Machine Learning Hybrid Approach for the Diagnosis of Parkinson’s Disease Using Electroencephalogram: A Comparative Analysis

Abstract

It has been challenging to develop efficient Machine Learning models due to the lack of electroencephalography (EEG) data. In this paper, a unique hybrid data augmentation pipeline strategy is proposed to improve the performance of a model that categorizes cognition in Parkinson’s disease (PD) using EEG signals. In this study, a hybrid method is implemented, which includes time-shifting, noise addition, subsampling, filtering and synthetic data synthesis using a GAN model. This technique is designed to augment the EEG data with realistic-looking synthetic EEG data. The experiment is carried out on an EEG dataset obtained from the North Shore Health Institute in Chicago. In this experiment, the hybrid data augmentation pipeline strategy is used to generate augmented data from the raw EEG data. After performing preprocessing using Independent Component Analysis (ICA), feature extraction is done using Debiased Weighted Phase Lag Index (dwPLI) and other connectivity measures. The classification is performed with a Random Forest classifier to classify four categories of cognition (healthy control (HC), PD-normal control (PD-NC), PD-mild cognitive impaired (PD-MCI) and PD-demented (PDD). The results show that the proposed approach is capable of generating authentic EEG data which improves the accuracy of the classification model. Furthermore, the mean accuracy of classification after data augmentation is increased by 7% for both training and testing accuracies. Additionally, sensitivity has increased significantly from 55 to 70%. Therefore, a hybrid data augmentation technique is helpful in overcoming data shortage concerns for EEG-based cognitive status categorization.