Parkinson's disease classification using optimized attention-based deep learning from EEG signals with interpretable...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Parkinson's disease (PD) diagnosis is challenging due to subtle neural changes not always detectable by clinical assessment.
• Resting-state electroencephalography (EEG) offers a non-invasive method to capture brain activity.
• Developing accurate and automated diagnostic tools for PD is crucial.

Purpose of the Study:

• To propose and evaluate an attention-based deep learning framework for classifying Parkinson's disease (PD) using resting-state EEG.
• To assess the model's performance on independent datasets with rigorous, leakage-safe cross-validation.
• To explore the neurophysiological interpretability of the model's findings.

Main Methods:

• Utilized raw EEG data, decomposed into sub-bands via discrete wavelet transform.
• Employed a ResNet-101 backbone with dual channel-spatial attention for time-frequency spectrogram classification.
• Implemented subject-wise nested leave-one-subject-out cross-validation and hyperparameter optimization using AH-CMA-ES.

Main Results:

• Achieved 96.77% subject-level accuracy on the UC San Diego dataset and 92.86% on the University of Iowa dataset.
• Demonstrated robust performance through leakage-safe evaluation, excluding held-out subjects from all training aspects.
• Topographical analysis revealed PD-related oscillatory patterns consistent with existing literature.

Conclusions:

• Attention-based deep learning on resting-state EEG shows potential for robust and interpretable Parkinson's disease classification.
• The proposed framework offers a promising avenue for improving PD diagnosis accuracy.
• Further validation on larger, diverse cohorts is recommended to confirm clinical generalizability.