Stroke Identification Based on EEG Convolutional Neural Network

Abstract

Stroke is a disease with high incidence and high disability rate. Stroke rehabilitation training technology based on motor imagery-brain computer interface (MI-BCI) has been widely studied. However, brain computer interface (BCI) technology can not realize auxiliary diagnosis. In this paper, the auxiliary diagnosis method of stroke disease is studied, and a stroke disease diagnosis experiment based on electroencephalogram (EEG) is designed. The EEG of 7 healthy subjects and 7 stroke patients in resting state and motor imagery (MI) task were collected and studied as follows: Step 1: the resting state and task state data are analyzed by power spectral density (PSD) according to different rhythms, and the frequency domain characteristics of all EEG signals are observed. Step 2: common spatial pattern (CSP) is performed on resting state and MI task data according to different rhythms, and the spatial features of all EEG signals are observed. Step 3: The EEG-CNN deep learning network architecture is designed, and the labels are designed according to whether they are stroke or not. Six healthy subjects and six patients are randomly selected as the training set, and the rest as the test set for 7-fold cross-validation training and testing. After the results of the above experiments and analysis, it is found that the EEG of the α rhythm of the MI task has the highest degree of discrimination, and the average correct rate on the EEG-CNN network model is 79.0%. Compared with other feature extraction and deep learning algorithms, EEG-CNN prediction accuracy has obvious advantages. The auxiliary diagnosis method and analytical modeling algorithm proposed in this paper provide a reliable research basis for stroke disease diagnosis and severity diagnosis.