Work

Advancing motor imagery based brain-computer interfaces through deep learning

Year: 2026
Type: dissertation
Abstract: Brain-computer interface (BCI) technology, especially those based on Electroencephalography (EEG), has revolutionized communication and rehabilitation for individuals with severe spinal cord injuries and stroke. Among various non-invasive EEG signals, motor imagery (MI) based BCI is widely used by patients to communicate with the external world. Therefore, considering the potential of MI-based BCI, this thesis aims to improve the existing BCI technology to achieve better motor imagery classification performance. Recent advancements in deep learning, particularly with convolutional neural networks (CNNs), have shown significant promise in classifying MI tasks using EEG data. However, deploying CNN-based BCIs in real-world applications presents challenges, particularly in mobile and embedded environments constrained by computational and memory resources. Addressing these challenges is crucial for the practical implementation of BCI technologies. Firstly, this thesis proposes two novel methodologies to reduce the model complexity while maintaining high performance. The first approach involves using a magnitudebased weight pruning technique to reduce the number of parameters in pre-trained CNN models. This method successfully achieves up to 90% sparsity, offering a compression ratio of 4.77× without significantly compromising the classification accuracy, which remains at 84.44% with only a negligible loss of 0.02% compared to the baseline model. The second approach leverages Shapley values for the optimal selection of EEG channels, reducing the number of channels from 62 to 20 while maintaining a comparable classification accuracy of 84.07% (± 11.84), closely matching the baseline accuracy of 84.46% (± 11.56). This SHAP-based channel selection method not only enhances model efficiency but also provides better interpretability and precision in identifying the most critical EEG channels for MI tasks compared to other explainable artificial intelligence (XAI) methods. Moving ahead, MI-EEG signals are recorded when the subject performs an imagery motor task. Each subject has specific patterns because of the spatial origin, change in amplitude, and brain signals thus, a calibration process is necessary. In present BCI systems, calibration is often a tedious process that causes mental fatigue for the subject. The extended calibration time has thus become a significant drawback of BCI systems. To address this, the thesis introduces a genetic algorithm-based transfer learning approach to transfer knowledge from known subjects to new ones, reducing calibration time. This method significantly boosts the baseline model’s average classification accuracy from 84.46% to 87.29%. Importantly, it demonstrates substantial improvements in cases where other fine-tuning approaches perform poorly and offers deeper neurophysiological insights through interpretability analysis of the adapted models. Classification accuracy in MI-BCI systems remains a critical challenge due to the complex nature of brain signals. To enhance accuracy, this thesis proposes a novel EEG Channel-Wise Attention Module (ECWAM) within a deep Convolutional Neural Network (CNN). The ECWAM method calculates and amplifies the scores of significant mu band EEG channels, improving the baseline model’s average classification accuracy from 68.96% to 72.78%. A comparison of the EEG channel rankings reveals that ECWAM more effectively highlights channels in the motor cortex, the primary region associated with MI activity, than conventional methods. Further, to address the performance degradation caused by challenging or "hard" trials, the thesis presents two innovative strategies for their identification and mitigation. The first strategy uses model prediction scores to identify hard trials, while the second employs a quantitative XAI approach for a more transparent and interpretable process. Removing hard trials and retraining the model improves the average classification accuracy from 63.77% to 68.70%. The quantitative XAI method outperforms the prediction-score method, providing better insights and more effective identification of hard trials, as evidenced by scalp map analysis comparing correctly and misclassified trials. In summary, this thesis presents several novel methodologies to enhance MI-BCI systems, focusing on improving calibration efficiency, classification accuracy, and model complexity reduction, thereby making them more practical for deployment in real-world scenarios.
Language: English
Open Access status: green
OpenAlex ID: W7127042172
Cites:
Cited by:
Related to: 10