Xinjie He

Objective: Inter-session and inter-subject variability in electroencephalography (EEG) signals, resulting from individual differences and environmental factors, poses a major challenge for neural decoding in brain-computer interface (BCI) applications. Methods: To address this issue, we propose RUNet, a zero-calibration motor imagery EEG decoding framework based on Riemannian manifold learning and unsupervised representation learning. RUNet incorporates a multi-scale spatiotemporal convolutional module that jointly captures local global spatial and multi-resolution temporal dynamics features. To enhance the robustness of EEG features against non stationarity, a polysynergistic covariance optimization module is employed, which strengthens the covariance matrix representation through multiple regularizations and adaptive fusion. In addition, RUNet integrates the Riemannian Affine Log Mapping layer, based on Affine-Invariant Transformation and Log-Euclidean Mapping, in an end-to-end manner to mitigate cross-domain covariance drift and enhance domain-invariant feature learning. A transfer learning framework is further integrated into RUNet: during pre-training, an unsupervised contrastive loss is applied to resting-state EEG data to learn domain-invariant spatiotemporal features; during retraining, task-specific data are used to enhance discriminability and feature disentanglement. Conclusion: Experimental results on the BCI Competition IV 2a, 2b datasets and a self-collected laboratory dataset show that RUNet achieves average cross-session accuracies of 87.19%, 88.03% and 85.45%, and cross-subject accuracies of 68.09%, 78.29% and 87.25%, respectively. On the PhysioNet dataset, a cross-subject accuracy of 78.14% is achieved. These results demonstrate the effectiveness of RUNet's unified pipeline and its robust cross-domain generalization.

Lightweight networks that include depthwise-separable convolution are widely used in motor imagery (MI) electroencephalogram (EEG) decoding of brain–computer interface (BCI). Many established MI classification networks are relatively shallow, preventing them from benefiting from the hierarchical feature extraction capabilities of deeper structures. Due to suboptimal residual connection structures, the mismatched residual baseline layer design, and the poor compatibility between data preprocessing and residual modules, the deepening of networks cannot be effectively combined with residual structures. This creates a depth barrier that hinders further performance improvements. To address these challenges, we propose a novel method, residual depthwise-separable deep neural network (ResDSNet), built upon an unraveled view-path analysis of residual connection structures. The analysis reveals that the residual mechanism achieves optimal performance when the layer distribution across different paths approximates a binomial distribution. Furthermore, we design a residual depthwise-separable convolution module and a tailored data-preprocessing module that effectively integrate with the residual structure, filtering noise and retaining MI task features. We evaluate ResDSNet on three publicly available datasets, including the BCI Competition IV Dataset IIa, the BCI Competition IV Dataset IIb, and the PhysioNet dataset, which collectively contain EEG signals recorded from 127 human subjects. ResDSNet achieves accuracies of 79.36%, 84.95%, and 64.13%, outperforming state-of-the-art methods by 3.16%, 1.59%, and 8.40% with statistical significance. Experimental results indicate that ResDSNet fully unlocks the hierarchical representation capabilities of deep networks for MI-EEG decoding, achieving robust performance and demonstrating substantial potential to overcome the inherent challenges in BCIs.

Motor imagery (MI) is a popular noninvasive brain computer interface (BCI) paradigm, yet its decoding accuracy remains hindered by the inherent nonstationarity and low signal-to-noise ratio of electroencephalogram (EEG) signals. Current decoding frameworks often fail to fully exploit the intricate spatial-temporal dependencies, leading to suboptimal feature representation and the omission of latent discriminative cues. To address these challenges, we introduce a deep neural network-powered multifaceted strategy (DPMS-Net) model, a novel approach that employs dynamic convolution to unearth effective discriminative cues across multiple dimensions, including the temporal, spatial, and frequency domains. This model synergizes channel and temporal attention mechanisms to adeptly capture the salient features of EEG signals across diverse spatial-temporal dimensions, thereby mitigating the risk of omitting critical information. Furthermore, we introduce a spectral-domain analysis component that unearths subtle oscillatory signatures hidden within the EEG spectrum, providing enriched evidence for classification. We evaluated the performance of DPMS-Net on two publicly available datasets and a self-collected dataset from stroke patients. On the BCI Competition IV 2a and BCI Competition IV 2b datasets, DPMS-Net achieved subject-dependent classification accuracies of 83.93% and 88.38%, respectively, alongside subject-independent classification accuracies of 65.88% and 76.01%. In the stroke patient dataset, DPMS-Net attained a subject-dependent classification accuracy of 67.67% and a subject-independent classification accuracy of 57.58%. Experimental results indicate that DPMS-Net possesses efficient decoding capabilities and robust stability, reflecting its potential for deployment in neurorehabilitation BCI systems.

Motor imagery (MI) is one of the most widely used paradigms in brain–computer interfaces (BCIs), known for its ability to trigger changes in brain activity without the need for an external “cue” stimulus. This unique characteristic has attracted significant attention from neuroscientists and researchers in fundamental science. However, compared to $P300$ and steady-state visual evoked potential (SSVEP), neural activity related to MI tends to be less stable and exhibits substantial variability between individuals. Consequently, accurately decoding MI, using both traditional machine learning and deep learning, has proven to be a considerable challenge. Moreover, given the difficulty of acquiring electroencephalography (EEG) data and the high data demands of deep learning, enhancing the accuracy of MI decoding with limited sample sizes remains a pressing issue that urgently needs to be addressed. This article addresses the challenges mentioned above by introducing a novel deep neural network designed for accurate MI decoding, which is designed to be effective with both small-sample sizes and larger datasets. This network, named the multiscale pooling spatial–temporal attention network (MPSTANet), integrates mix pooling techniques with spatial–temporal attention mechanisms. MPSTANet first employs local and global spatial attention, along with multiscale temporal attention, to thoroughly extract spatial–temporal information from EEG signals. Next, MPSTANet utilizes feature fusion and the proposed mix pooling technique to preserve as much of the extracted spatial–temporal information as possible. Finally, channel interaction attention (CIA) and 3-D weight attention (3-DWA) are employed to recalibrate the weights of the fused channels and spatial–temporal features, respectively. To validate the performance of our proposed MPSTANet model, we conducted experiments on four public datasets, including both small-sample sizes and subject-independent scenarios. MPSTANet achieved cross-session decoding accuracies of 84.82%, 72.92%, 88.20%, and 46.54% on the BCI Competition IV 2a dataset, the Open BMI dataset, the BCI Competition IV 2b dataset, and the PhysioNet dataset, respectively. Furthermore, MPSTANet demonstrated a significant lead compared to other deep learning models in both small-sample and subject-independent experiments. These results demonstrate the robustness of MPSTANet in MI decoding and its promising potential for BCI applications.

Steady-state visual evoked potentials (SSVEPs) are widely used in Brain–computer interfaces (BCIs) due to their stability and high signal-to-noise ratio. For Internet of Things (IoT) applications, however, achieving plug-and-play usability without subject-specific calibration remains a critical challenge. To address this, we propose FlexFusionNet, a deep learning framework composed of four main modules: the inception block for multiscale temporal feature extraction, the spatio-temporal feature extractor (STFE)—which integrates the hybrid temporal-spatial convolution block (HTSCB) and the multiscale fusion residual block (MFRB)—for capturing local interactions and preserving global dependencies, and the temporal refinement module (TRM) for fine-grained temporal enhancement. Together, these modules ensure robust subject-invariant feature learning. The core lies in MFRB, which preserves global dependencies while integrating multiscale temporal features, and in the overall integration strategy that ensures robustness across both intra- and cross-subject scenarios. The framework is validated on two public datasets (Benchmark and BETA) under both cross-subject and subject-specific settings. In cross-subject experiments with a 1.2 s window, it achieves 74.57% accuracy and 122.28 bits/min ITR on Benchmark, and 62.60% accuracy and 89.86 bits/min on BETA, substantially outperforming classical and deep learning baselines. In particular, the framework achieves outstanding performance in both scenarios, demonstrating its strong generalization ability. This study not only highlights the technical potential of SSVEP signal classification in IoT systems but also provides a theoretical and practical foundation for developing efficient, robust, and user-friendly BCI systems that can be easily integrated into diverse IoT ecosystems.

High-frequency steady-state visual evoked potential-based brain-computer interface (SSVEP-BCI) systems offer improved user comfort but suffer from reduced performance compared to their low-frequency counterparts, limiting their practical application. To address this issue, we propose a transfer learning-based method that leverages low-frequency SSVEP data to enhance high-frequency SSVEP performance. A filtering mechanism is designed to extract informative components from low-frequency signals, and the least squares algorithm is employed to generate high-quality synthetic high-frequency data. Experiments conducted on two public datasets using TDCA, eTRCA, and advanced TRCA-based algorithms demonstrate significant performance improvements. Our approach requires only two calibration trials, achieving 9.03% and 14.49% accuracy increases for eTRCA and TDCA in Dataset 1, and 13.91% and 14.53% improvements in Dataset 2, all within 1.5 s. Moreover, our approach effectively addresses the issue of single calibration data for high-frequency SSVEP-BCI systems. These results support the feasibility of fast calibration and improved performance in real-world high-frequency BCI applications.

Brain–computer interfaces (BCIs) provide a direct communication pathway between the central nervous system and external environments, enabling human–machine interaction control. Among them, event-related potential (ERP)-based BCIs are among the most accurate and reliable BCI systems. However, current mainstream classification algorithms struggle to eliminate calibration requirements and rely heavily on costly labeled data, limiting the practical usability of ERP-based BCIs. To address this, the development of unsupervised algorithms is critical for advancing real-world BCI applications. In this study, we propose the spatio-temporal equalization sliding-window distribution distance maximization (STE-sDDM) algorithm, which introduces spatio-temporal equalization (STE) to unsupervised ERP classification for the first time and integrates it with a novel unsupervised classification method, sliding-window distribution distance maximization (sDDM). STE estimates and removes colored noise interference in background noise to enhance the signal-to-noise ratio of inputs for sDDM. Meanwhile, sDDM leverages an enhanced inter-class divergence metric based on the ergodic hypothesis theory, utilizing sliding windows to emphasize temporally discriminative features, thereby improving unsupervised classification accuracy. The experimental results demonstrate that the integration of STE and sDDM significantly enhances ERP feature separability, outperforming state-of-the-art unsupervised online classification algorithms in spelling accuracy and the information transfer rate (ITR), facilitating more accurate and faster plug-and-play real-time control for BCI systems. Additionally, static spatio-temporal equalizer architectures were found to outperform dynamic architectures when combined with this framework.

Brain-Computer interface (BCI) technology enables the control of external devices by recognizing user intentions. Steady-state visual evoked potential (SSVEP)-based BCI technology has been widely applied in the field of Internet of things (IoT) device control, including smart healthcare, smart homes, and robotics, and has achieved significant results. However, as the field of BCI-based IoT device control is still in its development stage, there remains considerable room for improvement in terms of accuracy, efficiency, and cost. Therefore, enhancing the classification accuracy of SSVEP decoding using a short time window, reducing both human and material costs, and improving work efficiency are crucial for the theoretical research and engineering applications of BCI technology in IoT device control. Based on this, we propose a novel approach to address the challenge of high accuracy feature extraction within brief timeframes. Our approach integrates a multi-scale convolutional neural network with a squeeze excitation module (SEMSCNN). This fusion leverages CNNs’ local feature learning capacity and the advantageous feature importance distinction offered by the squeeze excitation mechanism. First, the EEG signals are band-pass filtered into distinct frequency bands and frequency band and channel features are extracted by a two-layer convolution. Then, temporal features are extracted via a multi-branch convolution of different scales. Finally, the squeeze and excitation (SE) module is introduced to learn the interdependence between features to improve the quality of the extracted features. The first stage of training exploits statistical commonalities across research participants by learning the global model, and the second stage fine-tunes each participants features separately by exploiting participant-specific differences in features. We evaluate our SEMSCNN model on two large public datasets, Benchmark and BETA, and we compare our model to other state-of-the-art models in order to evaluate the effectiveness of our proposed network.Our experimental results indicate that our method effectively improves the accuracy of target recognition and information transfer rate under short-duration stimuli, showing a significant advantage compared to other baseline methods. This provides a broad prospect for the practical application of BCIs in the field of IoT.

The decoding approach of steady-state visual evoked potentials (SSVEP) based on supervised learning have achieved remarkable results. However, these approaches require extensive calibration efforts to train the mode parameters for each stimulus. To facilitate the calibration process, we proposed a cross-stimulus transfer learning framework using the common periodic repetition components (CSTLF-CPRC) in fast calibration scenario. Firstly, a source stimulus mode was constructed, which can use periodic repetition components to obtain a source synthetic SSVEP template and source ensemble spatial filter. Secondly, leveraging the common information between period repeated component templates across multi-stimulus periods, the common source aliasing matrix was further estimated. Finally, leveraging the commonality between target and source stimuli, a cross-stimulus transfer learning mode was constructed for SSVEP cross-stimulus recognition. Offline tests on public datasets show that the CSTLF-CPRC outperforms the state-of-the-art (SOTA) method such as FBCCA, tlCCA, and CIRCSTL in a fast calibration scenario. Our method only needs 16s to calibrate 40 targets on two public datasets and achieves an average ITR of 227.86 1 106.47 bit/min and 162.41 1 124.29 bit/min, respectively. The study has the requirement of a few calibration data to achieve high performance recognition and to promote effective development of the practical system.