Continual-Learning-Enhanced CNN-Transformer Framework for Real-Time Motor-Imagery BCI in Virtual Environments

Motor imagery (MI)-based brain-computer interfaces (BCIs) provide an intuitive pathway for neural interaction and rehabilitation, yet their practical deployment remains constrained by long calibration requirements, substantial inter-subject variability, and the non-stationary nature of EEG signals. These challenges are amplified when using dry-electrode EEG, which offers superior convenience for real-world systems but produces noisier and less stable recordings than traditional wet electrodes. As a result, online or real-time four-class MI detection-especially with dry electrodes-has been explored only in a limited number of studies, underscoring an important gap in the field and the need for adaptive, intelligent models capable of coping with continuous signal drift. In this study, we propose a real-time MI-BCI framework that integrates immersive action observation (AO) in virtual reality with a continual learning strategy to manage the evolving nature of dry-EEG features. A CNN-Transformer hybrid model is first initialized through AO-enhanced pre-training and subsequently refined via online continual adaptation during user interaction. This continual learning mechanism enables the classifier to incrementally assimilate new MI patterns while preserving previously acquired knowledge, thereby mitigating the performance degradation that typically arises in extended MI-BCI sessions. Experimental results across four motor classes demonstrate improved decoding accuracy and strengthened sensorimotor activation over time, confirming the system's capacity for user-specific and session-to-session adaptation. By addressing the rarely studied combination of dry electrodes, online four-class MI decoding, and continual learning, the proposed approach enhances MI-BCI robustness, reduces calibration burden, and supports sustainable long-term deployment in intelligent neurotechnology applications.