Interactive Local and Global Feature Coupling for EEG-Based Epileptic Seizure Detection

Abstract

Automatic seizure detection based on scalp electroencephalogram (EEG) can accelerate the progress of epilepsy diagnosis. Current seizure detection models based on deep learning usually rely on single convolutional neural network (CNN) or RNN models. In terms of feature extraction, single model often has limitations. Inspired by the recent progress of CNN and Transformer, we put forward a seizure detection model based on interactive local and global feature coupling. Local feature and global representation of the EEG are respectively extracted by convolution operation and selfattention mechanism. To make up of the shortcomings of local feature and global representation, a feature coupling block (FCB) is utilized to fuse the two kinds of information in an interactive way. The enhanced feature representation is fed to the classifier for seizure and normal EEG classification. Extensive experiments are conducted on the CHB-MIT dataset. The experimental results demonstrate that the model can effectively detect epileptic seizures from the original EEG signal without extra feature extraction. The accuracy, sensitivity and specificity are 98.55%, 97.24% and 97.38%, respectively.

Note:
Funding Information: This work was made possible through support from the China Postdoctoral Foundation (No.2017M612335), National Natural Science Foundation of China (NO.81871508, No.61773246, No.61701270), the program for Youth In- novative Research Team in University of Shandong Province (NO. 2019KJN010).

Conflict of Interests: The authors declare that they have no conflict of interest.

Keywords: Electroencephalography (EEG), Seizure Detection, Convolutional Neural Networks(CNN)；Transformer；Feature Fusion