Over the past decade, attention mechanisms have experienced significant advancements. SENet (Hu et al., 2018) introduced the principle of channel-wise attention, which has subsequently influenced the development of numerous architectures, including A²-Net (Chen et al., 2018), CBAM (Woo et al., 2018), ECANet (Wang et al., 2020), and FcaNet (Qin et al., 2021). These mechanisms have also been extensively applied in MI-EEG classification (Song et al., 2023), where their integration into decoding frameworks frequently results in marked performance enhancements. Inspired by these developments, we integrate attention mechanisms within DSSICNN.

CFC denotes the modulation of neural oscillatory activity in one frequency band by activity in another band. Empirical studies have demonstrated that during MI tasks, CFC occurs across multiple frequency bands (Feng et al., 2020). Moreover, distinct MI tasks induce characteristic CFC patterns across both spectral and spatial dimensions. Consequently, CFC constitutes a vital source of discriminative features relevant to MI decoding.

The core concept of DSIGSTFE is to incorporate principles from network neuroscience (Bassett and Sporns, 2017) into the MI-EEG decoding pipeline to improve the extraction of spatial semantic features. Specifically, DSIGSTFE captures spatial features across both Euclidean and non-Euclidean domains simultaneously. While numerous studies (Liu et al., 2023) have utilized spatial convolutions to extract spatial information with considerable decoding accuracy, insights from network neuroscience indicate that EEG signals possess graph-structured characteristics, and spatial convolution alone is insufficient to comprehensively model interactions among brain regions, often referred to as spatial interactions. To address this limitation, a dynamic GNN is employed to model these spatial interactions and extract Spatial Guidance Feature, thereby enhancing the spatial representations derived from convolutional operations. Furthermore, in comparison with previous approaches, the proposed dynamic GNN is integrated into a parameter-efficient MI-EEG decoding framework, resulting in a novel architecture that demonstrates notable strengths in both methodological innovation and computational efficiency. The parameter-efficient MI-EEG decoding framework we designed effectively mitigates overfitting during training, while simultaneously ensuring that DSSICNN possesses strong potential for deployment in MI-BCI.

At present, most DL-based MI-EEG decoding algorithms tend to introduce, in the shallow layers of the network, a convolutional layer with multiple filters to transform raw input signals into more expressive latent representations (e.g., the spatial convolution layer in DSSICNN). Such latent representations typically consist of multiple spectral (frequency) channels. Inspired by the phenomenon of CFC, we posit that explicitly modeling the interactions among spectral channels can facilitate the extraction of MI-related discriminative features. However, the majority of existing DL-based MI-EEG decoding approaches largely overlook the modeling of interactions among spectral channels. Given that prior studies have rarely calibrated features according to the relative importance of individual spectral channels, this work designs an attention-based module inspired by CFC to further enhance decoding performance. This module adaptively recalibrates spectral features, thereby promoting the effective representation of cross-spectral interactions. In particular, spectro-wise features are explicitly adjusted via attention mechanisms to capture interactions across spectral bands. Achieving this recalibration necessitates modeling the correlations among distinct frequency components.

To characterize the relationships among spectral bands, global spectro-wise information is initially aggregated. Both max pooling and average pooling are considered effective for capturing critical global spectro-wise features. The computation of this spectro-wise global information is expressed as follows:

where zmaxpool∈Rm is the spectro-wise global information generated by max pooling, F_maxpool(·) is the max pooling layer, zavgpool∈Rm is the spectro-wise global information generated by average pooling, and F_avgpool(·) is the average pooling layer.

Subsequently, the aggregated spectro-wise global information is utilized to systematically capture correlations across spectral bands. The spectro-wise global information obtained in the previous step serves as the initial set of coefficients, which are then employed to model cross-spectral interactions through coefficient reconstruction. Inspired by cross-correlation methods, a procedure is devised for reconstructing these coefficients. The computation for obtaining the reconstructed coefficients is defined as follows:

where umaxpool∈Rm and uavgpool∈Rm are the spectral context descriptors, weightmaxpool∈Rl and weightavgpool∈Rl are the trainable weight vectors, and l is the size of the cross-correlation window.

Once the spectral context descriptors have been obtained, they are integrated through element-wise addition.

where u∈R^m is the fused spectral context descriptors, and ϕ(·) is the sigmoid activation function.

Finally, spectro-wise feature recalibration is accomplished by performing element-wise multiplication between the ACSIM input features and the spectral context descriptors. This operation further enhances cross-spectral interactions, as illustrated below:

where ⊗ represents the spectro-wise multiplication, and the xacsif∈Rm×t represents the output features of ACSIM.

It is worth noting that several recent studies have begun to introduce attention-based spectral feature recalibration modules. Among them, M-FANet (Qin et al., 2024) is a representative example, which directly employs SENet (Hu et al., 2018) to achieve spectral feature recalibration. However, the two linear layers inherent in SENet, which perform dimensionality reduction followed by dimensionality expansion, may lead to the loss of information in the original spectral channels. In contrast, ACSIM models interactions among spectral channels directly via cross-correlation without relying on fully connected layers, thereby avoiding information loss in the original spectral channels. Moreover, SENet extracts spectral channel information solely through average pooling, whereas ACSIM jointly utilizes average pooling and max pooling, enabling the capture of richer and potentially more informative spectral channel representations.

Mental fatigue can lead to significant variations in the quality of MI-related temporal fluctuation patterns within the time windows described in Section 3.2.1. This phenomenon occurs because fatigue may compromise a participant's ability to consistently perform MI tasks at certain stages of the experiment. Accordingly, accurately modeling the quality of temporal fluctuation patterns throughout the MI process, in order to assess the contributions of different MI stages to decoding performance, is crucial for enhancing overall decoding accuracy.

To address this challenge, this study proposes a novel attention-based framework that models the quality of MI fluctuation patterns across different stages of MI task execution by subjects, thereby aggregating feature representations from multiple time windows along the temporal axis. This strategy enhances the model's capacity to capture dynamic temporal dependencies. Specifically, attention weights are assigned to quantify the relative significance of features within each time window, and a weighted sum of features across all windows is subsequently computed based on these weights. Furthermore, because the spatial convolution in the Temporal-Spectral-Spatial Branch encodes spectral information within each time window (as detailed in Section 3.2.1), temporal fluctuation patterns may vary across different spectral bands. To account for this variability, independent attention weights are employed to aggregate temporal features for each spectral band individually.

In summary, the computation of ASDTA can be formally expressed as follows:

where xasdtf∈Rm is the output feature of ASDTA, α_ij is the attention weight of the j^th time window in the i^th spectral channel, and weightasdtf∈Rm×t is the trainable matrix.

Finally, a linear layer is employed to compute the probability distribution over the classes. The cross-entropy loss function is used to guide the training process. In addition, the Adam optimizer (Kingma and Ba, 2014) is adopted to update the parameters of DSSICNN. The hyperparameters are configured as follows: the number of spatial filters m is set to 64, the time window size w is set to 250, the dropout rate is set to 0.1, the weight decay is set to 0.001, and the learning rate is set to 0.001.

The effectiveness of DSSICNN was evaluated using two publicly available MI-EEG datasets, namely BCIC-IV-2a (Dataset A) (Tangermann et al., 2012) and OpenBMI (Dataset B) (Lee et al., 2019). The MI-EEG data were preprocessed following the methodology described in Liu et al. (2023). No data augmentation was employed to increase the size of the two datasets mentioned above.

DSSICNN was evaluated against a range of baseline models, including Deep ConvNet (Schirrmeister et al., 2017), Shallow ConvNet (Schirrmeister et al., 2017), EEGNet (Lawhern et al., 2018), FBCNet (Mane et al., 2021), SHNN (Liu et al., 2022), EEGConformer (Song et al., 2023), FBMSNet (Liu et al., 2023), IFNet (Wang J. et al., 2023), ADFCNN (Tao et al., 2024), MI-BMInet (Wang et al., 2024), and CLTNet (Gu et al., 2025). To ensure the rigor and fairness of the comparison, most baseline models were implemented using publicly available open-source code.

To systematically evaluate the performance of DSSICNN, we conducted experiments under both subject-dependent and subject-independent settings. The subject-dependent setting further includes session-dependent and session-independent configurations. Specifically, the subject-dependent setting indicates that both the training and testing sets are drawn from the same subject, whereas the subject-independent setting denotes that the testing set is obtained from a single subject and the training set is constructed from all remaining subjects. The session-dependent setting refers to the case where both the training and testing sets are derived from the first session, with a training-to-testing ratio of 9:1, whereas the session-independent setting indicates that all data from the first session are used as the training set and all data from the second session are used as the testing set.

The Wilcoxon signed-rank test was applied to evaluate the statistical significance of performance differences between DSSICNN and the baseline models. The Wilcoxon signed-rank test is the non-parametric counterpart of the paired t-test. In addition, the Wilcoxon signed-rank test was conducted at the subject level. All models were trained on an NVIDIA Tesla A100 GPU.

The highest decoding performance is indicated in bold, with * denoting p < 0.1 and ** denoting p < 0.05. It should be noted that the accuracy of Dataset A refers to the classification accuracy across the four classes.

As presented in Table 1, under the session-independent setting, DSSICNN achieves the highest accuracy and F1-score, attaining 80.09% and 79.79%, respectively, and significantly outperforming most baseline models (p < 0.05). Similarly, in the session-dependent setting, DSSICNN records the highest accuracy and F1-score, reaching 86.71% and 86.19%, respectively, significantly surpassing all baseline models. Moreover, in the session-independent setting, DSSICNN outperforms the leading baseline model (ADFCNN) by 2.2% in accuracy and 2.29% in F1-score. In the session-dependent setting, it achieves gains of 4.65% in accuracy and 4.81% in F1-score relative to the best baseline model (FBMSNet). Importantly, DSSICNN delivers these substantial performance improvements while maintaining a parameter-efficient architecture. It is worth noting that CLTNet achieves superior decoding performance compared to DSSICNN. However, CLTNet relies on data augmentation during practical deployment, whereas DSSICNN does not introduce any data augmentation, which to some extent compensates for the slightly lower decoding performance of DSSICNN relative to CLTNet. In addition, CLTNet employs stacked Transformer modules and LSTM modules. The inherent self-attention mechanism within Transformers substantially increases both the spatial and temporal computational complexity of CLTNet, while the intrinsically non-parallelizable computation of LSTM further leads to a significant increase in time complexity. These factors impose certain limitations on the practical deployment of CLTNet in MI-BCI.

The comparative results in Table 2 demonstrate that DSSICNN achieves superior performance in both accuracy and F1-score under session-independent and session-dependent settings on Dataset B. In the session-independent setting, DSSICNN attains an accuracy of 77.87% and an F1-score of 77.54%, significantly outperforming all baseline models (p < 0.05). Similarly, in the session-dependent setting, DSSICNN achieves an accuracy of 89.00% and an F1-score of 88.72%, significantly surpassing all baseline models (p < 0.05). Furthermore, in the session-independent setting, DSSICNN improves upon the leading baseline model (IFNet) by 4.56% in accuracy and 5.88% in F1-score. In the session-dependent setting, it outperforms the top-performing baseline model (IFNet) by 12.53% in accuracy and 13.39% in F1-score.

The outcomes of the comparative experiments indicate that integrating network neuroscience priors into the MI-EEG decoding framework substantially enhances decoding performance relative to conventional CNN and Transformer architectures. Notably, DSSICNN exhibits a more pronounced improvement on Dataset B, which may be attributable to the relatively high incidence of “BCI illiteracy” within this dataset. These findings suggest that DSSICNN is particularly effective in improving decoding performance for individuals with lower inherent BCI proficiency compared with the baseline models.

In addition, as a supplement to the comparison of subject-dependent decoding performance, ADFCNN, FBCNet, EEGNet, Multi-branch 3D CNN (Zhao et al., 2019), CCNN (Amin et al., 2019), CRAM (Zhang et al., 2019) and FBCSPCNN (Bhambare and Jain, 2024) were selected as baselines to evaluate subject-independent decoding performance on Dataset A. The experimental results are presented in Table 3. DSSICNN also achieved the best decoding performance under the subject-independent setting. Its decoding performance was significantly superior to that of most baselines, except for ADFCNN and EEGNet. Moreover, the accuracy and F1-score of DSSICNN exceeded those of the strongest baseline (ADFCNN) by 1.52% and 3.23%, respectively. These results demonstrate that DSSICNN exhibits strong generalization capability and has substantial potential for deployment in MI-BCI. It is worth noting that FBCSPCNN achieves superior decoding performance compared to DSSICNN. However, FBCSPCNN relies on data augmentation in practical deployment, whereas DSSICNN does not introduce any data augmentation, which to some extent compensates for the slightly lower decoding performance of DSSICNN relative to FBCSPCNN.

To further discuss the feasibility of deploying DSSICNN in MI-BCI, Table 4 presents a comparison between DSSICNN and the baselines on Dataset A in terms of the number of trainable parameters, the average training and testing time per epoch, the memory footprint, and the FLOPs. Analysis indicates a strong correlation between the number of trainable parameters and memory footprint: models with a larger number of trainable parameters generally exhibit a larger memory footprint, and vice versa. Based on the comprehensive analysis of Tables 1, 4, DSSICNN achieves a relatively favorable balance among decoding accuracy, spatial complexity, and temporal complexity. Therefore, DSSICNN can be regarded as “parameter-efficient.” Although ADFCNN and IFNet have smaller memory footprints than DSSICNN, DSSICNN exhibits faster training/inference speeds and superior decoding performance compared with both ADFCNN and IFNet. While FBCNet and EEGNet require less memory and achieve faster training/inference than DSSICNN, the decoding performance of DSSICNN is substantially superior to that of FBCNet and EEGNet. Notably, despite having one of the smallest memory footprints, ADFCNN demonstrates the slowest training/inference speeds, which significantly limits its applicability to MI-BCI. We hypothesize that this limitation arises from the inclusion of multi-scale CNN components that cannot be efficiently parallelized, as well as a computationally intensive self-attention mechanism. The FLOPs of DSSICNN are second only to those of EEGNet and EEGConformer. Notably, DSSICNN exhibits higher FLOPs than FBMSNet, whose spatial and temporal complexities are greater than those of DSSICNN. We attribute this difference to the use of spatial convolution layers in DSSICNN with a relatively large number of output channels and full connectivity between input and output channels. In addition, the Spatial Guidance Branch of DSSICNN involves matrix multiplications with relatively high dimensionality, which further increases the FLOPs. Owing to the optimizations for convolution operations and matrix multiplications in the PyTorch framework adopted in this study, DSSICNN is able to maintain favorable training and inference speeds. The number of trainable parameters in FBCSPCNN is substantially higher than that of DSSICNN. This indicates that FBCSPCNN relies, to a certain extent, on data augmentation to expand the training data in order to effectively train a network with a large number of trainable parameters.

By achieving a favorable balance among decoding accuracy, spatial complexity, and temporal complexity, DSSICNN exhibits strong potential for deployment in MI-BCI. First, prior studies have demonstrated that excessively long recalibration times can reduce user engagement with MI-BCI and thereby degrade MI-BCI performance (Edelman et al., 2025). DSSICNN does not employ data augmentation and thus does not significantly increase data preprocessing time, and it further exhibits favorable training and inference speeds. Consequently, DSSICNN can effectively shorten recalibration time and improve MI-BCI performance. Second, as reported in Forenzo et al. (2025), MI-BCI is often implemented on hardware with limited computational resources, and DSSICNN can effectively reduce hardware memory consumption.

To evaluate the sensitivity of DSSICNN to hyperparameter selection, experiments were conducted on two datasets under the session-independent setting, investigating the impact of varying hyperparameter configurations on decoding performance.

To evaluate the individual contributions of DSSICNN components to decoding performance, ablation experiments were performed on two datasets under the session-independent setting. The configurations of the DSSICNN variants employed in these studies are detailed in Table 6. In this framework, Branch 1 refers to the Temporal-Spectral-Spatial Branch, while Branch 2 corresponds to the Spatial Guidance Branch.

The results reported in Table 7 demonstrate that DSSICNN significantly outperforms its variants (p < 0.05) across both datasets, underscoring the substantial contributions of each component to overall decoding performance. Figure 3 displays the confusion matrices of DSSICNN and its variants for Subject 3 from Dataset A, showing that DSSICNN consistently achieves higher classification accuracy across all classes. These observations further highlight the essential role of each DSSICNN component in MI-EEG decoding. Examination of the DSSICNN confusion matrix indicates that the foot and tongue classes exhibit the highest misclassification rates across most tasks. This finding aligns with the results reported in Liu et al. (2025) and suggests that DSSICNN still has a limited capacity to extract sufficiently discriminative MI-related features for these two classes.

For Dataset A, ASDTA emerges as the component exerting the greatest impact on DSSICNN's decoding performance. This effect is likely due to substantial variability in subjects' attention levels during MI task execution, which induces pronounced changes in the temporal fluctuation patterns associated with MI. Further examination of the confusion matrix in Figure 3 elucidates the specific contributions of each DSSICNN component to the classification of distinct MI tasks. In particular, the Spatial Guidance Branch of DSIGSTFE enhances the discriminative capacity of high-level semantic features associated with left-hand movements, whereas ACSIM amplifies the representational power of high-level features related to right-hand movements. In contrast, ASDTA primarily influences decoding accuracy for MI-EEG signals corresponding to regions other than the tongue. Notably, for Dataset B, all DSSICNN components contribute approximately equally to improvements in decoding performance.

To evaluate the feasibility of deploying DSSICNN in real-world applications, and inspired by Forenzo et al. (2024) and Forenzo et al. (2025), a simplified MI-BCI based robotic arm control system was developed in this study. The system comprised three principal components: a signal acquisition module, a laptop computer, and a robotic arm. MI-EEG signals were recorded using the NeuSen W Wireless EEG Acquisition System [Neuracle Technology (Changzhou) Co., Ltd., Changzhou, China] as the signal acquisition module. The laptop was responsible for cue presentation, reception and decoding of the MI-EEG signals transmitted from the acquisition module, and generation of decoding outputs for robotic arm control. The graphical user interface (GUI) and the MI-BCI experimental paradigm are illustrated in Figure 4. The cues presented on the GUI included both textual and graphical instructions, where the textual cues specified the actions to be performed by the participants and the graphical cues indicated the type of MI to be executed. Within the paradigm, the text displayed along the time axis (e.g., “Ready”) corresponded to the textual instructions, whereas the images aligned with the axis represented the graphical instructions. Five participants were recruited for the online experiments, and each participant completed 60 MI trials. The study protocol was approved by the Ethics Committee of Sichuan Provincial Rehabilitation Hospital, China (Approval No. CKLL 2018008), and written informed consent was obtained from all participants. The MI tasks consisted of two classes, left hand and right hand MI, which were mapped to the control of the left and right robotic arms, respectively. A 1s time window was used to segment the subjects' MI-EEG signals as control inputs. Given that the MI-EEG acquisition module operates at a sampling rate of 250 Hz, we set w = 50. In addition, the command delivery cadence was reported in terms of the average inter-command interval. For a given subject, a small amount of subject-specific data was first used to calibrate the MI-BCI. Subsequently, online experiments were conducted, and the decoding accuracy, F1-score, and command delivery cadence, reported in terms of the average inter-command interval, were evaluated.

The online experimental results are presented in Table 8. DSSICNN achieved satisfactory decoding performance and command delivery cadence in the online experiments, thereby validating its potential for deployment in MI-BCI.

Although the proposed DSSICNN demonstrates excellent decoding performance, several limitations remain. First, while DSSICNN can be effectively trained on small-scale datasets without the need for data augmentation, MI-EEG data are inherently graph-structured, and future work will investigate graph-based data augmentation techniques to increase the effective size of MI-EEG datasets. Additionally, although the brain is modeled as a complex network, the current approach does not explicitly account for its partitioned organizational structure. As a direction for future research, we aim to decompose the whole-brain graph into multiple subgraphs based on established functional brain partitions.