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

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Transformers show promise for Electroencephalogram (EEG)-based neurological disorder classification.
• A key limitation is the difficulty in capturing cross-dimension dependency interactions in EEG data.
• Hierarchical encoding of brain node states and global adjacency associations across channels is challenging.

Purpose of the Study:

• To introduce a novel Cross-Dependency Spatiotemporal Interactive Network (CD-STIN) framework.
• To enhance the classification of neurological disorders using EEG data by addressing cross-dimension dependencies.
• To improve the explicit capture of hierarchical brain node states and channel adjacency.

Main Methods:

• Employed a temporal-wise Convolutional Neural Network (CNN) for local feature extraction.
• Utilized a graph processing layer for spatial aggregation of channel information and topological connections.
• Applied a Multi-head Self-Attention (MSA) layer to capture long-range temporal dependencies across brain nodes.

Main Results:

• The CD-STIN framework achieved high F1 scores of 98.54% on the CHB-MIT dataset.
• The framework obtained an F1 score of 98.84% on the DEAP dataset.
• These results indicate superior performance in EEG-based neurological disorder classification.

Conclusions:

• The proposed CD-STIN framework demonstrates superior performance in EEG-based neurological disorder classification.
• Extensive experiments confirm the generalization capabilities of the CD-STIN framework across different datasets.
• The novel approach effectively addresses limitations in capturing cross-dimension dependencies in EEG signals.