Prediction and Classification of Parkinson's Disease Severity In Patients Using Electroencephalogram Signals

Abstract

Parkinson's Disease (PD) is a chronic neurological disorder that has been placed second behind Alzheimer's disease worldwide. Parkinson's disease is characterized by a wide range of symptoms, including tremors (shaking) in the hands, arms, legs, and face; rigidity; sluggish movement (bradykinesia); and difficulties with balance and coordination. Over 10 million individuals worldwide are affected with PD. In this study, ten electroencephalographic (EEG) channels were used to examine changes in brain connectivity within the default mode network (DMN) that occurs during rest. Changes in the default mode network were linked to the severity of PD, and their coherence was utilized to evaluate the causal effects across regions. Much of the brain's default mode network consists of the posterior cingulate cortex, medial prefrontal cortex, or precuneus, and the lateral parietal cortex. Connectivity in the default mode network were estimated using EEG data from 54 patients (27 controls, 27 individuals using PD medication, and 27 individuals not taking medication) over five frequency bands (delta (1-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), low beta (13-21 Hz), and high beta (22-30 Hz). The features are estimated using coherence of EEG channels ( ‘F3’, ‘Fz’, ‘F4’, 'F7', 'F8', 'P4', 'Pz', 'P3', 'P7', 'P8' ) at different frequencies which are then used as input to Artificial Neural Network (ANN) for binary classification of PD and HC cases and also for the linear prediction of severity (UPDRS score) in PD cases. Using leave-one-out cross-validations (LOOCV), Adam optimizer, ReLu and sigmoid activation function. Using the proposed artificial neural network for PD Off medication vs. HC gives an average accuracy of 90.01% ± 4.24 %, specificity of 89.01%,sensitivity of 89.35% and ON medication vs HC gives an average accuracy of 89.36% ± 4.12 % , specificity of 89.03%,sensitivity of 89.70% and ON-OFF PD vs HC gives an average accuracy of 90.52% ± 4.37 % , specificity of 94.57%,sensitivity of 89.90%. For severity (UPDRS score) prediction leave-one-out cross-validations (LOOCV), Adam optimizer, activation function ReLu and linear was used on Parkinson’s disease patients. We found that the average mean absolute error was 2.688 ± 1.4. Index Terms— Parkinson’s Disease, Default mode network, Artificial Neural Network, Leave-One-Out Cross-Validation, Mean Absolute Error.