A Hierarchical Diffusion-Convolutional Network with Node-wise Localization for EEG-NIRS-based Brain-Computer Interface

In this work, we propose a lightweight Hierarchical Node-wise Localized Diffusion-Convolutional Network (HNLDCNet) for motor imagery (MI) and mental arithmetic (MA) classification tasks based on EEG-fNIRS data. The proposed HNLDCNet utilizes a layer-adaptive agglomerative clustering algorithm to construct a graph hierarchy of spatial information from EEG-fNIRS channels, enhancing the spatial feature extraction of EEG-fNIRS signals. By incorporating the philosophy of node-wise localized feature mapping and DiffPooling, we employ a learnable directed graph structure for efficient message passing between nodes. Different from undirected graph neural networks, HNLDCNet captures the effective connectivity, improving the extraction of discriminative information. In subject-dependent experiments, HNLDCNet achieves mean accuracies of 98.87% and 99.11% for MI and MA, respectively. Additionally, we provide visualizations to enhance the interpretability of the proposed model.