Adaptive filter of frequency bands based coordinate attention network for EEG-based motor imagery classification

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCI) enable device control via neural signals.
• Motor imagery (MI) classification faces challenges due to low signal-to-noise ratio, multiple channels, and non-linearity.
• Accurate decoding of MI is crucial for effective BCI applications.

Purpose of the Study:

• To improve motor imagery (MI) classification performance in brain-computer interfaces (BCI).
• To investigate the role of adaptive frequency band selection and spatial-temporal feature learning for decoding MI.
• To address challenges of low signal-to-noise ratio and non-linearity in Electroencephalogram (EEG) signals.

Main Methods:

• Proposed an Adaptive Filter of Frequency Bands based Coordinate Attention Network (AFFB-CAN).
• Developed an adaptive frequency band selection method to avoid manual limitations.
• Incorporated a Coordinate Attention Network (CAN) for emphasizing key brain regions and temporal segments, with a multi-scale module for enhanced temporal context learning.

Main Results:

• Achieved an average accuracy of 0.7825, kappa of 0.7104, and Macro F1-Score of 0.7486 on the BCI Competition IV-2a dataset.
• Obtained an average accuracy of 0.8879, kappa of 0.7427, and F1-Score of 0.8734 on the BCI Competition IV-2b dataset.
• Demonstrated significant improvements in MI classification performance using the AFFB-CAN method.

Conclusions:

• The AFFB-CAN method effectively enhances motor imagery classification accuracy.
• Confirmed the association of µ and β rhythms with motor imagery.
• Identified a significant role for γ rhythms in motor imagery classification.