Introduction
The motor imagery (MI) technology based on electroencephalography (EEG)1 boasts high spatiotemporal resolution, non-invasiveness, and real-time capability, making it extensively applicable in fields such as rehabilitation medicine, neuroscience, human-computer interaction, and virtual reality2–4. However, MI technology still faces several challenges. For instance, the issue of channel redundancy may increase the burden during processing and analysis, while individual differences and environmental variations may compromise the generalization ability of models, limiting their effectiveness in practical applications. Therefore, further research and improvement are necessary to overcome these challenges and enhance the reliability and practicality of MI technology.
The signal acquisition in brain-computer interface (BCI) systems typically requires multi-channel EEG data due to their excellent performance. However, certain channels may carry redundant information and noise. Consequently, channel selection methods are essential to improve the performance of MI-based BCI systems5. These methods help eliminate channels with noise and redundant signals, thereby reducing data dimensionality, improving signal-to-noise ratio, and mitigating the occurrence of overfitting. Moreover, considering the variability in brain electrical activity among different individuals, selecting signal recording locations relevant to the task can effectively reduce preparation time and enhance the comfort of non-clinical BCI applications6. By meticulously selecting the optimal channels, noise interference can be minimized and the computational costs associated with processing high-dimensional data can be reduced.
Channel selection methods generally fall into four categories: filtering, wrapping, embedding, and manual selection. Manual selection requires a significant amount of time and effort, while filtering and embedding methods lack consideration for the spatial distribution of electrodes, thus failing to select the optimal channel configuration adequately. Wrapping methods rely on various subset search strategies, typically utilizing classifier accuracy as a reference. Aydemir et al.7 proposed a robust and individual-specific sequential forward search method (RSS-SFSM) for channel selection. It performs sequential search within candidate channels during the channel selection process to find channels that maximize validation set classification performance. Wei et al.8 employed a genetic algorithm to evaluate each selected channel subset and chose channels based on the obtained classification accuracy, with Fisher discriminant analysis used as the classifier. Some studies select channels based on feature weights obtained from the evaluator. Handiru et al.6 introduced an iterative multi-objective channel selection (IMOCS) method, which iteratively finds the most relevant channels based on the anatomy and functional relevance of EEG channels for a given MI classification task. Additionally, recursive feature elimination (RFE) algorithms iteratively determine the optimal channels. Lal et al.9 employed the RFE algorithm to find the most relevant channels for MI without relying on prior knowledge of neurophysiology. Building on this, Schröder et al.10 further explored using the RFE algorithm to transfer a set of participant channel rankings to new participants, selecting relatively good channels. Currently, there is no algorithm that combines the weights and accuracy obtained from evaluators, and evaluators are relatively singular, typically exhibiting good fitting effects for only certain types of data, thus failing to consider the complex and diverse EEG data information. Building upon this, our study integrates random forest (RF), gradient boosting machine (GBM), and logistic regression (LR) with RFE to propose a hybrid recursive feature elimination-based channel selection method.
Deep learning is widely used in the field of BCIs11, surpassing traditional machine learning methods by a large margin. In contrast to traditional machine learning models, deep learning employs neural network architectures with multiple complex layers, each performing distinct functions. Many studies have applied methods such as convolutional neural network (CNN), recurrent neural network (RNN), and deep boltzmann machine (DBM) to motor imagery classification, achieving good recognition results12–14. However, these sample objects are all regular data in Euclidean space, without considering the topological relationships between electrodes. In order to address the inability to convolve non-euclidean spatial data, graph convolutional networks (GCN) have been proposed, as they can take into account the topological information between nodes, and have been widely applied in recent years15. In EEG signal analysis, nodes represent individual electrodes, and the correlations between different channels constitute the internal topological relationships of the data. Traditional CNN and RNN can only handle euclidean spatial data, whereas brain electrode-recorded data is essentially non-euclidean, and spatial information is crucial for recognizing motor imagery patterns. Graph convolutional networks, due to their unique algorithmic principles, reconstruct the spatial relationships of the brain. Sun et al.16 proposed an end-to-end deep learning framework based on GCN, achieving an 82% classification accuracy on the PhysioNet dataset. However, to fully utilize the topological relationships between electrode channels, all channels’ data are typically used as network inputs, leading to a considerable amount of redundant data17.
Therefore, this paper proposes a method that integrates recursive feature elimination-based channel selection and residual graph convolutional neural networks for classifying MI data. The channel selection method integrates three feature evaluation methods, RFE-RF, RFE-GBM, and RFE-LR, synthesizing the feature weight functions of three classifiers. Different weights are used to aggregate channel weights, resulting in the final channel importance ranking and selection of the optimal channel combination. Subsequently, a residual graph neural network is used to decode EEG-MI signals, fully exploiting the data’s feature information. Our method is validated on the Shanghai University motor imagery dataset and the PhysioNet dataset.
The remainder of this paper is organized as follows. The second part is the core content of the paper, which introduces the channel selection algorithm and the Graph ResNet model in detail. The third part tests the proposed algorithm, and provides experimental results and analysis. The fourth part discusses the proposed method, and the fifth part summarizes the overall research methodology of this paper.
Method
This paper proposes a method that integrates recursive feature elimination and residual-based graph convolutional neural networks18 for classifying MI data. The channel selection method combines random forest, gradient boosting, and logistic regression algorithms, realizing recursive feature elimination based on feature importance, which automatically ranks the importance of channels and comprehensively considers the distribution characteristics of EEG data. Subsequently, the residual graph neural network deeply explores the EEG data, making full use of the electrode characteristics reflected by EEG. It extracts information from the spatial domain to achieve classification, compensating for the low spatial resolution of EEG. Figure 1 shows the overall framework of the proposed method. First, channel selection is performed. Then, the multi-channel EEG data undergoes correlation computation to obtain the graph Laplacian representation, resulting in the EEG spatial topology representation. Additionally, the data stream is augmented, and finally, the data is fed into the residual graph convolutional network for classification and recognition.
Fig. 1 [Images not available. See PDF.]
Overall framework of the H-RFE combined with Res-GCN classification model.
Channel selection algorithm
Recursive feature elimination
Feature selection is crucial in deep learning, aiming to find the optimal subset of features while maintaining model performance as much as possible. It helps to remove weakly correlated and redundant features from large datasets, thereby reducing the number of features, improving model accuracy, and reducing the time required for model fitting. Recursive feature elimination (RFE) is a type of wrapper method, with the implementation involving heuristic search through sequential selection and backward search. The process of the RFE algorithm is illustrated in Fig. 2.
Fig. 2 [Images not available. See PDF.]
Flowchart of the recursive feature elimination algorithm.
RFE is a type of greedy algorithm that starts its search from the entire feature set and selects subsets through a feature ranking method19. This algorithm repeatedly constructs machine learning models to rank the importance of features. By fitting the data to the estimator, it eliminates one feature with the lowest feature weight at each iteration. After completion of iterations, it generates feature weights for all features. RFE generates feature subset rankings based on certain evaluation criteria, typically the predictive accuracy of classifiers. Common classifiers include random forests, gradient boosting, and support vector machines, among others. In this paper, we apply the RFE algorithm to compare the importance of EEG channels, thereby eliminating minor factors to improve model accuracy. The specific algorithm process is shown in Algorithm 1.
Overall, RFE iteratively eliminates unimportant features and evaluates model performance through cross-validation to select the best-performing feature subset on the given data. This aids in enhancing the model’s capacity for generalization while mitigating the likelihood of overfitting. In this paper, we apply the feature recursive elimination algorithm to compare the importance of EEG channels, subsequently eliminating minor factors to reduce computational resource consumption and enhance model accuracy.
Hybrid-recursive feature elimination (H-RFE) algorithm
In machine learning endeavors, employing a fusion of multiple machine learning techniques could yield superior performance in contrast to utilizing a solitary method. Therefore, this paper proposes a fusion recursive feature elimination algorithm that combines RF, GBM, and LR. Random forest (RF) uses the Bootstrap sampling method, which involves resampling n samples from the original dataset to construct n different training sets, and each decision tree is trained on these different training sets20. Gradient boosting machine (GBM) is an ensemble learning model based on the CART algorithm21, which iteratively trains multiple decision trees and stacks them to improve prediction accuracy. Logistic regression (LR) is a statistical learning method used to solve binary classification problems by weighting the input features and applying a logistic function to map the data to probabilities between 0 and 1 for classification22. The channel selection is performed by integrating the channel weights and accuracy obtained from the three RFE models. The algorithm schematic is shown in Fig. 3.
Fig. 3 [Images not available. See PDF.]
Schematic diagram of the fusion recursive feature elimination algorithm.
、 and are the channel weights obtained from the training of the RFE-RF, RFE-GBM, and RFE-LR models, respectively. Due to the differing scales of the channel weights obtained from the three models, normalization of the weights is required. The algorithm is as follows:
1
Thus, 、 and represent the normalized weights of 、 and , respectively. To obtain the ranking of channel weights, the fusion cross-validation recursive feature elimination algorithm integrates channel weights and model accuracy to rank the importance of all EEG channels.
2
Where RF.acc, GBM.acc, and LR.acc represent the model accuracies obtained by training the data using the RF, GBM, and LR algorithms, respectively, after obtaining the optimal channel combinations from the three RFE models and dividing the EEG data into training and testing sets.
Residual graph convolutional neural network
Graph convolutional networks can be divided into spectral domain graph convolutional networks and spatial domain graph convolutional networks23. Spectral methods leverage the convolution theorem in the spectral domain to define graph convolution, whereas spatial methods commence from the node domain, aggregating each central node and its neighboring nodes through the definition of aggregation functions. Spectral-domain graph convolutional networks define convolution operations based on the eigenvalue decomposition of the graph Laplacian matrix. They use the spectral information of graph signals for convolution operations, performing convolution operations through the Fourier transform of graph signals. For EEG signals, spectral domain graph convolutional networks can effectively capture the complex spatial dependency relationships and spectral features between brain regions, extracting spatial and frequency domain features from EEG signals, and making them suitable for processing EEG signal data with complex spatial structures. Therefore, this paper chooses the spectral domain graph convolution method.
Graph representation
For an undirected graph, we denote it as , where indicates there are n nodes, represents the edge set, and represents the adjacency matrix, defining the connection relationships between nodes. , where denotes the absolute Pearson matrix between nodes, and is the identity matrix. represents the graph Laplacian matrix, where is a diagonal matrix, denotes the degree of the i-th node, and . The normalized Laplacian matrix [24] is defined as:
3
Because is a real symmetric matrix, performing eigendecomposition on yields , where represents the eigenvector matrix composed of n mutually orthogonal eigenvectors of L, and is the diagonal matrix of eigenvalues, and corresponds to the eigenvalue of . The correlation matrix, Pearson matrix, absolute Pearson coefficient matrix, adjacency matrix, degree matrix, and Laplacian matrix for subject 1 in the SHU dataset are shown in Fig. 4.
Fig. 4 [Images not available. See PDF.]
Illustration of the matrices for Subject 1 in the SHU dataset. (a) Shows the correlation matrix (b) displays the Pearson matrix (c) represents the absolute Pearson coefficient matrix (d) depicts the adjacency matrix (e) illustrates the degree matrix (f) presents the Laplacian matrix.
Residual graph convolutional neural Network Architecture
With the increase in the number of layers and neurons, the non-linear fitting capability of deep neural networks is enhanced. Nevertheless, merely stacking network layers may result in issues like gradient vanishing, gradient explosion, and network degradation. Therefore, we introduce residual structures into graph convolutional neural networks.
In the spectral domain, the graph convolution of signals x1 and x2 is defined as:
4
represents the graph convolution operator, and denotes the Hadamard product. For the input signal , the graph convolution operation with the convolution kernel filter is defined as:
5
If is represented as , then the graph convolution operation on x can be simplified as:
6
The Chebyshev network (ChebyNet)[25] parameterizes the convolutional kernel instead of the spectral domain kernel. is defined as , where , and the Chebyshev polynomial is defined as . and . Therefore, the ChebyNet graph convolution operation is:
7
Let , then . The ChebyNet graph convolution operation can be simplified as:
8
ChebyNet graph convolution does not require eigenvalue decomposition of the Laplacian matrix, and the convolution kernel has only K + 1 learnable parameters. The complexity of parameters is greatly reduced, thereby improving computational speed.
ChebyNet utilizes a complete binary tree for pooling operations. Firstly, the input feature tensor is partitioned into several equally sized blocks based on the coarsening stage of the Graclus multilevel clustering algorithm24. Then, these blocks are arranged according to the properties of a complete binary tree, where each non-leaf node has exactly 2 children. At each non-leaf node, the most compatible feature blocks are pooled based on a greedy criterion, merging two nodes into one. At each leaf node, the corresponding feature block is directly passed to the next layer of the network. If the number of feature blocks at the last layer is not a multiple of 2, zero-padding blocks are added on the rightmost side to make the total number of blocks a multiple of 2.
Figure 5 illustrates the residual graph convolutional neural network model. The model adopts 12 layers of graph convolution, with a graph maximum pooling layer connecting every two layers to reduce the dimension by half. Chebyshev polynomials of order three are used to approximate the graph convolution filters in the experiments. Cross-entropy with L2 norm (0.001 regularization parameter) is employed as the loss function in this study, and the residual learning framework facilitates the convergence of deeper models.
Fig. 5 [Images not available. See PDF.]
Residual Graph Convolutional Neural Network Model.
We utilize the data from each sampling time point as input samples, shuffling the entire EEG data collected during the motor imagery period of each subject to create training data. The dimension of the graph Laplacian matrix is N*N, where N represents the number of sampling channels. For instance, in the PhysioNet dataset, which employs the international 10–10 system, it includes 64 EEG electrodes. Therefore, the input dimension is 64 × 64. Subsequently, convolutional operations are performed, involving neighbor information aggregation, weight learning, local feature extraction, and multi-layer stacking. For each node, graph convolution aggregates information from the node and its neighboring nodes using dynamically learned weight matrices, achieving operations similar to traditional convolutions. Through the learned weights, graph convolution extracts local features while preserving the local connectivity of nodes. Multi-layer stacking of graph convolutions gradually extracts higher-level feature representations. Additionally, the residual mechanism reduces issues such as gradient vanishing and explosion caused by multi-layer stacking. Fully connected layers unfold the features extracted by graph convolution into one-dimensional data, incorporating Dropout to reduce parameters. Softmax transforms the raw outputs of the network into a probability distribution. Specifically, for a motor imagery classification problem with multiple categories, Softmax converts the raw output for each motor imagery category into the corresponding class probability, ensuring that the sum of probabilities for all categories is 1.
Evaluation Metrics
This paper evaluates the performance of the model using accuracy and F1-score based on the confusion matrix. The expressions for accuracy and F1-score are as follows:
9
10
11
12
where TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.
Experimental results
Datasets
In this experiment, we used two datasets: the SHU dataset and the PhysioNet dataset. The SHU dataset is a cross-session, cross-subject motor imagery dataset publicly released by Shanghai University in 2022 25. This dataset consists of 25 subjects, each of whom participated in 5 sessions over 2–3 days. Each session comprises 100 trials. EEG signals were automatically labeled using EEGLAB, with segments exceeding 100uV considered as bad data, further confirmed and removed by experts. Consequently, some sessions may have a small number of missing trials. The motor imagery tasks include left hand (L) and right hand (R) movements. Data were recorded using 32 electrodes based on the International 10–10 system, with a sampling rate of 250 Hz. Each trial recorded data for 4 s, resulting in 500 trials per subject (5 sessions * 100 trials), with each trial containing 1000 sampling points.
The PhysioNet dataset comprises 109 subjects performing motor imagery tasks involving left hand (L), right hand (R), both fists (B), and both feet (F). EEG data were recorded using 64 electrodes based on the International 10–10 system, with a sampling rate of 160 Hz. Data were recorded in 4-second time windows, with each subject having 84 trials (3 runs * 7 trials * 4 tasks) per task, and each trial containing 640 sampling points.
Runtime environment and experimental parameter settings
The work was conducted on a remote server based on the Ubuntu 18.04 operating system. TensorFlow 1.15 deep learning framework was utilized within a Python 3.6 environment. Model training and inference were performed using 1 NVIDIA Tesla V100-PCIE-32GB GPU and 1 Intel Xeon Processor (Skylake) 2.4 GHz CPU. The system had 147GB of RAM and a hard disk capacity of 6 TB.
Each experiment employed 10-fold cross-validation, with 90% of the dataset used for training and the remaining 10% for testing. Training proceeded with batches of 100 samples per epoch, followed by evaluation on the test set after each epoch. The Leaky ReLU activation function was used, along with the Adam optimizer with a learning rate of 0.0001, and a dropout rate of 0.5.
Channel distribution results based on the fusion recursive feature elimination algorithm
In this work, the data from the first ten subjects of both datasets were initially fitted using the RFE-RF, RFE-GBM, and RFE-LR models. The weight information of each channel was obtained, and the optimal channel combinations were determined using the RFE method. Subsequently, the optimal channel combinations were employed for classification using RF, GBM, and LR machine learning models, respectively. Five-fold cross-validation was employed to reduce randomness. The optimal channel combination, classification results, and F1-score are shown in Table 1.
Table 1. Results of three RFE models applied to SHU dataset and PhysioNet dataset.
Dataset | Subject | Three RFE models | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
RF | GBM | LR | ||||||||
Channels | Accuracy (%) | F1 | Channels | Accuracy (%) | F1 | Channels | Accuracy (%) | F1 | ||
SHU | Sub-1 | 26 | 85.38 | 86.18 | 31 | 58.81 | 57.91 | 1 | 50.18 | 51.46 |
Sub-2 | 28 | 88.628 | 89.75 | 24 | 65.469 | 64.29 | 5 | 51.8 | 52.78 | |
Sub-3 | 27 | 86.23 | 86.97 | 32 | 56.875 | 57.15 | 1 | 50.206 | 51.34 | |
Sub-4 | 26 | 93.661 | 94.26 | 31 | 61.192 | 62.18 | 9 | 50.769 | 51.87 | |
Sub-5 | 28 | 88.684 | 89.13 | 32 | 63.196 | 64.75 | 28 | 50.737 | 51.96 | |
Sub-6 | 29 | 86.27 | 87.74 | 30 | 55.16 | 55.98 | 15 | 49.16 | 49.47 | |
Sub-7 | 30 | 85.48 | 86.20 | 32 | 64.98 | 65.16 | 24 | 50.89 | 51.48 | |
Sub-8 | 29 | 75.29 | 75.98 | 28 | 51.24 | 51.48 | 1 | 48.26 | 48.75 | |
Sub-9 | 27 | 73.24 | 74.02 | 23 | 59.18 | 60.87 | 13 | 50.98 | 51.74 | |
Sub-10 | 29 | 80.12 | 80.07 | 32 | 65.93 | 66.34 | 27 | 50.07 | 50.88 | |
Mean | 27.9 | 84.2 | 85.03 | 29.5 | 60.2 | 60.61 | 12.4 | 50.31 | 51.17 | |
PhysioNet | Sub-1 | 51 | 87.44 | 88.01 | 55 | 69.646 | 70.13 | 63 | 40.74 | 41.37 |
Sub-2 | 56 | 73.68 | 74.51 | 61 | 70.35 | 71.23 | 61 | 55.91 | 56.96 | |
Sub-3 | 63 | 92.7 | 93.57 | 64 | 76.6 | 77.23 | 62 | 52.77 | 53.78 | |
Sub-4 | 61 | 99.69 | 98.57 | 55 | 90.61 | 91.24 | 62 | 54.0 | 55.12 | |
Sub-5 | 63 | 65.92 | 67.13 | 57 | 62.04 | 61.28 | 63 | 36.62 | 37.46 | |
Sub-6 | 58 | 88.69 | 89.06 | 54 | 70.23 | 70.21 | 65 | 50.13 | 50.74 | |
Sub-7 | 62 | 81.39 | 81.47 | 53 | 70.26 | 70.34 | 67 | 47.03 | 47.69 | |
Sub-8 | 53 | 89.67 | 90.12 | 64 | 80.34 | 80.46 | 61 | 39.16 | 40.16 | |
Sub-9 | 57 | 90.14 | 90.75 | 61 | 82.46 | 82.67 | 62 | 52.14 | 52.97 | |
Sub-10 | 54 | 90.13 | 90.56 | 62 | 80.34 | 80.69 | 65 | 46.89 | 47.14 | |
Mean | 57.8 | 85.95 | 86.38 | 58.6 | 75.29 | 75.55 | 63 | 47.54 | 48.34 | |
Table 1 presents the results of applying different RFE models to SHU dataset and PhysioNet dataset, including the optimal channel combinations and classification accuracies for 20 participants in total. The results indicate that the RF algorithm performs well in processing EEG signals, while the LR shows the lowest performance. For the SHU dataset, the classification accuracies for ten participants under different classifiers are 84.2% (RF), 60.2% (GBM), and 50.31% (LR) respectively. As for the PhysioNet dataset, the average classification accuracies for ten participants using different evaluators are 85.95% (RF), 75.29% (GBM), and 47.54% (LR) respectively.
After obtaining the channel weights and classification accuracies from the RFE models, the feature weights of all channels are sorted by weighting the channel weights and classification accuracies obtained from the three RFE models. For the data of the first five participants in both datasets, the top ten channels ranked by feature weights after H-RFE channel sorting are shown in Fig. 6.
Fig. 6 [Images not available. See PDF.]
Displays the top ten channels ranked by feature weights obtained from the H-RFE algorithm.
Motion imagery recognition results based on H-RFE and Res-GCN
First, to verify the effectiveness of the channel selection method proposed in this paper, two datasets were classified using Res-GCN, and the experimental results are shown in Table 2.
Table 2. Experimental results of classification using Res-GCN.
Dataset | SHU | PhysioNet | ||
|---|---|---|---|---|
Subject | Accuracy (%) | F1 | Accuracy (%) | F1 |
Sub-1 | 95.67 | 95.67 | 92.60 | 92.58 |
Sub-2 | 96.44 | 96.44 | 93.28 | 92.24 |
Sub-3 | 95.87 | 95.87 | 96.10 | 95.12 |
Sub-4 | 69.1 | 69.1 | 96.74 | 95.75 |
Sub-5 | 67.13 | 67.08 | 92.18 | 92.14 |
Sub-6 | 97.38 | 97.38 | 91.74 | 91.68 |
Sub-7 | 95.76 | 95.76 | 87.31 | 87.18 |
Sub-8 | 58.39 | 58.38 | 97.38 | 97.40 |
Sub-9 | 55.74 | 55.38 | 95.08 | 95.12 |
Sub-10 | 74.2 | 74.18 | 96.44 | 96.44 |
Mean | 80.57 | 80.52 | 93.88 | 93.57 |
After channel weighting sorting with the H-RFE algorithm for ten participants from both datasets. For the SHU dataset, there are a total of 28 channel combinations ranging from 5 channels to 32 channels according to the sorting. Similarly, for the PhysioNet dataset, there are a total of 60 channel combinations ranging from 5 channels to 64 channels according to the sorting. The graph residual convolutional neural network was then used for training and classification. The experimental results are shown in Fig. 7.
Fig. 7 [Images not available. See PDF.]
Training and classification results of the graph convolutional neural network.
For the five participants in the SHU dataset, the highest classification accuracies are as follows: 94.41%, 96.22%, 95.26%, 96.44% and 86.2%, respectively. The highest accuracy for participants Sub-1 to Sub-4 is achieved when approximately 25 channels of data are used as input to the network. Subsequently, the classification accuracies for Sub-1 to Sub-3 stabilize, while there is a significant drop in the classification results for Sub-4. The training results for Sub-5 fluctuate considerably, possibly due to suboptimal data collection. For the five participants in the PhysioNet dataset, the highest classification accuracies are as follows: 97.82%, 90.38%, 97.73%, 97.28%, and 85.6%, respectively. For participants Sub-1, Sub-3, and Sub-4, the training data converge and stabilize when approximately 30 channels of data are used as input to the network. The classification accuracy for Sub-2 generally improves steadily with minor fluctuations. However, the classification accuracy for Sub-5 shows significant fluctuations after a gradual increase. This could be attributed to the presence of noise in some channels for this participant, leading to fluctuations when lower-ranked channel data is included in the network input. The highest accuracy and highest F1-score achieved by twenty subjects across the two datasets, along with the number of channels corresponding to the highest accuracy, are shown in Table 3.
Table 3. Number of channels at highest accuracy.
Dataset | SHU | PhysioNet | ||||
|---|---|---|---|---|---|---|
Subject | Channels | Accuracy (%) | F1 | Channels | Accuracy (%) | F1 |
Sub-1 | 26 | 94.41 | 95.41 | 44 | 97.82 | 98.67 |
Sub-2 | 25 | 96.22 | 97.54 | 56 | 90.38 | 91.45 |
Sub-3 | 32 | 95.26 | 96.32 | 59 | 97.73 | 98.97 |
Sub-4 | 23 | 96.44 | 97.89 | 24 | 97.28 | 98.67 |
Sub-5 | 15 | 86.20 | 86.97 | 54 | 85.60 | 86.74 |
Sub-6 | 18 | 94.23 | 94.67 | 41 | 89.65 | 90.04 |
Sub-7 | 20 | 96.57 | 96.9 | 36 | 90.14 | 90.45 |
Sub-8 | 27 | 89.14 | 89.67 | 49 | 96.34 | 96.67 |
Sub-9 | 19 | 76.24 | 76.37 | 51 | 97.31 | 97.64 |
Sub-10 | 30 | 75.63 | 75.81 | 50 | 97.68 | 97.8 |
Mean | 23.5 | 90.03 | 90.44 | 46.4 | 93.99 | 94.17 |
As shown in Table 3, the highest classification accuracy for the SHU dataset was 96.44% for Sub-4, with the highest F1-score also achieved under the H-RFE algorithm, utilizing an average of 75.63% of the total 32 channels. For the PhysioNet dataset, the highest classification accuracy was 97.82% for Sub-1, with the highest F1-score also achieved, utilizing an average of 68.75% of the total 64 channels under the H-RFE algorithm. Compared to the results of using Res-GCN alone for classification, the average classification accuracy increased by 9.46% on the SHU dataset and by 0.11% on the PhysioNet dataset. Overall, the training data for most subjects converged and stabilized with only a small number of channels used as input, demonstrating the necessity and effectiveness of the channel selection algorithm proposed in this paper.
Comparing different channel selection methods
To validate the effectiveness of the proposed channel selection method, we investigated other channel selection methods in the field. Under the same dataset and experimental conditions, we conducted testing experiments on these methods. The classification method uniformly adopted a residual graph convolutional neural network. The experimental results are shown in Table 4.
Method 1 (3 C)26: The three channels most relevant to motor imagery, namely C3, C4, and Cz, were selected.
Method 2 (CCS)27: Based on correlation, channel selection was performed using Pearson correlation coefficients.
Method 3 (NMI)28: Channel selection was conducted based on normalized mutual information.
Method 4 (Ada)18: Adaptive channel selection method based on Pearson correlation coefficient and spatial location.
Method 5 (H-RFE): Channel selection based on the fusion recursive feature elimination algorithm.
Table 4. Comparison of the performance of various channel selection algorithms across two datasets.
Dataset | Subject | Method | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
3 C | CCS | NMI | Ada | H-RFE | |||||||
Acc(%) | F1 | Acc(%) | F1 | Acc(%) | F1 | Acc(%) | F1 | Acc(%) | F1 | ||
SHU | 1 | 54.49 | 55.76 | 93.27 | 93.29 | 93.67 | 93.88 | 95.67 | 95.78 | 94.41 | 94.76 |
2 | 57.57 | 58.37 | 96.02 | 96.57 | 93.69 | 94.06 | 93.89 | 94.06 | 96.22 | 96.8 | |
3 | 56.64 | 56.87 | 94.95 | 95.14 | 93.79 | 94.13 | 95.87 | 96.11 | 95.26 | 95.76 | |
4 | 55.92 | 56.1 | 61.93 | 62.16 | 95.53 | 95.87 | 93.32 | 93.67 | 96.44 | 96.73 | |
5 | 56.98 | 57.14 | 66.22 | 66.87 | 90.26 | 90.94 | 88.67 | 88.97 | 86.2 | 86.97 | |
6 | 55.36 | 55.87 | 91.13 | 92.37 | 92.34 | 92.54 | 94.34 | 94.66 | 94.23 | 94.67 | |
7 | 58.12 | 58.79 | 92.14 | 92.36 | 91.24 | 91.67 | 95.76 | 96.13 | 96.57 | 96.9 | |
8 | 57.31 | 57.98 | 60.12 | 60.78 | 65.24 | 65.84 | 84.06 | 84.68 | 89.14 | 89.67 | |
9 | 50.13 | 50.68 | 59.36 | 59.67 | 66.65 | 66.97 | 55.74 | 55.89 | 76.24 | 76.37 | |
10 | 51.37 | 51.81 | 84.12 | 84.69 | 87.34 | 87.84 | 74.20 | 74.98 | 75.63 | 75.81 | |
Mean | 55.39 | 55.94 | 79.23 | 80.39 | 86.78 | 87.37 | 87.15 | 87.49 | 90.03 | 90.44 | |
PhysioNet | 1 | 46.35 | 46.35 | 96.08 | 96.31 | 96.47 | 96.55 | 96.33 | 96.31 | 97.82 | 97.88 |
2 | 43.68 | 43.98 | 86.07 | 86.14 | 87.98 | 88.04 | 88.11 | 88.5 | 90.38 | 90.38 | |
3 | 50.04 | 50.24 | 95.33 | 95.64 | 96.13 | 96.71 | 97.17 | 97.6 | 97.73 | 97.8 | |
4 | 49.63 | 49.87 | 96.47 | 96.77 | 96.67 | 96.8 | 96.99 | 96.74 | 97.28 | 97.28 | |
5 | 48.07 | 48.64 | 85.83 | 85.9 | 73.98 | 74.12 | 81.81 | 82.46 | 85.6 | 85.74 | |
6 | 41.03 | 41.97 | 82.1 | 82.67 | 82.34 | 82.67 | 85.01 | 85.7 | 89.65 | 90.04 | |
7 | 46.35 | 46.88 | 85.32 | 85.67 | 86.47 | 86.78 | 82.34 | 82.66 | 90.14 | 90.45 | |
8 | 45.2 | 45.87 | 83.4 | 83.64 | 82.15 | 82.96 | 95.64 | 95.89 | 96.34 | 96.67 | |
9 | 49.71 | 50.03 | 88.6 | 88.09 | 87.3 | 87.4 | 95.99 | 96.47 | 97.31 | 97.64 | |
10 | 50.2 | 50.46 | 90.31 | 90.75 | 92.34 | 92.8 | 97.32 | 97.9 | 97.68 | 97.8 | |
Mean | 47.03 | 47.45 | 88.95 | 89.16 | 88.18 | 88.48 | 91.67 | 92.02 | 93.99 | 94.17 |
Methods 2 and 3 selected the same number of channels per participant as the optimal number achieved by the proposed H-RFE method for each participant. For both datasets, the proposed H-RFE method achieved the highest classification accuracy. Specifically, for all 10 participants in the SHU dataset, the average accuracies were 55.39% (using 3 C), 79.23% (using CCS), 86.78% (using NMI), 87.15% (using Ada), and 90.03% (using H-RFE). For all 10 participants in the PhysioNet dataset, the average accuracies were 47.03% (using 3 C), 88.95% (using CCS), 88.18% (using NMI), 91.67% (using Ada), and 93.99% (using H-RFE). Table 4 demonstrates the advantages of the proposed H-RFE algorithm. Both on an individual basis and on average, the H-RFE method achieved relatively higher classification accuracies compared to other methods. The results applied to the two datasets confirm the superiority of H-RFE over other methods.
Discussion
There are two crucial reasons for channel selection in motor imagery tasks. Firstly, from the perspective of motor imagery mechanisms, specific brain regions are associated with particular limb movements, and different channels correspond to electrodes located at different positions on the scalp. Therefore, by selecting channels corresponding to brain regions specifically related to certain movements, the specificity and accuracy of the signals can be enhanced, aiding in capturing task-related neural activity. Secondly, influenced by factors such as the participant’s skills, physiological and mental states, significant differences may exist in EEG signals across different sessions. Moreover, environmental factors such as electrode quality and hair interference can also contribute to signal variability. Therefore, channel selection can identify the optimal channel configuration for specific individuals and environmental conditions, thereby enhancing the performance and stability of brain-computer interface systems. In summary, channel selection is crucial in motor imagery tasks as it improves signal specificity and accuracy while accommodating individual differences and environmental variability, thereby providing essential support for the effective operation of brain-computer interface systems.
Based on the experimental results in Fig. 6, we plotted the channel heatmap in Fig. 8. From Fig. 8, it can be observed that the channels with higher weights for the five participants in the SHU dataset are mainly concentrated in F7, F8 and FP1, rather than in the channels related to motor imagery such as C3, C4, and Cz. This result could be attributed to various factors. Firstly, the SHU dataset is a large-scale dataset with high individual and session variability, aimed at evaluating the robustness of motor imagery algorithms; Secondly, based on EEG brain mechanisms, motor imagery involves activities in multiple brain regions, not limited to the regions covered by channels C3, C4, and Cz. Other regions such as the frontal lobe (F7, F8) and frontal-temporal lobe (FP1) may also participate in motor imagery tasks. Additionally, individual differences and task design may also lead to differences in channel weights. The experimental results indicate that the H-RFE algorithm proposed in this paper can adaptively and accurately capture the cortical excitation areas in the motor imagery activities of these five participants. As for the PhysioNet dataset, the channel distribution varies among the five participants. However, for the majority of participants, the channel weights are concentrated around C3, C4, and Cz, indicating that individual differences in the PhysioNet dataset are not very high, and these channels still carry important information closely related to individuals’ motor imagery activities. This also demonstrates that the H-RFE algorithm can fully consider the variability of EEG data and effectively capture the key channels related to motor imagery tasks.
Fig. 8 [Images not available. See PDF.]
Depicts the heatmap of channels selected by the H-RFE algorithm for the top five participants.
To assess the individual adaptability of the proposed method, we surveyed and summarized the channel selection results of several methods for different participants. Figure 9 displays the common channels selected by five channel selection methods for different participants. From the figure, it can be observed that our proposed method selects the fewest common channels among the ten participants, with 9 common channels for the SHU dataset and 17 common channels for the PhysioNet dataset. In contrast to other methods, our proposed H-RFE method selects the fewest common channels, indicating greater individual adaptability. This is because our method comprehensively considers the feature weights of multiple evaluators and dynamically adjusts channel weights based on their fitting effects on the data, thus achieving more accurate channel selection. Furthermore, when combined with the classification accuracy shown in Table 4, the superiority of the fusion recursive feature elimination algorithm is further highlighted.
Fig. 9 [Images not available. See PDF.]
Common channels selected by different channel selection methods.
At the same time, to verify the performance of the proposed method, we compared our experimental results with other publicly published findings. The comparative results are shown in Table 5.
Table 5. Performance comparison of different algorithms on two datasets.
Dataset | Work | Method | Channels Strategies | Average Acc |
|---|---|---|---|---|
SHU | This work | ResGCN | H-RFE | 90.03% |
Jia et al.29 | ResGCN | All Channels | 82.92% | |
Ma et al.30 | Deep ConvNets | All Channels | 78.86% | |
Xia et al.18 | ResGCN | Ada | 89.34% | |
PhysioNet | This work | ResGCN | H-RFE | 93.99% |
Jia et al.29 | ResGCN | All Channels | 92.61% | |
Hou et al.31 | GCN | All Channels | 89.67% | |
Ma et al.30 | RNN | All Channels | 68.20% | |
Hou et al. 32 | CNN | All Channels | 92.50% | |
Huang et al.33 | RP-BCNNs | All Channels | 90.54% | |
Sun et al.16 | EEG-ARNN | All Channels | 82.00% | |
Xia et al.18 | ResGCN | Ada | 92.51% | |
Huang et al.34 | 3DCNN-LSTM | All Channels | 85.88% | |
Gong et al.35 | SGLNet | All Channels | 84.15% |
In the case of the ten participants from the SHU dataset, the combination of our proposed H-RFE channel selection algorithm with Res-GCN yielded an average classification accuracy of 90.03%. This result slightly outperforms Jia’s29 achievement of 82.92% using the same classifier and significantly surpasses the benchmark method25 utilized in the paper presenting the dataset, which achieved a maximum classification accuracy of 78.86% under cross-session adaptation conditions.
For the ten participants from the PhysioNet dataset, our proposed method demonstrates significant superiority in classification accuracy, reaching 93.99%. This performance is slightly higher than Hou’s31 usage of the entire dataset with graph convolutional neural networks (GCN) achieving 89.67%, Hou’s32 CNN-based approach with 92.50% accuracy, a classification accuracy of 92.61% achieved using a residual graph convolutional neural network29, and Huang’s novel model combining recursive graphs and Bayesian convolutional neural networks reaching 90.54% accuracy. Additionally, our method significantly outperforms the average classification accuracy of 68.20% obtained by Ma et al.30 using Spatial and Temporal Recurrent Neural Networks, the 82% classification accuracy achieved by Sun et al.16 with their hybrid deep framework EEG-ARNN based on CNN and GCN. Moreover, this method also surpasses the 85.88% accuracy achieved by the 3DCNN-LSTM model proposed by Huang et al.34, and the 84.15% accuracy obtained by the SGLNet network proposed by Gong et al.35.
The superiority of GCN over other classification methods is evident from the comparison results. GCNs are effective in handling graph-structured data and capturing spatial and temporal features in EEG signals. In GCN, each node represents an EEG channel, and the connections between nodes represent relationships between channels. By performing convolution operations on the graph and considering the information from neighboring nodes, GCN effectively capture the spatial dependencies and local structures in EEG signals. Furthermore, integrating residual blocks into GCN further enhances network performance. Residual blocks alleviate the problem of gradient vanishing, accelerate network convergence, and improve training stability. Therefore, adding residual blocks to GCN enables better feature extraction and utilization of EEG signal characteristics, resulting in improved classification accuracy and model generalization.
Overall, validation on both datasets demonstrates that the proposed H-RFE channel selection algorithm, combined with residual graph convolutional neural networks, achieves higher accuracy with fewer channel data compared to some studies utilizing all channel data.
Conclusions
MI based on EEG is one of the essential paradigms in BCIs. Due to individual and environmental variabilities, personalized and accurate identification, as well as effective interaction, still pose significant challenges in the practical application of EEG-based MI-BCI. Therefore, this paper proposes a novel channel selection method and combines it with residual graph convolutional neural networks (Res-GCN). The channel selection method integrates RFE and a multi-feature evaluation fusion strategy to evaluate the feature selection effectiveness based on the decision fusion results of different classifiers, obtaining the final channel importance ranking and selecting the optimal channel combination. Subsequently, classification is performed using graph convolutional neural networks embedded with residual blocks. The residual graph convolutional neural network is a deep learning model based on graph structure, which extracts and learns spatial and temporal features from EEG signals through graph convolutional layers and residual connections. In this study, the residual graph convolutional neural network utilizes the selected channel combination as input and performs classification on EEG signals through multi-layer graph convolution operations and residual connections. Experimental results demonstrate that the channel selection strategy based on fused recursive feature elimination achieves effective personalized channel selection for different subjects across sessions. The residual-based graph convolutional neural network fully exploits and learns spatial and temporal features from EEG signals, significantly improving the robustness of MI recognition algorithms. Using 73.44% of the total channel count on the SHU dataset achieves a classification result of 90.03%, while on the PhysioNet dataset, using only 72.5% of the channel data achieves a classification result of 93.99%. Compared with existing research results, the proposed method significantly enhances classification accuracy, demonstrating high generalization and application value.
Author contributions
All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by D.L., K.Y.L and Y.Q.X. The first draft of the manuscript was written by K.Y.L and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.
Funding
This work was supported by the Young Teacher Foundation of Henan Province [No.2021GGJS093], the Key Science and Technology Program of Henan Province [No.242102211058], the Key Science Research Project of Colleges and Universities in Henan Province of China [No.22A520046], the National Natural Science Foundation of China [62303427], the Doctor Natural Science Foundation of Zhengzhou University of Light Industry [2022BSJJZK13].
Data availability
The dataset used in our study is available at the following link: https://figshare.com/articles/software/shu_dataset/19228725/1 and https://www.physionet.org/content/eegmmidb/1.0.0/.
Declarations
Conflict of interest
All author declares that they have no conflict of interest.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
1. Raza, H; Rathee, D; Zhou, SM; Cecotti, H; Prasad, G. Covariate shift estimation based adaptive ensemble learning for handling non-stationarity in motor imagery related EEG-based brain-computer interface. Neurocomputing; 2019; 343, pp. 154-166. [DOI: https://dx.doi.org/10.1016/j.neucom.2018.04.087] [PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7086459]
2. Lazarou, I., Nikolopoulos, S., Petrantonakis, P. C., Kompatsiaris, I. & Tsolaki, M. E. E. G. -Based brain-computer interfaces for Communication and Rehabilitation of people with Motor Impairment: <i > a Novel Approach of the < i > 21st Century</i >. Front. Hum. Neurosci.12https://doi.org/10.3389/fnhum.2018.00014 (2018).
3. Palumbo, A., Gramigna, V., Calabrese, B., Ielpo, N. & Motor-Imagery EEG-Based BCIs in Wheelchair Movement and Control: a systematic literature review. Sensors. 21. https://doi.org/10.3390/s21186285 (2021).
4. Saibene, A., Caglioni, M., Corchs, S. & Gasparini, F. EEG-Based BCIs on Motor Imagery Paradigm using Wearable technologies: a systematic review. Sensors. 23. https://doi.org/10.3390/s23052798 (2023).
5. Liang, W et al. Novel channel selection model based on graph convolutional network for motor imagery. Cogn. Neurodyn.; 2023; 17, pp. 1283-1296. [DOI: https://dx.doi.org/10.1007/s11571-022-09892-1]
6. Handiru, VS; Prasad, VA; Optimized Bi-Objective, EEG. Channel Selection and Cross-subject Generalization with Brain-Computer interfaces. Ieee Trans. Human-Machine Syst.; 2016; 46, pp. 777-786. [DOI: https://dx.doi.org/10.1109/thms.2016.2573827]
7. Aydemir, O; Ergün, E. A robust and subject-specific sequential forward search method for effective channel selection in brain computer interfaces. J. Neurosci. Methods; 2019; 313, pp. 60-67. [DOI: https://dx.doi.org/10.1016/j.jneumeth.2018.12.004]
8. Wei, QG; Lu, ZW; Chen, K; Ma, YH. Channel selection for optimizing feature extraction in an Electrocorticogram-based brain-computer interface. J. Clin. Neurophysiol.; 2010; 27, pp. 321-327. [DOI: https://dx.doi.org/10.1097/WNP.0b013e3181f52f2d]
9. Lal, TN et al. Support vector channel selection in BCI. IEEE Trans. Biomed. Eng.; 2004; 51, pp. 1003-1010. [DOI: https://dx.doi.org/10.1109/tbme.2004.827827]
10. Schröder, M et al. Robust EEG channel selection across subjects for brain-computer interfaces. EURASIP J. Appl. Sig. Process.; 2005; 2005, pp. 3103-3112. [DOI: https://dx.doi.org/10.1155/asp.2005.3103]
11. LeCun, Y; Bengio, Y; Hinton, G. Deep learning. Nature; 2015; 521, pp. 436-444.
12. Luo, J. et al. Improving the performance of multisubject motor imagery-based BCIs using twin cascaded softmax CNNs. J. Neural Eng.18. https://doi.org/10.1088/1741-2552/abe357 (2021).
13. Luo, T. J., Zhou, C. L. & Chao, F. Exploring spatial-frequency-sequential relationships for motor imagery classification with recurrent neural network. BMC Bioinform.19. https://doi.org/10.1186/s12859-018-2365-1 (2018).
14. Dai, G. H., Zhou, J., Huang, J. H. & Wang, N. HS-CNN: a CNN with hybrid convolution scale for EEG motor imagery classification. J. Neural Eng.17. https://doi.org/10.1088/1741-2552/ab405f (2020).
15. Jia, J. J. et al. CR-GCN: Channel-relationships-based graph Convolutional Network for EEG emotion recognition. Brain Sci.12. https://doi.org/10.3390/brainsci12080987 (2022).
16. Sun, B; Liu, ZK; Wu, ZX; Mu, CX; Li, T. Graph Convolution Neural Network Based End-to-End Channel Selection and Classification for Motor Imagery Brain-Computer Interfaces. IEEE Trans. Industr. Inf.; 2023; 19, pp. 9314-9324. [DOI: https://dx.doi.org/10.1109/tii.2022.3227736]
17. Qiu, ZY et al. Improved SFFS method for channel selection in motor imagery based BCI. Neurocomputing; 2016; 207, pp. 519-527. [DOI: https://dx.doi.org/10.1016/j.neucom.2016.05.035]
18. Xia, YQ et al. An Adaptive Channel Selection and Graph ResNet Based Algorithm for Motor Imagery classification. Int. J. Adv. Comput. Sci. Appl.; 2023; 14, pp. 240-248.
19. Ding, X. J., Yang, F. & Ma, F. M. An efficient model selection for linear discriminant function-based recursive feature elimination. J. Biomed. Inform.129. https://doi.org/10.1016/j.jbi.2022.104070 (2022).
20. Speiser, JL; Miller, ME; Tooze, J; Ip, E. A comparison of random forest variable selection methods for classification prediction modeling. Expert Syst. Appl.; 2019; 134, pp. 93-101. [DOI: https://dx.doi.org/10.1016/j.eswa.2019.05.028] [PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7508310]
21. Natekin, A; Knoll, A. Gradient boosting machines, a tutorial. Front. Neurorobotics.; 2013; [DOI: https://dx.doi.org/10.3389/fnbot.2013.00021]
22. Budimir, MEA; Atkinson, PM; Lewis, HG. A systematic review of landslide probability mapping using logistic regression. Landslides; 2015; 12, pp. 419-436. [DOI: https://dx.doi.org/10.1007/s10346-014-0550-5]
23. Wu, ZH et al. A Comprehensive Survey on Graph neural networks. Ieee Trans. Neural Networks Learn. Syst.; 2021; 32, pp. 4-24.MathSciNet ID: 4205495[DOI: https://dx.doi.org/10.1109/tnnls.2020.2978386]
24. Dhillon, IS; Guan, YQ; Kulis, B. Weighted graph cuts without eigenvectors: a multilevel approach. IEEE Trans. Pattern Anal. Mach. Intell.; 2007; 29, pp. 1944-1957. [DOI: https://dx.doi.org/10.1109/tpami.2007.1115]
25. Ma, J. et al. A large EEG dataset for studying cross-session variability in motor imagery brain-computer interface. Sci. Data. 9. https://doi.org/10.1038/s41597-022-01647-1 (2022).
26. Tabar, Y. R. & Halici, U. A novel deep learning approach for classification of EEG motor imagery signals. Journal of Neural Engineering 14. https://doi.org/10.1088/1741-2560/14/1/016003 (2017).
27. Jin, J et al. Correlation-based channel selection and regularized feature optimization for MI-based BCI. Neural Netw.; 2019; 118, pp. 262-270. [DOI: https://dx.doi.org/10.1016/j.neunet.2019.07.008]
28. Wang, ZM; Hu, SY; Song, H. Channel selection method for EEG emotion Recognition using normalized mutual information. Ieee Access.; 2019; 7, pp. 143303-143311. [DOI: https://dx.doi.org/10.1109/access.2019.2944273]
29. Jia, S., Hou, Yimin, Shi, Y. & Li, Y. Attention-based Graph ResNet for Motor Intent Detection from Raw EEG signals. (2020). https://doi.org/10.48550/arXiv.2007.13484
30. Ma, X., Qiu, S., Du, C., Xing, J. & He, H. Improving EEG-Based Motor Imagery Classification via Spatial and Temporal Recurrent Neural Networks. Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference 1903–1906 (2018). https://doi.org/10.1109/embc.2018.8512590 (2018).
31. Hou, YM et al. GCNs-Net: a graph convolutional neural Network Approach for Decoding Time-resolved EEG Motor Imagery signals. Ieee Trans. Neural Networks Learn. Syst.; 2022; [DOI: https://dx.doi.org/10.1109/tnnls.2022.3202569]
32. Hou, Y. M., Zhou, L., Jia, S. Y. & Lun, X. M. A novel approach of decoding EEG four-class motor imagery tasks via scout ESI and CNN. J. Neural Eng.17. https://doi.org/10.1088/1741-2552/ab4af6 (2020).
33. Huang, W. Q., Yan, G. H., Chang, W. W., Zhang, Y. C. & Yuan, Y. T. EEG-based classification combining bayesian convolutional neural networks with recurrence plot for motor movement/imagery. Pattern Recogn.144. https://doi.org/10.1016/j.patcog.2023.109838 (2023).
34. Huang, W. Q., Chang, W. W., Yan, G. H., Zhang, Y. C. & Yuan, Y. T. Spatio-spectral feature classification combining 3D-convolutional neural networks with long short-term memory for motor movement/imagery. Eng. Appl. Artif. Intell.120. https://doi.org/10.1016/j.engappai.2023.105862 (2023).
35. Gong, PL; Wang, PP; Zhou, YY; Zhang, DQ. A spiking neural Network with Adaptive Graph Convolution and LSTM for EEG-Based brain-computer interfaces. IEEE Trans. Neural Syst. Rehabil. Eng.; 2023; 31, pp. 1440-1450. [DOI: https://dx.doi.org/10.1109/tnsre.2023.3246989]
You have requested "on-the-fly" machine translation of selected content from our databases. This functionality is provided solely for your convenience and is in no way intended to replace human translation. Show full disclaimer
Neither ProQuest nor its licensors make any representations or warranties with respect to the translations. The translations are automatically generated "AS IS" and "AS AVAILABLE" and are not retained in our systems. PROQUEST AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY, ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Your use of the translations is subject to all use restrictions contained in your Electronic Products License Agreement and by using the translation functionality you agree to forgo any and all claims against ProQuest or its licensors for your use of the translation functionality and any output derived there from. Hide full disclaimer
© The Author(s) 2024. This work is published under http://creativecommons.org/licenses/by-nc-nd/4.0/ (the "License"). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Abstract
In the field of brain-computer interface (BCI) based on motor imagery (MI), multi-channel electroencephalography (EEG) data is commonly utilized for MI task recognition to achieve sensory compensation or precise human-computer interaction. However, individual physiological differences, environmental variations, or redundant information and noise in certain channels can pose challenges and impact the performance of BCI systems. In this study, we introduce a channel selection method utilizing Hybrid-Recursive Feature Elimination (H-RFE) combined with residual graph neural networks for MI recognition. This channel selection method employs a recursive feature elimination strategy and integrates three classification methods, namely random forest, gradient boosting, and logistic regression, as evaluators for adaptive channel selection tailored to specific subjects. To fully exploit the spatiotemporal information of multi-channel EEG, this study employed a graph neural network embedded with residual blocks to achieve precise recognition of motor imagery. We conducted algorithm testing using the SHU dataset and the PhysioNet dataset. Experimental results show that on the SHU dataset, utilizing 73.44% of the total channels, the cross-session MI recognition accuracy is 90.03%. Similarly, in the PhysioNet dataset, using 72.5% of the channel data, the classification result also reaches 93.99%. Compared to traditional strategies such as selecting three specific channels, correlation-based channel selection, mutual information-based channel selection, and adaptive channel selection based on Pearson coefficients and spatial positions, the proposed method improved classification accuracy by 34.64%, 10.8%, 3.25% and 2.88% on the SHU dataset, and by 46.96%, 5.04%, 5.81% and 2.32% on the PhysioNet dataset, respectively.
You have requested "on-the-fly" machine translation of selected content from our databases. This functionality is provided solely for your convenience and is in no way intended to replace human translation. Show full disclaimer
Neither ProQuest nor its licensors make any representations or warranties with respect to the translations. The translations are automatically generated "AS IS" and "AS AVAILABLE" and are not retained in our systems. PROQUEST AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY, ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Your use of the translations is subject to all use restrictions contained in your Electronic Products License Agreement and by using the translation functionality you agree to forgo any and all claims against ProQuest or its licensors for your use of the translation functionality and any output derived there from. Hide full disclaimer
Details
1 School of Computer Science and Technology, Zhengzhou University of Light Industry, Zhengzhou, Henan, China (ROR: https://ror.org/05fwr8z16) (GRID: grid.413080.e) (ISNI: 0000 0001 0476 2801)