BACKGROUND:Motor imagery (MI) brain-computer interfaces (BCI) rely on precise electroencephalogram (EEG) classification. However, issues such as the reliance on extensive manual experience for MI-EEG model design, parameter tuning, and optimization directions, along with the poor task flexibility of foundation models and state degradation during long-term multi-agent iterations, severely restrict the state-of-the-art (SOTA) efficiency of MI-EEG. NEW METHOD:To address these challenges, we propose AutoMI, a novel framework that uses multi-agent automated rapid iterations to construct SOTA MI-EEG models. AutoMI introduces a hybrid decision mechanism that tightly couples Q-learning strategies with deterministic rules. By integrating planning, execution, and output agents with predefined tools, AutoMI ensures broad general applicability across various hyperparameter optimizations and structural improvements. Furthermore, AutoMI integrates experience tracking and rollback mechanisms to prevent ambiguous optimization. RESULTS:In evaluations on the IV2a, OpenBMI, and ECUST-MI datasets, the SOTA models finally constructed through AutoMI iterations achieve accuracies of 77.62%, 78.08%, and 83.02%, with maximum improvement reaching 24.69%, 23.35%, and 23.28% respectively. Furthermore, the average time per iteration for a single subject on the OpenBMI dataset is approximately 500 s. COMPARISON WITH EXISTING METHODS:Compared with automated optimization algorithms, the accuracies increase by 18.42%, 9.27%, and 19.25% respectively, demonstrating the effectiveness of the proposed AutoMI framework and proving that its optimization capability reaches SOTA. CONCLUSION:Experimental results indicate that AutoMI provides a novel perspective and framework design reference for future BCI model optimization.

Brain–computer interface (BCI) technology, which controls external devices by directly decoding brain activities, has made important progress and practical applications in recent years in many fields. However, the domain bias issue in cross-domain applications remains a significant challenge in the practical implementation of BCI technology. This is particularly acute in scenarios where target data are unavailable, largely because of the noise sensitivity and acquisition limitations inherent in electroencephalography (EEG) signal data. When processing nonstationary EEG signals, existing domain generalization methods face limitations: Adversarial training may compromise model stability, while global feature alignment approaches struggle to sufficiently decouple category-dependent and category-independent features, thereby constraining generalization performance. Therefore, in this paper, we propose a hybrid approach based on domain-invariant feature learning and data enhancement. We introduce a “fixed” structure enhancement method that combines domain-invariant feature learning with data enhancement strategies, decouples domain-invariant features from other features, optimizes cross-domain feature extraction, and reduces the effect of noise in data. Through extensive experimental validation on multiple publicly available datasets, the model proposed in this paper outperforms the existing state-of-the-art methods, providing a novel and effective solution to the domain bias problem in BCI.

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.

Training-based algorithms significantly outperform training-free methods in terms of recognition performance for steady-state visual-evoked potential (SSVEP)-based brain-computer Interfaces (BCIs). However, collecting training data requires calibration experiments that are effort-intensive and often costly. These calibration demands limit the practicality of BCI, as users (and even system operators) may experience fatigue or lose interest in continued use. Transfer learning (TL) offers an effective solution, but it typically relies on either a certain amount of target domain data or extensive source domain data. To address this limitation, we introduce the concept of cross-dataset TL in SSVEP for the first time to extract transfer knowledge from other datasets. During this process, we identified a data mismatch problem that severely compromises the generalizability of transfer knowledge. To overcome this challenge, we propose a TL-SSVEP decoding algorithm calibrated with single-trial data (TL-CSTD). Specifically, we use 2 s of 8 Hz single-trial calibration data from the target domain to obtain matched transfer templates from the source domain. These templates are then corrected to extract holistic and single-period transfer knowledge, which are subsequently employed to construct an efficient TL-SSVEP decoding model for the target subject. Experimental results on three large SSVEP datasets demonstrate that TL-CSTD effectively addresses the data mismatch problem and achieves excellent SSVEP recognition performance using only 2 s of single-trial calibration data, showing its significant application potential and practicality.

Objective: Brain-computer interfaces (BCIs) based on event-related potentials (ERPs) are among the most accurate and reliable BCIs. However, current mainstream classification algorithms struggle to eliminate the need for calibration and rely on expensive labeled data, limiting the practical usability of ERP based BCIs. The development of fully unsupervised algorithms is essential for the advancement of practical applications of BCI systems. Methods: In this study, we propose a novel unsupervised classification method called sliding-window distribution distance maximization (sDDM). This algorithm utilizes sliding windows to highlight important temporal features and transforms the metric of inter-class differences from absolute distances to relative distribution distances in Mahalanobis space, while incorporating information on target event similarity from the BCI paradigm. Additionally, our proposed spatial dimensionality reduction strategy ensures smaller spatial dimensions and more prominent spatial features. Results: We compare our proposed method to other state of-the-art unsupervised classification methods and evaluate it offline on our self-collected dataset, a public dataset recorded during the use of a P300 Speller by patients with ALS, and the BCI Competition III Dataset II. Our results demonstrate that our proposed method achieves the best spelling accuracy across all datasets, surpassing other unsupervised algorithms. We further explore its improvement effectiveness through ablation experiments. Conclusion: Our proposed method enhances the performance of unsupervised classification in ERP-based BCIs.

Steady state visual evoked potential (SSVEP)-based brain-computer interfaces have been widely studied for their fast response speeds and high information transfer rates. However, how to fully utilize the potential information of existing subjects to realize the mining of common information among different subjects and then realize the information migration in a small amount of data scenarios is a difficult problem faced by current research. In order to solve the above problems, this study proposes a deep neural network based on the pyramid squeeze attention (PSA-DNN) mechanism to enhance the performance of SSVEP-BCI through common information migration. Specifically, the band-pass filtered EEG signals were first Fourier transformed to obtain the frequency domain information; subsequently, the frequency domain information is input into a deep neural network, followed by a spatial convolution step to extract spatial domain information. In order to further enhance the quality of information extraction, a pyramid attention module is introduced into the network to realize the enhancement of frequency domain and spatial domain information. Time domain information from the EEG signals is then mined using temporal convolution. Finally, the full connectivity layer is used to output the recognition results. The model is trained in a three-stage stepped approach for SSVEP target recognition. The first stage uses data from all participants in the training set for common information learning and transfers the model parameters trained in the first stage to the network model in the second stage. In the second stage, some of the information from participants in the test set is used for fine-tuning and to mine personalized information from these new participants. The third stage uses the remaining data from participants in the test set to produce classification results. The proposed method is systematically evaluated using the Benchmark and BETA datasets, where it demonstrates favorable performance compared to established baselines. These findings contribute theoretical insights and methodological References for the application of SSVEP-based brain-computer interfaces in real-world scenarios.

Hybrid brain-computer interfaces (hBCIs) integrate multiple neuroimaging modalities and utilize their complementary information to address the inherent limitations of single-modality neural signal decoding. For electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) hybrid BCIs, advanced fusion algorithms are crucial to fully exploit the superior spatial localization capability of fNIRS and the millisecond-level temporal resolution of EEG. This work proposes an end-to-end spatial calibration-based multi-scale EEG-fNIRS fusion network named EFS-Net, which organically integrates EEG and fNIRS signals through a multi-scale spatio-temporal fusion architecture. The network consists of three complementary functional branches: a multi-scale temporal convolution branch for capturing rapidly changing cortical electrophysiological features of EEG, an EEG spatial branch for constructing latency-compensated cortical topographies to adapt to the delayed hemodynamic response of fNIRS, and a spatially calibrated fNIRS spatial branch for dynamically fusing spatial feature maps with EEG counterparts to generate temporally aligned and spatially enhanced neural representations. This three-branch fusion structure constructs abundant spatio-temporal feature embeddings and improves the discriminability of neural features. Evaluated on two public datasets including Word Generation (WG) and Mental Arithmetic (MA) with a rigorous subject-specific leave-one-session-out cross-validation protocol, EFS-Net achieves classification accuracies of 77.71 ± 8.23

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.

BACKGROUND:Motor imagery signals encompass a broad range of frequency components, and frequency band decomposition can improve the precision of frequency-domain features, helping the model focus on task-relevant information. However, existing methods often treat signals from different frequency bands uniformly, overlooking their heterogeneity and coupling, which leads to redundant features and loss of cooperative information. NEW METHOD:We propose a HCFNet that explores heterogeneous feature extraction and coupling across frequency bands. HCFNet first separates the raw signal into high and low-frequency bands, extracting spatiotemporal features through specialized modules. A cross-frequency coupling module then fuses these features, using data augmentation for regularization to capture robust spectral-spatiotemporal features and high-low frequency coupling. RESULTS:We evaluated our model on the BCIC-IV-2a and OpenBMI benchmark datasets, and our model achieves average accuracies of 82.41 % and 76.52 %. Notably, HCFNet maintains excellent performance even with shorter time windows. COMPARISON WITH EXISTING METHODS:HCFNet outperforms all the state-of-the-art methods we benchmark against. Compared with traditional multi-band isomorphic methods, frequency-band heterogeneous coupling performs better in capturing task-related features and significantly reduces redundancy during feature fusion. CONCLUSIONS:This study significantly advances the decoding technology of motor imagery signals through an innovative frequency-band heterogeneous coupling method. Its substantial potential for rapid responses brings tangible improvements to brain-computer interface systems and is expected to be further applied in domain adaptation, cross-domain alignment, and cross-subject contexts in the future.

Temporal misalignment and inherent latency variability in electroencephalography (EEG) signals pose significant challenges for motor imagery (MI) brain-computer interface (BCI) decoding. While deep learning methods have achieved remarkable progress, they typically assume temporal consistency and often fail to generalize across sessions with substantial phase shifts. To address this limitation, we propose a unified framework that integrates a hybrid convolutional neural network (CNN) and temporal convolutional network (TCN) backbone with a novel differentiable phase shifting mechanism, which formulates temporal synchronization as a fully differentiable spectral phase-correction module. Specifically, a lightweight guidance network predicts sample-specific phase offsets from frequency-domain representations, which parameterize a complex-valued rotation operation performing continuous phase adjustment while preserving amplitude information, enabling end-to-end gradient-based optimization of phase alignment. A variance regularization term further constrains the corrected latent representations to form a compact and consistent feature manifold. Experiments on the benchmark MI BCI Competition IV-2a dataset demonstrate that the proposed method significantly enhances the extraction of classdiscriminative features and achieves an average classification accuracy of 80.94 %, outperforming standard baselines. These results indicate the effectiveness of explicit phase modeling in mitigating temporal jitter and indicate that autonomously learned shift-invariant representations provide a robust solution for decoding complex EEG dynamics, supporting the development of more reliable BCIs in real-world scenarios.

Precision brain stimulation is an emerging paradigm in clinical neuroscience, offering the potential to modulate dysfunctional circuits with individualized protocols for disorders ranging from neurological to psychiatric disorders. The convergence of large-scale neuroimaging datasets, physics-based computational models, and artificial intelligence (AI) is rapidly accelerating this vision. This review integrates developments across four domains. First, we examine the availability and harmonisation of population-level neuroimaging repositories such as the UK Biobank, Human Connectome Project (HCP), Adolescent Brain Cognitive Development (ABCD) study, and ENIGMA consortium, which provide unprecedented insight into structural and functional variability relevant to stimulation targeting biomarkers. Second, we trace the evolution of physics-based and computational head models that simulate electric, magnetic, and acoustic field distributions across non-invasive modalities—including transcranial electrical stimulation (tES), temporal interference stimulation (TIS), transcranial magnetic stimulation (TMS), and transcranial focused ultrasound (tFUS)—as well as invasive techniques such as cortical and deep brain stimulation. Third, we highlight the emergence of AI-driven brain imaging foundation models, such as NeuroSTORM, which leverage large-scale pre-training to enable predictive modeling of brain states, treatment response, and individualized stimulation optimization. Fourth, we present the concept of individualized digital twin brains—computational avatars that integrate neuroimaging, physics-based simulations, and AI foundation models to simulate and test stimulation strategies tailored to each patient’s connectivity and activation profile. Finally, we illustrate translational potential through two clinical use-case scenarios: precision targeting for seizure suppression in epilepsy and individualized stimulation for craving reduction and relapse prevention in addiction. By bridging databases, physics, and AI, this review outlines a roadmap for precision brain stimulation, emphasizing not only technological advances but also the clinical, ethical, and translational implications for real-world practice.

Motor imagery signal decoding is an important research direction in the field of brain-computer interfaces, which aim to judge the motor imagery state of an individual by analyzing electroencephalogram (EEG) signals. Deep learning technology has been gradually applied to EEG classification, which can automatically extract features. However, when processing complex EEG signals, the existing decoding models cannot effectively fuse features at different levels, resulting in limited classification performance. This study proposes a multi-scale spatiotemporal efficient feature fusion network (MS-TSEFNet), which learns the dynamic changes in EEG signals at different time scales through multi-scale convolution modules and combines the spatial attention mechanism to efficiently capture the spatial correlation between electrodes in EEG signals. In addition, the network adopts an efficient feature fusion strategy to deeply fuse features at different levels, thereby improving the expression ability of the model. In the task of motor imagery signal decoding, MS-TSEFNet shows higher accuracy and robustness. We use the public BCIC-IV2a, BCIC-IV2b and ECUST datasets for evaluation. The experimental results show that the average classification accuracy of MS-TSEFNet reaches 80.31%, 86.69% and 71.14%, respectively, which is better than the current state-of-the-art algorithms. We conducted an ablation experiment to further verify the effectiveness of the model. The experimental results showed that each module played an important role in improving the final performance. In particular, the combination of the multi-scale convolution module and the feature fusion module significantly improved the model's ability to extract the spatiotemporal features of EEG signals.

Motor imagery is a non-invasive process that operates independently of external stimuli, and can be used to establish a direct connection between the brain and external devices solely through the imagination of a specific movement. Nonetheless, the complexity and variability of neural patterns pose substantial challenges, as accurately decoding motor imagery from electroencephalography signals remains a significant obstacle. This paper introduces an enhanced dynamic spatiotemporal -frequency attention convolutional neural network (EDSF-Net) for the precise decoding of motor imagery. EDSF-Net employs a refined spatiotemporal attention mechanism, grounded in enhanced dynamic convolution (EDConv), to emphasize localized spatial features alongside high and low-frequency temporal characteristics. Subsequently, EDConv is utilized for global spatial feature extraction. Following this, group convolutions formed by EDConv are implemented to fuse the extracted features effectively. Ultimately, a synchronized channel-frequency attention mechanism is employed to capture critical channel and frequency domain information, facilitating the model's focus on features most pertinent to the task throughout the learning process. We conducted a comprehensive evaluation of the performance of EDSF-Net on two public datasets, BCI Competition IV 2a and OpenBMI. In the hold-out session experiments, EDSF-Net achieved decoding accuracies of 84.26% and 75.14%, respectively. In the leave-one-subject-out experiments, EDSF-Net attained decoding accuracies of 66.78% and 82.24%, respectively. These results show that EDSF-Net has robust generalization capabilities, affirming its efficacy in addressing complex pattern recognition tasks, with significant potential for diverse applications.

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.

In recent years, artificial neural networks have been effectively used to improve the target recognition performance of steady-state visual evoked potential (SSVEP) based Brain-Computer interfaces (BCIs). However, these models require the collection of a large number of calibration trials from users, which typically results in a poor user experience. When fewer calibration trials are acquired this leads to insufficient training of model parameters and weak recognition performance. To tackle these issues, this study proposes a two-branch multi-scale convolutional correlation network (TBMSCCN) in which a correlation network framework is introduced to reduce the model training parameters and prior knowledge of the SSVEP is used to enhance the model representation ability and convergence. First, a multi-scale temporal convolution module is designed to learn local temporal dependencies in a parallel two-branch feature extraction module. Next, a contrastive loss function is constructed in the latent feature space, which can guide the model to learn the intra-class consistent features while speeding up model convergence. Finally, a group convolution module is used as a decision layer to reduce the network parameters, while learning distinguishability features between targets and non-targets. Our offline tests on two public datasets show that proposed TBMSCCN method outperforms TRCA, eTRCA, DNN, Conv-CA and Bi-SiamCA in individual calibration scenarios, which can achieve an average information transform rates (ITRs) of 378.03 ± 139.18 bit/min and 198.92 ± 111.27 bit/min on the “Benchmark” dataset and the “Beta” dataset respectively. Additionally, proposed TBMSCCN method outperform FBCCA, ttCCA, EEGNet, and TST-CFSR in calibration-free scenarios. Furthermore, an online Chinese spelling experiment confirmed the real-world effectiveness of the proposed method. The proposed model has the characteristics of low parameter and strong robustness, which can facilitate the practical engineering application of SSVEP-Based-BCI system. The code is available at https://github.com/xinjieHe123/TBMSCCN.

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.

Objective: The motion trajectory prediction (MTP) based brain-computer interface (BCI) leverages electroencephalography (EEG) signals to reconstruct the three-dimensional trajectory of upper limb motion, which is pivotal for the advancement of prosthetic devices that can assist motor-disabled individuals. Most research focused on improving the performance of regression models while neglecting the correlation between the implicit information extracted from EEG features across various frequency bands with limb kinematics. Current work aims to identify key channels that capture information related to various motion execution movements from different frequency bands and reconstruct three-dimensional motion trajectories based on EEG features. Methods: We propose an interpretable motion trajectory regression framework that extracts bandpower features from different frequency bands and concatenates them into multi-band fusion features. The extreme gradient boosting regression model with Bayesian optimization and Shapley additive explanation methods are introduced to provide further explanation. Results: The experimental results demonstrate that the proposed method achieves a mean Pearson correlation coefficient (PCC) value of 0.452, outperforming traditional regression models. Conclusion: Our findings reveal that the contralateral side contributes the most to motion trajectory regression than the ipsilateral side which improves the clarity and interpretability of the motion trajectory regression model. Specifically, the feature from channel C5 in the Mu band is crucial for the movement of the right hand, while the feature from channel C3 in the Beta band plays a vital role. Significance: This work provides a novel perspective on the comprehensive study of movement disorders.

Brain-computer interface (BCI) system offers an alternative or supplementary means of interaction for individuals with disabilities. P300 speller is a commonly utilized BCI system due to its high stability, and reliability and without intensive user training. Nevertheless, the inherent class imbalance within P300 datasets predisposes the system to overfit, potentially impacting the classification performances. Existing class rebalancing methods mainly rely on resampling or adjusting the class weight with a fixed value, thus it is still tricky to ensure that the output is evenly balanced. To mitigate the above class imbalance issue, this study proposes a dual branch learning (DBL) method that concurrently considers feature representation and class imbalance. This approach involves the ingestion of two distinct sample types-uniformly sampled and reverse-sampled data-into the feature extraction and classification modules during the training phase. Furthermore, a dynamic learning mechanism is implemented to incrementally emphasize minority class samples (specifically the P300 component) as training progresses. The effectiveness of the proposed DBL method is proved using both publicly accessible and self-collected datasets in a subject-dependent scheme. The proposed DBL method can achieve an accuracy of 97.37 % and 88.72 % in the above datasets. Besides, it provides superior and more reliable results compared with several deep learning and rebalancing methods. These findings highlight the promising potential of the proposed DBL framework in P300-based BCI.

Recent advancements in large language models (LLMs) provide a more effective pathway for upgrading brain-computer interface (BCI) technology in terms of user interaction. The widespread adoption of BCIs in daily application scenarios is still limited by factors such as their single functionality, restricted paradigm design, weak multilingual support, and low levels of intelligence. In this paper, we propose an innovative BCI system that deeply integrates a steady-state visual evoked potential (SSVEP) speller with an LLM application programming interface (API). It allows natural language input through the SSVEP speller and dynamically calls large models to generate SSVEP paradigms. The command prompt, blinking frequency, and layout position are adjustable to meet the user's control requirements in various scenarios. More than ten languages are compatible with the multilingual support of LLM. A variety of task scenarios, such as home appliance control, robotic arm operation, and unmanned aerial vehicle (UAV) management are provided. The task interfaces of the system can be personalized according to the user's habits, usage scenarios, and equipment characteristics. By combining the SSVEP speller with an LLM, the system solves numerous challenges faced by current BCI systems and makes breakthroughs in functionality, intelligence, and multilingual support. The introduction of LLM not only enhances user experience but also expands the potential applications of BCI technology in real-world environments.