EEGProgress: A fast and lightweight progressive convolution architecture for EEG classification.

Area of Science:

• Neuroscience and Artificial Intelligence
• Computational Neuroscience
• Machine Learning for Brain Signal Analysis

Background:

• Extracting deep spatial features from electroencephalograph (EEG) signals is complex due to the brain's intricate topology.
• Effective topological spatial information extraction is vital for accurate EEG classification.
• Convolutional neural network (CNN) architectures significantly impact the performance and complexity of EEG classification.

Purpose of the Study:

• To propose a progressive convolution CNN architecture, EEGProgress, for efficient extraction of topological spatial information from EEG signals.
• To enable multi-scale feature extraction (electrode, brain region, hemisphere, global) with improved speed.
• To validate the performance and effectiveness of the proposed EEGProgress and topological permutation method.

Main Methods:

• Developed the EEGProgress CNN architecture with a progressive feature extractor.
• Employed an empirical topological permutation rule to integrate EEG data with topological properties.
• Utilized prior, electrode, region, and hemisphere convolution blocks for progressive spatial feature extraction.

Main Results:

• EEGProgress demonstrated superior feature extraction capabilities, achieving an average accuracy increase of 4.02% over other CNN models.
• The model performed effectively in both cross-subject and within-subject EEG classification scenarios.
• EEGProgress showed improved model complexity, outperforming comparison models in testing time, FLOPs, and parameter count.

Conclusions:

• The proposed EEGProgress architecture significantly enhances the extraction of topological spatial information from EEG signals.
• Topological permutation effectively integrates spatial properties, improving classification performance.
• EEGProgress offers a computationally efficient and accurate solution for EEG classification tasks.