Multi-channel EEG-based neurological disorder classification using Cross-Dependency Spatiotemporal Interactive...

Recently, the application of Transformers to Electroencephalogram (EEG)-based neurological disorder classification tasks has garnered significant attention. However, a notable limitation lies in the difficulty of explicitly capturing cross-dimension dependency interactions, which involves hierarchically encoding brain node state information and global adjacency associations across brain channels.

To address this challenge, we introduce a novel Cross-Dependency Spatiotemporal Interactive Network (CD-STIN) framework for EEG-based neurological disorder classification. Specifically, a temporal-wise Convolutional Neural Network (CNN) extractor is employed to encode local patterns and extract low-level features. Next, to account for the diverse connectivity patterns within brain regions, a tailored graph processing layer is utilized to manage the varying topological connections across different channels and spatially aggregate information. Following this spatial aggregation step, we leverage a Multi-head Self-Attention (MSA) layer to address temporal relationships within brain nodes, capturing long-range temporal dependencies by processing the time sequences of each channel. Subsequently, an aggregation module is employed to generate a refined representation of the input features through iterative aggregation of spatially and temporally connected components.

This enhanced representation is then fed into a classification head to produce the final result, which has obtained the best F1 of 98.54% and 98.84% on two available CHB-MIT and DEAP datasets, respectively.

Extensive experiments conducted on these EEG-based databases demonstrate the superiority and generalization capabilities of the proposed CD-STIN framework.