EEG Decoding Research Articles

A protocol for trustworthy EEG decoding with neural networks

Deep learning solutions have rapidly emerged for EEG decoding, achieving state-of-the-art performance on a variety of decoding tasks. Despite their high performance, existing solutions do not fully address the challenge posed by the introduction of many hyperparameters, defining data pre-processing, network architecture, network training, and data augmentation. Automatic hyperparameter search is rarely performed and limited to network-related hyperparameters. Moreover, pipelines are highly sensitive to performance fluctuations due to random initialization, hindering their reliability. Here, we design a comprehensive protocol for EEG decoding that explores the hyperparameters characterizing the entire pipeline and that includes multi-seed initialization for providing robust performance estimates. Our protocol is validated on 9 datasets about motor imagery, P300, SSVEP, including 204 participants and 26 recording sessions, and on different deep learning models. We accompany our protocol with extensive experiments on the main aspects influencing it, such as the number of participants used for hyperparameter search, the split into sequential simpler searches (multi-step search), the use of informed vs. non-informed search algorithms, and the number of random seeds for obtaining stable performance. The best protocol included 2-step hyperparameter search via an informed search algorithm, with the final training and evaluation performed using 10 random initializations. The optimal trade-off between performance and computational time was achieved by using a subset of 3-5 participants for hyperparameter search. Our protocol consistently outperformed baseline state-of-the-art pipelines, widely across datasets and models, and could represent a standard approach for neuroscientists for decoding EEG in a trustworthy and reliable way.

Non-invasive brain-machine interface control with artificial intelligence copilots.

Motor brain-machine interfaces (BMIs) decode neural signals to help people with paralysis move and communicate. Even with important advances in the last two decades, BMIs face key obstacles to clinical viability. Invasive BMIs achieve proficient cursor and robotic arm control but require neurosurgery, posing significant risk to patients. Non-invasive BMIs do not have neurosurgical risk, but achieve lower performance, sometimes being prohibitively frustrating to use and preventing widespread adoption. We take a step toward breaking this performance-risk tradeoff by building performant non-invasive BMIs. The critical limitation that bounds decoder performance in non-invasive BMIs is their poor neural signal-to-noise ratio. To overcome this, we contribute (1) a novel EEG decoding approach and (2) artificial intelligence (AI) copilots that infer task goals and aid action completion. We demonstrate that with this "AI-BMI," in tandem with a new adaptive decoding approach using a convolutional neural network (CNN) and ReFIT-like Kalman filter (KF), healthy users and a paralyzed participant can autonomously and proficiently control computer cursors and robotic arms. Using an AI copilot improves goal acquisition speed by up to 4.3 × in the standard center-out 8 cursor control task and enables users to control a robotic arm to perform the sequential pick-and-place task, moving 4 randomly placed blocks to 4 randomly chosen locations. As AI copilots improve, this approach may result in clinically viable non-invasive AI-BMIs.

A transient high-dimensional geometry affords stable conjunctive subspaces for efficient action selection

Flexible action selection requires cognitive control mechanisms capable of mapping the same inputs to different output actions depending on the context. From a neural state-space perspective, this requires a control representation that separates similar input neural states by context. Additionally, for action selection to be robust and time-invariant, information must be stable in time, enabling efficient readout. Here, using EEG decoding methods, we investigate how the geometry and dynamics of control representations constrain flexible action selection in the human brain. Participants performed a context-dependent action selection task. A forced response procedure probed action selection different states in neural trajectories. The result shows that before successful responses, there is a transient expansion of representational dimensionality that separated conjunctive subspaces. Further, the dynamics stabilizes in the same time window, with entry into this stable, high-dimensional state predictive of individual trial performance. These results establish the neural geometry and dynamics the human brain needs for flexible control over behavior.

A Spatio-Temporal Capsule Neural Network with Self-Correlation Routing for EEG Decoding of Semantic Concepts of Imagination and Perception Tasks.

Decoding semantic concepts for imagination and perception tasks (SCIP) is important for rehabilitation medicine as well as cognitive neuroscience. Electroencephalogram (EEG) is commonly used in the relevant fields, because it is a low-cost noninvasive technique with high temporal resolution. However, as EEG signals contain a high noise level resulting in a low signal-to-noise ratio, it makes decoding EEG-based semantic concepts for imagination and perception tasks (SCIP-EEG) challenging. Currently, neural network algorithms such as CNN, RNN, and LSTM have almost reached their limits in EEG signal decoding due to their own short-comings. The emergence of transformer methods has improved the classification performance of neural networks for EEG signals. However, the transformer model has a large parameter set and high complexity, which is not conducive to the application of BCI. EEG signals have high spatial correlation. The relationship between signals from different electrodes is more complex. Capsule neural networks can effectively model the spatial relationship between electrodes through vector representation and a dynamic routing mechanism. Therefore, it achieves more accurate feature extraction and classification. This paper proposes a spatio-temporal capsule network with a self-correlation routing mechaninsm for the classification of semantic conceptual EEG signals. By improving the feature extraction and routing mechanism, the model is able to more effectively capture the highly variable spatio-temporal features from EEG signals and establish connections between capsules, thereby enhancing classification accuracy and model efficiency. The performance of the proposed model was validated using the publicly accessible semantic concept dataset for imagined and perceived tasks from Bath University. Our model achieved average accuracies of 94.9%, 93.3%, and 78.4% in the three sensory modalities (pictorial, orthographic, and audio), respectively. The overall average accuracy across the three sensory modalities is 88.9%. Compared to existing advanced algorithms, the proposed model achieved state-of-the-art performance, significantly improving classification accuracy. Additionally, the proposed model is more stable and efficient, making it a better decoding solution for SCIP-EEG decoding.

CTNet: a convolutional transformer network for EEG-based motor imagery classification

Brain-computer interface (BCI) technology bridges the direct communication between the brain and machines, unlocking new possibilities for human interaction and rehabilitation. EEG-based motor imagery (MI) plays a pivotal role in BCI, enabling the translation of thought into actionable commands for interactive and assistive technologies. However, the constrained decoding performance of brain signals poses a limitation to the broader application and development of BCI systems. In this study, we introduce a convolutional Transformer network (CTNet) designed for EEG-based MI classification. Firstly, CTNet employs a convolutional module analogous to EEGNet, dedicated to extracting local and spatial features from EEG time series. Subsequently, it incorporates a Transformer encoder module, leveraging a multi-head attention mechanism to discern the global dependencies of EEG's high-level features. Finally, a straightforward classifier module comprising fully connected layers is followed to categorize EEG signals. In subject-specific evaluations, CTNet achieved remarkable decoding accuracies of 82.52% and 88.49% on the BCI IV-2a and IV-2b datasets, respectively. Furthermore, in the challenging cross-subject assessments, CTNet achieved recognition accuracies of 58.64% on the BCI IV-2a dataset and 76.27% on the BCI IV-2b dataset. In both subject-specific and cross-subject evaluations, CTNet holds a leading position when compared to some of the state-of-the-art methods. This underscores the exceptional efficacy of our approach and its potential to set a new benchmark in EEG decoding.

A Method to Extract Task-Related EEG Feature Based on Lightweight Convolutional Neural Network.

Unlocking task-related EEG spectra is crucial for neuroscience. Traditional convolutional neural networks (CNNs) effectively extract these features but face limitations like overfitting due to small datasets. To address this issue, we propose a lightweight CNN and assess its interpretability through the fully connected layer (FCL). Initially tested with two tasks (Task 1: open vs closed eyes, Task 2: interictal vs ictal stage), the CNN demonstrated enhanced spectral features in the alpha band for Task 1 and the theta band for Task 2, aligning with established neurophysiological characteristics. Subsequent experiments on two brain-computer interface tasks revealed a correlation between delta activity (around 1.55 Hz) and hand movement, with consistent results across pericentral electroencephalogram (EEG) channels. Compared to recent research, our method stands out by delivering task-related spectral features through FCL, resulting in significantly fewer trainable parameters while maintaining comparable interpretability. This indicates its potential suitability for a wider array of EEG decoding scenarios.

HiRENet: Novel convolutional neural network architecture using Hilbert-transformed and raw electroencephalogram (EEG) for subject-independent emotion classification

Background and objectivesConvolutional neural networks (CNNs) are the most widely used deep-learning framework for decoding electroencephalograms (EEGs) due to their exceptional ability to extract hierarchical features from high-dimensional EEG data. Traditionally, CNNs have primarily utilized multi-channel raw EEG data as the input tensor; however, the performance of CNN-based EEG decoding may be enhanced by incorporating phase information alongside amplitude information. MethodsThis study introduces a novel CNN architecture called the Hilbert-transformed (HT) and raw EEG network (HiRENet), which incorporates both raw and HT EEG as inputs. This concurrent use of HT and raw EEG aims to integrate phase information with existing amplitude information, potentially offering a more comprehensive reflection of functional connectivity across various brain regions. The HiRENet model was developed using two CNN frameworks: ShallowFBCSPNet and a CNN with a residual block (ResCNN). The performance of the HiRENet model was assessed using a lab-made EEG database to classify human emotions, comparing three input modalities: raw EEG, HT EEG, and a combination of both signals. Additionally, the computational complexity was evaluated to validate the computational efficiency of the ResCNN design. ResultsThe HiRENet model based on ResCNN achieved the highest classification accuracy, with 86.03% for valence and 84.01% for arousal classifications, surpassing traditional CNN methodologies. Considering computational efficiency, ResCNN demonstrated superiority over ShallowFBCSPNet in terms of speed and inference time, despite having a higher parameter count. ConclusionOur experimental results showed that the proposed HiRENet can be potentially used as a new option to improve the overall performance for deep learning-based EEG decoding problems.

EEG decoding reveals task-dependent recoding of sensory information in working memory

Working memory (WM) supports future behavior by retaining perceptual information obtained in the recent past. The present study tested the hypothesis that WM recodes sensory information in a format that better supports behavioral goals. We recorded EEG while participants performed color delayed-estimation tasks where the colorwheel for the response was either randomly rotated or held fixed across trials. Accordingly, observers had to remember the exact colors in the Rotation condition, whereas they could prepare for a response based on the fixed mapping between the colors and their corresponding locations on the colorwheel in the No-Rotation condition. Results showed that the color reports were faster and more precise in the No-Rotation condition even when exactly the same set of colors were tested in both conditions. To investigate how the color information was maintained in the brain, we decoded the color using a multivariate EEG classification method. The decoding was limited to the stimulus encoding period in the Rotation condition, whereas it continued to be significant during the maintenance period in the No-Rotation condition, indicating that the color information was actively maintained in the condition. Follow-up analyses suggested that the prolonged decoding was not merely driven by the covert shift of attention but rather by the recoding of sensory information into an action-oriented response format. Together, these results provide converging evidence that WM flexibly recodes sensory information depending on the specific task context to optimize subsequent behavioral performance.

Distinct basal ganglia contributions to learning from implicit and explicit value signals in perceptual decision-making

Metacognitive evaluations of confidence provide an estimate of decision accuracy that could guide learning in the absence of explicit feedback. We examine how humans might learn from this implicit feedback in direct comparison with that of explicit feedback, using simultaneous EEG-fMRI. Participants performed a motion direction discrimination task where stimulus difficulty was increased to maintain performance, with intermixed explicit- and no-feedback trials. We isolate single-trial estimates of post-decision confidence using EEG decoding, and find these neural signatures re-emerge at the time of feedback together with separable signatures of explicit feedback. We identified these signatures of implicit versus explicit feedback along a dorsal-ventral gradient in the striatum, a finding uniquely enabled by an EEG-fMRI fusion. These two signals appear to integrate into an aggregate representation in the external globus pallidus, which could broadcast updates to improve cortical decision processing via the thalamus and insular cortex, irrespective of the source of feedback.

A systematic evaluation of Euclidean alignment with deep learning for EEG decoding

Objective: Electroencephalography signals are frequently used for various Brain–Computer interface (BCI) tasks. While deep learning (DL) techniques have shown promising results, they are hindered by the substantial data requirements. By leveraging data from multiple subjects, transfer learning enables more effective training of DL models. A technique that is gaining popularity is Euclidean alignment (EA) due to its ease of use, low computational complexity, and compatibility with DL models. However, few studies evaluate its impact on the training performance of shared and individual DL models. In this work, we systematically evaluate the effect of EA combined with DL for decoding BCI signals. Approach: We used EA as a pre-processing step to train shared DL models with data from multiple subjects and evaluated their transferability to new subjects. Main results: Our experimental results show that it improves decoding in the target subject by 4.33% and decreases convergence time by more than 70%. We also trained individual models for each subject to use as a majority-voting ensemble classifier. In this scenario, using EA improved the 3-model ensemble accuracy by 3.71%. However, when compared to the shared model with EA, the ensemble accuracy was 3.62% lower. Significance: EA succeeds in the task of improving transfer learning performance with DL models and, could be used as a standard pre-processing technique.

MSHANet: a multi-scale residual network with hybrid attention for motor imagery EEG decoding

MSHANet: a multi-scale residual network with hybrid attention for motor imagery EEG decoding

TFormer: A time–frequency Transformer with batch normalization for driver fatigue recognition

Within the framework of the advanced human-cybernetic interfaces (HCI), Cross-subject electroencephalogram (EEG)-based driver fatigue recognition is emerging as a pivotal application in the paradigm of Industry 5.0. Recognizing the importance of ensuring driver safety through proactive monitoring, it is essential to offer a general EEG decoding system to improve road safety. This work investigated the use of Transformers for the challenging cross-subject EEG decoding task due to the great success the Transformers have achieved in various applications. Previous research focused on using Transformers to capture global temporal information, but less work targeted global frequency-domain patterns. Furthermore, in order to leverage a standard Transformer architecture grounded in natural language processing for EEG decoding, it is imperative to account for inherent characteristics in EEG and make pertinent adjustments accordingly. In this work, we proposed a time–frequency Transformer (TFormer) that can automatically learn the global time–frequency patterns from raw EEG data. TFormer consisted of three components: convolutional stems for input embedding, time–frequency multi-head cross-attention (TF-MCA) for integrating time-domain patterns into frequency points, and self-attention to further learn global time-frequency patterns. Moreover, we analyzed TFormer’s internal settings and found batch normalization (BN) more suitable for cross-subject EEG decoding than layer normalization (LN). The experiment results demonstrated the superiority of the proposed model compared to existing methods. Overall, our work contributes to the development of Transformer models in EEG decoding and illustrates a different way to leverage Transformers for decoding raw EEG data.

Exploring an EM-algorithm for banded regression in computational neuroscience

Abstract Regression is a principal tool for relating brain responses to stimuli or tasks in computational neuroscience. This often involves fitting linear models with predictors that can be divided into groups, such as distinct stimulus feature subsets in encoding models or features of different neural response channels in decoding models. When fitting such models, it can be relevant to allow differential shrinkage of the different groups of regression weights. Here, we explore a framework that allows for straightforward definition and estimation of such models. We present an expectation-maximization algorithm for tuning hyperparameters that control shrinkage of groups of weights. We highlight properties, limitations, and potential use-cases of the model using simulated data. Next, we explore the model in the context of a BOLD fMRI encoding analysis and an EEG decoding analysis. Finally, we discuss cases where the model can be useful and scenarios where regularization procedures complicate model interpretation.

Multi-source deep domain adaptation ensemble framework for cross-dataset motor imagery EEG transfer learning

Objective. Electroencephalography (EEG) is an important kind of bioelectric signal for measuring physiological activities of the brain, and motor imagery (MI) EEG has significant clinical application prospects. Convolutional neural network has become a mainstream algorithm for MI EEG classification, however lack of subject-specific data considerably restricts its decoding accuracy and generalization performance. To address this challenge, a novel transfer learning (TL) framework using auxiliary dataset to improve the MI EEG classification performance of target subject is proposed in this paper. Approach. We developed a multi-source deep domain adaptation ensemble framework (MSDDAEF) for cross-dataset MI EEG decoding. The proposed MSDDAEF comprises three main components: model pre-training, deep domain adaptation, and multi-source ensemble. Moreover, for each component, different designs were examined to verify the robustness of MSDDAEF. Main results. Bidirectional validation experiments were performed on two large public MI EEG datasets (openBMI and GIST). The highest average classification accuracy of MSDDAEF reaches 74.28% when openBMI serves as target dataset and GIST serves as source dataset. While the highest average classification accuracy of MSDDAEF is 69.85% when GIST serves as target dataset and openBMI serves as source dataset. In addition, the classification performance of MSDDAEF surpasses several well-established studies and state-of-the-art algorithms. Significance. The results of this study show that cross-dataset TL is feasible for left/right-hand MI EEG decoding, and further indicate that MSDDAEF is a promising solution for addressing MI EEG cross-dataset variability.

Online continual decoding of streaming EEG signal with a balanced and informative memory buffer

Electroencephalography (EEG) based Brain Computer Interface (BCI) systems play a significant role in facilitating how individuals with neurological impairments effectively interact with their environment. In real world applications of BCI system for clinical assistance and rehabilitation training, the EEG classifier often needs to learn on sequentially arriving subjects in an online manner. As patterns of EEG signals can be significantly different for different subjects, the EEG classifier can easily erase knowledge of learnt subjects after learning on later ones as it performs decoding in online streaming scenario, namely catastrophic forgetting. In this work, we tackle this problem with a memory-based approach, which considers the following conditions: (1) subjects arrive sequentially in an online manner, with no large scale dataset available for joint training beforehand, (2) data volume from the different subjects could be imbalanced, (3) decoding difficulty of the sequential streaming signal vary, (4) continual classification for a long time is required. This online sequential EEG decoding problem is more challenging than classic cross subject EEG decoding as there is no large-scale training data from the different subjects available beforehand. The proposed model keeps a small balanced memory buffer during sequential learning, with memory data dynamically selected based on joint consideration of data volume and informativeness. Furthermore, for the more general scenarios where subject identity is unknown to the EEG decoder, aka. subject agnostic scenario, we propose a kernel based subject shift detection method that identifies underlying subject changes on the fly in a computationally efficient manner. We develop challenging benchmarks of streaming EEG data from sequentially arriving subjects with both balanced and imbalanced data volumes, and performed extensive experiments with a detailed ablation study on the proposed model. The results show the effectiveness of our proposed approach, enabling the decoder to maintain performance on all previously seen subjects over a long period of sequential decoding. The model demonstrates the potential for real-world applications.

Contribution of low-level image statistics to EEG decoding of semantic content in multivariate and univariate models with feature optimization

Spatio-temporal patterns of evoked brain activity contain information that can be used to decode and categorize the semantic content of visual stimuli. However, this procedure can be biased by low-level image features independently of the semantic content present in the stimuli, prompting the need to understand the robustness of different models regarding these confounding factors. In this study, we trained machine learning models to distinguish between concepts included in the publicly available THINGS-EEG dataset using electroencephalography (EEG) data acquired during a rapid serial visual presentation paradigm. We investigated the contribution of low-level image features to decoding accuracy in a multivariate model, utilizing broadband data from all EEG channels. Additionally, we explored a univariate model obtained through data-driven feature selection applied to the spatial and frequency domains. While the univariate models exhibited better decoding accuracy, their predictions were less robust to the confounding effect of low-level image statistics. Notably, some of the models maintained their accuracy even after random replacement of the training dataset with semantically unrelated samples that presented similar low-level content. In conclusion, our findings suggest that model optimization impacts sensitivity to confounding factors, regardless of the resulting classification performance. Therefore, the choice of EEG features for semantic decoding should ideally be informed by criteria beyond classifier performance, such as the neurobiological mechanisms under study.

Attention-based convolutional neural network with multi-modal temporal information fusion for motor imagery EEG decoding

Convolutional neural network (CNN) has been widely applied in motor imagery (MI)-based brain computer interface (BCI) to decode electroencephalography (EEG) signals. However, due to the limited perceptual field of convolutional kernel, CNN only extracts features from local region without considering long-term dependencies for EEG decoding. Apart from long-term dependencies, multi-modal temporal information is equally important for EEG decoding because it can offer a more comprehensive understanding of the temporal dynamics of neural processes. In this paper, we propose a novel deep learning network that combines CNN with self-attention mechanism to encapsulate multi-modal temporal information and global dependencies. The network first extracts multi-modal temporal information from two distinct perspectives: average and variance. A shared self-attention module is then designed to capture global dependencies along these two feature dimensions. We further design a convolutional encoder to explore the relationship between average-pooled and variance-pooled features and fuse them into more discriminative features. Moreover, a data augmentation method called signal segmentation and recombination is proposed to improve the generalization capability of the proposed network. The experimental results on the BCI Competition IV-2a (BCIC-IV-2a) and BCI Competition IV-2b (BCIC-IV-2b) datasets show that our proposed method outperforms the state-of-the-art methods and achieves 4-class average accuracy of 85.03% on the BCIC-IV-2a dataset. The proposed method implies the effectiveness of multi-modal temporal information fusion in attention-based deep learning networks and provides a new perspective for MI-EEG decoding. The code is available at https://github.com/Ma-Xinzhi/EEG-TransNet.

A learnable continuous wavelet-based multi-branch attentive convolutional neural network for spatio–spectral–temporal EEG signal decoding

Deep learning methods, particularly convolutional neural networks (CNNs), have advanced the field of motor imagery (MI) by effectively extracting spatio-spectral–temporal (SST) features from electroencephalography (EEG) signals. Image-based CNNs using time–frequency analysis such as wavelet transform have outperformed signal-based CNNs in MI-EEG classification. However, the spatial characteristics of EEG signals can be neglected due to the high dimensionality of spectral–temporal feature space. Moreover, the performance hinges on whether the predefined transformation can capture the class-discriminative spectral–temporal representations of EEG signals well. Despite attempts to alleviate these challenges by selecting EEG electrodes or restricting spectral–temporal feature space, large inter-subject variabilities have impeded successful MI-EEG decoding. In this paper, we propose a learnable continuous wavelet-based multi-branch attentive CNN framework for decoding MI-EEG signals. Our method is designed to automatically generate class-discriminative SST representations of EEG signals within subject-specific sub-bands by utilizing learnable wavelet-based convolutions. Additionally, it employs multi-branch attentive CNNs for the effective and efficient extraction of local and global SST features. Our comprehensive evaluation on two public datasets demonstrates that the proposed method significantly outperformed state-of-the-art methods. We also conduct an ablation study under various configuration settings within the proposed framework to demonstrate the effectiveness of our method. The proposed method provides a powerful and innovative approach to MI-EEG signal decoding, with implications for personalized EEG decoding.

A fused multi-subfrequency bands and CBAM SSVEP-BCI classification method based on convolutional neural network.

For the brain-computer interface (BCI) system based on steady-state visual evoked potential (SSVEP), it is difficult to obtain satisfactory classification performance for short-time window SSVEP signals by traditional methods. In this paper, a fused multi-subfrequency bands and convolutional block attention module (CBAM) classification method based on convolutional neural network (CBAM-CNN) is proposed for discerning SSVEP-BCI tasks. This method extracts multi-subfrequency bands SSVEP signals as the initial input of the network model, and then carries out feature fusion on all feature inputs. In addition, CBAM is embedded in both parts of the initial input and feature fusion for adaptive feature refinement. To verify the effectiveness of the proposed method, this study uses the datasets of Inner Mongolia University of Technology (IMUT) and Tsinghua University (THU) to evaluate the performance of the proposed method. The experimental results show that the highest accuracy of CBAM-CNN reaches 0.9813 percentage point (pp). Within 0.1-2s time window, the accuracy of CBAM-CNN is 0.0201-0.5388 (pp) higher than that of CNN, CCA-CWT-SVM, CCA-SVM, CCA-GNB, FBCCA, and CCA. Especially in the short-time window range of 0.1-1s, the performance advantage of CBAM-CNN is more significant. The maximum information transmission rate (ITR) of CBAM-CNN is 503.87bit/min, which is 227.53bit/min-503.41bit/min higher than the above six EEG decoding methods. The study further results show that CBAM-CNN has potential application value in SSVEP decoding.

A robust multi-branch multi-attention-mechanism EEGNet for motor imagery BCI decoding

BackgroundMotor-Imagery-based Brain-Computer Interface (MI-BCI) is a promising technology to assist communication, movement, and neurological rehabilitation for motor-impaired individuals. Electroencephalography (EEG) decoding techniques using deep learning (DL) possess noteworthy advantages due to automatic feature extraction and end-to-end learning. However, the DL-based EEG decoding models tend to show large variations due to intersubject variability of EEG, which results from inconsistencies of different subjects’ optimal hyperparameters. New methodsThis study proposes a multi-branch multi-attention mechanism EEGNet model (MBMANet) for robust decoding. It applies the multi-branch EEGNet structure to achieve various feature extractions. Further, the different attention mechanisms introduced in each branch attain diverse adaptive weight adjustments. This combination of multi-branch and multi-attention mechanisms allows for multi-level feature fusion to provide robust decoding for different subjects. ResultsThe MBMANet model has a four-classification accuracy of 83.18% and kappa of 0.776 on the BCI Competition IV-2a dataset, which outperforms other eight CNN-based decoding models. This consistently satisfactory performance across all nine subjects indicates that the proposed model is robust. ConclusionsThe combine of multi-branch and multi-attention mechanisms empowers the DL-based models to adaptively learn different EEG features, which provides a feasible solution for dealing with data variability. It also gives the MBMANet model more accurate decoding of motion intentions and lower training costs, thus improving the MI-BCI’s utility and robustness.