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

Resting-state electroencephalography (EEG) is an easily obtainable and noninvasive signal source for Parkinson's disease (PD) research. However, automatic PD classification from EEG is not a simple task, because EEG signals are noisy, non-stationary, and vary considerably among individuals. This article presents a computationally lightweight two-stage offline framework for window-based PD and healthy-control (HC) classification. In the first stage, a Random Forest classifier is used to generate the initial PD/HC prediction for each EEG window, along with the corresponding prediction confidence. In the second stage, SARSA is applied as a temporal decision-refinement method. Instead of treating each EEG window as an isolated decision, SARSA uses the classifier confidence and the previously refined label to reduce sudden changes between neighboring window predictions. During training, the SARSA reward encourages correct PD/HC classification and applies a small penalty for unnecessary label switching. During testing, the learned SARSA policy is kept fixed and is used only to refine the prediction sequence of unseen subjects. The proposed framework was evaluated using the OpenNeuro ds002778 resting-state EEG dataset. The Stage-1 Random Forest model achieved 74.0%-fold-aggregated window-level accuracy, while the final RF + SARSA framework improved the accuracy to 78.6%. For subject-level evaluation, majority voting was applied to the refined window-level predictions of each participant. The RF + SARSA framework achieved 77.4% subject-level accuracy, with 24 correct predictions from 31 subjects and a 95% confidence interval of 60.2%-88.6%. McNemar's exact test showed no statistically significant subject-level difference between Random Forest and RF + SARSA. Overall, the results suggest that SARSA-based refinement can improve temporal consistency in offline EEG-based PD/HC classification.