A robust method for parkinson's disease diagnosis: Combining electroencephalography signal features with reconstructed phase space images

Abstract

Parkinson's disease (PD) is a neurodegenerative disease. Since the diagnosis of the PD is mainly made based on the symptoms and after the disease progression, early diagnosis can play a crucial role in delaying the passage of the PD. There have been many methods focusing on disease diagnosis using electroencephalography (EEG) signals, where most of the proposed methods are data-dependent. Here, the study aims to propose a technique that, despite its high accuracy, is robust. Various features including fractal dimension, approximate entropy, largest Lyapunov exponent, and the energy of different frequency sub-bands were extracted from EEG signals. Multi-layer perceptron neural networks were used for classification based on these features. Additionally, 2D phase space images reconstructed from EEG signals were classified using convolutional neural networks. Finally, a combination of these features and images was used for classification using ResNets. During 10 rounds of training and testing, the mean accuracies were calculated for three cases: using only features, only images, and a combination of both. The mean accuracies were 84.67 %, 76.5 %, and 90.2 % respectively. The variances for each case were 35.6 %, 19.5 %, and 13.97 %. The lower variance when using a combination of features and images indicates a more accurate and robust classification.