Abstract
This investigation endeavours to formulate an automated classification system for identifying individuals with PD using electroencephalography recordings. An open-access dataset comprising EEG signals from 82 subjects, 41 Individuals clinically characterised by PD neurodegeneration and 41 healthy controls, was utilised. Spectral analysis of the EEG signals revealed a statistically significant reduction in power within the delta and theta frequency bands among PD patients. Three deep learning architectures were implemented and evaluated: a convolutional neural network, a hybrid model combining CNN with a long short-term memory layer, and a bidirectional long short-term memory method combined with a deep feedforward convolutional framework. The CNN-BiLSTM model achieved the highest classification performance, with an accuracy of 98.69% and an area under the receiver operating characteristic curve of 97.97%. In contrast, the CNN model showed limited discriminative capability, with an accuracy of 65.18% and an AUC of 50.02%. These findings underscore the importance of integrating both spatial and temporal information in EEG-based classification tasks. Additionally, spectral and statistical preprocessing contributed to improved pattern recognition. The model's stability and reproducibility were confirmed on independent test samples, indicating its potential utility for early-stage neurological identification of Parkinsonian symptomatology.