SPTU-Lite: An Efficient Auxiliary Diagnostic Approach Combining Spiking Neural Networks And Large Kernel Extractors For ECG Signals

Electrocardiogram (ECG) is essential for the early prevention and diagnosis of cardiovascular diseases. The rise of wearable ECG devices has facilitated high-precision heart rate detection, enabling early interventions. Advances in Artificial Neural Networks (ANNs), particularly One-Dimensional Convolutional Neural Networks (CNNs) and Transformers, have significantly improved ECG signal classification. However, their computational and parametric demands limit their application, especially in small-scale devices. Spiking Neural Networks (SNNs), which operate on an event-driven principle, offer a solution with their resource efficiency. In this study, we propose a U-shaped hybrid CNN-Transformer architecture combining Spiking Fully Connected (FC) Layers and large kernel extractor (LKE) modules, named Spiking TransU-Lite (SPTU-Lite). The LKE module enhances the network's receptive field and improves parameter efficiency, reducing model size without significant precision loss. Additionally, spiking neurons replace the Fully Connected classification layer, further reducing the model's complexity. Empirical results on the MIT-BIH dataset show that SPTU-Lite reduces parameters by approximately 95%, while maintaining an accuracy of 99.62% and a loss rate below 0.5%, demonstrating its potential for ECG analysis in resourceconstrained environments.