Abstract
Parkinson's disease (PD) is a progressive neurodegenerative disorder that affects both motor and cognitive functions, often resulting in misdiagnosis during its early stages. The condition severely impacts daily living, diminishing an individual's ability to work and carry out routine tasks independently. Consequently, the development of automated methods for reliable PD detection has gained growing research interest. Among the available approaches, Electroencephalography (EEG) has emerged as a promising non-invasive and cost-effective tool. Nevertheless, most existing studies have predominantly focused on resting-state EEG, which constrains the generalizability and robustness of the proposed detection models. This study introduces a cross-stimulation evaluation framework to assess its impact on Parkinson's disease detection algorithms and conducts channel-wise analysis to identify the most discriminative brain regions for accurate diagnosis. To support this research, we present the newly introduced Parkinson's disease EEG (ParEEG) database, comprising 203,520 EEG samples from 60 subjects recorded based on Resting-State Visual Evoked Potential (RSVEP) and Steady-State Visually Evoked Potential (SSVEP) stimuli. In this study, we evaluate the performance of individual EEG channels using two handcrafted and two deep learning-based methods, employing a 10-fold cross-validation strategy, to ensure statistical reliability and establish benchmark results. Experimental results show that CRC and LSTM consistently achieved high accuracies (95-100%) with low variability (standard deviation < 2%). The analysis indicates that EEG channels in the frontal, fronto-central, and central-parietal regions consistently yield higher classification accuracy in Parkinson's disease detection. Our findings offer valuable insights into channel-specific neural alterations for better interpretability in PD, and the cross-stimulation evaluation enhances the generalizability of EEG-based PD detection for practical diagnostic purposes.