A novel feature extraction method using chemosensory EEG for Parkinson's disease classification

Parkinson's disease (PD) is an incurable nervous system disease that affects millions of people all around the world. The loss of smell is one of the first symptoms that come into prominence in the early diagnosis of PD. The main motivation of this study is to provide a more accurate diagnosis in the early period of the disease using chemosensory electroencephalography (EEG) signals, which are difficult to study and also less studied. For this purpose, we proposed a hybrid feature extraction method called EEMD_VAR that combines Ensemble Empirical Mode Decomposition (EEMD) and Vector Autoregressive Model (VAR). In contrast to conventional feature extraction methods, the proposed method is to prevent arbitrary selection of features and to determine the number of features. The pre-processed EEG signals were decomposed using EEMD and the obtained intrinsic mode functions (IMFs) used as independent variables in VAR. The coefficients of the VAR model were employed as features in frequently used supervised classification algorithms. The performance metrics of the EEMD_VAR were compared to the performance metrics of the autoregressive (AR) model and Hjorth parameters. The maximum classification accuracy of the proposed method was 100% using artificial neural networks (ANN) in C2 electrode, while the AR method and Hjorth parameters only obtained a maximum of 72%. The other metrics also corroborate the proposed method's ability to perform well in the classification. In addition, the higher results from right side electrodes may lead to the conclusion that the right side of the brain is more sensitive to odor stimuli.