Deep Learning based Classification of Motor Imagery Electroencephalography Signals

Brain-Computer Interface (BCI) is a technology that enables direct communication between the brain and external devices. BCI systems often use Electroencephalography (EEG) to measure the electrical fields produced by brain activities, serving as a prominent brain mapping and neuroimaging technique utilized extensively within and beyond clinical settings. Motor imagery (MI), a prevalent BCI paradigm, enables individuals, particularly those with disabilities, to regulate brain signals voluntarily, bypassing the need for external stimuli. By decoding MI-EEG signals, the gap between motor intention and sensory feedback in motor movements disrupted by brain disorders is bridged, thereby facilitating swift motor functional recovery. However, the non-linear and nonstationary nature of MI-EEG signals poses a challenge to MI intention recognition, leaving room for potential
classification enhancement. Moreover, factors such as subject variability, experimental conditions, and EEG recording devices impact the adaptability and robustness of models, consequently constraining the practicality of MI-EEG applications. The primary objective of this thesis is to investigate efficient deep learning models for decoding EEG signals and classifying MI tasks. The specific contributions of the thesis are outlined as follows:
1) A multi-view convolutional neural network (CNN) encoding approach for MI-EEG signals is proposed in Chapter 3. First, multiple frequency sub-band MI-EEG signals are created as the CNN model inputs through bandpass filters based on brain rhythms. Then, temporal and spatial features are captured based on the whole frequency band and the filtered sub-band signals, respectively. Further, utilizing two dense blocks with multi-CNN layers enhances model learning capabilities and strengthens information
propagation. The proposed method achieves an average accuracy of 75.16% on the public Korea University EEG dataset which consists of 54 healthy subjects for the two-class motor imagery tasks.
2) Chapter 4 introduces a local and global convolutional transformer-based MI-EEG classification model. To make up for the shortcomings of the CNN model, a local transformer encoder is employed to dynamically extract temporal features. The global transformer encoder and densely connected network are combined to improve the information flow and reuse. The spatial features from all channels and the difference in hemispheres are obtained to improve the robustness of the model. In the experiment, three scenarios including within-session, cross-session, and two-session are designed. Results show that the proposed model achieves up to 1.46%, 7.49%, and 7.46% accuracy improvement respectively in the three scenarios for the public Korean dataset compared with Tensor-CSPNet. For the BCI-IV-2a dataset, the proposed model also achieves a 2.12% and 2.21% improvement for the cross-session and two-session scenarios respectively.
3) Chapter 5 presents a cross-subject MI-EEG decoding method with domain generalization. In this study, the domain-invariant features from source subjects are extracted. The knowledge distillation framework is adopted to obtain the internally invariant representations based on spectral features fusion. Then the correlation alignment approach aligns the mutually invariant
representations between each pair of sub-source domains. In addition, we use distance regularization on two kinds of invariant features to enhance generalizable information. The results demonstrate that the proposed model achieves 8.93% and 4.4% accuracy improvements on the public Korean dataset and BCI-IV-2a dataset respectively compared with the ConvNet and Dyanmic
EEGInception model.
4) Chapter 6 proposes a graph convolutional network (GCN) based on transfer learning for cross-device MI-EEG decoding. Leveraging multi-channel information, the GCN module is employed to aggregate topological features. The pre-trained model is guided with few-channel signals as inputs through a knowledge distillation framework and adapted to the few-channel dataset
using a transfer learning strategy with minimal data training. Experimental results show that the proposed model achieved an accuracy of 71.19% based on across-dataset, 7.04% higher than filter bank common spatial pattern (FBCSP) and EEG-ARNN model, demonstrating the effectiveness of our approach in cross-dataset MI-EEG decoding and enhancing the practicality of
MI-BCI applications.
In summary, this study endeavors to decode MI-EEG signals using deep learning methods, improve the accuracy of motor intention recognition, and enhance the practicality of MI-based systems by enhancing model performance and robustness on cross-session, cross-subject, and cross-dataset scenarios.