Epileptic Seizure Detection Using a Lightweight Network based on Style-Controlling Normalization

Intelligent detection of epileptic seizure in electro-encephalography (EEG) is a tedious task for neurologists. While numerous automated seizure detection methods exist, deep learning techniques have garnered the most attention. However, recent deep learning models have not focused on the differences in EEG patterns across patients, which may lead to poorer detection performance of the model for some patients. Additionally, the computational demands of deep networks hinder their real-time clinical application. In this paper, an automatic single-channel EEG seizure detection model that has good generalization and real-time performance is proposed by using style-controlling normalization (SCNorm) and partial convolution (PConv) operator. Moreover, a frequency domain divide-and-conquer operation is proposed to further reduce the computational expenses of the model. The seizure detection model has been evaluated on two epileptic EEG datasets. And the results show that the proposed model achieves a good balance between detection accuracy and calculation speed across various patients.