Deep learning-based EEG motor imagery signal classification for brain–computer interface applications

Brain-Computer Interfaces (BCIs) are new technologies that allow people to communicate directly with devices using their brains. It is beneficial for those who have difficulty moving. There are two main types of BCI that are invasive, which involve surgery, and noninvasive systems. Invasive brain-computer interfaces (BCIs) can give obvious signals, but come with risks from surgery, making them more difficult to use in hospitals. On the other hand, noninvasive systems like EEG-based BCIs are safer and easier for real-time neural interfacing. This study presents a robust framework for classifying EEG signals using advanced deep learning models. The goal is to improve the control of assistive devices during motor imagery (MI) tasks. Four different models are tested: Long-Short-Term Memory (LSTM) networks have the highest accuracy at 96.51%, Gated Recurrent Units (GRU) followed by 96. 31%, Convolutional Neural Networks (CNN) achieved 94.42%, and Graph Isomorphism Networks (GIN) constitute a novel graph-based approach with 92.00% accuracy. The results demonstrate that advanced deep learning architectures can effectively decode motor imagery-related brain activity and provide a reliable foundation for real-time brain–computer interface (BCI) applications. Unlike earlier studies that looked at models separately, our research compares them all using the same processing and evaluation methods. This work allows for a fair comparison of how well each model performs. Furthermore, to address the limitations of static graph-based modeling, this study introduces a dynamic graph attention-based extension that captures time-varying inter-channel connectivity, leading to improved EEG motor imagery classification performance. To strengthen the experimental evaluation, additional benchmark methods, including EEGNet, Common Spatial Pattern with Support Vector Machine (CSP + SVM), and Filter Bank Common Spatial Pattern with Support Vector Machine (FBCSP + SVM), were implemented using the same preprocessing and training protocol. All models were evaluated over five independent runs, and statistical significance was examined using the Wilcoxon signed-rank test, which confirmed that the proposed Dynamic Graph Attention Network (DGAT) consistently outperformed most of the benchmark methods. This work presents a unified comparative study of recurrent, convolutional, and graph-based learning models for motor imagery EEG classification under a consistent preprocessing and evaluation framework. In addition, a Dynamic Graph Attention Network (DGAT) is introduced to capture time-varying relationships among EEG channels and to improve classification performance, providing vital information for future hybrid adaptive MI-BCI systems.