Decoding Multi-Class Motor Imagery From Unilateral Limbs Using EEG Signals

The EEG is a widely utilized neural signal source, particularly in motor imagery-based brain-computer interface (MI-BCI), offering distinct advantages in applications like stroke rehabilitation. Current research predominantly concentrates on the bilateral limbs paradigm and decoding, but the use scenarios for stroke rehabilitation are typically for unilateral upper limbs. There is a significant challenge to decoding unilateral MI of multitasks due to the overlapped spatial neural activities of the tasks. This study aims to formulate a novel MI-BCI experimental paradigm for unilateral limbs with multitasks. The paradigm encompasses four imagined movement directions: top-bottom, left-right, top right-bottom left, and top left-bottom right. Forty-six healthy subjects participated in this experiment. Commonly used machine learning techniques, such as FBCSP, EEGNet, deepConvNet, and FBCNet, were employed for evaluation. To improve decoding accuracy, we propose an MVCA method that introduces temporal convolution and attention mechanism to effectively capture temporal features from multiple perspectives. With the MVCA model, we have achieved 40.6% and 64.89% classification accuracies for the four-class and two-class scenarios (top right-bottom left and top left-bottom right), respectively. Conclusion: This is the first study demonstrating that motor imagery of multiple directions in unilateral limbs can be decoded. In particular, decoding two directions, right top to left bottom and left top to right bottom, provides the best accuracy, which sheds light on future studies. This study advances the development of the MI-BCI paradigm, offering preliminary evidence for the feasibility of decoding multiple directional information from EEG. This, in turn, enhances the dimensions of MI control commands.


I. INTRODUCTION
I N RECENT years, motor imagery-based brain-computer interface (MI-BCI) has gained prominence as effective tools for rehabilitation and communication in stroke patients [1].MI-BCI is a technology that measures human brain activities related to motor imagery (MI) and translates them into commands for controlling external devices [1], [2], [3].This process involves capturing oscillatory neural activity within specific frequency bands of the motor cortex, without the need for external stimulation [4].Furthermore, MI-BCI systems have shown promising results in stroke motor rehabilitation and the control of external devices [5], [6].Their noninvasive nature, as opposed to invasive surgical alternatives, enhances patient acceptability.Moreover, non-invasive BCI systems, particularly those utilizing electroencephalography (EEG), offer a balance of safety, broad applicability, and a high temporal resolution, making them a preferred choice in clinical settings [7], [8].
The development of EEG-based MI-BCI confronts two principal challenges.The first is the prevalent research focus on the bilateral limbs paradigm [9], which contrasts sharply with the unilateral upper limb scenarios typically encountered in stroke rehabilitation [10].This discrepancy impedes the effective application of MI-BCI in the scenarios of stroke rehabilitation.The second challenge is the data scarcity for the unilateral paradigm, compounded by the overlapping spatial neural activities during different tasks present significant hurdles in decoding the EEG data of the unilateral with multitasks [7].Additionally, the inherent characteristics of EEG, such as its susceptibility to noise and the limited availability of training data, further complicate the decoding process [11], [12].
MI-BCI research predominantly focuses on motor patterns that are distinguishable at a spatial scale, such as tasks involving left and right hands or feet [9], [13], [14].Despite recent advancements, the MI paradigm still tends to emphasize the bilateral limbs.In 2019, M. H. Lee and colleagues contributed significantly by releasing a comprehensive dataset involving fifty-four participants across three sessions, focusing on grasping movements with left and right hands [15].Following this, J. Ma and team provided an MI dataset that encompassed bilateral limb movements in a two-task format, with twenty-five participants over five sessions [16].However, the evolution of bilateral multi-class paradigms has been limited.The most typical example is the 2008 BCI Competition (BCIC) IV-2a EEG dataset [17], which included nine subjects performing four MI tasks: left hand, right hand, foot, and tongue.Such MI-BCI paradigms, primarily involving singular tasks, may not adequately induce neural plasticity in the motor areas for effective rehabilitation.Furthermore, they might not meet the growing demands for diverse MI control commands [18].
Most of the stroke patients are hemispheric, that means they need to exercise one side of the upper limb to gain motor recovery.Traditional two-side MI-BCI systems are not usually directly useful in this case.As MI paradigms advance and the demands in stroke rehabilitation increase, research has progressively pivoted towards MI-BCI experimental paradigms that focus on unilateral limbs engaging in multitasks.For instance, Ma et al. developed a paradigm involving three distinct tasks, moving the right hand, moving the right elbow, and resting, specifically for unilateral upper limbs, with data collected from twenty-five healthy individuals [18].Complementing this, Jeong et al. provided a dataset encompassing 11 intuitive movement tasks for a single upper limb, which includes diverse activities such as arm-reaching (6 tasks), hand-grasping (3 tasks), and wrist-twisting (2 tasks) [19].These research efforts primarily concentrate on MI tasks associated with the elbow, wrist, fingers, and other joints, thereby providing more detailed granularity than bilateral limb tasks.However, these paradigms have not yet demonstrated superior effectiveness in generating higher-dimensional control commands.Moreover, the complexity of these tasks raises concerns regarding their suitability for rehabilitation training in stroke patients.
Current research on EEG signal recognition for MI primarily focuses on different bodies, such as the tongue, hands, and feet [20], [21], [22], [23].These deep learning models are adept at autonomously extracting temporal, spectral, and spatial features from EEG signals, adeptly managing the complexity of large and high-dimensional datasets [24], [25], [26], [27], [28], [29].Deep learning networks designed based on EEG signal features have achieved excellent results [30], [31], [32], [33].However, there is limited research on the decoding of MI-EEG involving tasks of unilateral limb.One of the major challenges in decoding MI-EEG for unilateral limbs is the high overlap of brain-activations by multi-class activities [34].For most existing neural networks, it is very difficult to extract effective features from overlapping activation regions.Qiu et al. optimized CNN based on FBCNet to decode unilateral upper limb tasks which are hand gripping and releasing [35].A classification accuracy of 67.8% was achieved on the two-class MI-EEG data of ten subjects.Y. Zhang and team introduced wasserstein generative adversarial network based domain adaptation network to decode unilateral MI-EEG [36].The algorithm showed 59.69% on the two-class of MI-EEG data of 25 subjects.However, due to the proximity of brain activation regions for unilateral limb movements, the classification of different tasks becomes more difficult and complex.
The progression of MI-BCI is supported by a variety of publicly accessible datasets, which predominantly focus on the bilateral limbs activities [14], [15], [17].This emphasis has influenced the trajectory of algorithmic development in the field, with a significant portion of MI-EEG research dedicated to decoding algorithms for bilateral limb tasks.This trend, however, presents a substantial challenge in the advancement of the unilateral upper limbs MI paradigms.The primary difficulty lies in the decoding of EEG signals associated with these paradigms, attributed to the inherently limited spatial information available from unilateral limb activities.This limitation necessitates more nuanced approaches to effectively interpret EEG data corresponding to unilateral upper limb movements.
In addressing the inherent challenges associated with the unilateral limbs paradigm and its associated decoding complexities, this study introduces a novel experimental paradigm specifically designed for the unilateral limbs, encompassing four imagined movement directions.The four motor directions included "left to right (left-right)", "top to bottom (topbottom)", "top right to bottom left (top right-bottom left)" and "top left to bottom right (top left-bottom right)".Forty-six healthy subjects participated in this experiment, underscoring the potential of this paradigm to advance the unilateral limbs MI research.Additionally, the data derived from this study serves as a valuable resource for the development of innovative deep learning algorithms aimed at decoding complex multi-task movements of unilateral limbs.We undertake a comprehensive analysis of the EEG signals corresponding to these motor directions, identifying and delineating their unique characteristics.Our primary contributions can be summarized as follows: • We formulate a novel MI-BCI experimental paradigm for the unilateral limbs with multitasks and establish a large dataset from 46 healthy subjects.The paradigm encompasses four imagined movement directions: top-bottom, left-right, top right-bottom left, and top left-bottom right.
• Commonly used machine learning techniques were employed for evaluating this dataset.We demonstrate for the first time that MI of multiple directions in unilateral limbs can potentially be decoded.
• We propose a novel method, the multi-view convolutional self-attention network (MVCA), which demonstrated superior performance in decoding motor directions compared to existing techniques.
• Our results particularly highlight that decoding two specific directions, top right-bottom left, and top leftbottom right, yields the highest accuracy.This finding provides valuable insights for future research in the field.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

II. MATERIALS AND METHODS A. Participants and Equipment
A total of forty-six healthy subjects, aged between 22 and 28 (25 females, all right-handed), participated in our experiment.Of these, twelve subjects are naive to MI-BCI use, while the remainder had prior experience with MI-BCI experiments.None of the participants had any history of neurological, psychiatric, or other medical conditions that might influence the outcomes of MI experiments.Prior to the experiment, each participant provided informed consent and completed a comprehensive questionnaire detailing their age, gender, handedness, previous BCI experience, and health status.Postexperiment, we conducted health checks to ensure participants' well-being and collected their feedback on the experiment through a structured questionnaire.The experiments were conducted in a controlled, quiet, and enclosed laboratory setting, supervised by trained personnel to ensure data reliability.This study received ethical approval from the Ethics Committee of Shanghai University (approval number: ECSHU 2021-233) and adhered to the principles outlined in the Declaration of Helsinki.
In this experiment, we utilize the NeuSen W series 64-channel wireless EEG acquisition system from Neuracle for data acquisition.The electrode placement conform to the international 10-20 system, as depicted in Fig. 1(a).The system's sampling rate is configured at 1000 Hz to ensure high-resolution data capture.We rigorously maintain the impedance of all electrodes below 5 K throughout the experiment, as illustrated in Fig. 1(b).To facilitate optimal EEG recording, participants are comfortably seated in a chair, positioned approximately 80±5cm from the LCD monitor.We provide clear instructions to the participants to remain relaxed, with an emphasis on minimizing any unnecessary eye or muscle movements during the EEG recording.

B. Experimental Paradigm
Our experimental paradigm encompasses four imagined movement directions for unilateral limbs.These four imagined movement directions are as follows: left-right, top-bottom, top right-bottom left and top left-bottom right.The structure of the experiment, as depicted in Fig. 2  The MI task phase is organized into five blocks, with each block separated by a mandatory one-minute rest period to ensure the participants' well-being and maintain optimal performance levels.Each block contained 80 trials, where the four MI tasks were presented in a random sequence, culminating in a total of 400 trials per participant (100 trials per MI task).As shown in Fig. 2(c), spanned 8 seconds, initiated with a 2-second period of video and audio cues.This is followed by an MI stage lasting 2-4 seconds, during which participants were required to mentally rehearse the imagined tasks 2-4 times, based on the cues provided.The trial concluded with a 2-second break period.
The EEG data collected from the 46 subjects under this paradigm is collectively referred to as the "motor direction dataset", providing a rich resource for subsequent analysis.

C. Data Preprocessing
To mitigate the inherent challenges associated with low signal-to-noise ratio (SNR) and artifacts in EEG signals [37], [39], we implement a rigorous preprocessing protocol on the EEG data from our 'motor direction dataset'.This preprocessing aimed to refine data quality and eliminate extraneous artifacts or noise, as illustrated in Fig. 3. Utilizing EEGLAB (https://sccn.ucsd.edu/eeglab/index.php), a widely recognized toolbox in Matlab, the preprocessing encompassed the following main steps: (1) Select data: The NeuSen W series 64-channel EEG acquisition system used in our experiment includes 59 EEG channels, four electrooculography (EOG) channels, and one electrocardiography (ECG) channel.Initially, we eliminate the unrelated EOG and ECG channels to focus solely on EEG data.
(2) Re-reference: We apply the common average reference (CAR) method for re-referencing, aiming to diminish the influence of the reference electrode and enhance the independence and comparability of signals from each electrode [40].
(3) Filtering: A finite impulse response (FIR) bandpass filter ranging from 0.5 to 40 Hz is employed to remove highfrequency noise, along with notch filtering to eliminate 50 Hz power line interference.These steps constitute our batch preprocessing routine.Subsequently, an independent component analysis (ICA) is conducted, necessitating individualized analysis for each participant.
(7) ICA: ICA is employed to extract and remove components attributable to non-brain sources such as eye movements and muscle artifacts, further purifying the EEG signals [41].
The effectiveness of our preprocessing protocol is visually apparent in Fig. 3, where the enhanced purity of the EEG signals post-preprocessing is displayed.

D. Benchmark Algorithms
To preliminarily evaluate the effectiveness of our motor direction dataset, we employed four benchmark machine learning algorithms widely recognized in the field of MI.These include FBCSP [41], EEGNet [23], deepConvNet [24] and FBCNet [30].Each of these algorithms has been extensively validated in MI research and offers a robust framework for assessing the dataset's utility in decoding MI tasks.
• FBCSP: The FBCSP algorithm is renowned for its proficiency in MI decoding tasks.Its strengths lie in its capacity to capture relevant information across multiple frequency sub-bands and optimize spatial filtering.• deepConvNet: Utilizing deep network structures, deep-ConvNet excels in learning complex hierarchical features from EEG data.
• FBCNet: FBCNet introduces a novel variance layer that efficiently aggregates time-domain information from EEG.Each algorithm brings unique strengths to EEG data analysis and motor imagery decoding, providing a comprehensive assessment framework for the motor direction dataset.

E. Proposed Architecture: Multi-View Convolutional Self-Attention Network
Inspired by the architecture of FBCNet, we propose a new network structure, termed the MVCA.MVCA is specifically designed to effectively extract temporal information across different frequency bands from EEG data.The architecture of MVCA, as illustrated in Fig. 4, is fundamentally comprised of two primary components: the multi-view convolutional block and the multi-head self-attention module (MSA).The multiview convolutional block is a crucial part of the MVCA architecture, drawing inspiration from FBCNet's architecture.But we replace the temporal variance layer with a temporal convolutional layer.This modification aims to enhance the network's capacity to process temporal dynamics in EEG data.
Additionally, we have fine-tuned the parameters of this block to tailor it to better suit our specific requirements for EEG signal analysis.Moreover, by leveraging the selfattention mechanism, the MSA is capable of extracting and emphasizing critical temporal information within the EEG signals [41].The MVCA architecture, through these stages, is designed to effectively process and analyze EEG data, particularly focusing on the extraction of meaningful temporal patterns and dynamics.At its core, the MVCA architecture is composed of the following stages: (1) Multi-View Convolutional Block: The network process preprocessed single-trial data, denoted as x ∈ R C×T , alongside its corresponding label y ∈ {0, 1, . . ., N c }, where, C represents number of EEG channels, T represents time points and N c represents number of distinct classes.It is obtained from prior knowledge that MI related information in EEG data primarily resides within the mu (8-12Hz) and beta (13-30Hz) frequency bands [3], the network initially creates a multi-view representation of the EEG data using N b filter banks.This results in the data entering the spatial convolution stage with a size of (N b × C × T ).
Subsequent to the filter bank operation, the data undergoes spatial convolution to extract relevant spatial information.This involves a deep convolution layer with a kernel size of (C, 1), tailored to learn spatial patterns specific to each frequency band.Here, each frequency band is convolved independently, with m spatial filters applied, resulting in an output feature size of (N b × m × T ).Batch normalization follows, applied along the feature map dimension.
The final component in this block is the temporal convolution, which addresses the critical task of extracting temporal features from EEG data.For this purpose, a convolution layer  (2) Multi-Head Self-Attention Block: The MSA block within the MVCA framework plays a pivotal role in extracting discriminant information from EEG data.The MSA leverages location coding to capture essential temporal information, making it particularly effective for analyzing motor direction EEG data.This mechanism allows the model to understand the positional relationships between different time points better than traditional CNN approaches.Moreover, the MSA's ability to establish flexible weight connections across various time steps enables it to capture long-term dependencies more effectively [42].
In our study, the features obtained post-spatial convolution (with size (N b × m × T )) are fed into the MSA block.This process is executed separately for each frequency band to extract temporal features.The core components of the MSA, query (Q), key (K), and value (V), are derived through linear mappings of the input vectors, as follows: (1) where, represent the weights for the Q, K, and V respectively, with SC ( out) denoting the output features post-spatial convolution.The number of heads, h, is set to 8 based on empirical considerations.Consequently, the channel features extracted from each frequency band are processed through an 8-head self-attention mechanism to extract temporal features.Attention scores are calculated using the Q and K matrices: The output of the scaled dot-product attention is then derived by multiplying the attention scores with the value matrix: Through this structured approach, the MSA block efficiently processes the EEG data to distill crucial temporal features, enhancing the overall efficacy of the MVCA network in EEG signal analysis.
(3) Classification block: Finally, the output of multi-view convolutional block and multi-head self-attention blocks are connected and input to the full connection (FC) layer.The output of the FC is then fed to the softmax layer to get the output probability for each class.

F. Training Procedure
In this study, we employed the PyTorch library for implementing all machine learning methods, utilizing Python version 3.7.The computational processing was accelerated using a Geforce 3090 GPU.To ensure consistency in our evaluation, the same training procedure was adopted for all deep learning algorithms.The log-cross-entropy loss was used for gradient updates.A twostage training strategy was used [25]

III. RESULTS
This section presents a detailed analysis of the motor direction dataset classification results obtained using benchmark algorithms in the MI-BCI field.Our objective is to validate the discriminability of the four tasks in the motor direction dataset and the effectiveness of the proposed architecture.

A. Multi-Class MI Classification
1) Benchmark Algorithms: The present study validates the motor directioin dataset using four advanced algorithms in the MI-BCI field, including FBCSP, EEGNet, deepConvNet, and FBCNet.Each algorithm was configured according to the optimal settings recommended by their respective authors.These algorithms are utilized to assess the performance and effectiveness of decoding MI-EEG in the specific tasks and have been demonstrated to be effective.This is exemplified in Fig. 5, which displays the 10-fold CV accuracies for all subjects.Additionally, Fig. 6 provides box plots illustrating the average classification accuracies for each of the four algorithms.Our analysis indicates that all four algorithms demonstrated commendable performance on the motor direction dataset.Notably, FBCNet consistently outperformed the other three in terms of classification accuracy.
A detailed observation reveals that FBCNet shows the best classification performance for some participants, as shown in Fig. 5.However, it's important to note that EEGNet, while effective, exhibits comparatively lower performance for certain participants.This leads to an overall reduced classification performance, especially when contrasted with the traditional FBCSP+LDA method.These results underscore the varying strengths and limitations of each algorithm in decoding MI-EEG data, contributing valuable insights into the suitability of these methods for specific MI tasks.

TABLE I AVERAGE CLASSIFICATION ACCURACY
The performance analysis of FBCNet, which achieved the highest average classification accuracy (39.18%) on the motor direction dataset, revealed notable results.Among the 46 subjects, 34 attained a classification accuracy exceeding 28.5%, the established chance level [43].A deeper investigation into the eight participants who achieved classification accuracies above 55% yielded intriguing insights.According to the data gathered from pre-and post-experiment questionnaires and feedback, a common factor among these participants was their MI strategy during MI tasks.All eight reported imagining their own hands executing the directional movements.Notably, they consistently repeated this imagined task 2 to 4 times within each 4-second MI trial.This pattern of consistent and specific mental engagement seems to have played a significant role in enhancing their classification accuracies, highlighting the critical influence of mental strategy and consistency in MI task performance.This pattern of consistent and specific MI strategy seems to have played a significant role in enhancing their classification accuracies, highlighting the critical influence of MI strategy and consistency in MI task performance.
These findings not only underscore the scientific validity of our paradigm design but also affirm the effectiveness of the motor direction dataset.The high classification accuracy achieved by FBCNet on this dataset suggests potential for further refinement of network architectures based on FBC-Net, aiming to elevate classification performance even more.Additionally, the robust performance demonstrated by the three deep learning algorithms, especially in their ability to decode multiple directions information in unilateral limbs from EEG signals, further validates the efficacy of CNN-based approaches in this domain.
2) Multi-View Convolutional Self-Attention Network: Table I presents complete classification results for motor direction datasets using all the methods.In addition to the benchmark algorithms (FBCSP, EEGNet, deepConvNet, and FBCNet), the MVCA is compared with several state-of-the-art methods.These include EEG-TCNet, introduced in 2020 to address resource-intensive requirements in BCI systems [44], the attention-based temporal convolutional network (ATCNet) which is an interpretable model with fewer parameters [45], and EEG-conformer, which has demonstrated promising results in MI and emotion recognition paradigms [46].
Following the analysis of classification accuracies averaged over all subjects, the performance of every subject is investigated.Table II presents complete classification results for all data (46 subjects) using FBCNet (the best classification accuracy in baseline methods) and MVCA.Here, 42 subjects achieved classification accuracies higher than the 28.5% chance level [43].Notably, MVCA matches or surpasses the best-performing method for most subjects, leading to the highest average classification accuracy.These results demonstrate MVCA's suitability for decoding the unilateral MI of multitasks.

B. Binary MI Classification
Our work provides a large dataset and analytical basis for research in decoding multiple directional information from EEG.
Fig. 7 shows the classification accuracies for pairwise combinations of the four imagined movement directions, as determined by the MVCA method.Notably, most subjects achieved classification accuracies above 56%, thereby validating the effectiveness of the motor direction dataset, given that 56% is identified as the chance level.An intriguing observation from this analysis is the contrast between our initial expectations and the actual results.While we initially hypothesized that the left-right and top-bottom directions would be the most distinguishable, the data revealed that the top right-bottom left and top left-bottom right directions are more effectively differentiated.This trend was not only evident in the MVCA method but also consistently observed across all methods, as detailed in Table III.
These findings prompt further investigation into the underlying reasons for this hypothesis.One plausible explanation is the significant distinction in the left-right dimension exhibited by these two tasks, potentially contributing to their higher classification accuracies.This leads us to hypothesize that EEG signals may exhibit increased sensitivity to MI in the left-right dimension.Future research will aim to explore this hypothesis in greater depth, seeking to understand the neurophysiological underpinnings of this phenomenon and its implications for the design of MI-BCI systems.

C. Ablation Study
In this section, we perform an ablation analysis to measure the effectiveness of each block in the MVCA model.Table IV presents the impact of removing one block in the MVCA model on the performance of MI classification using the motor direction dataset.The results showed that the MSA block increases the overall accuracy by 1.35%, temporal convolution by 1.04%, and spatial convolution by 0.45%.All introduced blocks contributed to the improvement in classification accuracy, especially the MSA.This preliminarily demonstrates the importance of temporal features for the motor direction dataset.

D. Parameter Sensitivity
To further validate the effectiveness of the MSA block in MI-EEG decoding, we analyze the module performance on the motor direction dataset with different hyperparameters.Fig. 8 compares the performance of the MSA block with a different number of heads using dimension sizes of 8 and 16.The results showed that using two heads each of size 8 gave the best results.With the increase of the number and dimension of heads, the training time of the model also becomes longer,  8. Accuracy on the motor direction dataset as a function of the number of attention heads using head sizes of 8 and 16.Reducing head size as well as the number of heads showed better performance due to the small size of the dataset and its light representation.
but the classification performance is not improved.This maybe due to the MI-EEG has a limited number of samples, which requires a light MSA layer to converge well.The number of convolution kernels has a significant effect on the performance of CNN models.If the number of convolution kernels is too many, the complexity and computational cost of the model will be increased, which will easily lead to overfitting.In contrast, if the kernel is too few, it may not be possible to adequately capture all useful features in the data.By adjusting the number of convolution kernels of the two convolution modules, the model performance and training process are further optimized.As shown in Fig. 9, the model has the best classification performance when the number of convolution kernels of two convolution kernels is equal and both are 32 (m = n = 32).Additionally, as can be seen from Fig. 9, when the number of convolution kernels increases, the classification performance of the model shows a declining trend.The experiments demonstrate that too many convolution  kernels do not improve the classification performance of the model due to the small sample size of EEG data.

E. Visualization
1) Feature Distribution: The t-distributed stochastic neighbor embedding (t-SNE) is a popular visualization method.To further illustrate the benefits of the MVCA method, we have provided a feature distribution plot for S5 after adequate training using both FBCNet and MVCA.This is depicted in Fig. 10.We can observe that under the FBCNet network, the features of different classes are relatively close to each other, as in Fig. 10(a).In contrast, under the MVCA network, the distances between different classes are more pronounced, as in Fig. 10(b).In other words, the features extracted by the MVCA show greater differences between different classes, effectively capturing discriminative information and enhancing the separability between classes.
2) Topography Map: In our study, we conduct a comprehensive analysis of the spatial patterns derived from the CSP feature pairs of all subjects' data, as illustrated in Fig. 11.This figure presents the CSP brain topography maps: the first row depicts the average maps across all participants, while the second and third rows focus on individual subjects with varying classification accuracies.
From the first row of Fig. 11, encompassing data from all subjects, we observe notable activations around the left motor cortex.This is in line with the expected neurophysiological response during right upper limb MI tasks, where an increase in electrical activity intensity in the left brain region is typically observed.The CSP maps reflect this through positive values in these areas, whereas the right brain region shows relatively lower activity intensity, indicated by negative values on the maps.Interestingly, our analysis revealed task-specific differences in the activation areas, aligning with findings from previous studies [47], [48].These spatial patterns thus reinforce the neurophysiological validity of the selected CSP features in our dataset.
In a more detailed examination, as shown in the second and third rows of Fig. 11, we compare the topographic maps of two participants -one with high classification accuracy (S08, 96.5%) and another with lower accuracy (S14, 29.25%).For subject S08, the spatial patterns display distinct and discriminative activations for each of the four imagined movement directions, underscoring their high classification performance.In contrast, subject S14's topographic maps lack this level of discriminative clarity in the spatial patterns.Additionally, our analysis, as corroborated by the findings in Fig. 11, reveals that all participants exhibited the event-related desynchronization/synchronization (ERD/ERS) phenomenon during the right upper limb MI tasks.This phenomenon further highlights the variability in brain region activations among different tasks and between participants.Such variability underscores the significant differences in classification accuracy across subjects and algorithms, emphasizing the complexity of decoding multiple directions in unilateral limbs from EEG signals.

IV. DISCUSSION
In this paper, we design a novel MI-BCI experimental paradigm for unilateral limbs with four imagined movement directions.Through this paradigm, we successfully acquired MI-EEG data from 46 participants, thereby compiling a large dataset for in-depth analysis.To our knowledge, this dataset is the first of its kind to focus specifically on unilateral limbs with four different imagined movement directions.A significant finding of this study is the demonstration that MI of multiple directions in unilateral limbs can be effectively decoded, which marks a notable advancement in the MI-BCI field.Furthermore, we develop a new network architecture, specifically optimized for the motor direction dataset.

A. Advantages of Motor Direction Dataset
The emergence of ERD/ERS features induce by bilateral limbs MI tasks exhibit distinct spatial distribution patterns in the brain, and they align with the functional cortical motor areas corresponding to the involved limbs.Therefore, traditional MI-BCI systems focus on recognizing MI patterns for Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
bilateral limbs.However, as MI-BCI research evolves, incorporating unilateral limb MI tasks, challenges arise, including action complexity, similarity, and overlapping spatial neural activities.To address these issues, we design a unilateral limbs MI paradigm involving four imagined motor directions.Through visualization and classification accuracy analysis, we demonstrate the distinctiveness among these four tasks.Compared to the traditional paradigm, the motor direction paradigm focuses on unilateral limb, addressing rehabilitation needs.It is known that differentiating multiple MI tasks from one side of the limb is technical challenging and this study provide an attempt to show the technical feasibilities.Moreover, this new paradigm creates opportunities for future research to explore and understand the intricate relationship between EEG signals and directional information.The motor direction dataset serves as a valuable tool for developing decoding algorithms targeting motor direction from non-invasive neural signals, potentially advancing MI-BCI paradigms and enhancing MI control command dimensions.

B. Performance of Classical Machine Learning Methods and State-of-the-Art Deep Learning Architectures
To better design decoding algorithms suitable for the motor direction dataset, a further comparison of the benchmark algorithms was performed from the perspective of network architecture.From Fig. 6, FBCNet emerged as a frontrunner, with an average classification accuracy of 39.18% on the motor direction dataset.Compared to other networks, the first advantage of FBCNet is the incorporation of a multi-view data representation module.This structure of multi-view data representation is similar to the concept of FBCSP.The main idea is to locate the most relevant mu and beta frequency bands associated with MI from all frequency bands.The most crucial aspect is utilizating of the novel variance layer to extract temporal information from the signals.Indeed, such a design facilitates the extraction of comprehensive signal features, enabling better capture of both temporal and frequency domain information from the EEG signals, consequently improving the model's performance.As illustrated in Table I, EEG-TCNet demonstrates superior classification performance on this dataset.Compared to typical CNN, TCNet has larger receptive fields, allowing them to learn more information.And, it's a network architecture that's particularly well-suited for handling time-series data.Both FBCNet and EEG-TCNet have achieved excellent accuracy, it can be concluded that the motor direction dataset contains valuable information in both temporal and spectral domains.This discovery provides a foundation for the development of straightforward and effective decoding algorithms for this dataset in future research.

C. Directions Information From EEG
The motor direction dataset shows consistently better classification results for the top right-bottom left and top left-bottom right tasks across different algorithms (Fig. 7 and Table III).Therefore, we propose the hypothesis that the EEG signal is more sensitive to information in the left-right dimension.Additionally, the differences in classification results for leftright and top right-bottom left tasks further emphasize this sensitivity.These findings demonstrate the potential of EEG signals to decode a rich array of directional information, providing novel insights for future MI paradigm and control command design.

D. Future Directions
Although we achieves promising classification results on the motion direction dataset, there are still several rooms for further study.Enhancing decoding algorithm performance is paramount, potentially through combining CNN and transformer models, which have shown success in various fields [46].Thus, in the future study, we will focus on decoding MI of multiple directions in unilateral by this model.Secondly, in this study, we provide preliminary evidence and propose the hypothesis that EEG is more sensitive to information in the left-right dimension.We will explore the strong evidences of decoding multiple directions from EEG and provide explanations for the proposed hypothesis from neurobiological perspective.Finally, in the context of stroke rehabilitation, developing experimental paradigms encompassing a broader range of directions and enhanced control commands for external devices remains a crucial objective.

V. CONCLUSION
In this research, we successfully developed and implemented a novel MI-BCI experimental paradigm focusing on the unilateral limbs with multiple movement directions.This paradigm was validated through an experiment involving forty-six subjects.We employed widely recognized machine learning techniques to provide a comprehensive assessment of the paradigm's effectiveness.A significant advancement in this study is the introduction of the MVCA model, specifically designed to enhance the decoding accuracy of the motor direction dataset.Further analysis of the data, particularly in two-class scenarios, revealed that the right top-left bottom and left top-right bottom directions yielded the highest accuracy.This finding offers valuable insights for future MI-BCI research and applications.The study's results contribute significantly to advancing the MI-BCI field, providing preliminary yet compelling evidence for the feasibility of decoding complex unilateral limb movements from EEG signals.

Fig. 1 .
Fig. 1.Electrode Placement Diagram.(a) 10-20 International electrode placement system.(b) Diagram of electrode placement positions on an EEG cap used in this experiment.Where, the numbers inside the circles represent the impedance values before the experiment (unit: KΩ), The green box represented electrodes in the motor area, while the red box represented electrodes in the visual area.
(b), comprised two main phases: a resting phase and a MI task phase.The resting phase alternated between one minute of eyes open and one minute of eyes closed, intended to provide a baseline for

Fig. 2 .
Fig. 2. Experimental flowchart and paradigm.(a) is the main process included pre and post-experiment questionnaires and MI experiments.(b) presents the two stages of MI experiment.(c) illustrates the timeline of a single trial.

Fig. 4 .
Fig. 4. Network architecture of the MVCA.(C: number of EEG channel, T: number of time points, N b : number of frequency bands, m: number of convolution kernels per frequency band in spatial convolution, n: number of convolution kernels per frequency band in temporal convolution, w: the dimension of the temporal convolution).
with a kernel size of (1, w) is employed.The convolutional process in the block transforms the input single-trial data size from = (C ×T ) to an output feature size of = (N b ×m×T /w).In our study, we have configured the parameters as follows: C = 59, T = 1000, N b = 9, m = n = 32, w = 64.

Fig. 5 .
Fig. 5. Classification accuracy for each subjects in 10-fold CV setting (sorted by FBCSP+SVM accuracy).Here, in contrast, FBCNet matches the performance of the best performing method for most of the subjects resulting in the best subject averaged classification accuracy.
, and the training data is further split into training and validation sets.Specifically, the training, validation, and test sets are divided in an 8:1:1 ratio.Specifically, during the training process, 80% of the data is used for training the model, 10% is used for validation to tune hyperparameters and monitor performance, and the remaining 10% is reserved for testing to assess the final model's generalization capability.In the first stage of training, the model is trained only on the training set, and the validation set accuracy is monitored.If the validation set accuracy does not increase for consecutive 200 epochs, the training is stopped and the network parameters with the best validation set accuracy are restored.Then, starting from this model, proceed to the second stage.In this stage, the model is trained using the training set and the validation set (training data).The second stage training is stopped when the validation set loss reduced below the first stage training set loss.And the maximum number of training epochs are limited to 1500 and 600 for training stage 1 and 2 respectively.Cross-validation (CV) is used to evaluate the performance of all algorithms.

Fig. 7 .
Fig. 7.The MVCA cross validation classification accuracies about two different tasks in different combinations.From left to right: left-right and top-bottom, left-right and top right-bottom left, left-right and top leftbottom right, top-bottom and top right-bottom left, top-bottom and top left-bottom right, top right-bottom left and top left-bottom right.The red line represented the chance level of the two-classes: 56%, and the gray diamonds represented the classification result of each subject.

Fig. 9 .
Fig. 9.The influence of the number of convolution kernels on the accuracy for the motor direction dataset.m: number of convolution kernels per frequency band in spatial convolution, n: number of convolution kernels per frequency band in temporal convolution.

Fig. 11 .
Fig. 11.The topography map of EEG signals for the four imagined movement directions.Where, red represented positive values (activated), while blue represented negative values (inhibited).

TABLE II CLASSIFICATION
ACCURACIES FOR EACH SUBJECT IN MOTOR DIRECTION DATASET

TABLE IV CLASSIFICATION
PERFORMANCE IN % OF MVCA COMPARED TO MVCA-WITHOUT SPATIAL CONVOLUTION, MVCA-WITHOUT TEMPOTAL CONVOLUTION AND MVCA-WITHOUT MSA Fig.