Publication

Abstract : Graph signal processing (GSP), an emerging field, provides a flexible framework to model and analyze electroencephalogram (EEG) sensor data that exhibit intricate relationships and dependencies that traditional signal processing methods may not effectively capture. This study aims to develop a novel approach for converting single-channel EEG sensor data to graph signals and deriving GSP-based features that effectively combine functional and structural neurological information. We present a novel algorithm combining visibility graph (VG) algorithms and pooling techniques to generate graph signals from single-channel EEG data. Leveraging GSP techniques, we generate vertex-frequency representations (VFRs) from graph signals that act as optimal feature sets for Parkinson’s disease (PD) detection. We evaluate the performance of these newly derived GSP-based features on PD datasets of the University of New Mexico (UNM) and the University of Iowa (UI) using deep learning (DL) architectures, such as vision transformer (ViT) and ResNet-50. We introduce the WNVG-Max algorithm, a novel graph signal modeling approach optimally combining VG and pooling techniques that generate superior classification results. Furthermore, we benchmark the performance of non-GSP features based on classical signal processing (CSP) and network-based features on the same datasets using the aforementioned DL models and observe that the VFR of graph signals derived using the WNVG-Max algorithm outperformed the other techniques. In addition, our work suggests a set of optimized channels to aid in the detection of PD. Our novel methodology, integrating GSP with DL techniques, offers promising results with increased accuracy and facilitates early intervention strategies.