A Hybrid Random Forest-SARSA Framework for Resting-State EEG-Based Parkinson's Disease Detection With Temporal...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Resting-state electroencephalography (EEG) offers a noninvasive method for Parkinson's disease (PD) research.
• Automatic PD classification from EEG is challenging due to signal noise, non-stationarity, and inter-individual variability.
• Existing methods often treat EEG windows independently, potentially missing temporal dynamics crucial for accurate classification.

Purpose of the Study:

• To develop and evaluate a computationally lightweight, two-stage offline framework for classifying Parkinson's disease (PD) from resting-state EEG.
• To enhance classification accuracy and temporal consistency by integrating a reinforcement learning approach (SARSA) with a Random Forest classifier.
• To reduce sudden changes in classification predictions between adjacent EEG windows.

Main Methods:

• A two-stage offline classification framework was implemented for window-based PD and healthy-control (HC) classification.
• Stage 1: A Random Forest classifier was used for initial PD/HC prediction and confidence estimation for each EEG window.
• Stage 2: SARSA (State-Action-Reward-State-Action) reinforcement learning was applied to refine temporal decision-making, using classifier confidence and previous predictions to ensure smoother transitions.

Main Results:

• The Stage-1 Random Forest model achieved 74.0% window-level accuracy.
• The final Random Forest + SARSA framework improved window-level accuracy to 78.6% and subject-level accuracy to 77.4% (24/31 subjects correctly classified).
• SARSA-based refinement demonstrated improved temporal consistency in the classification sequence without significant statistical difference in subject-level accuracy compared to Random Forest alone.

Conclusions:

• The proposed two-stage RF + SARSA framework effectively improves temporal consistency for offline EEG-based Parkinson's disease classification.
• SARSA reinforcement learning offers a valuable method for refining sequential predictions in noisy, non-stationary biological signals like EEG.
• This computationally lightweight approach shows promise for enhancing the reliability of automated diagnostic tools for neurological disorders using EEG data.