EEG-Based Early Detection of Parkinson’s Disease Using Deep Learning

Abstract

This thesis introduces a unique approach using neural networks to enhance Parkinson’s disease (PD) classification and predict its severity through novel analyses of the Default Mode Network (DMN) via electroencephalogram (EEG) data. We employed a 2D convolutional neural network (CNN) such as EEGNet, that was enhanced through our custom modification. In addition to it, we performed connectivity analysis through measures such as mean squared coherence and partial directed coherence in the DMN area. Our results demonstrate the enhanced capability of the modified EEGNet, which significantly outperforms existing models in both classification accuracy and predictive validity. Specifically, we also applied GRAD-CAM in conjunction with EEGNet on EEG signals, highlighting critical time points and signal regions vital for accurate PD classification. Our connectivity analysis has successfully delineated the direction of disrupted connectivity within the DMN, offering a ground-breaking insight into the neural disruptions in the DMN region of PD patients. By pinpointing specific time point, signal regions and connectivity disruptions, our method provides a robust framework for early detection and offers a substantial improvement over traditional diagnostic tool. It advances our understanding of PD’s neural mechanisms and opens avenues for the development of non-invasive, precise diagnostic techniques utilising results and changes observed during this thesis