Parkinson's disease is characterized by sub-second resting-state spatio-oscillatory patterns: A contribution from...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Deep convolutional neural networks (DCNNs) offer advanced multivariate analysis for spatio-oscillatory patterns in electroencephalography (EEG) data.
• Traditional DCNN applications face limitations due to low interpretability, hindering clinical relevance beyond predictive accuracy.
• Existing EEG analysis methods often struggle with the common reference problem.

Purpose of the Study:

• To develop a generalizable and clinically relevant DCNN model for Parkinson's disease (PD) detection using EEG.
• To enhance the interpretability of DCNN models in neurophysiological research.
• To identify specific EEG oscillatory patterns associated with PD and medication status.

Main Methods:

• A minimalistic DCNN architecture with large penalized terms was designed for enhanced generalizability and interpretability.
• The DCNN was trained on an open-access EEG dataset comprising healthy controls and PD patients (with/without medication).
• Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize and interpret the network's key predictive features.

Main Results:

• The DCNN successfully differentiated PD patients from healthy controls across two independent datasets.
• Key predictive patterns included beta band oscillations (occipitoparietal), gamma band (left motor cortex), and theta band (frontoparietal).
• Specific oscillatory patterns correlated with off-medication motor symptoms, disease duration, and on-medication symptom improvement.

Conclusions:

• The developed DCNN approach effectively characterizes PD patho-electrophysiology through multivariate topographic analysis, integrating spatial and frequency aspects.
• The method provides interpretable insights into EEG biomarkers for PD, overcoming common EEG reference issues.
• This approach holds promise for advancing objective diagnostic tools for neurodegenerative diseases.