Improving EEG-Based Motor Imagery Classification via Spatial and Temporal Recurrent Neural Networks.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) Brain-Computer Interfaces (BCIs) are crucial for assisting severely disabled individuals.
• Traditional MI-BCI methods often rely on handcrafted EEG features, limiting end-to-end performance.
• Convolutional Neural Networks (CNNs) capture spatial EEG information, while Recurrent Neural Networks (RNNs) excel at temporal data processing.

Purpose of the Study:

• To develop a pure RNNs-based parallel method for simultaneously encoding spatial and temporal raw EEG data.
• To improve the accuracy and efficiency of movement intention recognition in MI-BCI systems.
• To offer an alternative to traditional feature extraction and CNN-based deep learning approaches.

Main Methods:

• A novel parallel architecture using bidirectional Long Short-Term Memory (bi-LSTM) and standard LSTM was proposed.
• EEG electrode indices were rearranged based on spatial relationships.
• A sliding window technique was applied to raw EEG data for sample augmentation, followed by simultaneous spatial and temporal encoding.

Main Results:

• The proposed pure RNNs-based method achieved an average accuracy of 68.20% in multi-class, trial-wise movement intention classification.
• The method significantly outperformed traditional (CSP+LDA, FBCSP+LDA) and a hybrid (CNN-RNN) approach, showing an 8.25% relative accuracy improvement.
• Experimental validation was conducted on the public MI-based eegmmidb dataset.

Conclusions:

• The pure RNNs-based parallel method demonstrates superior performance for MI-BCI.
• This approach offers a feasible and effective solution for real-world BCI systems.
• Simultaneous encoding of spatial and temporal EEG data using RNNs is a promising direction for BCI research.