Diagnosis of mild Alzheimer's disease by EEG and ERP signals using linear and nonlinear classifiers

Abstract

A new method for the diagnosis of Alzheimer's disease in the mild stage is presented due to its association with the characteristics of the EEG brain signal. In the proposed method, three channels Pz, Cz, and Fz of brain signals were recorded from 40 participants in the four modes of closed eyes, open eyes, recall, and stimulation. After preprocessing, wide various features of EEG signals were extracted. Results have shown that after stimulation, the amplitude and the latency of the P300 component change with increasing age and the stages of Alzheimer's disease. In addition, the first changes in brain signal in mild Alzheimer's patients are increased activity in the theta band and decreased activity in the beta band, which is associated with a decrease in alpha-band activity. After selecting the proper features, linear discriminant analysis, and Elman and convolutional neural networks were used to classify the participants' features. The results showed that the Pz channel among three EEG signals and the stimulation mode among the four recording steps had greater accuracy than the others. By using features of the Pz channel, the accuracy of LDA was 59.4% and 66.4% in the recall and excitation modes, accuracy of Elman neural network was 92.3% and 94.1%, and CNN was 97.5% and 99% respectively. Results have shown that extracting appropriate linear and nonlinear features has increased the accuracy of the classifiers. In addition, due to the dynamic nature of the brain signal, the convolutional neural network had better performance than LDA and Elman.