A Comparison of Machine Learning Models for Classification of Parkinson’s Disease During a Working Memory and Sustained Attention Task

Background and Objectives: Individuals with Parkinson’s disease (PD) experience deficits in working memory (WM) and sustained attention (ATTN), but diagnosing and monitoring these deficits remains challenging. This study compares machine learning (ML) classification models trained on EEG and pupillometry data from WM and ATTN tasks to identify task-specific and shared cognitive biomarkers of PD. Methods: EEG and pupillometry were recorded from PD patients and healthy controls (HC) during a visual change detection WM task and a continuous performance ATTN task. A standardized toolbox extracted 108 features, reduced via PCA and recursive elimination (RE) and classified using an SVM-RBF within a nested, 5-fold cross-validated pipeline, with class balancing (SMOTE) and feature selection performed strictly within training folds to prevent leakage. Results: On internal test folds, WM achieved 71% accuracy (F1 = 0.701) and ATTN achieved 73% (F1 = 0.699); on an independent hold-out set, WM achieved 63% accuracy (F1 = 0.626) and ATTN achieved 70% (F1 = 0.623). ATTN showed higher accuracy and precision, WM showed higher recall, and univariate analyses independently supported several top-ranked features (e.g., theta/beta and alpha/theta ratios); PAI and FAA, though top features in both tasks, reached univariate significance only in ATTN. These results indicate WM and ATTN yield complementary, task-linked neurophysiological signatures relevant to PD classification. Conclusion: At its current stage, this pipeline functions as a research tool for biomarker discovery rather than a clinical diagnostic, though larger, externally validated samples could support future screening and monitoring applications.