Decoding multi-class motor attempt from the affected unilateral limbs in chronic stroke patients

Abstract

Current brain–computer interface (BCI) research predominantly focuses on decoding bilateral limb movements, whereas practical stroke rehabilitation typically involves unilateral upper limb control. Decoding unilateral multi-task motor attempts remains challenging due to overlapping cortical activations across different movement categories. To address this, we propose a Multi-View Cortical Muscle Graph Network (MVCMGNet), a novel architecture grounded in neurophysiological principles. MVCMGNet enhances decoding through three integrated approaches: multi-view graph convolutional block extracts discriminative spectral-spatial features; cortical muscular connection module identifies distinctive connectivity signatures for different movements; and a mixture of experts module robustly fuses these multi-modal features. Evaluated on a dataset comprising 45 chronic stroke patients, MVCMGNet demonstrated strong performance in the dual-task scenario, achieving a classification accuracy of 78.52%, while exhibiting moderate performance in the more challenging quad-task scenario, attaining a classification accuracy of 52.79%. This study demonstrates that cortical-muscular connectivity patterns can effectively decode multi-category motor attempts in unilateral limbs. Our findings confirm that the distinctiveness of action-specific neuromuscular patterns enhances decoding accuracy, providing valuable insights for future BCI research and supporting the feasibility of complex unilateral decoding tasks for clinical rehabilitation.

Introduction

Motor impairment is one of the most common and persistent functional deficits following stroke, significantly affecting patients’ independence and quality of life [1]. Although existing rehabilitation interventions can promote functional recovery to some extent, they often require lengthy treatment courses and demonstrate limited efficacy [2]. Therefore, how to further enhance rehabilitation efficiency and outcomes remains a critical issue in stroke rehabilitation that urgently requires resolution [3]. In recent years, brain-computer interface (BCI) technology has demonstrated substantial potential in post-stroke motor function rehabilitation due to its versatility and broad applicability, with evidence suggesting improvements in motor recovery and rehabilitation outcomes [4].

Nevertheless, BCI technology faces significant challenges in clinical translation [5, 6]. On the one hand, the prevailing emphasis on decoding bilateral tasks stands in stark contrast to the unilateral upper-limb impairments commonly observed in stroke rehabilitation [7]. This discrepancy hinders the practical application of BCI technology in stroke recovery. On the other hand, the scarcity of unilateral data from stroke patients, together with the high spatial similarity of cortical activation patterns across different motor tasks, poses major challenges for decoding unilateral multi-task electroencephalography (EEG) data [8].

For most existing EEG-based decoding models, particularly those based on convolutional neural networks (e.g., EEGNet [9], EEG-Conformer [10]), extracting effective features from overlapping cortical activation regions remains challenging [11]. These models often struggle to distinguish subtle spatial and frequency-domain differences among unilateral motor tasks, especially when the underlying neural representations exhibit strong bilateral overlap [8]. The lack of sufficiently large and task-diverse unilateral EEG datasets further exacerbates this problem, as most publicly available datasets—such as BCI Competition IV Dataset 2a [12], PhysioNet Motor Movement/Imagery Dataset [13], and the Graz Dataset [14]—primarily focus on bilateral or symmetric motor paradigms. Consequently, model generalization and decoding performance in stroke-related unilateral applications remain limited.

Regarding spatial features, recent studies have explored more refined feature extraction strategies by adopting more expressive network architectures, such as Transformer-based models and attention mechanisms, to capture higher-level and long-range spatial dependencies beyond those represented by traditional handcrafted or shallow learning approaches [15–17]. However, for bilateral decoding tasks, increasing the complexity of spatial modeling can substantially elevate model complexity, which may reduce sensitivity to hemispheric asymmetries and consequently hinder effective discrimination between left and right brain regions [18]. In contrast, the core of unilateral multi-class decoding lies in extracting subtle spatial differences between tasks within the same brain region, imposing greater demands on model design. Building on this approach, Rong et al. proposed a multi-view convolutional self-attention network to decode unilaterally multi-directional tasks, achieving a classification accuracy of 40.6% [8]. Similarly, Wang et al. combined self-attention modules with temporal convolutional blocks to capture EEG’s intricate spatiotemporal features and learn intrinsic differences, reaching a classification accuracy of 51.58% on data from 13 healthy volunteers [19].

Regarding frequency-band features, different frequency bands contain complementary information [20, 21]. Building on this insight, some studies have attempted to enhance unilateral multi-class decoding performance by extracting multi-frequency features. For example, Qiu et al. optimized a convolutional neural network based on FBCNet to decode unilateral upper-limb motor imagery (MI) tasks, including hand grasping and releasing, achieving a classification accuracy of 67.8% on two-class MI-EEG data from 10 healthy volunteers [22]. Ma et al. achieved the highest accuracy using FBCNet in a binary classification decoding task for the left and right hands of 25 stroke patients [23].

Each muscle movement exhibits a unique pattern of activated muscle fibers, enabling the classification of movement attempts through multi-channel electromyography recordings [24]. Soto-Florido et al. implemented a combined EEG and electromyography (EMG) classifier to detect four attempts in chronic stroke patients: rest, grasp, hold, and release [25]. Li et al. proposed a sequential learning model incorporating graph isomorphism networks (GINs) to process sequential graph-structured data derived from EEG and EMG signals [26].

Regarding task activation patterns, distinct motor tasks exhibit differences in both brain region activation and muscle activation [27]. Investigating these differences through functional connectivity represents another effective strategy to improve unilateral multi-class decoding accuracy [28]. Yang et al. proposed a functional magnetic resonance imaging (fMRI)-weighted convolutional neural network, in which channel weights were adaptively redistributed according to task-related brain activation patterns derived from fMRI scans, thereby enhancing classification performance [29]. Their visualization results revealed similarities in brain activation areas across grasping, extension, and rotation tasks, while highlighting distinct differences in the most prominent activation centers, providing direct evidence for the critical role of brain region activation in improving unilateral multi-class decoding.

Notably, several scholars have proposed that functional connectivity patterns between EEG and EMG signals could facilitate more effective decoding. Xi et al. constructed a brain–muscle functional network based on human motor mechanisms, integrating brain functional networks with corticomuscular coherence (CMC) to validate the information exchange between the brain and muscles during movement [30]. This approach demonstrated significant advantages and strong potential for decoding human motor attempts (MAs). Similarly, Yang et al. developed various graph-structured networks by integrating EEG and EMG functional connectivity [31]. Their results showed that graph networks improved robustness against anomalous or fatigued data and that multimodal fusion of EEG and EMG yielded higher decoding accuracy compared with unimodal approaches.

However, despite these advances, based on our systematic examination of existing studies, two important limitations can be identified: (1) most reported EEG–EMG datasets for motor decoding are small in scale and are primarily collected from healthy subjects, with very limited unilateral datasets from chronic stroke patients involving multiple MAs [8, 32], and (2) existing approaches typically focus on either spatial, spectral, or cortico–muscular connectivity features in isolation, with few studies attempting to integrate these aspects into a unified framework for unilateral multi-task decoding [23, 33]. Addressing these limitations is crucial for promoting the clinical translation of BCI-based stroke rehabilitation.

Building upon these findings, and to address the inherent challenges of unilateral limb multi-task decoding, this study introduces the Multi-View Cortical Muscle Graph Network (MVCMGNet)—an algorithm specifically designed for decoding affected limb motor tasks in chronic stroke patients. The MVCMGNet architecture comprises three key components: a multi-view graph convolutional block, a cortical–muscle connection module, and a Mixture of Experts (MoE) module. The multi-view graph convolutional block enhances the extraction of frequency-band information while preserving the spatial feature representation capabilities of graph convolutions [34]. The cortical–muscle connection module improves classification performance by decoding brain–muscle activation patterns for different actions via CMC [30, 35]. The MoE module aggregates multi-feature representations to reduce information redundancy and enhance overall decoding [36, 37]. Experimental results demonstrated that MVCMGNet effectively improved unilateral multi-task decoding accuracy, underscoring its significance in exploring motor control mechanisms in the central nervous system.

Equally noteworthy, this study established a large-scale dataset by synchronously collecting EEG and EMG signals from 45 chronic stroke patients performing four unilateral MA tasks—representing one of the largest sample sizes to date in EEG-based decoding research. The main contributions of this work are summarized as follows:

1. We established a unilateral upper-limb MA dataset comprising four MA tasks from 45 chronic stroke patients.

2. We proposed MVCMGNet, a novel framework that enhances decoding performance for unilateral multi-class tasks by integrating frequency-band features, spatial features, and brain–muscle activation patterns.

3. Our visualization results revealed distinct differences in action-specific activation patterns across unilateral multi-class tasks, providing valuable insights for future research in MA decoding.

The remainder of this paper is organized as follows: Sect. 2 describes the experimental paradigm and data processing; Sect. 3 details the proposed model architecture; Sect. 4 presents the experimental results; Sect. 5 provides a deeper analysis of MVCMGNet’s outcomes and visualizations; and Sect. 6 concludes the paper.

Materials

Subjects

A total of 50 post-stroke individuals aged between 28 and 72 years participated in this study. The cohort included 49 males and 1 female, with 29 participants being left-handed and the remainder right-handed. The summary table of patient characteristics is shown in Table 1. The complete patient information form is presented in the appendix. Muscle strength level is assessed using manual muscle testing (MMT) [38]. The inclusion criteria are as follows: (1) diagnosis of cerebral hemorrhage or cerebral infarction, with disease duration greater than 6 months and in a stable stage; (2) clear consciousness and ability to comprehend all experimental procedures; (3) no prior experience with BCI; (4) absence of severe epilepsy, major complications, or contraindications; and (5) provision of informed consent by the participants or their legal caregivers. The experiment was conducted in a quiet, enclosed, and controlled laboratory environment under the supervision of trained personnel to ensure data quality and reliability. The study protocol was reviewed and approved by the Institutional Ethics Committee of Beijing Tiantan Hospital (Approval No. KY2023-205-03).

Table 1.

Demographic and clinical characteristics of the chronic stroke patients. Q1 and Q3 represent the first (25th percentile) and third (75th percentile) quartiles, respectively

Characteristics	Values (n = 50)
Age (years)	28–72 (range)
Age distribution (Median [Q1, Q3])	47 [41, 56]
Sex (male/female)	49 / 1
Handedness (left/right)	29 / 21
Lesion side (left/right)	26 / 24
Stroke type (hemorrhagic/ischemic)	35 / 15
Time since stroke (months)	6–36 (range)
Time since stroke distribution (Median [Q1, Q3])	9.5 [7, 12]
Patients within 1 year post-stroke	40 (80%)
Lesion location	basal ganglia (88%), brainstem (6%), cerebellum (4%), thalamus (2%)
Muscle strength level (Median [Q1, Q3])	2 [1, 2]
Clinical status	Stable stage
Cognitive or perceptual deficits	None severe enough to affect task performance

EEG data were collected using a 32-channel EEG acquisition system (Neuracle), and EMG data were recorded from the affected limb via a 4-channel EMG system (Neuracle). Both systems were integrated within the same platform, enabling synchronous EEG–EMG data acquisition. EEG electrodes were positioned according to the international 10–10 system (Fig. 1A) [39], with electrode impedance strictly maintained below 5 kΩ throughout the experiment. To ensure optimal recording quality, participants were seated comfortably in a chair and instructed to remain relaxed, minimizing unnecessary eye movements during EEG acquisition. Both EEG and EMG signals were sampled at 1000 Hz to ensure high temporal resolution.

Fig. 1.

Schematic representation of the experiments. A. Electrode placement for EEG and EMG. B. Example of action presentation, in which fist clenching is shown as a looping image sequence. C. Experimental paradigm flow

Experimental protocol

Research has demonstrated the presence of two distinct and structurally defined motor systems, responsible for the control of proximal and distal upper-limb movements, respectively [40]. The present experimental paradigm is therefore designed to evaluate upper-limb motor function across multiple joint levels, ranging from distal to proximal, which is particularly pertinent for the assessment of post-stroke motor impairment and recovery. Accordingly, four unilateral upper-limb movements were selected: fist clenching (Fc), pinch grip (Pg), wrist lift up (Wlu), and elbow flexion (Ef). Together, these movements encompass the finger, wrist, and elbow joints, enabling a hierarchical evaluation of motor control from fine distal actions to gross proximal movements.

Fist clenching and pinch grip are selected to represent distal hand functions, which are known to be more severely impaired and slower to recover after stroke compared with proximal joints. These actions are fundamental for activities of daily living (ADLs), such as grasping, holding objects, and performing precision tasks, and are core components of standardized upper-limb functional assessments [41]. In particular, pinch grip requires precise coordination of intrinsic and extrinsic hand muscles and is highly sensitive to corticospinal tract damage, making it a discriminative task in both EEG spectral and cortico-muscular coupling analyses [42].

Wrist lift up (wrist extension) is selected as an intermediate joint movement bridging distal hand control and proximal arm function. Wrist extension plays a critical biomechanical role in stabilizing the hand during grasping and manipulation tasks and is therefore functionally relevant for ADLs. Moreover, wrist movements engage large superficial forearm muscles that are well suited for surface EMG recording and have been widely adopted in EEG–EMG coupling and CMC studies [43].

Elbow flexion is included to represent proximal upper-limb function, which is typically less impaired and recovers earlier than distal movements in stroke survivors [44]. Evaluating elbow flexion allows assessment of proximal motor pathways and complements distal joint analysis, enabling a comprehensive characterization of motor hierarchy and neural control strategies across joint levels.

Correspondingly, EMG signals were recorded from four representative muscles: brachioradialis (EMG1), wrist flexors (EMG2), wrist extensors (EMG3), and biceps brachii muscle (EMG4). These muscles are selected based on their primary functional roles in the chosen movements and their clear anatomical correspondence to the elbow–wrist–hand joint hierarchy. Specifically, the brachioradialis and biceps brachii are major contributors to elbow flexion [45], while the wrist flexor and extensor groups are essential for wrist stabilization and hand function during grasping and precision tasks [46–48].

The experimental structure consists of two main phases: a rest phase and an MA task phase (Fig. 1C). The rest phase involves 3 min of daydreaming, aimed at providing a baseline for EEG measurements. During the MA task phase, the movements are presented in a looping sequence of multiple images (Fig. 1B), and participants are instructed to concentrate and attempt to perform the affected-side movement as much as possible under the guidance of synchronized visual and auditory cues.

Each paradigm experiment consists of 10 trials, each lasting 10 s. Data collection lasts 5 s. After each data collection, participants have a 5-second rest period to maintain their physical condition throughout the experiment.

Data preprocessing

The EEG and EMG signals were processed separately, with the objective of enhancing data quality while preserving the original signal characteristics as much as possible. EEG signals were processed using EEGLAB (Version 2023.1), a widely recognized MATLAB toolbox (Version R2019a), while EMG signals were processed using the pyemgpipeline library (Version 1.0.0) in Python 3.10.0.

EEG signal processing

EEG signals were preprocessed using a standardized pipeline. First, a band-pass filter ranging from 1 to 40 Hz was applied to suppress slow drifts and high-frequency noise [49]. In addition, a 50 Hz notch filter was applied to remove power-line interference.

Subsequently, EEG data were segmented into epochs based on experimental event markers corresponding to individual trials. To further reduce low-frequency fluctuations and inter-trial offsets caused by fatigue and muscle tension, a baseline correction was performed using the 3 s pre-stimulus interval preceding each trial onset.

All EEG signals were then re-referenced to a bilateral mastoid reference, which has been shown to reduce the bias introduced by single-reference electrodes and to provide a more balanced representation of bilateral sensorimotor cortical activity.

Finally, independent component analysis (ICA) was conducted to identify and remove non-neural artifacts, including ocular and residual muscle-related components [50]. ICA was performed on 32 EEG channels, which provides sufficient spatial information for reliable source separation. On average, 4.1 ± 1.2 components per subject were removed, resulting in improved signal quality while preserving task-related neural activity.

EMG signal processing

EMG signals were preprocessed following a standard pipeline for cortico–muscular coherence analysis. First, direct current (DC) offsets arising from electrode–skin impedance mismatch or device-related noise were removed to eliminate baseline bias. Subsequently, a band-pass filter ranging from 20 to 450 Hz was applied to suppress motion-related low-frequency noise and high-frequency interference, while preserving the dominant frequency content of surface EMG signals associated with muscle activation [51].

The filtered EMG signals were then subjected to full-wave rectification, converting negative amplitudes to positive values. This step enhances the representation of muscle activation intensity and facilitates subsequent envelope-based analysis. To extract the temporal modulation of muscle activity relevant to neural control, a secondary low-pass filtering step with a cutoff frequency of 40 Hz was applied to obtain the EMG envelope [52].

Notably, the frequency components of the extracted EMG envelope are predominantly distributed within the 0–40 Hz range, which exhibits substantial spectral overlap with the preprocessed EEG signals (1–40 Hz). As a result, both EEG and EMG envelope signals submitted for coherence analysis share synchronized power within the same frequency band, ensuring that cortico–muscular coherence calculations are both methodologically feasible and physiologically meaningful.

Normalization

Following preprocessing, z-score normalization was applied to each dataset (EEG, EMG and CMC) independently to ensure that all features contributed to the model with comparable magnitudes. This step mitigates bias from amplitude variations, enabling balanced contributions of EEG and EMG features during model training [53].

1

where denotes the data before processing, denotes the mean of , denotes the standard deviation of , and denotes the data after processing.

After preprocessing and quality control, data from 45 chronic patients collected in this experimental paradigm were deemed usable, while data from 5 patients were excluded. Specifically, datasets were excluded if excessive motion artifacts or lapses in attention resulted in fewer than 80% of the planned trials remaining usable after preprocessing. Based on this criterion, data from 5 patients did not meet the minimum trial requirement and were therefore excluded. The processed dataset was named the “chronic stroke patient dataset,” providing a reliable resource for subsequent model development and analysis.

Methods

Inspired by the MSVTNet architecture [54], we propose a novel network structure termed the MVCMGNet. MVCMGNet is specifically designed to extract task-specific information from cortical-muscle connections, as illustrated in Fig. 2. The overall architecture comprises three main components: the multi-view graph convolutional block, the cortical muscle connection module, and the MoE module [36, 37]. The multi-view graph convolutional block, as a crucial component of MVCMGNet, draws inspiration from MSVTNet but replaces conventional convolutions with graph convolutions to enhance the network’s spatial representation capabilities across multiple views. Decoding multi-class unilateral limb movements primarily relies on the cortical–muscle connection module, which is specifically designed to capture and decode distinct cortical–muscular activation patterns associated with different limb movements. Finally, the MoE module aggregates diverse feature representations obtained from the preceding components, thereby reducing information redundancy and strengthening the overall discriminative power of the network.

Fig. 2.

The framework of MVCMGNet. Dot elements indicate feature omissions, “Mean” denotes averaging, and “Feature Fusion” represents feature merging

Multi-view graph convolutional block

Multi-view filtering block

The original EEG and EMG signals are represented as and , respectively, and the corresponding labels are represented as , where and denote the number of EEG and EMG channels, respectively, denotes the number of time points, and denotes the total number of categories.

Considering that task-relevant EEG information related to motor control is primarily distributed within the 4–40 Hz frequency range, we focus on canonical motor-associated rhythms for spectral decomposition [55]. In particular, theta (4–8 Hz), alpha (8–12 Hz), beta (12–30 Hz), and low-gamma (30–40 Hz) bands are selected due to their well-established neurophysiological relevance to sensorimotor processing [33, 56–59]. Four bandpass filters are applied to generate a multi-view representation of the EEG signals, enabling the model to capture complementary motor-related information across frequency domains. This filter bank design follows the conventional spectral band division adopted in the FBCSP framework and has been widely validated in prior EEG-based motor decoding studies for improving classification performance [60, 61]. The filtering process can be expressed as:

2

3

where denotes the bandpass filtering operation.

Following filtering, the multi-view data is processed using a Graph Attention Network (GAT) to extract spatially relevant information. Unlike bilateral classification tasks, unilateral multi-class movement decoding often requires distinguishing subtle activation differences across cortical regions on the same side. The GAT is therefore employed to adaptively learn spatial features from the underlying connectivity structure by assigning learnable, edge-specific attention weights to neighboring nodes, enabling the model to capture heterogeneous spatial dependencies [62].

Connectivity matrix

To construct the functional connectivity matrix, the weighted phase lag index (wPLI) is employed to quantify phase-based synchronization between EEG channels. wPLI is chosen because it effectively suppresses spurious zero-lag correlations caused by volume conduction and common reference effects, which are prevalent in multichannel EEG recordings [63–66]. It is defined as:

4

where represents the cross spectral density between signal i and signal j; denotes the imaginary part, and represents the expectation across time windows.

Graph attention network

The GAT adaptively aggregates neighborhood information by assigning learnable importance weights to adjacent nodes [67–69]. Given node features , the attention coefficient between nodes i and j is computed as:

5

where is a learnable weight matrix, is the learnable attention vector, denotes concatenation, is the neighborhood set of node , denotes the dimension of node features, and denotes the number of nodes. The definition of the set of neighbors is as follows:

where is the WPLI-based adjacency matrix and is the threshold (set to the top 20% strongest connections).

The updated node representations are then computed as:

6

where denotes the ELU activation function.

Cortical muscular connection module

From the perspective of action pattern connectivity, we propose a cortical–muscular connectivity module to capture differences in cortical control of muscle activity across various actions (Fc, Pg, Wlu and Ef). For unilateral multi-classification tasks, this network provides a more intuitive and effective way to distinguish same-side movements compared with traditional EEG-based approaches (e.g., EEGNet [9], EEG-Conformer [10]).

Corticomuscular coherence calculation

Corticomuscular connectivity reflects the degree of functional coupling between cortical activity and muscle movement output, and is a key indicator for measuring central motor drive capacity [70]. CMC refers to the spectral relationship between cortical and muscle electrophysiological signals, reflecting the degree of synchronization between cortical and spinal cord activity during motor execution [71].

Constructing sparse connection matrices

CMC indicators across different frequency bands often exhibit distinct trends [33]. In the cortical–muscular connectivity network, these differences are represented as connection matrices corresponding to different bands [30]. Specifically, the mean CMC value within each band is selected as the connection coefficient to construct four band-specific matrices, consistent with the multi-view convolutional structure.

To construct the functional connectivity matrix, a proportional thresholding approach is adopted in which only the strongest connections exceeding a predefined percentile are retained. The specific threshold level for the top 20% of connections is determined based on prior research, as this threshold effectively captures the dominant structure of functional connectivity while minimizing spurious weak edges in EEG/MEG network analysis paradigms [30, 72, 73]. Specifically, connections between channels are retained only when , where is the threshold determined by the top 20% of correlations. This yields four sparse adjacency matrices corresponding to the four frequency bands, denoted as , , , and .

Node feature extraction

The CMC matrices capture connectivity between EEG and EMG in the frequency domain. However, as graph convolutional networks (GCNs) rely primarily on node-level features, directly averaging the CMC values of each node would cause severe information loss [74]. To address this, we introduce a mapping operation to derive representative node features from connectivity:

8

where is the mapping matrix between node connections and node features.

Node feature update

After constructing the adjacency matrices, the corresponding Laplacian matrix () is computed, and graph convolutional operations are performed to propagate information between neighboring nodes:

9

where and denote the output and input of the node, respectively. The initial feature representation corresponds to the combined EEG and EMG features.

MoE module

The MoE module in MVCMGNet consists of two sequential MoE groups: feature fusion and frequency-band fusion. Feature fusion adaptively integrates multimodal information from three sources: EEG features, EMG features, and CMC features. Frequency-band fusion aggregates features across four frequency bands to capture complementary spectral information. In MoE module, each modality is processed by a dedicated expert network composed of fully connected layers, enabling modality-specific representation learning. To adaptively determine the contribution of each modality, a gating network is employed [37, 75, 76]. The gating network takes the concatenated feature vector as input and generates a set of normalized weights through a softmax operation. These weights are then used to compute a weighted sum of the expert outputs, formulated as:

10

where denotes the gating weight for the i-th expert, and represents the corresponding expert output.

This architecture enables MVCMGNet to dynamically emphasize the most informative modalities, suppress modality-specific noise or redundancy, and facilitate more effective multimodal information integration for motor decoding tasks.

Classification

After obtaining the fused features from the MoE module, these features are fed into a two-layer fully connected network, which serves as the classifier for final MA recognition.

Training procedure

In this study, all experiments are performed in the Pytorch framework (Version 1.12.1 + cu116), using the Python 3.10.0 environment, and accelerated using Nvidia RTX 3090 (24 GB, manufactured by NVIDIA Corporation). To ensure the reliability and robustness of model evaluation, we adopt a 5-fold cross-validation scheme with 10 independent repetitions using different random seeds [10, 77]. For each fold, the model is trained for 200 epochs, and the final results are obtained by averaging over all folds and repetitions. Given the inherently low signal-to-noise ratio of EEG data and the relatively limited dataset size, we employ the AdamW optimizer for training [78, 79]. Its parameters are set as follows: learning rate of 1 × 10⁻⁴, batch size of 64, with a warm-up strategy to stabilize the convergence process. Model performance is quantitatively assessed using average accuracy (Accuracy) and standard deviation (Std) across the subjects, where each subject’s accuracy is the average of the 5-fold cross-validation results. All 5s data and information from 32 channels are used for classification. Accordingly, the dataset (40 trials) for each subject is divided into 32 training data and 8 testing data. Crucially, the connectivity matrix of the graph network is constructed entirely based on the training data from each cross-validation fold, with no involvement of the test data in the graph construction process. This strategy aims to prevent information leakage and ensure a fair evaluation of the proposed model.

T-distributed stochastic neighbor embedding visualization

T-distributed Stochastic Neighbor Embedding (t-SNE) is a popular nonlinear dimensionality reduction technique for visualizing high-dimensional data [80]. In the specific implementation, perplexity is set to 5, the number of iterations is 200, principal component analysis (PCA) is chosen as the dimensionality reduction method, and the t-SNE function in the sklearn library (Version 1.4.1.post1) is used for computation.

Ablation analysis

To evaluate the contribution of each module to the model’s final performance, key components are sequentially removed for ablation experiments. Each experiment modifies only one module while keeping all others unchanged [81].

Statistical analysis

All statistical analyses are performed using Python 3 (Version 3.10.0). For each subject, decoding performance is first averaged across all folds and repetitions of the 5-fold cross-validation. Preliminary normality assessment (Shapiro–Wilk test) indicates that the performance metrics did not follow a normal distribution. Therefore, a non-parametric Wilcoxon signed-rank test is employed instead of parametric analysis of variance (ANOVA) to assess paired differences in decoding performance [10, 82]. Statistical significance is defined as p < 0.05.

Results

This section provides a detailed analysis of the classification results obtained using ten algorithms on chronic stroke patient datasets in the BCI field. Our objective is to validate the distinguishability of the four tasks (Fc, Pg, Wlu and Ef) in the chronic stroke patient dataset and the effectiveness of the proposed architecture.

Multi-class MA classification

This study validates the chronic stroke patient dataset using classification algorithms from the BCI field. In addition to benchmark algorithms such as EEGNet [9], LMDA-Net [83], FBCNet [61], EEG-Conformer [10], ATCNet [84] and CNN-GNN [34], MVCMGNet is also compared against several state-of-the-art methods, including MSVTNet, Labram [85], and Sparcnet [86]. Each algorithm is configured according to recommended optimal settings. These compared algorithms are employed to evaluate the performance and effectiveness of decoding EEG signals for specific tasks and have been demonstrated to be effective.

Table 2 shows the classification accuracy of all methods across the four categories. MVCMGNet achieves the highest average classification performance in 5-fold cross-validation analysis, outperforming the best-performing comparison method Conformer by 9.15% (Wilcoxon signed-rank test, p = 2.4710-2). This improvement is statistically significant, validating the efficacy of MVCMGNet relative to the comparison methods. Our analysis indicates that all algorithms demonstrate commendable performance on the chronic stroke patient dataset.

Table 2.

Subject-dependent classification accuracy (Mean/Std)

Method	Accuracy	Std	p-value
EEGNet	35.99	14.73	3.21 ✕ 10-4
LMDA-Net	37.88	12.79	1.12 ✕ 10-3
FBCNet	41.35	15.54	7.86 ✕ 10-3
Conformer	43.64	13.25	2.47 ✕ 10-2
ATCNet	43.16	16.39	4.91 ✕ 10-3
CNN-GNN	42.74	15.09	1.77 ✕ 10-2
MSVTNet	43.57	14.14	3.59 ✕ 10-2
Labram	41.71	14.73	9.41 ✕ 10-3
Sparcnet	36.22	12.41	4.28 ✕ 10-4
MVCMGNet	52.79	12.09

The bolded text in the table indicates the accuracy of the model we proposed

Figure 3 presents box plots illustrating participants’ classification performance across different algorithms. A more detailed analysis shows that, compared with the competing models, MVCMGNet achieves both the highest maximum classification accuracy (100%) and the highest minimum accuracy (25%) across all evaluations, indicating not only superior peak performance but also improved robustness under the most challenging conditions. Moreover, multi-view approaches (FBCNet and MSVTNet) demonstrate relatively stable and high performance compared with traditional CNN-based models such as EEGNet [9]. These findings reveal the limitations of conventional bilateral classification models, which were originally designed for left–right symmetric decoding paradigms, when applied to unilateral decoding tasks. Representative architectures such as EEGNet, LMDA-Net, FBCNet, and EEG-Conformer primarily focus on inter-hemispheric discriminative patterns. In contrast, multi-view modeling strategies—including MSVTNet, and the proposed MVCMGNet—consistently demonstrate superior performance in unilateral stroke decoding, providing valuable insights into effective modeling strategies for stroke-related EEG data.

Fig. 3.

Average classification accuracies of algorithms. Scatter points represent each subject’s classification accuracy, while bars indicate the standard deviation among subjects

Figure 4 displays the confusion matrices for each model. Compared with the nine baseline models, MVCMGNet achieves superior overall performance. The confusion matrices indicate that the nine baseline models tended to exhibit substantial confusion in the Fc–Pg and Wlu–Ef classification tasks, with misclassification rates generally exceeding 20%. In contrast, MVCMGNet reduces the classification errors in these two tasks to approximately 15%.

Fig. 4.

Confusion matrices for each model

Binary MA classification

Our study provides a large dataset and analytical foundation for research on unilateral multimodal decoding of EEG signals in chronic stroke patients.

Figure 5 illustrates the classification accuracy for pairwise combinations of the four motor tasks using the MVCMGNet method. Notably, classification accuracy exceeds 70% for all combinations, validating the efficacy of the stroke dataset. Notably, Fc-Pg and Wlu-Ef are the most easily distinguishable task pairs, whereas Fc-Ef and Pg-Wlu remain the most challenging.

Fig. 5.

5-fold cross validation classification accuracies of MVCMGNet for two different tasks in different combinations. Scatter points represent each subject’s classification accuracy, while bars indicate the Std among subjects

Ablation study

In this section, we conduct an ablation study to evaluate the contribution of each block within the MVCMGNet model. Table 3 presents the impact of removing individual blocks on the classification performance using the stroke dataset. The results indicate that the multi-view graph convolutional block improves overall accuracy by 2.06%, the cortical muscular connection module by 6.14%, and the MoE module by 0.81%. All introduced modules contribute to improved classification accuracy, among which the cortical-muscular connection network yields the most substantial performance gain. These results indicate that incorporating explicit neurophysiology-based modeling of brain–muscle connectivity plays a critical role in achieving effective multi-class decoding for unilateral motor tasks.

Table 3.

Ablation study for MVCMGNet

Multi-view graph convolutional block	Cortical muscular connection module	MoE module	Accuracy (Mean ± Std)
✕	✓	✓	50.73 ± 12.58
✓	✕	✓	46.65 ± 14.37
✓	✓	✕	51.98 ± 15.13
✓	✓	✓	52.79 ± 12.09

To illustrate the impact of EEG and EMG signals on the model, we further perform ablation tests. Table 4 shows the classification accuracy when EMG signals are used alone, EEG signals are used alone, or combined EMG-EEG signals are used, without the MoE module and connection module. The results indicate that decoding using EEG signals alone achieves a 2.13% improvement over EMG signal decoding. However, without the MoE module, the fusion of EEG and EMG signals yields poorer results, with both achieving lower accuracy than either signal alone.

Table 4.

Ablation study based on utilized signals

Only EEG	Only EMG	EEG + EMG	Accuracy (Mean ± Std)
✓	✕	✕	44.95 ± 15.92
✕	✓	✕	42.82 ± 14.46
✕	✕	✓	40.37 ± 17.32

Visualization

t-SNE visualization

To illustrate the advantages of the proposed MVCMGNet approach, t-SNE is used to visualize the feature distributions of the fully trained MSVTNet and MVCMGNet models in Fig. 6. Features of the same category clustered together under the MSVTNet architecture, the boundaries between different categories remain indistinct (Fig. 6A). In contrast, the MVCMGNet architecture produces features with more pronounced separation between classes (Fig. 6B). This demonstrates that the features extracted by MVCMGNet exhibit greater inter-class disparity, effectively capturing discriminative information and enhancing overall class separability.

Fig. 6.

t-SNE visualization comparing feature distributions between different models. A The feature distributions of MSVTNet. B The feature distributions of MVCMGNet

Activation of the brain region and muscles

We conduct a comprehensive analysis of brain region activation across all subject data (Fig. 7A) and a comprehensive analysis of CMC connectivity patterns (Fig. 7B). The brain topography displays activation under different tasks, with values represented by the event-related desynchronization metric (negative values indicating activation intensity). To standardize the presentation of neural responses, the brain topographies of patients with right hemisphere involvement are flipped to achieve left-right symmetry. Consequently, the right side of each topography represents the unaffected hemisphere, and the left side represents the affected hemisphere.

Fig. 7.

Activation of the brain region and muscles. A Brain region activation. B CMC connectivity patterns

The impact of the time window on decoding results

Figure 8 illustrates the effect of time window length on decoding performance. As shown in Fig. 8A, the classification accuracy exhibited an initial upward trend followed by a slight plateau as the window length increased. When a 5-s window is used, the model achieves an accuracy of 52.79% ± 12.09%. The highest performance occurs with a 4-s window, reaching 55.14% ± 15.37%, followed by a moderate decrease to 50.65% ± 17.65% with a 3-s window. As the time window length is further reduced, a pronounced decline in classification accuracy is observed. Specifically, the accuracy drops to 35.27% ± 24.11% with a 2-s window and further decreases to 31.09% ± 22.84% with a 1-s window.

Fig. 8.

Accuracy and ITR metrics for different time window lengths. A Accuracy metrics. B ITR metrics

The Information Transfer Rate (ITR) followed a similar non-monotonic trend (Fig. 8B). The peak ITR of 4.45 bits/min was reached at the 4-s window. Notably, although the 5-s window yields an ITR of only 3.05 bits/min, it exhibits the lowest standard deviation of 3.22 bits/min. Meanwhile, the ITRs for the 1-s and 2-s windows are significantly smaller than those for other durations.

Discussion

This paper proposes a GCN structure based on cortical-muscular connectivity for decoding unilateral multi-classification tasks. We successfully acquired EEG and EMG data from 45 chronic stroke patients, compiling a large-scale dataset for in-depth analysis. The mean time since stroke onset was 11.48 ± 6.93 months. This study contributes to the decoding of unilateral motor tasks and lays a foundation for advancing the clinical application of BCIs.

Unilateral stroke dataset

While traditional BCI systems leverage the distinct spatial distributions of brain activation from bilateral limb movements, the rehabilitation of chronic stroke patients necessitates a focus on the flexible control of the affected limb. To address this need, we compiled a unilateral multi-class dataset from 45 chronic stroke patients. In contrast to conventional datasets (such as BCI Competition IV Dataset 2a [12], PhysioNet Motor Movement/Imagery Dataset [13], and the Graz Dataset [14]), this dataset is specifically designed to emphasize unilateral limb control, thereby directly addressing critical rehabilitation requirements. Distinguishing between multiple MA tasks on a single side of the body is a recognized technical challenge [87–89]. Through visualization and classification accuracy analysis, this study demonstrates the technical feasibility of this approach. This unilateral stroke dataset serves as a valuable resource for developing decoding algorithms for unilateral limb movements based on non-invasive neural signals. It has significant potential to advance the MA BCI paradigm and promote the development of effective rehabilitation technologies for chronic stroke patients.

Unilateral multi-class decoding

In conventional BCI tasks, such as left- versus right-hand MA tasks, decoding algorithms leverage distinct spatial patterns of brain activation to achieve high accuracy. However, in unilateral multi-class tasks, activations occur on the same hemisphere, and their spatial patterns exhibit crossover and high similarity, making it difficult for traditional methods to perform well (Fig. 7A). As shown in Table 2, multi-view methods and models incorporating attention mechanisms often yield better results. The core rationale is that analyzing more critical frequency bands (theta: 4–8 Hz, alpha: 8–12 Hz, beta: 12–30 Hz, gamma: 30–40 Hz) and employing complex structures allows the model to focus on key information, including spectral differences and subtle variations in ipsilateral activation. An interesting finding in Table 2 is that a simple CNN-GNN model matches the performance of more complex models (Labram, Sparcnet), likely because it effectively captures spatial connections between brain regions.

As illustrated in Fig. 7B, significant activation is observed around the left motor cortex during Fc and Pg movements, which is consistent with the expected neurophysiological response during right upper-limb motor tasks, where enhanced electrical activity typically emerges in the contralateral (left) hemisphere. In contrast, pronounced activation around the right motor cortex is observed during Wlu and Ef movements, which is indicative of compensatory neural mechanisms commonly reported in chronic stroke patients with motor impairment. Considering the discrepancy between cortical activation patterns and CMC connectivity—where CMC serves as a key biomarker for stroke rehabilitation—we further demonstrate that both contralateral activation and ipsilateral compensatory mechanisms coexist in chronic stroke patients. In addition, activation in visual regions supports the validity of the experimental paradigm. Nevertheless, the absence of clearly separable spatial activation patterns across tasks [8, 89, 90] underscores the inherent difficulty of decoding unilateral limb movements based solely on cortical activation features.

Compared with other baseline methods, the proposed MVCMGNet explicitly incorporates neurophysiologically motivated design principles. Given that motor execution arises from cortical control over muscles and that distinct tasks are mediated by task-specific brain–muscle pathways, we introduce a cortical–muscular connection module to capture these interactions. As shown in Fig. 4, the nine baseline models exhibit substantial confusion, particularly in distinguishing between Fc–Pg and Wlu–Ef tasks, whereas the connectivity-aware representations learned by MVCMGNet effectively alleviate this ambiguity (Fig. 5) and achieve the highest classification accuracy. These results indicate that explicitly modeling functional connectivity can substantially enhance decoding performance for unilateral multi-class data [91, 92], thereby providing a foundation for the development of compact and efficient decoding algorithms tailored to unilateral multi-category paradigms in future studies.

Furthermore, an interesting observation from the ablation experiments is that EMG-based decoding yields lower accuracy than EEG-based decoding. This outcome is consistent with the clinical observation that most patients were unable to execute the movements accurately, reflecting severe motor impairments, and aligns with findings reported in previous studies [93]. At the same time, this result is also influenced by algorithmic factors. The unified framework adopted in the ablation experiments, which was designed to validate the proposed model, may not fully exploit the modality-specific characteristics of EMG signals. Finally, the poorest performance observed with the EEG + EMG fusion approach can be attributed to the heterogeneous nature of EEG and EMG signals in terms of signal-to-noise ratio, temporal dynamics, and semantic content. In the absence of an adaptive fusion strategy, simple feature concatenation introduces cross-modal interference, allowing dominant EMG features or modality-specific noise to obscure discriminative EEG representations. As a result, complementary information is not effectively leveraged, leading instead to degraded decoding performance.

Neurophysiological challenges of unilateral motor decoding in stroke

Decoding unilateral motor intention in stroke survivors is of particular clinical importance, as post-stroke motor impairment predominantly affects one side of the body [94]. Rather than distinguishing left- versus right-sided movements, unilateral decoding aims to discriminate multiple actions of the affected limb, thereby enabling multi-dimensional and task-adaptive active rehabilitation training.

Compared with bilateral motor tasks, unilateral motor decoding is intrinsically more challenging from a neurophysiological perspective [8]. In bilateral paradigms, movements performed on opposite sides elicit spatially symmetric and well-separated activation patterns across the two hemispheres, providing discriminative spatial features that can be effectively exploited by conventional decoding algorithms [95]. In contrast, unilateral motor tasks engage overlapping functional regions within the same hemisphere, where the limited spatial resolution of EEG further amplifies inter-class similarity and reduces separability between different actions [28].

This challenge is substantially aggravated in stroke survivors due to stroke-induced neural reorganization. Damage to the affected hemisphere often leads to reduced cortical excitability, impaired corticospinal output, and disrupted corticomuscular coupling, resulting in attenuated and unstable task-related neural signatures [96]. Moreover, chronic stroke patients frequently recruit compensatory mechanisms involving ipsilateral or contralesional regions to support unilateral motor execution [97]. Such compensatory activation, together with altered interhemispheric inhibition, introduces pronounced inter-subject heterogeneity and atypical activation patterns that deviate from healthy motor control models, further complicating decoding.

Despite these challenges, functional connectivity between cortical regions and corresponding muscles in chronic stroke patients tends to stabilize over time [98]. Building upon traditional decoding methods that primarily leverage spatial features, our model incorporates task-specific cortical-muscle connectivity across different motor-related tasks. Importantly, this connectivity reflects intrinsic relationships between brain regions and muscles, exhibiting minimal sensitivity to fluctuations in signal amplitude or connection strength (Table 2). By integrating these stable corticocutaneous composite features, the model achieves superior discrimination of multiple movements in the affected limb, even in the presence of stroke-induced alterations in cortical activation patterns (as shown in Fig. 7).

Failure analysis of baseline models in unilateral stroke decoding

The inferior performance of several baseline models observed in Table 2 can be attributed to their underlying assumptions, which are not well aligned with the neurophysiological characteristics of unilateral motor tasks in stroke patients. Traditional CNN-based approaches (EEGNet, SparcNet) primarily rely on stable and discriminative spatial activation patterns across EEG channels. However, in unilateral multi-class scenarios, particularly under stroke-induced cortical reorganization, spatial patterns are highly variable and overlapping, leading to severe class confusion and degraded decoding performance [8].

Models incorporating attention mechanisms (LMDA-Net, ACTNet) or Transformer-like architectures (Conformer, LaBraM) demonstrate improved flexibility in modeling long-range dependencies but remain sensitive to inter-subject variability and limited training data. In chronic stroke populations, neural representations of motor actions differ substantially across individuals due to heterogeneous lesion locations, recovery stages, and compensatory strategies. As a result, attention weights learned by these models may become unstable or biased toward subject-specific noise rather than task-relevant neural features, limiting their generalization capability.

Multi-view and multi-frequency models (FBCNet, MSVTNet) achieve relatively better performance by exploiting complementary spectral information across canonical frequency bands. Nevertheless, these methods typically perform implicit feature fusion without explicitly modeling the functional relationships between cortical activity and muscular output. In unilateral stroke tasks, where cortical–muscular coupling is weakened and highly task-dependent, the absence of explicit connectivity modeling restricts the ability of such models to disambiguate actions with similar cortical signatures.

The CNN–GNN baseline exhibits competitive performance despite its relatively simple architecture, likely because graph convolution explicitly models spatial relationships among EEG channels, partially alleviating the reliance on fixed spatial activation patterns. By capturing topological dependencies between brain regions, the GNN component introduces a degree of structural robustness against spatial variability. Nevertheless, this model remains limited by its exclusive reliance on cortical information and the absence of explicit cortical–muscular interaction modeling, restricting its ability to disambiguate unilateral motor tasks with similar cortical signatures.

In contrast, the proposed MVCMGNet explicitly integrates cortical–muscular connectivity into the decoding framework, aligning model architecture with the underlying neurophysiological control pathways of unilateral motor actions. By modeling functional interactions between EEG and EMG signals, the network reduces reliance on stable spatial patterns and mitigates the impact of stroke-induced variability. This design enables more robust discrimination between unilateral motor categories and explained the superior performance of MVCMGNet relative to baseline approaches in this challenging decoding scenario.

Similarities and differences between actions

Similarities between movements often produce overlapping patterns of cortical activation, which poses a significant challenge for distinguishing between multiple unilateral motor categories [8, 30]. Our results indicate that Fc-Pg and Wlu-Ef are the most distinguishable action pairs, whereas Fc-Ef and Pg-Wlu are the most challenging to differentiate (Fig. 5).

These findings prompt further investigation into the underlying reasons for this hypothesis. A plausible explanation is that kinematic similarities between actions lead to analogous neuromuscular activations, while kinematic differences enhance decoding performance. These two tasks exhibit significant differences in the flexion-extension dimension, while finger opposition demonstrates a degree of extension strategy, which may contribute to their higher classification accuracy. This leads us to hypothesize that when studying the upper limbs, EEG signals may exhibit increased sensitivity to MAs along the flexion-extension dimension. Future research will aim to explore this hypothesis more deeply, seeking to understand the neurophysiological basis of this phenomenon and its implications for MA-BCI system design.

Limitation

This study has several limitations that should be acknowledged. First, data from 5 participants were excluded after preprocessing due to insufficient signal quality. Excessive head motion or lapses in attention in these participants resulted in fewer than 80% of the planned trials remaining usable, which is inadequate for reliable EEG–EMG and corticomuscular coupling analysis. Although this exclusion was necessary to ensure data reliability, it reduced the final sample size and may limit the generalizability of the findings.

Second, the sex distribution of the cohort was highly imbalanced, with only one female participant compared to 44 male participants. This imbalance reflects the actual patient availability during the six-month data collection period, during which only one eligible female chronic stroke patient was recruited. Although this participant was retained to better reflect real-world clinical data, the skewed sex ratio may introduce bias and limits the applicability of the findings across sexes.

Third, the age range of the participants is relatively broad (28–72 years). Such age heterogeneity may contribute to inter-subject variability in neural and neuromuscular characteristics, potentially influencing the EEG–EMG and corticomuscular coupling patterns observed. Nevertheless, this age distribution represents the realistic demographic characteristics of chronic stroke patients encountered during the data collection period.

Fourth, this decoding framework relies on relatively long time windows (3–5 s) to achieve stable and reliable classification performance, limiting its direct application in real-time brain-computer interface systems. Practical online BCI applications, particularly in rehabilitation settings, require rapid feedback generation from the emergence of motor intent to system response. Our analysis indicates that beyond a 3-s window, decoding accuracy and ITR decline significantly as the time window shrinks, revealing a trade-off between classification reliability and system latency (Fig. 8). Notably, the steep performance decline under short windows may partly stem from severe motor impairment in chronic stroke patients—who typically complete only a single movement attempt within 5 s rather than repeated actions (the accuracy increase at 4 s arises from greater data volume in the sliding window). Consequently, shorter windows may fail to capture sufficient task-relevant neural and neuromuscular information, further exacerbating performance loss.

Conclusion

In this study, we proposed MVCMGNet, a novel model leveraging cortical-muscle connectivity patterns, to address the challenge of decoding unilateral multi-class movements in chronic stroke patients. Experimental validation involving 45 chronic stroke patients demonstrates that MVCMGNet achieves superior decoding performance and exhibits stronger robustness compared to existing state-of-the-art benchmark models. Data visualization further confirms that incorporating cortical-muscle connectivity pattern effectively mitigates the issue of overlapping spatial activation patterns across movement categories. Further analysis reveals that movement similarities result in correlated neuromuscular activations, while distinctive kinematic features enhance decoding effectiveness. These findings provide valuable insights for future MA BCI research and clinical applications. Overall, this study contributes significantly to the advancement of MA-BCI systems and offers compelling preliminary evidence for decoding complex unilateral movements from EEG signals.

Appendix

See Table 5.

Table 5.

Information sheet for patients

Number	Age	Gender	Duration of illness (month)	Etiology	Lesion	Muscle strength level
1	30	M	6	Hemorrhagic	Basal ganglia	2
2	30	M	10	Hemorrhagic	Basal ganglia	1
3	30	M	32	Hemorrhagic	Basal ganglia	0
4	72	M	11	Hemorrhagic	Basal ganglia	2
5	72	M	8	Hemorrhagic	Basal ganglia	2
6	63	M	13	Ischemic	Brainstem	2
7	57	M	10	Hemorrhagic	Basal ganglia	1
8	63	F	12	Hemorrhagic	Basal ganglia	2
9	47	M	19	Ischemic	Basal ganglia	1
10	38	M	8	Hemorrhagic	Basal ganglia	2
11	63	M	7	Hemorrhagic	Brainstem	0
12	28	M	8	Ischemic	Basal ganglia	2
13	70	M	10	Hemorrhagic	Basal ganglia	4
14	42	M	28	Hemorrhagic	Basal ganglia	2
15	61	M	10	Hemorrhagic	Brainstem	2
16	45	M	6	Ischemic	Basal ganglia	0
17	50	M	29	Ischemic	Basal ganglia	1
18	56	M	12	Hemorrhagic	Basal ganglia	1
19	56	M	9	Hemorrhagic	Basal ganglia	4
20	46	M	23	Ischemic	Basal ganglia	2
21	43	M	18	Hemorrhagic	Basal ganglia	2
22	32	M	8	Ischemic	Basal ganglia	3
23	58	M	9	Hemorrhagic	Basal ganglia	0
24	55	M	6	Ischemic	Basal ganglia	3
25	49	M	7	Hemorrhagic	Basal ganglia	1
26	37	M	15	Ischemic	Basal ganglia	2
27	44	M	9	Hemorrhagic	Basal ganglia	2
28	41	M	5	Ischemic	Basal ganglia	0
29	49	M	7	Hemorrhagic	Basal ganglia	1
30	41	M	9	Hemorrhagic	Basal ganglia	1
31	41	M	8	Hemorrhagic	Basal ganglia	0
32	28	M	8	Ischemic	Cerebellum	2
33	61	M	6	Hemorrhagic	Basal ganglia	3
34	57	M	12	Hemorrhagic	Basal ganglia	3
35	53	M	7	Ischemic	Basal ganglia	0
36	43	M	6	Hemorrhagic	Basal ganglia	1
37	43	M	11	Ischemic	Basal ganglia	2
38	47	M	12	Hemorrhagic	Basal ganglia	2
39	43	M	6	Hemorrhagic	Basal ganglia	3
40	60	M	14	Hemorrhagic	Basal ganglia	2
41	46	M	9	Hemorrhagic	Basal ganglia	3
42	52	M	10	Ischemic	Cerebellum	3
43	52	M	12	Hemorrhagic	Thalamus	0
44	35	M	7	Hemorrhagic	Basal ganglia	1
45	50	M	12	Ischemic	Basal ganglia	2
46	50	M	11	Hemorrhagic	Basal ganglia	1
47	48	M	6	Hemorrhagic	Basal ganglia	3
48	36	M	36	Hemorrhagic	Basal ganglia	2
49	40	M	6	Hemorrhagic	Basal ganglia	1
50	41	M	11	Hemorrhagic	Basal ganglia	0

M male, F female, Muscle strength level is assessed using manual muscle testing (MMT)

Author contributions

Conceptualization: Jiuxiang Song, Xiaoke Chai; Resources: Nan Wang, Zhaolin Li, Xuemin Zhang, Zeping Lv; Methodology: Jiuxiang Song; Writing - original draft preparation: Jiuxiang Song; Writing - review and editing: Xiaoke Chai; Funding acquisition: Yi Yang; Supervision: Xinying Shan, Jizhong Liu.

Funding

This study was supported by National Key R&D Program on Brain Science and Brain-Like Research: Precision Regulation Technology for Spinal Cord-Peripheral Nerve Integration Aimed at Motor Function Restoration (2025ZD0215200) and the Collaborative Research and Development Project “Deep Brain Stimulation System Development and Technical Research” (Grant No. HX202409200057).

Data availability

The data and code used in this study can be found at https://github.com/sjx66606/MVCMGNet.

Declarations

Ethics approval and consent to participate

Approval of all ethical and experimental procedures and protocols was granted by the Ethics Committee of Beijing Tiantan Hospital under Application No. KY2023-205-03.

Consent for publication

Subjects provided their consent to publish their data.

Competing interests

The authors declare no competing interests.