Hybrid Bi-LSTM-Spiking Neural Network for Energy-Efficient ECG Arrhythmia Classification

Abstract

Wearable ECG monitoring requires a classifier that can operate within a limited power budget. This study therefore combined a bidirectional long short-term memory (Bi-LSTM) layer, which processed the waveform in both temporal directions, with two leaky integrate-and-fire (LIF) spiking layers. The recurrent output was converted into sparse events, while a fast-sigmoid surrogate gradient enabled training through the nondifferentiable spike function using backpropagation through time. Twenty hyperparameter configurations were evaluated using 60-40, 70-30, and 80-20 training-testing partitions. The best configuration, Configuration 3 using the 80-20 partition, comprised one 64-unit Bi-LSTM layer, a membrane decay parameter of 0.85, and a learning rate of 0.00020. It achieved a training accuracy of 97.64% and a test accuracy 96.88%, exceeding the reporting accuracy of the standalone SNN by 5.21 percentage points. Thus, the model maintained high ECG classification performance while shifting its later-stage processing to event-driven spike activity.