A Hybrid Convolutional-Transformer Approach for Accurate Electroencephalography (EEG)-Based Parkinson's Disease Detection.

Abstract

Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by motor and cognitive impairments. Early detection is critical for effective intervention, but current diagnostic methods often lack accuracy and generalizability. Electroencephalography (EEG) offers a noninvasive means to monitor neural activity, revealing abnormal brain oscillations linked to PD pathology. However, deep learning models for EEG analysis frequently struggle to balance high accuracy with robust generalization across diverse patient populations. To overcome these challenges, this study proposes a convolutional transformer enhanced sequential model (CTESM), which integrates convolutional neural networks, transformer attention blocks, and long short-term memory layers to capture spatial, temporal, and sequential EEG features. Enhanced by biologically informed feature extraction techniques, including spectral power analysis, frequency band ratios, wavelet transforms, and statistical measures, the model was trained and evaluated on a publicly available EEG dataset comprising 31 participants (15 with PD and 16 healthy controls), recorded using 40 channels at a 500 Hz sampling rate. The CTESM achieved an exceptional classification accuracy of 99.7% and demonstrated strong generalization on independent test datasets. Rigorous evaluation across distinct training, validation, and testing phases confirmed the model's robustness, stability, and predictive precision. These results highlight the CTESM's potential for clinical deployment in early PD diagnosis, enabling timely therapeutic interventions and improved patient outcomes.