A Parallel Feature Fusion Network Combining GRU and CNN for Motor Imagery EEG Decoding

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning for motor imagery (MI) electroencephalography (EEG) decoding is promising for brain-computer interfaces (BCI).
• Existing methods are mature for two-class MI tasks, but four-class MI decoding requires further exploration.
• Limited EEG dataset sizes pose challenges like overfitting and poor robustness for deep learning models.

Purpose of the Study:

• To develop a data augmentation method for MI-EEG datasets.
• To construct a robust deep learning model for decoding four-class MI tasks.
• To improve the performance of BCI systems using limited EEG data.

Main Methods:

• A novel data augmentation technique involving sliding and reconstructing EEG data along the time axis.
• A parallel-structured feature fusion network combining gated recurrent unit (GRU) and convolutional neural network (CNN).
• Decoding four-class MI tasks using the proposed model on the BCI Competition IV 2a dataset.

Main Results:

• Achieved a global average classification accuracy of 80.7% and a kappa value of 0.74 on the BCI Competition IV 2a dataset.
• The proposed method demonstrated improved robustness for deep learning on small-scale EEG datasets.
• Effectively alleviated overfitting issues caused by insufficient training data.

Conclusions:

• The developed data augmentation and deep learning approach enhances MI-EEG decoding for BCI.
• The method is suitable for BCI applications requiring decoding from small, daily recorded EEG datasets.
• Addresses the critical need for robust decoding in multi-class MI tasks with limited data.