Interpretable Electrophysiological Features of Resting-State EEG Capture Cortical Network Dynamics in Parkinsons Disease

A research team led by Antonios G. Dougalis has published a study demonstrating how interpretable AI can decode brain activity to track Parkinson's Disease (PD). The system analyzes resting-state EEG data using a comprehensive set of 28 interpretable electrophysiological features. These features are grouped into two categories: 'Standard' descriptors like spectral power and phase synchronization, and 'Dynamical' descriptors that capture more complex network behaviors, including aperiodic activity, cross-frequency coupling, and neuronal avalanche statistics. A multi-head attention transformer classifier, rigorously validated using Leave-One-Subject-Out (LOSO) protocols, was trained to identify disease-related patterns.

The AI model revealed distinct neural signatures. Standard features were most effective at detecting changes in medication state (PD patients on vs. off medication), showing medication-sensitive reductions in delta power and voltage variance. In contrast, the novel Dynamical feature set proved highly competitive at distinguishing PD patients from healthy controls, indicating it captures broader, disease-related alterations in cortical network organization. Random feature ablation analysis confirmed that these Dynamical descriptors provide complementary, non-redundant information. The findings, detailed in the arXiv preprint (2604.01475), establish a promising multivariate framework for developing non-invasive, AI-powered biomarkers that could one day aid in diagnosis and treatment monitoring for Parkinson's.