Abstract

Objective Parkinson's disease (PD) is a neurodegenerative disorder characterized by motor and non-motor symptoms, including cognitive impairment. Its diagnosis, which used to be based on clinical assessment, increasingly relies on biomarkers. While electroencephalography (EEG) biomarkers are still at an experimental stage, they have been studied using deep learning models. Our aim was to determine whether a cognitive task could improve the accuracy of EEG-based disease detection by activating cortical regions affected by the disease. Approach We trained a deep learning model to discriminate PD patients from controls based on their high-density EEG recordings. Previous studies have employed a range of preprocessing techniques, models and, predominantly, resting state EEG. We also investigated different network architectures and hyperparameters, and the role of spatial and temporal resolution. Main results The best model gave a classification accuracy of 83% on the cognitive task EEG dataset and 76% on the resting state EEG dataset. Sensitivity analysis indicated that the model predominantly uses specific temporal and spatial components of the EEG in the cognitive task condition, differing from the resting state. Significance Our results suggest that cortical activation by the cognitive task unveils EEG features that are effective in distinguishing between PD and controls. These features can be used by the model, thereby improving its diagnostic accuracy.