A Decision Support System For Early Stage Parkinson's Diagnosis from EEG Data Using Symbolic Mutual Information and KAC Features

Parkinson's disease (PD) is a serious neurological disease that is threatening the whole world population. The devolution of the neurons located in the substanstia nigra of the brain causes, bradykinesia, rigidity and resting tremor, which are characteristic motor symptoms, occuring in advanced stages. Currently, there is not an effective treatment for PD, it is just controlled by some prescribtions. Early detection of this disease affects the choice of treatment. Recent studies on early diagnosis by analyzing electroencephalography (EEG) recordings have provided a glimmer of hope. Therefore, in this study, an efficient PD detection method from EEG data by using a new set of features is searched. An opensource resting state data of 28 subjects divided as Parkinson and control gorups were anlyzed. PSDs of the EEG frequency bands that are delta, theta, alpha, beta and gamma and Median Spectral Frequency (MSF), Spectral Entropy (SE), Kolmogorov Algorithmic Complexity (KAC) and Weighted Symbolic Mutual Information (wSMI) were extracted as features. The performance of the PD and control group was evaluated with Gradient Boosting (GB), Gaussian Naive Bayes (GNB), and K-nearest Neighbor (KNN), Support Vector Machines (SVM), Logistic Regression (LR), Categorical Boosting (CatBoost) and Extreme Gradient Boosting (XGBoost) Algorithms. A 85% accuracy was achieved with the XGBoost algorithm, using 31 channels and 13 features which outperforms the results of previous studies using this dataset in the literature.