Coherence-Based Graph Convolution Network to Assess Brain Reorganization in Spinal Cord Injury Patients

1. Introduction

Spinal cord injury (SCI) is a severe condition that leads to significant functional impairment of the limbs below the injured segment. It not only causes substantial physical and psychological suffering for patients but also brings a considerable economic burden to society.¹ Studies have shown^2,3,51 that brain–computer interface (BCI) technologies can assist SCI patients in reconstructing motor function and restoring mobility.

BCI enables direct communication between external devices and the brain without relying on muscles or peripheral nerves, enabling interaction with the external environment.^4,5,6 Electroencephalography (EEG) is a widely used neural signal acquisition method in BCI systems.⁷ Compared to other neuroimaging techniques, EEG offers advantages, including noninvasiveness, high temporal resolution, and cost-effectiveness, making it the most used recording technology in BCIs.

MI involves mentally simulating a specific action rather than physically performing it.⁸ MI is extensively used in motor skill learning, neurorehabilitation, and BCI applications.^9,63,64 MI-based BCIs (MI-BCIs) hold significant promise and have contributed significantly in medical and nonmedical domains. For instance, they have proven beneficial in rehabilitating patients with impaired motor function.^10,11 As illustrated in Fig. 1, a typical MI-BCI system comprises four main components: signal acquisition, signal processing, applications, and feedback.

The classification and recognition algorithms for EEG signals are pivotal in MI-BCI systems. Traditional machine learning methods have been extensively employed for MI-EEG signal classification. In recent years, deep learning, pioneered by Geoffrey Hinton, has gained widespread adoption across various fields, including MI-BCI systems.¹² Techniques such as long short-term memory networks (LSTMs),¹³ deep neural networks (DNNs),¹⁴ deep belief networks (DBNs),^15,16 reinforcement learning (RL)⁵² and others have successfully enhanced MI-BCI performance. Convolutional neural networks (CNNs), introduced in 2012, have also become a popular tool for MI classification and recognition. For instance, Li et al.¹⁶ combined CNN and LSTM to propose a neural network feature fusion algorithm, significantly improving MI intention recognition accuracy to an average of 87.68%.

While CNNs have been widely used in diverse fields, their application in MI-BCI systems has some limitations. Much of the CNN research on EEG focuses on individual channels, often overlooking the inter-channel correlation. To solve this problem, graph convolutional neural networks (GCNs) have become a key area of research, enabling extracting effective spatial features from EEG data for deep learning.¹⁷ Originally proposed by Scarelli et al.¹⁸ in 2009, GCNs integrate CNNs with spectral graph theory to effectively describe relationships between different nodes in the graph’s discrete spatial domain. Li et al.⁴⁹ in 2023 used GCNs to study the differences between patients in motion image EEG. Previous studies have shown that GCNs offer new insights into motion intention recognition in MI-BCI systems.¹⁰ However, GCNs prioritize spatial relationships between channels without fully exploring functional connections.

To address these limitations, this study proposes a coherence-based resGCN approach for MI intention recognition. This method uses EEG coherence connections to explore functional connectivity patterns relevant to MI tasks. The brain is a sophisticated network where information is processed among spatially distinct but functionally interconnected regions.^19,20,49 Current studies indicate that MI-BCI systems can support motor function rehabilitation in SCI patients,^3,35 although the precise mechanism of brain reorganization from resting to MI states remains incompletely understood. Brain network analyses, including coherence and phase-locked value measurements, can reveal underlying neural mechanisms and connectivity patterns between specific brain regions.^21,22

The brain continuously processes information, even at rest.²³ Spontaneous brain activity during rest reflects the brain’s readiness to process information during subsequent tasks.^24,25 Studies using resting-state functional magnetic resonance imaging (fMRI) have shown that efficient information integration across distributed brain regions correlates with human intelligence.²⁴ Furthermore, spontaneous brain activity provides valuable insights into the physiological mechanisms of diseases like epilepsy,²⁶ Alzheimer’s,^57,58 developmental dyslexia,^27,59 Parkinson’s,⁶⁰ and SCI.²⁸ Effective brain activity at rest improves task performance.²⁹ Resting-state research has also explored networks involved in MI tasks.⁹ During MI, areas such as the Brodmann area (BA4), supplementary motor area (SMA), parietal area, and premotor cortex (PMc) are activated, contributing to MI-related information processing.^20,30

In a previous study, Zhang et al. found that an efficient resting-state brain helps improve subjects’ performance in MI tasks,²⁵ suggesting that such activity provides a physiological foundation for processing relevant information during MI. Other studies have shown large-scale brain network reorganization can affect MI task performance.

Thus, resting brain activity may be a biomarker for predicting information processing during MI tasks. Understanding how SCI patients’ brains reorganize from the resting to the MI state and exploring the relationship between this reorganization and MI, is crucial.

Studies have shown that the default mode network (DMN)^33,34,61 is negatively correlated with task performance, meaning the DMN deactivates when the brain engages in tasks.

This study aims to elaborate on the mechanism of brain function reconstruction in SCI patients as they transition from a resting state to an MI task state. To address the limitations of graph convolutional networks (GCNs) in representing functional connectivity across various brain regions, this paper makes the following key contributions:

(1)	Introducing an MI-recognition method based on a coherence-based residual graph convolutional neural network that leverages EEG signals’ characteristics in time, frequency, and spatial domains. Achieving an accuracy rate of 96.25%, this method presents a novel approach to recognizing movement imagery in MI-BCI systems.
(2)	Exploring the brain functional mechanisms of SCI patients using EEG-based coherence analysis.
(3)	Modeling univariate EEG features as multivariate features of the graph signal within the coherence functional network to enhance our ability to characterize EEG’s spatial and functional intrinsic connections. This approach provides new insights into methods for recognizing motor intention.

The remainder of this paper is arranged as follows. Section 2 introduces the experimental dataset, data analysis, and relevant theories. Section 3 reports the experimental results and discusses the findings. Section 4 concludes this study. Figure 2 shows the overall experimental process.

2. Experimental Data

2.1. Participants

The experiment received approval from the Medical Ethics Committee of Qilu Hospital, Qilu Medical College of Shandong University (Approval No. KYLL-202308-008). Before the experiment, all participants read and signed the informed consent form. This study recruited 18 patients with SCI, all with normal hearing and vision and no history of drug abuse or cognitive impairment.

2.2. Experimental procedures

Throughout the experiment, all subjects were asked to remain relaxed and sit comfortably in front of the computer. Each subject performed 80 independent MI experiments, including 40 left-hand MI (LMI) and 40 right-hand MI (RMI). Before performing MI, baseline data of resting state will be recorded for 2min with both open and closed eyes. Each MI experiment lasted 7s. The first 4s were the rest period, and a cross would appear in the middle of the screen to remind the subject to get ready; the last 3s were the task period, and arrows pointing left and right would appear on the screen to prompt the subject to imagine waving with the left or right hand. The order of the left- and right-hand MI experiments was randomly presented to the subjects. The whole process of a trial as shown in Fig. 3.

2.3. EEG data recording

All subjects’ EEG signals were collected using a Boricon 64-lead EEG cap, and the 64 Ag/AgCl electrodes were placed according to the international 10–20 system standard. In all experiments, the electrode impedance was kept below 10K$\mathrm{\Omega }$, and the sampling rate was set at 1000Hz.

2.4. EEG data analysis

2.4.1. EEG data preprocessing

Before the construction of the functional network, the raw EEG data were zero-referenced using the reference electrode standardization technique (REST), band-pass filtered between 8 and 45Hz, segmented into 3s time window, and artifacts were removed ($\pm 75$  $\mu$V as the threshold) to obtain artifact-free data trials. We applied the ICA algorithm to identify and remove components related to artifacts (such as those caused by eye movement or muscle activity). Each MI trial comprised a 4s resting phase and a 3s task phase. We applied a 3s window to the resting EEG to ensure consistency in data analysis between the resting state and the task state. To reduce the impact of volume conduction between electrodes, 21 electrodes (i.e. Fp1, Fpz, Fp2, F3, F4, F7, Fz, F8, C3, Cz, C4, T7, T8, P3, Pz, P4, P7, P8, O1, Oz, and O2) of the 64 channels were selected for subsequent processing, as shown in Fig. 4.

2.4.2. Resting-state EEG

We recorded spontaneous brain activity during the resting state with eyes closed and open. Each condition lasted for 2min and was repeated twice. We randomly presented the two conditions. For the recorded resting-state EEG, we excluded the first and last 30s of EEG and segmented the remaining data into 3s windows. Then, we used coherence to construct a functional network for each 3s segment. Finally, we averaged the adjacency matrices of all segments to evaluate the resting-state connection matrix for each subject.

2.4.3. Motor imagery EEG

For the MI task-state EEG, at the onset of recording left- or right-hand MI (i.e. when the left or right arrows showed on the screen), we extracted the EEG from the last 3s of each trial. Then, we constructed the weighted adjacency matrix for each MI trial using coherence and averaged the trials to obtain the brain functional network of LMI or RMI for each subject.

3. Methods

3.1. Brain network analysis

All preprocessed resting/MI test data contributed to the construction of the brain functional network of each subject. This study used coherence to build brain networks that characterize the functional connectivity between multiple brain regions. Coherence is the degree of linear correlation between two signals $x(t)$ and $y(t)$, which measures the strength of interaction between each pair of electrodes at specific frequencies. The coherence coefficients of $x(t)$ and $y(t)$ are defined as

where $P_{xy}(f)$ is the cross-spectral density of $x(t)$ and $y(t)$, $P_{xx}(f)$ and $P_{yy}(f)$ are the autospectral densities of signals $x(t)$ and, respectively, $y(t)$. The coherence coefficient ranges from 0 to 1; a value closer to 1 indicates stronger connectivity between the two electrodes, while a lower value indicates weaker connectivity. Then, a $21\times 21$ weighted adjacency matrix is obtained by averaging the results from the relevant frequency band, leading to the final functional network.

Based on the final coherence matrix, we calculated the network properties of each participant, including global efficiency (Ge), characteristic path length ($L)$, cluster coefficient (CC) and local efficiency (Le).^9,21,29 These network properties quantitatively evaluate the brain efficiency of patients. The calculation formulas of network properties are as follows :

where $W_{ij}$ represents the coherence value between nodes, $d_{ij}$ represents the shortest path length between node j and node i, A represents the number of nodes, and T represents the sets of nodes in the network.

3.2. GCN theory

Graph convolutional networks can be categorized into two categories: spectral convolution and spatial domain convolution.³¹ Spectral convolution involves transforming graph signals into the Fourier domain before processing them with a convolutional network, facilitating effective filtering and feature learning in the frequency domain. The product of graph input $x\in R^N$ and the filter $g_{\theta }=\mathrm{\text{diag}}(\theta )$ is defined as the spectral convolution of graph signal, with the filter parameterized by $\theta \in R^N$ in the Fourier domain :

where $g_{\theta }$ represents the eigenvalue of the Laplacian matrix, U represents the eigenvector of the normalized Laplacian matrix, defined as where $\mathrm{\Lambda }$ is a diagonal matrix consisting of the eigenvalues of the Laplacian matrix. $U^{T}x$ represents the graph Fourier transform of the graph input.

In the spatial domain, a truncated expansion of the K-order Chebyshev polynomial approximates the filter to reduce computational complexity. The Chebyshev polynomial³² is defined as $T_k(x)=2x{T}_{k-1}(x)-T_{k-2}(x)$, where $T_0(x)=1,$ $T_1(x)=x$. After filtering the signal input x with the K-order filter, it is expressed as follows :

where $\tilde{L}=\frac{2L}{{\lambda }_{max}{N}_{max}}$ represents the maximum eigenvalue of Laplacian matrix, leading to the derivation of the spatial eigenvalue.

3.3. C-ResGCN model

Traditional CNN has several limitations, such as gradient disappearance. To address traditional CNNs’ shortcomings, we use a residual network (ResNet), which adds residual connections between the input and the output layers, significantly improving CNN performance.

Figure 5 illustrates the residual block structure. In addition, to solve the problem of GCN not representing the functional connections of different brain regions, this study proposes a C-ResGCN network that leverages the time–frequency and spatial characteristics of EEG signals. The proposed network architecture comprises four residual blocks and a fully connected layer. Each residual block comprises two convolutional layers, two ReLU layers, and a graph pooling layer.

In the initial graph convolution frameworks, we need to learn the underlying residual mapping function, F. We derive another underlying mapping function F by fitting it into another mapping function Y. The feature map F is transformed into the graph representation $G_l$ for the current layer. Then, $G_l$ is updated by adding the graph representation from the next layer $G_{l+1}$, thereby forming an accumulative process. The residual mapping function F learns the re-represented output graph $G_{l+1}^{\mathrm{\text{res}}}$ by taking the graph as input, where ${\omega }_l$ represents the weight set of the lth layer. We define the model of ResGCN as

C-ResGCN is a model that combines the coherence function network and ResGCN. The model comprises two parts: (1) construction based on the coherence graph signal and (2) task recognition using ResGCN. Figure 6 shows the overall framework.

The initialized model uses the ResGCN structure to perform pattern recognition on MI. The ResGCN layer extracts generalized features from 21 channels. Dimensionality reduction was achieved through the pooling layer, and the prediction results were ultimately generated by the softmax function in the fully connected (FC) layer. To find the optimal model parameters, we conducted extensive hyperparameter searches using both grid search and random search, systematically exploring the effects of different parameter combinations. The final selected parameters were validated through a series of independent test sets to ensure stability and reliability. Table 1 lists the model parameters.

In our proposed C-ResGCN model, the optimization of network parameters using cross-entropy loss function is as follows :

where $p(x)$ represents the true value of the training data, $q(x)$ represents the predicted value, the $\lambda J(w)$ prevents the model from overfitting. Finally, the subject’s EEG data are trained and tested on the C-ResGCN, and cross-validation is conducted. Algorithm 1 outlines the detailed training steps of the model. The proposed algorithm processed the data on a computer with the configuration shown in Table 2.

4. Results and Discussion

4.1. Network reorganization during MI

In this study, we focused on the classification results of the 8–30Hz ($\alpha$ and low $\beta$ bands) and 31–45Hz (high $\beta$ and low $\gamma$ bands) frequency ranges. The selection of these specific frequency bands was based on extensive research and support from previous literature.^{41,42,43,44,45,46,47}

Figures 7 and 8 show the reorganization of network topology in SCI patients as they transition from resting state to left- and right-hand MI ($p<0.05$, FDR corrected). Figure 7(a) shows the network shifting from a resting state to an MI task state within the 8–13Hz frequency range. During this transition, the connectivity between the frontal and parietal lobes weakened. In the 14–30Hz frequency band, the long-range connectivity between the occipital and frontal lobes decreased, as illustrated in Fig. 7(b).

As shown in Fig. 8, within the $\gamma$ frequency band of 31–45Hz, the default mode network in the frontal lobe was reduced, and the brain functional connectivity decreased during the MI stage. Meanwhile, the functional connectivity between the bilateral motor areas and the occipital lobe strengthened during the reorganization process.

4.2. Restructuring network attributes

To further investigate the brain network reorganization of SCI patients during MI, we calculated the network properties of each stage, such as Characteristic path length (CPL), CC, LE, and GE. Figure 9 shows that when the brain transitions from a resting state to an MI task state, its efficiency in processing information increases. Therefore, the statistical results indicate that the CC, GE, and LE of the MI group are significantly higher than those of the Rest group; moreover, the CPL of the Rest group is significantly higher than that of the MI group. These results demonstrate significant differences in network efficiency between the Rest and MI groups.

4.3. Classification results

Classification performance serves as a crucial metric for evaluating data quality. Figure 10 shows a line graph depicting classification accuracy and loss.

We have established clear inclusion criteria based on factors such as disease diagnosis, course of illness, and severity of symptoms to ensure that the selected patients meet our research objectives and ensure sample homogeneity and representativeness. 18 patients were selected from 40 participants as experimental subjects, as shown in Table 3.

Table 4 shows the results of the classification and identification of EEG data in different frequency bands of 18 SCI patients using the proposed C-ResGCN model. In the C-ResGCN model, the highest classification accuracy reached 96.25%. The average classification accuracy in the 8–30Hz band was 93.2%, slightly lower than the 93.97% achieved in the 31–45Hz band.

Overall, regardless of the frequency bands of 8–30Hz or 31–45Hz, the majority of participants scored very high on these four indicators, indicating that the C-ResGCN model demonstrated good classification ability in both frequency bands.

For all participants, there is usually a high degree of consistency between accuracy, precision, recall, and F1 score in both frequency bands, indicating that the model can maintain a high balance in prediction.

4.4. Model evaluation of C-ResGCN

We trained, validated, and statistically analyzed the model to ensure its generalizability and practicality.

Cross-validation is a widely used model evaluation method in machine learning. This study implements a 10-fold cross-validation method for evaluating the model. During the experiment, 70% of the dataset is used for the training set, 15% for the verification set, and the remaining 15% for the test set. At the same time, to further enhance the robustness of the results, we used k-fold cross-validation ($k=10$) in the training phase. This method ensures that each observation value in the original sample is used for training and verification, so as to provide more accurate model performance estimation. Additionally, it provides a more accurate evaluation of performance, enhances generalization, and identifies the best model.

To validate the advancement of the proposed method, the study compares it with several research outcomes. The complexity of the computational model is one of the criteria for measuring performance. FLOPs represent the number of floating-point operations and determine the training time of the model. Table 5 illustrates the classification accuracy, FLOPs of different methods. The proposed algorithm performs better than previous methods ($p<0.05$).

We further assessed the performance of the C-ResGCN algorithm by validating on a publicly available Datasets, the BCI Competition IV 2a (2008),³⁶ that encompasses four MI samples corresponding to left-hand, right-hand, foot, and tongue movements gathered from nine subjects. The data in this dataset was recorded using 22 electrodes and sampled at 250Hz. Our proposed algorithm demonstrated impressive classification accuracy on this dataset, reaching 92.59% on the BCI Competition IV 2a dataset. Table 6 compares the classification results achieved by our method with various previously published methods when applied to the Competition IV 2a dataset.

The t-test is commonly used to evaluate the difference between two means. The p-value in t-test is the probability of the null hypothesis that the means of two samples are equal. In statistical terms, probability relates to the hypothesis that no difference exists between the two groups of observations. When $p<0.05$ indicates a probability of occurrence less than 5%, allowing the null hypothesis to be rejected, indicating a significant difference between the groups. As shown in Table 7, the significance value from the independent sample test is less than 0.05. The difference between the two categories is statistically significant by comparing the values of left and right MI in categories 0 and 1 ($p<0.05$). This shows that there is a significant difference between the two categories of left- and right-hand MI. Research in SCI rehabilitation technology can, therefore, benefit from these features.

4.5. Ablation experiments

Ablation studies aim to understand the contribution of these components to the entire system, and usually use the method of removing certain components to evaluate the performance of the system. In the case of decreased or even missing performance of certain components, the system can still continue to work. Table 8 presents the evaluation metrics after the ablation study, highlighting the system’s performance with each module removed. The results indicate that incorporating residual blocks maximizes overall performance.

4.6. Discussion

This study examined the relationship between the brain’s resting state and the left- and right-hand MI networks in the 8–30Hz and 31–45Hz frequency bands in SCI patients. It analyzed the brain’s reorganization process from rest to task state and incorporated ResGCN for recognizing MI movement, providing valuable insights for MI-based BCI.

In individuals with spinal cord injury (SCI), the partial or complete interruption of sensorimotor pathways leads to a significant reduction or loss of sensory feedback from the limbs. This absence of sensorimotor feedback disrupts the closed-loop control system formed by normal motor execution and sensory feedback, compelling the brain to initiate a series of compensatory mechanisms to adapt to the new condition. Specifically, this manifests as a weakening of connectivity in the $\alpha$- and $\beta$-frequency bands, and an increase in $\gamma$-band connectivity. The latter reflects the brain’s attempt to compensate for the lost sensory input through internal simulation. Additionally, the decline in default mode network activity indicates that the brain shifts from internal thought processes to focus on external tasks, striving to maintain the execution of motor imagery (MI) tasks through the reallocation of internal resources. These changes reveal the brain’s remarkable adaptability and plasticity under such extreme conditions. They not only deepen our understanding of the mechanisms underlying motor imagery but also provide valuable theoretical foundations for developing rehabilitation strategies tailored to SCI patients. This insight into how the brain reorganizes itself in response to the loss of sensory-motor feedback can guide the design of more effective interventions, promoting functional recovery and enhancing the quality of life for these patients.

When considering the relevant brain networks, Figs. 7 and 8 show the necessary brain network reorganization in SCI patients as they transition from a resting state to the left- and right-hand MI task state in the 0–45Hz frequency band. As shown in Fig. 7(a), in the $\alpha$ band, during LMI and RMI, the long-range connectivity between the occipital and frontal lobes and the motor area is weakened in SCI patients. Similarly, in Fig. 7(b), the $\beta$ band showed weakened long-range connections in the same frontal–occipital regions. The spinal cord is the pathway for transmitting neural signals from the brain to the limbs. We infer that spinal damage in SCI patients leads to functional suppression of low-frequency EEG signals during MI tasks. In fact, MI resembles actual movement, as both activate brain regions, such as BA4, PMc, parietal area, and SMA. MI involves multiple high-level cognitive processes, including attention, working memory, and decision-making, which highlight the critical role of $\gamma$ activity in MI information processing. During MI, multiple brain regions collaborate and communicate to ensure the smooth execution of the task, enhancing the subjects’ performance. As shown in Fig. 8, we analyzed the brain network reorganization in the 31–45Hz range. When the brain transitions from a resting state to the MI task state, the network in the frontal lobe weakens, resembling the default mode network, while the connectivity of the primary motor area, with C3 and C4 electrodes as central nodes, strengthens to support MI-related information processing and ensure the smooth execution of MI. Research shows that the DMN and task performance negatively correlate, with the DMN deactivating during tasks. Our study confirms that this phenomenon also occurs in SCI patients.

Based on the reorganized network shown in Figs. 7 and 8, the calculated network properties appear in Fig. 9. In the resting state, the brain is idle, and the network efficiency is low (i.e. L is longer, LE, GE, and CC are smaller). In contrast, when performing MI tasks, the network efficiency gradually increases. As shown in Figs. 7 and 8, different stages require different brain functions, which manifest in the distinct network structures corresponding to different network efficiencies. The resting state requires the subject to remain relaxed and fully prepared for the MI task. There are no other complex functions or related information-processing processes involved. Therefore, the observed network efficiency remains low. When the subject begins to imagine the required action, various brain regions connect and work together, activating multiple high-level cognitive functions, resulting in higher network efficiency for processing MI information to ensure the subject can complete the MI task. Previous studies have shown that brain network reorganization is a common phenomenon that can be observed in both healthy individuals and patients with specific diseases. The phenomenon of brain network reorganization shown in Fig. 10 is not limited to patients with spinal cord injury (SCI), but also applicable to healthy individuals.

The proposed C-ResGCN algorithm extracts the time-frequency-spatial features and EEG signals’ brain functional connectivity information. The effective brain network mechanism provides a reliable physiological basis for MI-based BCI. Table 1 shows the parameters for the C-ResGCN model outlined in this study. Table 4 shows the classification results of 18 subjects, while Fig. 10(b) shows the loss. In Table 4, the highest classification accuracy reached 96.25%. The average classification accuracy in the 8–30Hz frequency band was 93.2%, lower than the 96.68% achieved in the 31–45Hz frequency band. This result indicates that compared to the low-frequency band, the $\gamma$ band in SCI patients contains more relevant information, and brain activity is more pronounced during MI. This finding aligns with our previous experimental conclusions that for SCI patients, the brain activity of the $\gamma$ band is more relevant to rehabilitation research. For all participants, there is usually a high degree of consistency between accuracy, precision, recall, and F1 score in both frequency bands, indicating that the model can maintain a high balance in prediction.

In addition, the enhanced coherence within the gamma band indicates that it is a potential effective frequency band for decoding the intent signal of SCI patients. Therefore, in the future, BCI system design for SCI patients can consider prioritizing the use of $\gamma$ band information to improve system performance and accuracy. Considering that some subjects perform better on specific frequency bands, future research can explore the possibility of personalized frequency band selection to enhance classification performance in personalized applications.

Our research has some limitations, such as insufficient sample data. We have implemented strict methods to ensure the robustness and adaptability of the model. Future research should aim for larger sample sizes and consider including healthy controls to further validate our findings. In summary, future work will focus on expanding the dataset and conducting external validation to enhance the generalizability of our findings.

5. Conclusion

This study developed brain functional networks based on coherence to investigate the effective brain network reorganization of SCI patients during MI. It proposed a C-ResGCN algorithm based on EEG time –frequency –space processing and analysis. This method combines coherence and ResGCN for MI pattern recognition using multi-channel EEG signals. Through coherence analysis across different frequency bands, SCI patients showed more effective brain reorganization in the $\gamma$ band. In other words, effective brain network reorganization from the resting state to the MI task state ensures the successful completion of the MI task. The highest classification and recognition accuracy of the proposed C-ResGCN algorithm reached 96.25%. In comparison, the average accuracy in the $\alpha +\beta$ band (8–30 Hz) was 93.2%, which is lower than the average classification accuracy of $\gamma$ band (31–45Hz). Higher classification performance indicates that the brain processes MI-related information more effectively, which aligns with the conclusion of the coherence network analysis. The findings of this study can provide valuable insights for the rehabilitation of SCI patients.

Acknowledgments

This work was supported in part by the Fundamental Research Funds for the Central Universities under Grant No. 2022JC013, in part by the National Natural Science Foundation of China under Grant Nos. 62271293 and 82330064, in part by the Natural Science Foundation of Shandong Province under Grant Nos. ZR2022MF289, ZR2021MH023, and ZR2019MA037, in part by the Qilu University of Technology (Shandong Academy of Sciences) Basic Research Project of Science, Education and Industry Integration Pilot under Grant No. 2023PY046, in part by the Introduce Innovative Teams of 2021 “New High School 20 Items” Project under Grant No. 2021GXRC071, in part by the Graduate Education and Teaching Reform Project of Qilu University of Technology (Shandong Academy of Sciences) in 2023, in part by the Talent Training and Teaching Reform Project of Qilu University of Technology in 2022 under Grant No. P202204, in part by the Research Leader Program of Jinan Science and Technology Bureau under Grant No. 2019GXRC061, in part by the Key Program of the National Natural Science Foundation of China under Grant No. 82330064, in part by the Major Innovation Project for the Science Education Industry Integration Pilot Project of Qilu University of Technology (Shandong Academy of Sciences) under Grant No. 2023JBZ03.

ORCID

Jiancai Leng  https://orcid.org/0000-0002-2427-7499

Jiaqi Zhao  https://orcid.org/0000-0003-1647-5548

Yongjian Wu  https://orcid.org/0009-0002-8436-8568

Chengyan Lv  https://orcid.org/0009-0009-5585-0866

Zhixiao Lun  https://orcid.org/0009-0000-6701-5934

Yanzi Li  https://orcid.org/0009-0006-9285-8646

Chao Zhang  https://orcid.org/0009-0009-6825-406X

Bin Zhang  https://orcid.org/0009-0005-0450-4462

Yang Zhang  https://orcid.org/0000-0002-3086-2432

Fangzhou Xu  https://orcid.org/0000-0001-7660-1206

Changsong Yi  https://orcid.org/0009-0008-2251-8827

Tzyy-Ping Jung  https://orcid.org/0000-0002-8377-2166