Early identification of Parkinson’s disease from EEG-based functional connectivity matrices

Parkinson’s disease (PD) disrupts large-scale brain networks, but most current diagnostics—reliant on late-appearing motor symptoms—miss the window for early intervention. While EEG captures neural dynamics, traditional connectivity metrics (e.g., linear correlation, coherence) overlook nonlinear dependencies critical to PD pathology. The goal is to detect early brain network changes linked to Parkinson’s disease by representing mutual information (MI) connectivity data as spatial images. Deep learning is then used subsequently for classifying PD and healthy controls, while using explainable AI to identify the possible electrodes underlying brain neural connectivity. Resting-state EEG recording data from PD patients and matched controls was transformed into whole-brain MI connectivity matrices, treated as 2D adjacency matrix images. The convolutional neural network (CNN) classified the matrices. Gradient-weighted Class Activation Mapping (Grad-CAM++) highlighted the key connections that drove the network’s decisions. These maps also visualized how the topology shifted. The CNN achieved strong classification performance, showing that MI matrices capture signatures of Parkinson’s disease. Grad-CAM localized decisions to pathological connections in frontal-central-temporal circuits core motor-execution networks degenerating earliest in PD. Further analysis validated compensatory strengthening of short-range intrahemispheric connections alongside degraded long-range integration, aligning with PD’s "network efficiency collapse" hypothesis. By exposing presymptomatic network reorganization via MI matrices and spatial deep learning, we offer an EEG-based signature for early detection.