Mapping Neural Differences in Parkinson’s Disease: A Source Reconstruction Approach to EEG Data

Abstract

Introduction: Parkinson’s disease (PD) affects millions worldwide, yet early and accurate diagnosis remains challenging. Electroencephalography (EEG) based identification of neural signatures may offer new insights for distinguishing PD from healthy individuals. Objective This study aimed to identify brain regions through EEG source reconstruction that could be used to differentiate between PD patients and healthy individuals. Materials and Methods: Publicly available EEG data, provided by OpenNeuro, from 14 healthy and 14 PD subjects was used for the study. Data underwent Finite Inpulse Response filtering, denoising with Principal Component analysis and Independent Component Analysis, and MinMax thresholding to remove noise. Source reconstruction, a method to trace EEG electrode signal origins to specific brain regions, was performed using OpenMEEG’s Boundary Element Method, registered to the Desikan-Killiany-Tourville (DKT) atlas, resulting in EEG signal reconstruction across 62 DKT-defined brain regions. Five power spectral densities (PSD) were calculated for a total of 310 features per recording. The Maximum Relevance Minimum Redundancy (MRMR) algorithm was used to select key brain regions and frequency bands. Machine learning (ML) algorithms Support Vector Machines, Logistic Regression, K-Nearest Neighbor, Naïve Bayes, Random Forest, and Deep Neural Network (DNN), with 10-fold cross-validation were used to discriminate PD patients from the healthy group Wilcoxon analysis was used to assess statistical significance of classification results among ML methods. Results: All ML methods successfully classified PD with the highest accuracy (96% ± 10%) achieved by DNN using five PSD features - left lateral occipital, right precuneus, and left cuneus in the Delta range, medial orbitofrontal in Theta range, and right lingual in Alpha range. ML methods used did not yield statistically significant differences in classification accuracy. The classification results demonstrate the potential of using EEG source reconstruction to distinguish between Parkinson’s disease and healthy individuals. However, due to the limited dataset size, these findings should be interpreted with caution. Future large-scale studies incorporating additional biomarkers and a broader parameter space are essential to validate and extend these preliminary insights.