FDCANET-SPIKE: EPILEPSY SEIZURE PREDICTION MODEL BASED ON FUSION OF DEEP CONVOLUTIONAL ATTENTION AND SPIKING NEURAL NETWORK

Epilepsy prediction has been a frequently debated subject. However, current methods typically demand substantial computational resources and high energy consumption, restricting their practical applicability. This study introduced fusion depthwise convolution attention network-spike (FDCANet-Spike), a method that integrated attention mechanisms with spiking neural network (SNN) for seizure prediction. By incorporating both channel and spatial attention mechanisms and utilizing a multi-scale deep convolution module, FDCANet-Spike effectively captured local spatial relationships, thereby enhancing prediction performance. To further enhance the model’s energy efficiency, temporal encoding was hereby adopted to convert artificial neural network (ANN) neurons into sparse spike neurons. This approach preserved high performance while significantly reducing computational load and energy consumption. FDCANet-Spike was validated on two publicly available datasets. Compared with existing state-of-the-art algorithms, FDCANet-Spike exhibited superior prediction accuracy and marked enhancement in energy efficiency, consuming merely 28% of the energy required by a comparable ANN model. The model offered high computational efficiency while maintaining low power consumption, highlighting its promising applications in electroencephalogram (EEG) analysis. Such advantages are essential for advancing efficient and user-friendly wearable devices.