SincNet-Based Hybrid Neural Network for Motor Imagery EEG Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI)-based brain-computer interfaces (BCIs) commonly use the Common Spatial Pattern (CSP) method.
• Optimal filter cut-off frequencies for CSP in MI-BCIs are challenging to determine, often leading to suboptimal spectral information utilization from electroencephalography (EEG).

Purpose of the Study:

• To propose a SincNet-based hybrid neural network (SHNN) to automatically optimize filter parameters for MI-based BCIs.
• To enhance the utilization of MI-related spectral information in EEG signals.

Main Methods:

• Raw EEG data were segmented and mapped to the CSP feature space.
• SincNets were employed as filter bank band-pass filters for automatic data filtering.
• Squeeze-and-excitation modules were used for sparse representation learning, followed by convolutional neural networks (CNNs) for deep feature extraction.
• Gated recurrent units (GRUs) were utilized to capture sequential dependencies, and a fully connected layer performed final classification.

Main Results:

• The SHNN method achieved mean classification accuracies of 0.7426 (kappa: 0.6648) on dataset 2a and 0.8349 (kappa: 0.6697) on dataset 2b from the BCI competition IV.
• Statistical analysis confirmed that the SHNN method significantly outperformed existing state-of-the-art methods on the tested datasets.

Conclusions:

• The proposed SHNN method effectively addresses the challenge of suboptimal filter selection in MI-BCIs.
• SHNN demonstrates superior performance in utilizing EEG spectral information for improved MI-based BCI accuracy.