide advantages such as improved energy efficiency, asynchronous event-driven processing, and a biologically inspired approach to signal classification. The method is evaluated through extensive experiments on the MIT-BIH Arrhythmia Database, following the Association for the Advancement of Medical Instrumentation (AAMI) recommendations. The heartbeat samples are grouped into four categories: Normal (N), supraventricular ectopic (SVEB), ventricular ectopic (VEB), and fusion of ventricular and normal beats (F). A binary classification experiment is also conducted to differentiate normal from arrhythmic beats. The training and testing data are divided according to inter-patient and intra-patient classification paradigms to ensure robust evaluation. Experimental results demonstrate that the proposed SNN-based model achieves an average classification accuracy of 83.15 % for binary classification in the inter-patient schema and 98.27 % in the intra-patient schema. For multi-class classification, the model achieves 86.38 % accuracy in the inter-patient schema and 98.23 % accuracy in the intra-patient schema. Finally, to reduce model complexity, a combined L1 regularization and pruning strategy was applied to the intra-patient multiclass paradigm, significantly lowering the energy consumption to an estimated 1.91 µJ per inference while maintaining a high accuracy of 98.23 %. These results highlight SNNs as an efficient and accurate approach for ECG-based arrhythmia detection.