A novel hybrid deep learning scheme for four-class motor imagery classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Classifying motor imagery electroencephalogram (MI-EEG) signals is crucial for brain-computer interfaces (BCIs).
• Achieving high classification accuracy with multiple classes and inter-subject variability remains a significant challenge.
• Understanding complex correlations within MI-EEG signals is key to improving performance.

Purpose of the Study:

• To develop an end-to-end deep learning framework for accurate classification of four-class MI-EEG tasks.
• To address the challenge of inter-subject variability by proposing a subject-independent shared neural network.
• To enhance the extraction and learning of spatial and temporal features from MI-EEG data.

Main Methods:

• A hybrid deep learning approach combining convolutional neural networks (CNNs) and long-term short-term memory (LSTM) networks.
• Preprocessing and feature extraction using a one-versus-rest filter bank common spatial pattern (FBCSP) method.
• Training a subject-independent shared network using data from all subjects to create a generalized model.

Main Results:

• The proposed hybrid deep learning framework achieved an accuracy of 83% and a Cohen's kappa value of 0.80 on the BCI competition IV dataset 2a.
• The subject-independent shared neural network demonstrated satisfactory classification accuracy when evaluated individually for each subject.
• The model effectively learns both spatial and temporal features simultaneously for MI-EEG decoding.

Conclusions:

• The developed hybrid deep learning scheme significantly improves the classification accuracy of four-class MI-EEG signals.
• The subject-independent nature of the shared network makes it robust to inter-subject variability, suitable for real-world applications.
• This approach holds considerable promise for advancing the development of practical and effective BCIs.