Machine Learning-based EEG Signal Classification of Parkinson's Disease

As the second most common neurodegenerative disease in the world, Parkinson's disease continues to affect the normal and healthy life of patients. In recent years, considerable progress has been made in studying the EEG of patients with Parkinson's disease. Many EEG data of patients with Parkinson's disease can be published, and more filtering algorithms and classification models suitable for EEG signals of Parkinson's disease have been proposed. However, studying channel redundancy of EEG signals in Parkinson's disease still faces challenges. The pathogenesis of Parkinson's disease is still uncertain in medicine, and it is difficult to propose a channel selection scheme suitable for all patients with Parkinson's disease. In this paper, the open UNM data set is used to extract multi-scale features based on the fourth-order Butterworth IIR filter and Wavelet Packet Transform. The channel selection is carried out by using single-channel verification. 12 and 25 channels with the relative best R2 scores were selected for the feature data set generated based on these two methods. The classification performance of data sets with and without channel selection was compared between the open and closed-eye data sets. The negative effect of open eye status on EEG classification of Parkinson's disease was found, and the channel selection was used to improve the AUC by 1% in the same data set. Results showed that the proposed channel selection scheme can alleviate the overfitting phenomenon that occurred in the training set in the testing set while maintaining the classification accuracy.