The uHD EEG data collection was approved by the Institutional Review Board of Korea Advanced Institute of Science and Technology (KH2018-127), and the gained dataset was published on OSF (Lee et al., 2022). The uHD EEG was collected from a finger movement experiment involving five healthy subjects (S1-S5). The participants used their dominant hands for testing, whereas S1 was left-handed and S2-S5 were right-handed. Each participant completed 10 sequential runs. Each run began with a 30-second baseline period, followed by alternating rest and finger extension periods. The rest period lasted between 3 and 4 s, while the extension (task) period lasted 5 s. During the task, participants extended the finger indicated by a colored cue displayed on a screen. Each finger was cued five times in a pseudo-random order, resulting in 25 finger extensions per run. Researchers use a novel g. Pangolin system to measure brain signal. The system achieves high spatial resolution, with 16 electrode grids (16 electrodes for each grid, 256 electrodes in total) placed over the contralateral sensorimotor cortex. The electrode grids were attached sequentially in the downward direction laterally of the head, starting with the first electrode grid at the Cz position. Dense distribution and small-sized electrodes result in an inter-electrode distance of 8.6 mm (uHD EEG). The sampling rate is 600 Hz, leading to 3,000 samples of each extension trail. The high number of electrodes and the high sample rate ensure that the data contain rich spatial and temporal information. This high sample rate also captures sufficient frequency bands for brain rhythms, which are key information carriers. Thus, the collected uHD EEG data encompass multi-domain information.

The data processing consists of two stages: preprocessing and feature extraction. In preprocessing, the raw data are initially standardized using a z-score, and then undergo common average referencing (CAR) to reduce interference from neighboring electrodes. Subsequently, a 4th-order Butterworth notch filter cascade is applied to eliminate interference of line power at 60 Hz and its harmonics (120 Hz, 180 Hz, and 240 Hz). We apply the FBCSP to extract spatial and frequency features. The FBCSP includes two steps:

Firstly, we utilize the Common Spatial Pattern (CSP) method to decompose the joint covariance matrices of the uHD EEG signals from two classes. CSP identifies spatial filters that lead to optimal discrimination between two classes:

Specifically, the CSP components refer to the eigenvector corresponding to the large eigenvalue of the joint covariance matrix (C1C2) associated with the first class, while the last CSP component corresponds to the small eigenvalue of the joint covariance matrix reflecting the large eigenvalue of the second class. To determine the optimal number of CSP components, experiments are conducted with the number of components ranging from 2 to 20. Ultimately, it is determined that a 12 (6 pairs) CSP components provides the best performance for the classification tasks, with 6 components corresponding to the first class and 6 to the second class for binary classification. Two CSP component patterns for movement pairs (thumb vs ring) are depicted in Figure 2 as an example.

Secondly, multiple 4th-order Butterworth band-pass filters are utilized to extract EEG rhythms: delta (1–4 Hz), alpha (8–12 Hz), beta (12–30 Hz), gamma (30–50 Hz), high gamma (75–115 Hz), and super high gamma (125–150 Hz). For CSP filtered rhythms, the centered mean power is calculated by subtracting the respective mean, squaring to get the power. The powers are log-transformed to enhance Gaussianity. Automated artifact rejection was performed to ensure signal integrity. For each channel, data segments with Z-scores greater than 6 were marked as artifacts. Entire experimental runs were rejected if more than 10% of the channels were classified as bad. All preprocessing parameters (z-score μ/σ), and all label-supervised spatial filters FBCSP were fitted using training runs only (runs 1-8) and then applied unchanged to test runs (runs 9–10). Hyperparameters were selected using validation within the training runs; test runs were used only once for final reporting. Notably, the percentage of data reduction from FBCSP is approximately 28%, and the output of the FBCSP will be fed into the proposed network.

In this study, we investigate the network performance by integrating convolutional layers with multiple current popular attention mechanisms, including self-attention mechanism, the Convolutional Block Attention Module (CBAM), and multi-head transformers. We define the network in a sandwich configuration, placing convolutional layers both before and after the attention modules to facilitate multi-scale feature extraction and summarization. For the CBAM module, which contains two sequential attention mechanisms, with the first attention mechanism deals with the filter domain information from the preceding convolutional layer, and the second attention mechanism processes 2D information from the spatial and rhythm domains. We then switch the two attention modules in CBAM to study the effect of the order by placing the first one to handle 2D spatial and rhythm information, and the second to process filter information. We also conduct a thorough investigation into the optimal number of convolutional layers surrounding the CBAM module by adding layers both before and after the CBAM. The classification results indicate that increasing the number of convolutional layers does not yield any significant benefits for final accuracy. We name the winning model as Sandwich enhanced CBAM (SCBAM). The flowchart of the SCBAM is indicated in Figure 3. The CBAM module is treated as a component in Figure 3a of the overall data flow, and the details of the CBAM are then elaborated in Figure 3b. Figure 4 presents the structure and computational process within the network. It should be noted that the input of the SCBAM network is the features extracted from FBCSP.

In this section, we provide details of the training process and the classification strategy. Given that the number of trials is relatively small for effective tuning of a deep neural network, we chop the trials from the 3000 samples into 600 samples per trial, with a 50% (300) overlap. We concatenate the chopped trials from 10 runs, resulting in a total of 2,250 trials, 80% of which are used for training and 20% for testing. To guarantee the separation of the training and testing data, the data from the first 8 runs are used for training, while the last 2 runs are used for testing. Furthermore, we utilize 10% of the training data for validation purposes. To improve the model's generalization capabilities, we will shuffle the training data along with the corresponding training labels. The proposed network architecture is configured to run for 150 epochs, allowing sufficient learning from the training data. It employs a batch size of 32, which balances memory efficiency and stability in gradient updates. A learning rate of 0.001 is chosen to facilitate effective learning and prevent overshooting during optimization. Additionally, the weights and biases of the filters in each layer are trained jointly. To ensure robust evaluation, we determined trial-wise accuracy using a majority voting scheme. For each 5-s trial, we aggregated the predictions of its nine constituent windows, assigning the most frequent class label to the entire event.

We initially conduct binary classification to assess the model's discrimination ability corresponding to each movement pair, leading to 10 classifiers in total, and subsequently balance memory efficiency and stability in gradient updates. The five-class classification is implemented using a One vs. One (OvO) strategy. The final results of the five-class classification are obtained through a majority voting mechanism involving 10 binary classifiers. Each binary classifier cast a vote for its predicted class, and the class with the most votes was declared the winner. The class receiving the highest number of votes is selected as the final prediction. If two or more classes receive an equal number of votes, the tie is first addressed by conducting a sub-vote among the tied candidates. If the tie remains unresolved, the final decision is determined by aggregating the softmax class probabilities from the corresponding pairwise classifiers; specifically, the class with the highest total cumulative probability across all relevant binary comparisons is assigned as the final label. Each binary classification is trained for 150 epochs, resulting in a total of 1,500 epochs (150 × 10 classifiers).

We use the accuracy metric for classification results evaluation:

where TP = True Positives, FP = False Positives, FN = False Negatives, and TN = True Negatives.

To provide a more detailed assessment of model performance, we also introduce the kappa statistic. Kappa contributes by quantifying the agreement between predicted and true classifications while accounting for chance agreement, thus offering insights into the model's reliability. Specifically, it highlights how well the model distinguishes between classes beyond random guessing, enabling a clearer evaluation of performance across both binary and multi-class scenarios.

where the P_o is the observed agreement, representing the proportion of instances where the raters agree. P_e is the expected possibility, indicating the proportion of instances where agreement will be expected by chance.

To evaluate classification significance against the empirical chance level, a non-parametric permutation test was performed with 5,000 iterations, where class labels were randomly shuffled at the trial level. The empirical p-value was calculated as the proportion of permutations where the shuffled accuracy was greater than or equal to the observed accuracy. To evaluate the performance difference between models, we conducted a paired t-test on the results obtained from the binary classification of 10 finger pairs at three significance levels (p < 0.05, p < 0.01, and p < 0.001). Furthermore, the stability of the results was quantified by calculating 95% confidence intervals via a non-parametric bootstrap procedure across subjects. For each model, subject-wise accuracies were resampled with replacement 1,000 times to estimate the distribution of the mean accuracy.

The binary and five-class classification results of individual fingers by the proposed SCBAM are presented in Table 1. The five subjects (S1 to S5) achieve accuracies of 77.38%, 79.40%, 79.69%, 78.57%, and 78.12% in binary classification, respectively, resulting in a mean accuracy of 78.63% (SD 1.56%; 95% CI [77.26, 80.00]%). Based on the results of binary classification, we observe that ten pairs of activities can be roughly categorized into three types: easy, moderate, and difficult.

The six pairs from Thumb vs. Index to Index vs. Ring in Table 1 are categorized as easy type, with accuracies above 80%. The Index vs. Little and Middle vs. Ring are recognized as moderate type, with accuracies between 70% and 80%. Meanwhile, the Middle vs. Little and Ring vs. Little are considered as difficult types, with accuracies ranging from 60% to 70%. The classification results indicate varying levels of performance among the subjects, with S2 achieving the best performance. Moreover, all classification pairs for all subjects performed significantly above chance (P < 0.05). The thumb vs. the Ring finger performs the best overall at 85.00 (1.16%) in mean accuracy across five subjects. The finger pairs Thumb vs. Index, Thumb vs. Middle, Thumb vs. Ring, and Index vs. Middle achieve a highly significant result at the level of P < 0.001. In binary classification, finger pairs yield distinct results based on the included fingers: pairs including the Thumb (Thumb vs. Index, Thumb vs. Middle, Thumb vs. Ring, Thumb vs. Little) achieve an average accuracy of 83.76% across five subjects. Additionally, pairs including the index, middle, ring, and little fingers achieve average accuracies of 81.83%, 76.87%, 76.78%, and 73.14%, respectively, across the same subjects.

The five subjects (S1 to S5) achieve accuracies of 60.73%, 62.36%, 60.00%, 61.81%, and 60.72% in five-class classification, respectively, resulting in a mean accuracy of 61.12% (SD 0.95%; 95% CI [60.29, 61.95]%). From the confusion matrix depicted in Figure 5, the SCBAM achieves average accuracies of 66.67% for Thumb, 64.71% for Index, 62.50% for Middle, 62.50% for Ring, and 59.46% for Little fingers across five subjects. From this confusion matrix, it is clear that adjacent fingers tend to confuse with each other. For example, the thumb is misclassified as the index at a rate of 13.33%, and the ring finger is classified as the little finger at a rate of 15.62%. The confusion matrix reveals distinct patterns, indicating that the thumb, index, and middle fingers form one group with significant confusion tendencies, while the middle, ring, and little fingers constitute another one. The confusion clusters observed in Figure 5 (orange boxes) align with the somatotopic overlap in the primary motor cortex. Specifically, the confusion between the middle, ring, and little fingers reflects the known muscular and neural coupling termed enslavement, where these fingers share common multi-tendon muscle groups. This makes the neural signatures for these fingers more similar than those of the thumb and index finger. The higher classification accuracy for the Thumb (66.67%) and Index finger (64.71%) compared to the Little finger (59.46%) is consistent with the principle of cortical magnification. The thumb and index finger have larger, more specialized representations in the motor homunculus. Our SCBAM model effectively captures this increased neural real estate through its attention mechanism, resulting in more robust feature extraction for these digits.

The Figure 6 depicts the loss and accuracy curves for binary and five-class classification tasks on training and validation sets. For binary classification, Figure 6a shows that the training loss and validation loss decrease and converge to 0.15 and 0.25, respectively. The training accuracy rapidly increases during the first 40 epochs, ultimately converging at 93%, indicating robust learning from the training dataset. In parallel, the validation accuracy rises quickly over the first 45 epochs, stabilizing at 80%, which suggests a commendable ability to generalize, albeit with a notable gap compared to training performance. For five-class classification, Figure 6b presents the training loss and validation loss exhibit a downward trend, but stabilize at around 0.20 for training loss and approximately 0.45 for validation loss. The training accuracy reaches about 78%, while validation accuracy is approximately 62%. Overall, the figure supports the assertion that 150 epochs are ample for ensuring sufficient learning of the model, as indicated by the convergence of both accuracy and loss metrics. Continuing training beyond this point is unlikely to enhance performance and may lead to overfitting.

As our study includes both binary classification and five-class classification, we employ the Kappa statistic as a complementary measure to accuracy, providing a more nuanced classification result by accounting for chance agreement. The results are shown in Table 2. It achieves an average Kappa of 0.7898 for binary classification across 10 pairs in five subjects. For five-class classification, the average Kappa is 0.6130.

In our network design, we conduct a thorough study of the parameter settings, including the number of convolutional layers, the number of attention modules, and the order of attention module placement. By comparing the average accuracy across five subjects from different model candidates reported in Section 2, we identify the CBAM as the most suitable attention mechanism for our classification task. First, the self-attention mechanism between the initial and the final convolutional layer (CSA) model achieves an average accuracy of 63.98% for binary classification and 58.95% for five-class classification. Next, two convolutional layers integrating with a CBAM module in sandwich shape (SCBAM) model achieve an accuracy of 78.63% for binary classification and 61.12% for five-class classification. Integrating with a multi-head transformer (CTF) achieves an average accuracy of 64.89% for binary classification and 59.34% for five-class classification. Results from these three models indicate that embedding the CBAM module yields the best classification accuracy for our task. Switching the two attention mechanisms results in a decrease of 23.00% in binary classification accuracy and a drop of 24.40% in five-class classification accuracy. These results reveal that the order of attention modules is significant in network design. Attempts to add extra convolutional layers before and after the SCBAM model (C-SCBAM and SCBAM-C) all resulted in a drop in classification accuracy as reported in Table 3. These findings demonstrate that having one layer before the attention modules and one layer after is sufficient for our task. Excessive layers led to a model that was too complex for the limited training data, resulting in overfitting. Besides, the significance test of the binary classification results for all variants in the network design is presented in Figure 7a, the ROC curves further illustrate the trade-offs between sensitivity and specificity for each model as depicted in Figure 7b.

The ablation studies of SCBAM performed across five subjects are reported in Table 4, involving a series of model variants. By removing two convolution layers, only the CBAM module achieves an average decrease in accuracy of 16.94% for the binary classification task and a decrease of 12.55% for the five-class classification task compared to the SCBAM. This significant decrease in accuracy (at a level of <0.001) illustrates that the two convolutional layers also play a crucial role in performance, although not as profound as that of the CBAM module. By removing the first convolutional layer, the CBAM model shows an average accuracy decrease of 15.12% for the binary classification task compared to the proposed model. Furthermore, for the five-class classification, removing the first convolutional layer results in an accuracy drop of 4.93%. The last convolutional layer also plays an important role, as its removal leads to a drop of 15.90% for the binary classification and a decrease of 1.12% for the five-class classification task. By conducting a t-test between the two model variants and the proposed SCBAM, the results indicate that both individual convolutional layers are significant contributors to the model's performance. By removing the attention modules, the two convolution layers (two convs) yield a value of 28.10% for five-class classification, whereas binary classification values range from 46.32% to 65.76%. This results in an average accuracy decrease of 21.62% for binary classification and 33.02% for five-class classification compared to the proposed model. These results demonstrate that the CBAM plays a crucial role in enhancing the model's performance. By removing the second attention module, the SA1 shows an average accuracy decline of 16.87% for the binary classification and a decrease of 9.57% for the five-class classification compared to SCBAM. By removing the first attention module, the SA2 model exhibits an average accuracy decrease of 9.77% for the binary classification and a decrease of 2.67% for the five-class classification compared to the proposed network. From the accuracy comparison, the first attention mechanism has the weakest effect on classification performance. This may be due to the fact that it operates in the filter domain generated by the preceding convolutional layer, meaning it refines the obtained features rather than extracting new features from the data. On the contrary, the second attention mechanism has a greater effect on the model than the other modules. This is because it is the key component for handling rhythm domain information, which is not addressed by any other modules.

Besides, ablation experiments are conducted to verify the effectiveness of FBCSP and CAR. Compared to the proposed method, if we remove the FBCSP from the data processing pipeline, the accuracy drops by 23.88% in binary classification and 12.03% in five-class classification. If we remove the CAR, the accuracy declines by 17.32% in binary classification and by 7.94% in five-class classification, respectively. These results confirm that both steps are indispensable for achieving optimal classification performance. It is worth noting that, through the significance test of the binary classification results for all variants within the proposed SCBAM data pipeline, as presented in Figure 8a, none of the modules or processing steps can be removed, as they all play significant roles in the final performance at a level of <0.01. Besides, the ROC curves further validate the trade-offs between sensitivity and specificity for each model as illustrated in Figure 8b.

We compare the classification performance of the proposed network with state-of-the-art models, including EEGNet, EEGLearn, Shallow ConvNets, Deep ConvNets, EEG Conformer, and Msvt Net with FBCSP features and raw data independently. The classification results with FBCSP features are presented in Table 5. For binary classification, SCBAM exceeds the performance of the EEGNet, EEG Learn, Shallow ConvNets, Deep ConvNets, EEG Conformer, and Msvt Net by 25.69%, 12.93%, 14.37%, 13.31%, 2.59%, and 13.25%, respectively. A paired t-test is conducted to assess the significance of the differences in binary classification accuracy between the proposed model and six benchmark models, as demonstrated in Figure 9a. Specipically, the SCBAM surpasses the EEG Conformer and Deep ConvNets at the 0.05 significance level. The SCBAM significantly outperforms EEGLearn, Shallow ConvNets, and Msvt Net at the 0.01 level. SCBAM exhibits a highly significant performance advantage over EEGNet at the 0.001 level. For five-class classification, SCBAM achieves an average classification accuracy of 61.12%, while the EEGNet, EEGLearn, Shallow ConvNets, Deep ConvNets, EEG Conformer, and Msvt Net exhibit accuracies in the range from 45% to 58%. The ROC curves further illustrate the trade-offs between sensitivity and specificity for each model as presented in Figure 9c. The accuracy and loss learning curves during training are depicted in Figures 10a, c. The accuracy of SCBAM rises sharply until it reaches 94.00% around the 45th epoch, surpassing benchmark models in both convergence speed and final accuracy. Subsequently, the accuracy of EEG Conformer and Deep ConvNets stabilizes around 90% at 120th epoch. While Shallow ConvNets, EEG Learn, and Msvt Net acheive approximately 80% accuracy at the 120th epoch. Among the six benchmark models, Shallow ConvNets converge most rapidly, reaching stability by the 80th epoch. The loss of SCBAM exhibits a sharp decline during the first 40 epochs, reaching a low value of approximately 0.25 by the 80th epoch. The loss of EEG Conformer and Deep ConvNets reaches a turning point around the 60th epoch, achieving a low value of approximately 0.4 by the 120th epoch. Shallow ConvNets converges more rapidly reaching its stability loss value of approximately 0.5 by the 100th epoch. The results indicate that SCBAM significantly surpasses the performance and efficiency of these six models.

The classification results with raw data are summarized in Table 6, SCBAM consistently outperforms benchmark models across both binary and five-class classification tasks. In binary classification, SCBAM achieves accuracy improvements of 20.06%, 14.42%, 9.95%, 5.03%, 0.99%, and 7.47% over EEGNet, EEGLearn, Shallow ConvNets, Deep ConvNets, EEG Conformer, and Msvt Net, respectively. In the five-class task, SCBAM surpasses these benchmark models by 18.03%, 16.50%, 13.13%, 9.12%, 4.61%, and 10.99%, respectively. A paired t-test is conducted to assess the significance of the differences in binary classification accuracy between the proposed model and six benchmark models, the results are demonstrated in Figure 9b. SCBAM achieves highly significant improvements over EEGNet and EEGLearn (P < 0.001), and significantly outperforms Shallow ConvNets and Msvt Net (P < 0.01) and Deep ConvNets at the 0.05 significance level. It is worth noting that the EEG Conformer demonstrates strong classification performance that is statistically comparable to our proposed SCBAM (P = 0.43). The ROC curves further illustrate the trade-offs between sensitivity and specificity for each model as presented in Figure 9d. Besides, the accuracy and loss learning curves during training are depicted in Figures 10b, d. SCBAM stabilizes at an accuracy of 81% around 130th epoch and a steady loss value of 0.4 around 150th epoch, outperforming other models.

In summary, the classification results of the SCBAM and benchmark models with FBCSP features significantly outperformed those obtained from raw data. This underscores the importance of integrating well-established features with deep learning networks, as it facilitates a balance between performance and efficiency. Consequently, we investigated the efficiency analysis of the seven models with FBCSP features. As described in Section 2.4, we conduct 2,250 trials from 10 experimental runs, where 1,800 trials from the first 8 runs are used for training, and 450 trials from the last 2 runs are used for testing. The assignment of training and test data is consistent for both binary and five-class classification scenarios. It is worth noting that the FBCSP step takes 314 s (around 5 min). The time consumption of seven models are indicated in Table 7. For binary classification, SCBAM exhibits the most efficient performance with a per-epoch time of 1 s, resulting in a total training time of 25 min and a testing time of just 10 s. Shallow ConvNets, though slightly slower with a per-epoch time of 1.8 s, maintain a competitive training time of 45 min and a testing time of 14 s. Deep ConvNets demonstrate a more pronounced increase in computational demands, featuring a per-epoch time of 3 s and a total training time of 74.5 min, with a testing time of 28 s. EEG Conformer performs a per-epoch time of 3.3 s and a total training time of 80 min, with a testing time of 24 s. EEGNet, Msvt Net, and EEGLearn exhibit even greater time consumption. For five-class classification, SCBAM exhibits the most efficient performance with a per epoch time of 1.5 s, resulting in a total training time of 37.5 min ((150 × 10 epochs × 1.5 s) / 60) and a testing time of just 17 s, which is lower compared to the other six models. This analysis underscores SCBAM's efficiency, making it the optimal choice for scenarios requiring quick training and inference times. It is worth noting that both the SCBAM and Shallow ConvNets have two convolutional layers. Deep ConvNets, EEG Conformer, EEGNet, EEGLearn, and Msvt Net rely on a larger number of convolutional layers and deeper networks for feature extraction, require significantly longer training times.

Comparing the time consumed in the binary and five-class classification scenarios of binary and five-class classification, it is evident that the five-class classification always consumes longer time in each epoch and total training and test time. This is due to the one-vs-one strategy employed in the five-class classification. In this scenario, we utilize 10 binary classifiers for each trial, leading to a cumulative processing time that may suggest a 10-fold increase relative to a single binary classification. In contrast, the binary classification task only requires distinguishing between two classes, which necessitates computational resources for just one binary classifier. Additionally, in the binary classification framework, each movement is represented in four pairs. For instance, the thumb is compared against the index, middle, ring, and little fingers. Consequently, each trial is read and classified four times overall. In summary, in five-class classification, each data trial is read once but classified 10 times using binary classifiers. In contrast, in the binary classification scenario, the data for each trial is read four times and classified through four different binary classifiers. These factors contribute to the observation that while the training time for five-class classification is longer than that for binary classification, it is not as extensive as 10 times the duration of binary classification.

First, we compare our results with those from recent literature that utilized the same uHD EEG dataset, including one study that employed an SVM classifier (Lee et al., 2022) and another that used an MLP (Nemes et al., 2024). The results are summarized in the Table 8, indicating varying levels of accuracy across different finger pairs. The SCBAM method consistently outperforms these two approaches, achieving the highest average accuracy of 78.63% across all finger pairs and subjects. In contrast, the SVM and MLP methods recorded average accuracies of 64.8% and 65.68%, respectively. As depicted in Figure 11, statistical analysis using the t-test reveals that SCBAM significantly improved performance compared to both SVM and MLP at a significance level of 0.001. These findings highlight the accuracy of the SCBAM achieved for individual finger binary classification, suggesting it as a superior choice for this task. One thing to note is that in both studies, the ring finger is reported to have greater accuracy and has been regarded as easier to distinguish from the others. This is attributed to the fact that the four movement pairs, including the ring finger, rank among the top four in accuracy of all ten finger pairs in their results, which was also inferred in Table 1 in (Lee et al. 2022) and Table 3 in (Nemes et al. 2024). However, our results indicate that the ring finger does not stand out significantly. Instead, it is more reasonable to assert that the thumb and index fingers demonstrate greater classification accuracy in our findings. The binary classification between five individual fingers with EEG data has also been reported in (Liao et al. 2014). Their findings indicate an average accuracy of 77.11%. In their results, even though the Ring vs. Little pair achieves the highest accuracy, the Ring vs. Thumb, Ring vs. Index, and Ring vs. Middle pairs are ranked 5th, 7th, and 9th in all ten pairs, respectively [as indicated in Table 2 in Liao et al. (2014)]. The overall classification performance of the ring finger pairs appears to be less pronounced.

A few studies have reported the five-class classification of individual fingers, particularly using ECoG signals. From the study of Kubanek (Kubánek et al., 2009), the five-class individual finger classification achieved 90.6%, 89.9%, 76.4%, 68.1%, and 76.7% for their five subjects, respectively, with an average of 80.3 (8.66)% across subjects. Chestek et al. (2013) employed a Naive Bayes classifier, which achieved classification accuracies of 74%, 98%, and 66% for each subject, with an average of 79.33 (13.49)% across three subjects. Their study emphasizes the effect of gamma band features. Both of these studies utilize macro ECoG signals, which have a spatial resolution of around 10 mm. For Hotson et al. (2016), who employed a micro ECoG with an electrode resolution of 2.3 mm, they achieved a five-class classification accuracy of 76% using Linear Discriminant Analysis (LDA) with only one subject. However, there are currently no literature reports on five-class classification results of individual finger flexion and extension tasks using uHD EEG data. In our study, we achieved five-class classification accuracies of 60.73%, 62.36%, 60.00%, 61.81%, and 60.72% for our five subjects, respectively. This results in an average accuracy of 61.12% with a small standard deviation of 0.95% across subjects. Compared to ECoG, our classification accuracy is lower; however, it is important to highlight that our results exhibit greater stability across subjects. The high variance observed in ECoG decoding is likely due to the cantotomy surgery, which is primarily designed to address clinical situations. This often results in electrode placements varying significantly among individuals, leading to configurations that may be either far from or close to the motor cortex. In contrast, uHD EEG collection does not depend on clinical surgery for electrode implantation, allowing for convenient electrode placements and calibrations. This property contributes to the stable results observed in our study. Xiao and Ding (2013) also demonstrated that with EEG signals, five-class classification reached an accuracy of 45.2% using three principal component features. The significant difference in accuracy between uHD EEG and EEG data underscores that spatial resolution is a crucial factor for individual finger decoding. What's more, from our confusion matrix and the confusion matrix derived from ECoG results, it is evident that adjacent fingers tend to confuse each other, while fingers that are farther apart are easier to distinguish. This aligns with the common understanding: fingers that are distanced from each other have less overlapping musculature and are less likely to interfere with one another during movement. Additionally, their corresponding cortical areas are situated farther apart, making them comparatively easier to differentiate in classification tasks. The observed grouping of the Thumb, Index, and Middle fingers in Figure 5 reflects both the common innervation by the median nerve and the functional synergies required for precision handling (Schieber and Santello, 2004). Conversely, the confusion between the Ring and Little fingers aligns with ulnar nerve distribution and the enslavement phenomenon, where peripheral mechanical and neural coupling limit independent digit control (Häger-Ross and Schieber, 2000). Our model's higher accuracy for the thumb and index is further supported by the principle of cortical magnification, as these digits possess significantly larger and more complex representations in the motor cortex (Schieber, 2001). This finding is crucial for understanding the underlying neural mechanisms of finger movements.

In our study, we design a Sandwich-enhanced CBAM (SCBAM) network that integrates convolutional layers with a CBAM module. The two attention mechanisms in CBAM process information from different domains for uHD EEG signals. The filter attention highlights time-frequency patterns, prioritizing the most informative frequency bands relevant to finger movements. Spatial attention focuses on the most active electrodes, enhancing sensitivity to subtle differences in motor patterns. The multidimensional feature extraction capability of the attentional mechanisms in CBAM aligns well with the multi-domain features of uHD EEG signals, leading to significant improvements in both performance and efficiency. This is also validated by the results in Section 3.4, the proposed SCBAM model shows significant improvements in performance over several state-of-the-art models, including EEGNet, EEGLearn, Shallow ConvNets, Deep ConvNets, EEG Conformer, and Msvt Net in both binary and five-class classification tasks, Tables 5, 6. It also requires much less inference time compared to these models in Table 7. This is due to the attention mechanisms in SCBAM, which are also validated in ablation studies.

The classification results indicate that while convolutional layers are highly effective at extracting multi-level features for movement discrimination, they eventually reach a performance saturation point. Simply increasing the depth of the model, as seen in architectures like EEGNet (Lawhern et al., 2018) or EEGlearn (Bashivan et al., 2016), does not necessarily lead to breakthroughs, creating a performance upper bound for both binary and five-class classification. The introduction of attention mechanisms, such as those in the EEG Conformer (Song et al., 2023), represents a strategic shift in model design. These mechanisms operate along the preceding convolutional filter domain to enhance both classification performance and computational efficiency. It is worth noting that the EEG Conformer demonstrates a strong classification ability that is statistically comparable to the proposed SCBAM (P = 0.43) when fed with raw data. However, the proposed SCBAM further refines this approach by embedding dual attention modules to manage innovative discriminative features from the filters, spatial, and rhythm domains. This dual attention module plays a crucial role in automatically extracting discriminative features from the input data in the proposed SCBAM network. The input of the proposed SCBAM network is the features extracted by the Filter Bank Common Spatial Pattern (FBCSP). These features have been thoroughly studied for decades (Lotte and Guan, 2011), and it is widely accepted that they are key features in neural decoding. We drew the inspiration of combining traditional features with neural networks from the Shallow ConvNets proposed by Tonnio Ball (Schirrmeister et al., 2017). This two-step design integrates established signal processing principles with the data-driven capabilities of deep neural networks, resulting in a lightweight architecture that achieves statistically significant results in individual finger classification. While the five-class accuracy of 61.3% notably above the 20% chance level, highlights the inherent difficulty of decoding fine-grained finger movements, our analysis of the error structure reveals that misclassifications are primarily concentrated among anatomically adjacent fingers. This performance provides a robust foundation for future dexterous BCI research, where these specific decoding patterns can be further refined through larger cohorts and online experiment.

While the current study utilizes an open-access dataset (Lee et al., 2022) consisting of five subjects. Our laboratory is currently initiating experiments using identical equipment and experimental paradigms with more than 20 subjects. This expansion will facilitate a more comprehensive investigation of inter-subject variability. Furthermore, we plan to bridge the gap between offline classification and practical utility by transitioning to closed-loop online experiments. Future research will also prioritize optimizing real-time performance metrics, such as Information Transfer Rate (ITR) and system latency, to meet the benchmarks necessary for intuitive and responsive BCI control.