Deep Learning Approaches for Subject-Independent EEG-Based Brain-Computer Interfaces

Abstract

Electroencephalography (EEG)-based brain-computer interfaces (BCIs) offer a noninvasive and promising pathway for facilitating direct communication between the human brain and external devices. However, traditional BCI systems often require subject-specific calibration, limiting their scalability and usability in real-world scenarios. This thesis addresses this limitation by developing and evaluating deep learning approaches for subject-independent EEG decoding, with a particular focus on two primary BCI paradigms: event-related potential (P300) classification and emotion recognition. The research presents a comprehensive investigation into how state-of-the-art deep learning architectures, including convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and Transformer-based models, can be designed and optimized to enhance cross-subject generalization in EEG decoding tasks. Central to this work is the development of a series of hybrid models that integrate spatial, temporal, and frequency-domain feature representations using attention mechanisms and lightweight network designs. A systematic review of existing deep learning-based EEG classification techniques was first conducted to identify gaps in subject-independent performance. Then, several model architectures were proposed and evaluated on publicly available datasets, including BCI Competition datasets, BNCI Horizon 2020, EPFL, ALS, and SEED-IV. A novel architecture combining CNN, LSTM, and a support vector machine (SVM)-enhanced attention module was developed for P300 classification, demonstrating improved interpretability and robustness across participants. Additionally, a residual CNN and Transformer hybrid model was proposed for emotion classification using the SEED-IV dataset, achieving state-of-the-art performance with an accuracy of 95.5. The experimental methodology involved subject-wise data division using leave-one-subject-out (LOSO) cross-validation to simulate true generalization scenarios. The findings show that incorporating domain-agnostic architectural elements such as multi-head self-attention, residual connections, and adaptive dropout strategies significantly improves classification consistency across subjects. The work further explores training dynamics, optimization techniques (e.g., AdamW), and regularization methods that contribute to performance gains without requiring subject-specific tuning.