Multiscale EEG biomarkers for Parkinson's disease: An MIL-inspired MEMD-SHAP channel selection framework

Abstract

Parkinson's disease (PD) is a progressive neurodegenerative disorder for which the early identification of robust neurophysiological biomarkers is essential for timely intervention. In this study, we propose a biomarker-driven EEG analysis framework that integrates Multivariate empirical mode decomposition (MEMD), physiologically meaningful feature extraction, and a Shapley Additive explanations (SHAP)-guided population-consensus channel selection strategy inspired by multiple instance learning (MIL). The framework is systematically compared with the conventional empirical mode decomposition (EMD) approach. MEMD enables adaptive multiscale decomposition of multichannel EEG into intrinsic oscillatory components across frequency ranges known to be affected in PD, and the results indicate that MEMD is a competitive alternative to conventional EMD, with performance depending on the dataset and evaluation setting rather than showing consistent superiority. From these components, statistical, entropy, and nonlinear features are extracted to directly quantify key PDrelated abnormalities, including cortical slowing, reduced neural complexity, and abnormal neural synchronization. Channel relevance is determined using a model-free, Monte Carlo-based SHAP-inspired analysis that quantifies each channel's marginal contribution to a predefined signal-energy target function, without training or explaining any supervised classifier. The selected channels are further validated through statistical testing with large effect sizes and correspondence to established PD EEG biomarkers. Machine learning classifiers are applied at the final stage as a validation tool to confirm the discriminative properties of the discovered biomarkers for application in automatic diagnostic tools. Using public EEG datasets, the framework is evaluated under eyes-open and eyes-closed conditions in both subject-dependent and subject-independent paradigms. In the subjectdependent setting, which probes within-subject separability and represents an optimistic upper bound rather than cross-subject generalization, peak accuracies of 98.14 (eyes-closed, MEMD-SVM), and 97.83 (eyesopen, MEMD-SVM) were observed. More practically relevant, the subject-independent analysis reached accuracies of up to 98.67 (eyes-closed, MEMD-RF) and 98.51 (eyes-open, EMD-RF) in one dataset and 98.89 (eyes-open, EMD-RF) in another dataset. Subject-level aggregation, used to approximate patient-level diagnostic performance under inter-subject variability, reached accuracies of up to 94 among datasets. These results highlight the potential of the proposed framework as a non-invasive, interpretable, and cost-effective tool for EEG-based PD assessment.