A Hybrid Covert Attention–Augmented Motor Imagery Paradigm for Brain–Computer Interfaces

Abstract

Motor imagery (MI) is widely used in brain–computer interfaces (BCIs) and has been applied in neurorehabilitation to support motor recovery in stroke patients. Although numerous decoding algorithms have been developed and achieve comparable performance, their effectiveness fundamentally depends on the availability of robust bio-markers elicited during MI. However, inexperienced users often evoke weak MI-related bio-markers, making reliable classification difficult even with advanced algorithms. Consequently, inaccurate classification lead to erroneous feedback to users, which further undermines the usability of MI. From the perspective of feature representation, augmenting MI with additional cognitive bio-markers can enrich neural features, improve feedback reliability, and facilitate more effective MI training and use. Therefore, a covert attention-augmented motor imagery (CAA-MI) paradigm is proposed as a hybrid BCI approach, which integrates covert spatial attention (CSA) with MI to introduce additional bio-markers and enrich the features used for classification. A transformer-based multi-branch EEG fusion network (TMEF-Net) is developed to fully leverage the diverse EEG features elicited by the proposed paradigm for decoding. Experiments involving 17 subjects demonstrate that CAA-MI consistently outperforms traditional MI (T-MI) in both intra-subject and inter-subject evaluations, particularly under short decoding windows. With a 3-s input, CAA-MI achieved intra-subject and inter-subject accuracies of 89% and 81%, outperforming T-MI (82% and 76%). Neurophysiological analysis further reveals that CAA-MI evokes broader activations across occipital–parietal and sensorimotor regions, providing more discriminative EEG features. These findings suggest that the proposed paradigm enables more robust neural responses and offers a promising strategy for neurorehabilitation-oriented BCIs.