Compact seizure detection based on spiking neural network and support vector machine for efficient neuromorphic implementation

Due to the rapid development of artificial intelligence, seizure detection has achieved great success in terms of accuracy and speed. However, low-power seizure detection algorithms remain a challenge. In this paper, a seizure detection method is presented based on spiking neural network (SNN) and support vector machine (SVM) for neuromorphic implementation. To be specific, the power threshold is applied in advance before SNN to avoid unnecessary computing on electroencephalogram (EEG) that is unlikely to be a seizure. A small-scale three-layer fully-connected SNN model is used to discriminate seizure efficiently. An ultra-low dimensional SVM is adopted to further improve accuracy at low computational cost. The computation complexity is reduced to 6 K add operations with power threshold, and no multiply operation is used. Experiments on the Children's Hospital Boston-Massachusetts Institute of Technology dataset show that the accuracy and sensitivity can achieve 95.07% and 88.44%, respectively. The energy consumption of the proposed method is smaller than other previous solutions, which is suitable to be integrated as near sensor intelligence for wearable long-term seizure detection.