A Survey of Methods and Architectures for Parkinson's Disease Recognition via Heterogeneous Split Attention-Based EEG and Siamese Graph Convolutional Attention Network

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that significantly affects motor and cognitive functions. Early and accurate diagnosis is essential for improving treatment outcomes, yet conventional diagnostic methods rely heavily on subjective clinical evaluations. Electroencephalography (EEG) has emerged as a promising non-invasive tool for PD detection due to its ability to capture neural activity with high temporal resolution. Recent advances in artificial intelligence (AI), particularly deep learning and graph-based models, have enabled automated PD recognition from EEG signals. Convolutional neural networks (CNNs) and hybrid CNN–LSTM architectures have demonstrated strong performance by capturing spatial and temporal dependencies in EEG data. Furthermore, graph convolutional networks (GCNs) have been widely adopted to model functional connectivity between EEG channels, improving classification accuracy. Emerging architectures such as heterogeneous split-attention mechanisms and Siamese graph convolutional attention networks further enhance feature extraction by focusing on relevant channels and learning inter-sample similarities. Attention-based sparse GCN models have shown improved performance by capturing channel relationships and emphasizing important EEG regions. Despite these advancements, challenges such as EEG noise, data scarcity, and computational complexity remain. This survey reviews recent developments, compares methodologies, and highlights future directions for robust and scalable PD recognition systems.