Optimized negative spikes in Spiking Neural Networks for low-latency ECG classification

Electrocardiogram (ECG) classification is a critical tool for diagnosing cardiac diseases. Recent advances have shifted from manual feature extraction to machine learning (ML), particularly deep learning (DL). Despite their effectiveness, DL models require large labeled datasets, substantial computational resources, and often lack interpretability, leading to overfitting with limited or unrepresentative data, hindering their use in low-power and real-time clinical applications. Spiking Neural Networks (SNNs), a third-generation neural network paradigm, provide a brain-like, event-driven approach that is particularly well-suited for processing spatial–temporal ECG signals. The inherent low power consumption of SNNs makes them highly advantageous for use in portable medical devices. However, practical deployment has been hampered by challenges that include the design of suitable network architectures and the reduction of quantization errors inherent in spiking neurons. This letter introduces an optimized shortcut path residual SNN architecture that fires negative spikes, effectively enhancing the performance in strong temporal variation ECG datasets, without compromising the event-driven characteristics of spikes while improving the representational capability of spiking neurons. Experiments on the UCR dataset demonstrate the efficacy of the proposed Negative Spike Spiking Neural Network (NS-SNN), thereby supporting the integration of SNNs into portable ECG monitoring devices.