Explainable framework to detect Parkinson’s disease related depression from EEG

Depression is a non-motor symptom inherent to Parkinson’s disease (PD). As an early manifestation of PD, PD-related depression is hard to diagnose, thereby contributing to morbidity. Recent endeavors have employed deep learning networks to assist in the diagnosis of PD and depression, achieving commendable levels of accuracy. However, little attention has been directed toward PD-related depression, and the decision process of the network lacks transparency and explainability. What has been learned by the network and whether pathological mechanisms contribute to the classifier’s result remain mysterious. In this study, we propose an explainable functional connectivity framework to recognize depression in PD. Specifically, the diagnosis feature extraction module learns high-dimensional information from functional connectivity features, followed by the diagnosis module making decisions. Furthermore, the explainable module provides interpretation and validation for the decisions on functional connectivity. Evaluation of the dataset demonstrates superb subject-wise predictive performance and provides visual evidence of the underlying pathology in EEG. The interpretation results bridge the gap between pathophysiological mechanisms and computer-aided diagnosis.