Multi-Band GAN-Augmented Deep Learning for Enhanced P300 Detection in RSVP EEG Signals for Assistive-Tech

Brain-Computer Interfaces (BCIs) leveraging Rapid Serial Visual Presentation (RSVP)-base Electroencephalography (EEG) signals offer promising avenues for direct neural communication and control. However, achieving high classification accuracy across diverse subjects and the class imbalance between sparse target and abundant non-target are challenging issues due to inter-subject variability and the complex spatiotemporal nature of EEG signals. In this study, we propose a novel approach by combining Generative Adversarial Networks (GANs) and a multi-band deep learning architecture to address the issue of class imbalance between target and non-target EEG events in RSVP paradigms. The preprocessing involves ‘multi-band’ filtering, dimensionality reduction, and class-specific augmentation using GANs trained on different frequency bands. The deep learning model consists of band-specific convolutional and LSTM layers along with an attention mechanism for spatiotemporal feature fusion. The evaluation is done on the basis of multiple key performance metrics including accuracy, recall, precision, F1 score, ROC-AUC and unweighted average recall (UAR). The results demonstrate the efficacy of deep learning architectures in decoding RSVP based EEG signals, offering a robust framework for real-time BCIs with reduced calibration demands. The findings in this paper contribute to the development of a scalable and an adaptive BCI system and enhance its applicability in realtime scenarios.