Motor imagery-based brain-computer interfaces (MI-BCIs) have significant potential for neurorehabilitation and motor recovery. However, most BCI systems employ multi-channel electroencephalogram (EEG) recording devices, during which the pre-experimental preparation and post-experimental hair cleaning are time-consuming and inconvenient for stroke patients, and potentially affect their motivation for rehabilitation training. In this paper, we introduced a wearable MI-BCI system for online MI classification using a wireless headband device with four EEG channels to reduce setup time while enhancing portability. To validate the performance of the system in decoding MI-EEG signals, extensive experiments and comparisons were performed on sixty-six healthy subjects. Specifically, an offline and an online experiment with forty-six subjects were conducted, with the system achieving average offline and online accuracies of 85.21% and 76.54%, respectively. Furthermore, a comparison experiment involving another twenty subjects showed that the online performance of our headband device (77.84%) was comparable to that of a mature commercial Neuroscan device (76.50%). Compared to several existing portable systems, our wearable system achieved superior performance with fewer channels and was validated on a larger number of subjects. These results demonstrated that our wearable BCI system can reduce preparation time, enhance portability, and meet the classification performance requirements for BCI-based rehabilitation intervention, indicating its substantial potential for large-scale clinical applications in enhancing motor recovery of stroke patients.

Brain-computer interfaces (BCIs) provide a means of translating neural activity into movement for stroke rehabilitation. Electroencephalography (EEG)-based motor imagery (MI) is a cognitive strategy to enhance motor recovery after stroke. However, traditional MI-BCI systems require extensive calibration before conducting online experiments, thus constraining their practicality. To enhance convenience, we propose a once-calibration strategy (ONCS) that allows each subject to perform only one calibration in continuous BCI interventions over one month. By using supervised and transfer learning to update the model with previous online data, repeated calibrations are eliminated. Furthermore, personalized channel selection (PCS) is designed to reduce the number of channels through the lowest event-related desynchronization (ERD). Compared to the traditional repeated calibration strategy (RECS), RECS for intra- and inter-subject models, the proposed ONCS for inter-subject (ONCS-inter) models achieves better classification performance using 28 channels. Wherein, the ONCS-inter shows statistically significant improvements ( p <0.05, one-tailed test). When using PCS for channel selection, ONCS-inter outperforms ONCS for intra-subject (ONCS-intra) ( p <0.01, for 16, 18, ..., 28 channels, two-tailed test) and surpasses RECS ( p <0.05 for all channels, two-tailed test). Remarkably, ONCS-inter exceeds the best results achieved with traditional RECS, even with only 2 channels. Extensive comparison and ablation studies demonstrate the effectiveness of our proposed ONCS combined with inter-subject models and a few channels in maintaining classification accuracy. The proposed ONCS with PCS holds promise for enhancing the convenience of continuous BCI interventions within one month for stroke patients.

Motor imagery brain-computer interfaces (MI-BCIs) remain challenging for practical use due to their reliance on multi-channel EEG devices and long calibration. To address these limitations, we proposed a wearable system for calibration-free online decoding using a lightweight, few-channel EEG headband, enabling portability, ease of use, and rapid setup. Specifically, we first built a large-scale wearable MI-EEG dataset from 100 healthy subjects to train a subject-independent model. We then developed a CNN-based temporal convolutional network (CTCNet) for online MI detection, which reduced computational complexity while maintaining high decoding performance. Furthermore, we introduced a supervised self-training (SST) strategy that leverages labeled online data and progressively fine-tunes a pre-trained subject-independent model, enabling calibration-free BCI operation without offline calibration. Four online experiments were conducted, involving 25 healthy subjects (Experiments II-IV) and 10 stroke patients (Experiment V). With the SST strategy, the accuracy of the subject-independent model improved from 69 % initially to 81 % after the first update and further increased to 86% after the second update, surpassing the subject-specific model (80%). Stroke patients exhibited a similar improvement trend. Moreover, simulated experiments confirmed the superiority of the subject-independent model compared to training from scratch. These findings demonstrate the effectiveness of the wearable MI-BCI system based on SST and CTCNet for online MI detection and highlight its substantial potential for motor recovery in stroke patients.

Brain–computer interfaces (BCIs) enable direct communication between the brain and external devices, providing a bio-inspired link between humans and artificial systems. However, electroencephalography (EEG)-based motor imagery (MI) decoding continues to pose challenges, due to limited temporal exploitation and insufficient spatio-temporal modeling. To alleviate these limitations, a complementary dual-branch receptive field network (DBRFNet) is proposed for cross-channel-density MI-EEG decoding. Specifically, two complementary branches with distinct temporal receptive fields are designed to capture multi-scale temporal dynamics. Each branch employs a progressive spatio-temporal module integrating temporal, depthwise spatial, and dilated convolutions for modeling short- and long-range temporal dependencies. The features from the two branches are fused using element-wise addition and further encoded for classification. The proposed method is evaluated on three public datasets representing low-, medium-, and high-density EEG configurations. DBRFNet achieves decoding accuracies of 80.48%, 84.37%, and 92.24% under cross-validation analysis, and 87.76%, 80.94%, and 94.64% under hold-out analysis, demonstrating competitive performance compared with state-of-the-art methods. Ablation studies further validate the effectiveness of each component within the proposed architecture. These results indicate that DBRFNet effectively learns discriminative spatio-temporal representations and provides a robust solution for MI-BCI systems.