Deep Multi-scale Feature Fusion Convolutional Neural Network for Automatic Epilepsy Detection Using EEG Signals

Abstract

Epilepsy is a brain neuropathy that can be diagnosed by electroencephalogram (EEG). Doctors can diagnose epilepsy by observing epileptic discharge (IED) or amplitude of EEG signals. But human eyes are prone to fatigue, and it is not accurate to rely only on human eyes to observe EEG signals for diagnosis of epilepsy. Automatic epilepsy detection based on EEG signals is necessary. Recently, more and more scholars have begun to classify EEG signals by deep learning. Among them, convolutional neural network (CNN) has achieved good results in feature extraction and classification of EEG signals. However, the CNN is different from the human eyes in learning methods and feature extraction methods of EEG signals. In this study, we designed a new feature fusion CNN model based on dilated convolution kernel to classify three states: normal, preictal, and seizure. Our model is trained and validated based on the epilepsy EEG database of Bonn University. Finally, our CNN model achieved accuracy of 98.67%, sensitivity of 99% and specificity of 98%. Then we use the method of power spectral density (PSD) to study the features of EEG signals extracted by convolution neural network and find it mainly extracts the features of EEG signals from the frequency and amplitude of EEG signals. Our method provides a new direction for automatic diagnosis of epilepsy and the study of feature extraction of EEG signals by deep learning.