Enhancing EEG Classification for Motor Imagery Control of a VR Game based on Deep Learning Techniques on Small Datasets

Enhancing EEG Classification for Motor Imagery Control of a VR Game based on Deep Learning Techniques on Small Datasets George Ramiotis and Katerina Mania∗ This is a pre-copyedit version of this article. The final authenticated version will be available at the proceedings of the 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’25) Abstract— Motor imagery-based Brain-Computer Interfaces (BCIs) suffer from limited accuracy when the EEG dataset is recorded from naive BCI users due to noisy components. Neural networks capture more robust representations of EEG features, but require large amount of data which is challenging to collect, due to long motor imagery training sessions. On the other hand, linearand Riemann-based machine learning algorithms achieve above chancelevel accuracy on small scale datasets, but, performance degrades on noisy datasets. To address this issue, we implemented a Wasserstein Generative Adversarial Network (WGAN) for data augmentation to prevent overfitting for the deep classifier, while reaching training convergence faster than existing models. For classification, we developed a Convolutional Neural Network (CNN) to eliminate noisy components caused by BCI illiteracy and extract robust temporal representations of EEG features. To evaluate our system, we designed a VR maze game utilizing the proposed BCI system to translate the EEG signal into movement for a playable character. We achieve increased accuracy, compared to conventional machine learning models, with minimal overfitting, on our own dataset, recorded from 16 naive BCI users. I. INTRODUCTION EEG-based Brain-Computer Interfaces (BCIs) showcase potential in medical rehabilitation [38], exoskeleton control [11] and neurogaming [29]. Motor Imagery (MI) is an active paradigm that involves the imagination of body limb movement to translate relevant EEG events into commands [14]. The EEG signal is filtered to remove artifacts from muscle movement, ocular and powerline noise and include only the alpha and beta frequency bands, relevant to Motor Imagery [26], [10]. Then, temporal, spectral or spatial EEG features are extracted for a better representation of MI events [35], [27]. To classify the EEG signal, past studies incorporated into the processing pipeline, linearand Riemann-based classifiers such as LDA, SVM and Minimum Distance to Mean (MDM), showcasing above chance-level performance on small scale datasets [24], [31], [3], [33]. However, motor imagery suffers from long training sessions that cause user fatigue, thus, producing unidentifiable EEG patterns [5], [6] degrading classification performance. Moreover, naive BCI users [1] lack experience in imagining motor movements further limiting accuracy [13]. To mitigate the problem of low accuracy in noisy datasets, past work utilized deep learning architectures to extract more robust representations of EEG features. CNNs [9] and Hybrid MLPs [22] were utilized to extract temporal and spatial EEG features, producing increased classification accuracy on the recorded dataset. CNNs [8] were also used as image classifiers for spectral images produced from a Short-Term Fourier Transform (STFT) and by incorporating auto-encoder models, the proposed classifiers achieved near 80% accuracy on benchmark dataset BCI Competition IV [36] that features data from 9 different participants using 22 EEG electrodes for 4 motor imagery classes. Generally, neural networks are trained ∗George Ramiotis and Prof. Katerina Mania are with the School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece (email: [email protected]; [email protected]) Fig. 1: EEG Headset placed under the Quest 2 HMD on datasets with data aggregated from multiple participants. Previous studies have presented lightweight deep architectures such as EEGNet [20], EEG-TCNet [17], LMDA-Net [25], EEG-Conformer [34] and FBCNet [23]. The aforementioned architectures perform subject-specific classification for a more personalized model that learns to eliminate subject-specific noise and extract more robust EEG features, since electrical brain activity varies for each user. Specifically, FBCNet was tested on the OpenBMI dataset [21] recorded from 54 participants using 62 electrodes, while other methodologies were tested on the BCI Competition IV datasets. By deploying a system with more electrodes, neural networks can yield higher accuracy, as the system allows for better localization of the EEG signal. However, neural networks require large datasets to prevent overfitting and achieve high classification results and combined with the long training sessions of motor imagery, recording a large dataset is proven to be challenging. To tackle the challenges of large datasets, research has been conducted on data augmentation using Generative Adversarial Networks (GANs) to artificially expand a recorded EEG dataset [39]. In past research [12], GANs were implemented to artificially expand an EEG dataset to double the size of the original within 10000 epochs. However, the original GAN framework suffers from the vanishing gradient problem, leading to unstable training [2]. To overcome this challenge, the WGAN was proposed that utilizes the Wasserstein distance as the loss function [2]. In past studies [28], [4], WGANs trained from the raw EEG signal are able to generate artificial EGG feature vectors that resemble the original sequence and within 4000 epochs [16]. However, the approach of training a WGAN on the raw EEG signal suffers from long training sessions to reach convergence due to the complexity and high dimensionality of the input signal. In this paper, we present an EEG-based BCI for motor imagery tasks, based on the OpenVibe [32] platform for signal acquisition; filtering; and extracting features on motor imagery events. OpenVibe’s classification system is replaced with our own implementation of a shallow Convolutional Neural Network (CNN). A dataset was recorded from 16 naive BCI users for four motor Fig. 2: Electrode placement of Enobio 8 imagery classes. To mitigate the problem of small EEG sample size, a Wasserstein Generative Adversarial Network (WGAN) was designed to train on the recorded features and augment the dataset with artificial ones. Unlike past work utilizing WGANs [28], [16], our WGAN reaches training convergence in less epochs. This set of deep networks is trained for each user separately. To test our proposed BCI, we developed a VR maze game where the movement of a playable character is directly controlled with motor imagery events. We conducted a performance analysis to determine the classification accuracy and training stability of our set of deep networks. The use of a convolutional network and a generative network presents increased classification accuracy compared to conventional machine learning algorithms. By training this set of networks on each user’s recorded EEG dataset, the need for complicated deep network architectures and aggregated datasets from multiple users is eliminated. Our specific contributions include: •A subject-specific Wasserstein Generative Adversarial Network to augment the recorded dataset with artificial EEG features and mitigate the problem of small datasets. Our WGAN reaches convergence in less training epochs than previous WGAN approaches. •A subject-specific Convolutional Neural Network to extract robust temporal representations of features from noisy samples, reaching increased classification accuracy on small datasets. •A comparative analysis of machine learning-based classification algorithms to determine the accuracy and generalization of our own methodology. Utilizing robust metrics such as RMSE, accuracy, AUC (Area Under the Curve) and F1 scores, our methodology showcases increased accuracy with minimal overfitting even on a small scale dataset recorded from 16 naive BCI users. II. METHODOLOGY The proposed processing pipeline uses the OpenVibe platform for acquiring the raw EEG signal relevant to motor imagery events; filtering ocular, muscular, powerline artifacts and noise; and extracting features for classification. Classification is performed by our WGAN model combined with a CNN model. The predicted motor imagery event is then transmitted as a command to the VR maze game to control the movement of a playable character. The system is tested for two, three and four motor imagery classes. A. Apparatus This study utilized the Enobio 8 EEG headset from Neurolectrics with 8 Ag/AgCl electrodes and a 500hz sampling frequency. We utilized the Meta Quest 2 VR headset due to the compact build, to not hinder EEG recording. The OpenVibe BCI platform is used for communication with the EEG headset. Layer Filters Kernel size Stride Output Activation function Input - - - [1, 100] - FC Layer 1 6272 - - [1, 6272] - Reshape - - - [128, 49] - ConvTranspose 1 128 [1, 1] 1 [128, 49] ReLU ConvTranspose 2 128 [1, 1] 1 [128, 49] ReLU ConvTranspose 3 128 [1, 1] 1 [128, 49] ReLU ConvTranspose 4 128 [1, 1] 1 [128, 49] ReLU ConvTranspose 5 8 [1, 1] 1 [8, 49] - TABLE I: The generator Layer Filters Kernel size Stride Output Activation function Input - - - [8, 49] - Conv 1 32 [1, 5] 1 [32, 45] ReLU MaxPool 1 32 [1, 2] 1 [32, 22] - Conv 2 32 [1, 5] 1 [32, 18] ReLU MaxPool 2 32 [1, 2] 1 [32, 9] - Conv 3 32 [1, 5] 1 [32, 5] ReLU MaxPool 3 32 [1, 2] 1 [32, 2] - Flatten - - - [1, 64] - FC Layer 1 1 - - [1, 1] - TABLE II: The critic The neural networks were implemented using the Pytorch framework. Training and testing of both the BCI system and the VR game, was conducted on a PC with an Intel i7 8700 CPU, 16GB of RAM memory and a Nvidia RTX 2060 graphics card with 6GB VRAM. B. EEG Data Collection Before the classification of EEG features, the user must perform a motor imagery training session to record the EEG signal when imagined motor movement occurs. Since the primary motor cortex and the primary somatosensory cortex are the main areas of the brain active during motor imagery events [30], the 8 available electrodes were positioned on the ’FC1’, ’C1’, ’CP1’, ’C3’, ’FC2’, ’C2’, ’CP2’ and ’C4’ channels, according to the 10-20 international system as depicted in Figure 2, along with 2 reference electrodes placed on the user’s ear lobe. During the training session, the user was initially asked to stay relaxed for 20sto stabilize the EEG fluctuations in amplitude across all channels. Then, an audio beep was played for 1ssignifying the start of a MI trial. Each trial lasted for 8.5s. A visual cue, indicating the MI class, was presented at t=1s. The participant performed the motor imagery movement for 3.75scontinuously, as depicted in Figure 3. The user performed feet MI and relaxation for two classes. For three classes, the right hand MI is added and for four classes the left hand MI. For each participant, a total of 40 MI trials were performed for each class [19]. C. Pre-processing During EEG acquisition, artifacts from minor muscle, ocular and cardiac activity are present in the lower spectrum of frequencies that overlap with the theta and delta frequency bands close to 0.1−5hz. Furthermore, in the range of 40-100hz the EEG signal displays artifacts derived from electromagnetic waves generated by the AC powerline, due to the insufficient or lack of wire shielding in EEG electrodes and noise from active muscle activity. For these reasons, we apply a band pass Butterworth filter of 4th order, with a low cut-off frequency of 8hz, a high cut-off frequency of 24hz and a pass band ripple of 0dB. Furthermore, we keep the frequency band of 8-24hz where the alpha and beta bands are active during a MI trial. A Butterworth filter was selected due to its maximally flat characteristic in the pass band, and the 4th order of the filter allows for better amplitude characteristics. D. Feature Extraction In this stage, the filtered EEG signal is epoched into 4stime windows starting from the visual cue activation, with a 0.5soffset, Stimulation based epoching (4s) Beep DisplayCue Feedback Rest Trial iTrial i+1 0s 1s 2.75s 6.5s 8.5s1.5s 5.5s Fig. 3: Motor Imagery training procedure Layer Filters Kernel size Stride Output Activation function Input - - - [8, 49] - Conv 1 64 [1, 3] 1 [64, 47] - BatchNorm 1 64 - - [64, 47] ReLU MaxPool 1 64 [1, 2] 1 [64, 23] - Conv 2 64 [1, 3] 1 [64, 21] - BatchNorm 2 64 - - [64, 21] ReLU MaxPool 2 64 [1, 2] 1 [64, 10] - Conv 3 64 [1, 3] 1 [64, 8] - BatchNorm 3 64 - - [64, 8] ReLU MaxPool 3 64 [1, 2] 1 [64, 4] - Flatten - - - [1, 256] - FC Layer 1 32 - - [1, 32] ReLU Dropout - - - [1, 32] - FC Layer 2 2 - - [1, 2] - TABLE III: CNN classifier model as shown in Figure 3. Then, the epoched 4ssignal is segmented further into overlapping 1ssignal chunks every 0.0625s. This epoching method is applied for each MI class separately, resulting in one signal stream of all epoched MI trials for every class. For each 1ssignal chunk the logarithmic bandpower is applied, to get the logarithmic squared amplitude distribution of the signal over the selected time windows. The result is one feature vector for each MI class with a shape (40 trials, 49 features, 8 channels). E. Classification 1) Data Augmentation: We implemented a WGAN to avoid the vanishing gradients’ problem present in the original GAN and stabilize the training process. The goal of the Wasserstein network is to learn the data distribution for each class and augment the dataset with artificial features. We trained the model on the EEG features instead of the raw EEG signal to lower the dimensionality and complexity of the input. For the architecture, a CNN was designed with fully connected layers, for both the generator and the critic. The model for the generator consists of one input layer of noise sampled from a normal distribution, a fully connected layer which projects the noise to a feature vector and five pointwise convolutional transpose layers with a 1D kernel of size 1x1. For each convolutional layer, the ReLU was used as the activation function. The critic’s model is based on three temporal convolutional layers with 1D kernel of size 1x5, combined with three max pooling layers of size 1x2 to decrease the dimensionality of the input feature vector and a fully connected layer as the output layer. The ReLU was also used as the activation function for the convolutional layers. The architecture of the WGAN’s generator and critic are presented in Tables I and II. The Adam algorithm was used as the optimizer; a learning rate of 5·10−5, a weight clip factor of 0.01 and a batch size of 4 was set for the hyperparameters, since our EEG dataset is small. The generator was trained once for every 5 training iterations of the critic. The WGAN was trained initially for 2000 epochs to observe the loss curve of the critic and generator. We set the training epochs to 1000 as the model performed the best around that mark. A different model was trained for each class resulting into NWGAN models for Nclasses. Based on the trained models, the recorded EEG feature dataset was augmented to double the size of the original, resulting in 80 Fig. 4: VR Training scene for Motor Imagery (Right MI) MI trial features per class from the original 40. An increase to the original dataset by 50% and 200% was also tested but this offered lower classification results compared to a 100% increase. We assume that a less than 100% increase is not enough generated data for robust classification, while a more aggressive data augmentation approach will yield a more contaminated sample population. 2) Deep Learning Classification: For the classification of EEG features a CNN was designed. The model consists of three temporal convolutional layers with a 1D kernel of size 1x3 and the ReLU for the activation function. For every convolutional layer, a batch normalization layer was used to stabilize training and a max pooling layer of size 1x2 was appended to reduce the dimensionality of the features. Then, the output is flattened and passed to two fully connected layers, with one Dropout layer appended to the first fully connected layer with a dropout rate of 0.5 to add regularization to the model. The fully connected layers will decode the convolved features to produce one logits vector of #classes size. The architecture of the CNN is shown in Table III. For the training procedure, the Adam optimizer was used with a learning rate of 5 ·10−5and a batch size of 32. L2 regularization was also added to the model with a weight decay of 10−5to reduce overfitting. The categorical cross entropy was chosen as the loss function since we have up to four classes. It should be noted that in the output layer of the network, a softmax layer was not added to produce a probability distribution of the predicted classes as it is handled internally in the loss function. The logits vector from the final fully connected layer is then passed as the predicted class logits vector in the loss function along with the target class label. The CNN was trained with the K-fold strategy to obtain a robust estimate of the network’s performance while ensuring that no augmented features were present in the validation set, and then trained on the whole dataset to produce the final parameters of the network. Finally, the CNN was trained for 500 epochs. III. GAME DESIGN The proposed game for testing the BCI is a VR maze game. The user is controlling a playable character inside the maze. The goal of the game is to move the character through the maze and reach the end by finding the treasure chest. To move the playable character, the user has to invoke MI commands. Fig. 5: WGAN loss for left hand MI Fig. 6: Artificial vs Real EEG features in one trial for left hand MI (RMSE = 0.3) A. Game Scenes 1) Training Scene: The training procedure of OpenVibe was replaced with an in-game VR scene to control the structure and color palette of the scene’s contents so that eye strain and fatigue are limited and motor imagery training is engaging. It was inspired by OpenVibe’s example of the Graz Motor Imagery Visualization. During training, the user remains still and concentrates on the central part of the screen. A randomized sequence of visual cues indicating the MI class, is displayed, as depicted in Figure 4. For the color palette of the training scene, we chose dim colors to limit the eye strain factor and avoid any training problems such as imagination of wrong body limb movement. 2) Maze Levels: The VR game features three levels facilitating the use of two, three and four MI commands, respectively. In the first level, the user moves the playable character in the forward direction only, by invoking the feet MI command. When the user is in a rest state the character does not move. In the second level, besides the forward movement, the use rotates the character to the right by invoking the right hand MI. Since rotation of the character is enabled, the second maze introduces corners. In the final level, the left hand MI is added, which enables left rotation of the character. This way the playable character is able to move at any direction. Each level features doors that block the passage in the maze, with which the user can interact with the feet MI to open them. Fig. 7: WGAN error for each MI class B. BCI Integration Integration of the VR game with the BCI is achieved with a TCP server running in the background in a separate thread from the main game. The server will listen for connections from the client that receives EEG features from OpenVibe. Since the BCI is trained with features generated from EEG samples with a duration of 4s, the client sends a new prediction from the trained classifier to the VR game every four seconds. When a new prediction is received, the motor imagery command is handled appropriately by the game to control the movement of the character or interact with a door. IV. USER STUDIES This study follows Helsinki Declaration principles (1975/2000). Informed consent was obtained, and ethical safeguards were implemented despite no institutional approval requirement. A. Setup 1) Participants: For the evaluation of our proposed BCI and VR game we recruited 16 participants, 11 male and 5 female in the age range of 24-35. All participants reported that they had no prior experience with BCI experiments including MI training. 2) Procedure: The experiment was divided into two parts. In the first part, each participant was seated comfortably in a chair and after being informed about the procedure, the EEG headset together with the VR headset were positioned on their head as depicted in Figure 1. Then, each participant performed one session of MI training for four MI classes. Following recording, the offline processing of the EEG signal is performed in OpenVibe. Then, the deep models are trained for data augmentation and classification. One set of these neural network models are trained for each level of the VR game, corresponding to two, three and four classes respectively. In each level, the participant is able to perform only the MI commands that the neural network is trained with. After all levels are completed, the participants were requested to complete one questionnaire investigating participants’ perceived performance, rating each scene of the game on a 5-point Likert scale (1= Completely Unacceptable, 5= Completely Acceptable) and one NASA-TLX questionnaire [15] rating the task of motor imagery on a 10-point Likert scale (1= Very Low, 10= Very High). Fig. 8: Mean accuracy (%) of each method across all number of classes (∗: 1 ·10−2<p≤5·10−2,∗∗ : 1 ·10−3<p≤1·10−2,∗∗∗: 1 ·10−4<p≤1·10−3,∗ ∗ ∗∗ :p≤1·10−4) (a) Two classes (b) Three classes (c) Four classes Fig. 9: Train/Validation Loss curve of proposed classification method (a) Two classes (b) Three classes (c) Four classes Fig. 10: Confusion Matrices of proposed classification method B. Results and Discussion The stability of WGAN training is displayed by a training loss curve of the generator for Left Hand MI. As depicted in Figure 5, the generator reaches training convergence in approximately 2000 training iterations (1000 epochs), less than existing WGAN models. Further, Figure 6 presents a visual comparison between artificial and real EEG features. To assess the efficacy of our WGAN model, the RMSE value was calculated between an artificially generated feature vector and a real feature vector for each motor imagery class, averaged across all channels. Figure 7 showcases low RMSE values for all MI classes. Feet MI recorded the lowest RMSE values (µ=0.4,σ=0.13), followed by Left Hand MI (µ=0.47,σ=0.22) and Right Hand MI (µ=0.49,σ=0.22). The Rest state displayed the highest RMSE (µ=0.52,σ=0.19). We chose to implement a WGAN model for data augmentation over simpler methods such as adding Gaussian noise to the input, due to the ability of WGANs to create a diverse feature distribution of a single trial rather than slightly altering the original. Conventional augmentation techniques, such as the injection of Gaussian noise or amplitude scaling, can adversely impact the signal-to-noise ratio and compromise the coherence of the EEG signal, thereby introducing Fig. 11: Questionnaire rating each scene of the game Fig. 12: AUC/F1 scores of proposed classifier for two classes variability in feature representations across motor imagery trials. Independent t-tests were conducted to compare mean classification accuracy (Figure 8) of our proposed methodology against linear and Riemann machine learning algorithms. For two classes, there was a significant difference in accuracy for the comparison of our own WGAN+CNN (µ=70.79,σ=8.30) with LDA (µ=61.15,σ=4.62,p<0.001), SVM with RBF Kernel (µ= 61.11,σ=3.42,p<0.001), MDM (µ=56.32,σ=13.27,p= 0.001), FgMDM (µ=55.93,σ=12.53,p<0.001) and K-Nearest Neighbors (KNN) (µ=54.37,σ=10.41,p<0.001). For three classes, there was a significant difference in accuracy for the comparison of our own WGAN+CNN (µ=49.33,σ=8.24) with LDA (µ=44.57,σ=3.40,p=0.04), SVM with RBF Kernel (µ=34.51,σ=3.21,p<0.001), MDM (µ=41.08,σ=8.56,p= 0.01), FgMDM (µ=38.95,σ=7.08,p<0.001) and KNN (µ= 39.26,σ=5.13,p<0.001). For four classes, there was a significant difference in accuracy for the comparison of our own WGAN+CNN (µ=39.82,σ=7.19) with LDA (µ=34.87,σ= 2.83,p=0.01), SVM with RBF Kernel (µ=23.48,σ=1.90,p< 0.001), MDM (µ=31.28,σ=6.99,p=0.002), FgMDM (µ= 30.78,σ=6.35,p<0.001) and KNN (µ=26.60,σ=4.18,p< 0.001). Examining AUC, F1 scores (Figure 12) and confusion matrices (Figure 10), classification performance for our method is further validated on correct predictions, while eliminating the case of predicting only one class. Training and validation loss curves (Figure 9) indicate that the model presents minimal overfitting for two and three classes, respectively. For four classes, early stopping was used, since the model presented overfitting after 100 training epochs, indicating worse generalization performance. Due to the variability in the combination of motor imagery classes and system hardware, direct comparison with other deep networks that utilize per subject classification, was not possible, as variations in the environment could impact classification performance [18]. Thus, considering our 8-electrode system, the combination of motor imagery actions and the number of MI trials, we chose to compare our methodology against linearand Riemannbased machine learning algorithms, achieving high accuracy in small scale datasets such as ours. Since we utilize per subject classification, the number of participants in our study offers a robust assessment of classification accuracy achieved by our proposed methodology, and, we expect that a larger pool of users would further validate the results. The 4sprediction window gives naive users enough time to imagine limb movement and then return to a neutral state. It is expected that the prediction window could decrease as the user becomes more familiar with motor imagery. From our questionnaire (Figure 11), we infer that the BCI responded correctly to all users’ motor imagery for two classes. For more than two classes, most of the participants experienced prediction errors due to lower accuracy results. Further, motor imagery was rated in the NASA-TLX questionnaire as a mentally demanding task (µ=5.07,σ=2.43), requiring a lot of effort (µ= 5.42,σ=1.50). Participants reported mediocre temporal demand (µ=3.57,σ=1.82) to invoke a motor imagery command leading to frustration (µ=4.0,σ=1.75) for some participants. V. CONCLUSION We developed a novel processing pipeline based on a Wasserstein generative network for data augmentation and a convolutional network for motor imagery classification on small datasets collected from 16 naive BCI users. We performed a qualitative and quantitative analysis and results showcase decreased training convergence time for our WGAN model and increased classification accuracy with no overfitting for all motor imagery classes compared to linearand Riemann-based classifiers. In the future, the incorporation of spatial filters [37] in the preprocessing stage, may enable the combination of electrode channels to generate a lower dimensional input with better isolation of motor imagery classes between brain regions. The use of incremental learning [7] for our convolutional network on multiple recording sessions for the same user, could potentially increase the classification accuracy of the network across subsequent attempts at using the BCI. ACKNOWLEDGMENT This work has received financial support by the Horizon Europe Research & Innovation Programme under Grant agreement N.101092612 (Social and hUman ceNtered XR - SUN project). Views and opinions expressed in this work are those of the author(s) only and do not necessarily reflect those of the European Union. Neither the European Union nor the European Commission can be held responsible for them. REFERENCES [1] B. Z. Allison and C. Neuper. Could Anyone Use a BCI?, page 35–54. Springer, London, 2010. [2] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein generative adversarial networks. In Proceedings of the 34th International Conference on Machine Learning, page 214–223. PMLR, July 2017. [3] L. Bonnet, F. Lotte, and A. L´ ecuyer. Two brains, one game: Design and evaluation of a multiuser bci video game based on motor imagery. IEEE Transactions on Computational Intelligence and AI in Games, 5(2):185–198, June 2013. [4] G. Bouallegue and R. Djemal. Eeg data augmentation using wasserstein gan. In 2020 20th International Conference on Sciences and Techniques of Automatic Control and Computer Engineering (STA), page 40–45, Dec. 2020. [5] L. Budach, M. Feuerpfeil, N. Ihde, A. Nathansen, N. Noack, H. Patzlaff, F. Naumann, and H. Harmouch. The effects of data quality on machine learning performance. arXiv preprint arXiv:2207.14529, Nov. 2022. arXiv:2207.14529 [cs]. [6] C. F. Caiafa, J. Sol´ e-Casals, P. Marti-Puig, S. Zhe, and T. Tanaka. Decomposition methods for machine learning with small, incomplete or noisy datasets. Applied Sciences, 10(2323):8481, Jan. 2020. [7] F. M. Castro, M. J. Marin-Jimenez, N. Guil, C. Schmid, and K. Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. [8] M. Dai, D. Zheng, R. Na, S. Wang, and S. Zhang. Eeg classification of motor imagery using a novel deep learning framework. Sensors, 19(33):551, Jan. 2019. [9] H. Dose, J. S. Møller, H. K. Iversen, and S. Puthusserypady. An endto-end deep learning approach to mi-eeg signal classification for bcis. Expert Systems with Applications, 114:532–542, Dec. 2018. [10] N. Elsayed, Z. Saad, and M. Bayoumi. Brain computer interface: Eeg signal preprocessing issues and solutions. International Journal of Computer Applications, 169(3):12–16, July 2017. [11] D. Elstob and E. L. Secco. A low cost eeg based bci prosthetic using motor imagery. arXiv preprint arXiv:1603.02869, Mar. 2016. [12] F. Fahimi, Z. Zhang, W. B. Goh, K. K. Ang, and C. Guan. Towards eeg generation using gans for bci applications. In 2019 IEEE EMBS International Conference on Biomedical & Health Informatics (BHI), page 1–4, May 2019. [13] C. D. Guerrero-Mendez, C. F. Blanco-Diaz, A. F. Ruiz-Olaya, A. L´ opez-Delis, S. Jaramillo-Isaza, R. M. Andrade, A. F. De Souza, D. Delisle-Rodriguez, A. Frizera-Neto, and T. F. Bastos-Filho. Eeg motor imagery classification using deep learning approaches in na¨ ıve bci users. Biomedical Physics & Engineering Express, 9(4):045029, 2023. [14] C. Guger, C. Kapeller, R. Ortner, and K. Kamada. Motor imagery with brain-computer interface neurotechnology. Motor imagery: emerging practices, role in physical therapy and clinical implications, pages 61–79, 2015. [15] S. G. Hart and L. E. Staveland. Development of NASA-TLX (Task Load Index): Results of Empirical and Theoretical Research, volume 52 of Human Mental Workload, page 139–183. North-Holland, Jan. 1988. [16] K. G. Hartmann, R. T. Schirrmeister, and T. Ball. Eeg-gan: Generative adversarial networks for electroencephalograhic (eeg) brain signals. arXiv preprint arXiv:1806.01875, June 2018. arXiv:1806.01875 [cs, eess, q-bio, stat]. [17] T. M. Ingolfsson, M. Hersche, X. Wang, N. Kobayashi, L. Cavigelli, and L. Benini. Eeg-tcnet: An accurate temporal convolutional network for embedded motor-imagery brain–machine interfaces. In 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 2958–2965. IEEE, 2020. [18] H. S. Kim, M. H. Chang, H. J. Lee, and K. S. Park. A comparison of classification performance among the various combinations of motor imagery tasks for brain-computer interface. In 2013 6th International IEEE/EMBS Conference on Neural Engineering (NER), pages 435– 438. IEEE, 2013. [19] Y.-T. Kim, S.-B. Lee, H. Kim, J.-H. Jeong, S.-W. Lee, and D.-J. Kim. Exploring the number of repetitions in trials for the performance convergence of classification in motor imagery task with handgrasping. In 2019 7th International Winter Conference on BrainComputer Interface (BCI), page 1–4, Feb. 2019. [20] V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance. Eegnet: a compact convolutional neural network for eeg-based brain–computer interfaces. Journal of neural engineering, 15(5):056013, 2018. [21] M.-H. Lee, O.-Y. Kwon, Y.-J. Kim, H.-K. Kim, Y.-E. Lee, J. Williamson, S. Fazli, and S.-W. Lee. Eeg dataset and openbmi toolbox for three bci paradigms: an investigation into bci illiteracy. GigaScience, 8(5):giz002, May 2019. [22] C. Li, H. Yang, X. Wu, and Y. Zhang. Improving eeg-based motor imagery classification using hybrid neural network. In 2021 IEEE 9th International Conference on Information, Communication and Networks (ICICN), page 486–489, Nov. 2021. [23] R. Mane, E. Chew, K. Chua, K. K. Ang, N. Robinson, A. P. Vinod, S.- W. Lee, and C. Guan. Fbcnet: A multi-view convolutional neural network for brain-computer interface. arXiv preprint arXiv:2104.01233, 2021. [24] I. Martiˇ sius and R. Damaˇ seviˇ cius. A prototype ssvep based real time bci gaming system. Computational Intelligence and Neuroscience, 2016:e3861425, Mar. 2016. [25] Z. Miao, M. Zhao, X. Zhang, and D. Ming. Lmda-net: A lightweight multi-dimensional attention network for general eeg-based braincomputer interfaces and interpretability. NeuroImage, 276:120209, 2023. [26] V. P. Oikonomou, K. Georgiadis, G. Liaros, S. Nikolopoulos, and I. Kompatsiaris. A comparison study on eeg signal processing techniques using motor imagery eeg data. In 2017 IEEE 30th International Symposium on Computer-Based Medical Systems (CBMS), page 781–786, June 2017. [27] N. Padfield, J. Zabalza, H. Zhao, V. Masero, and J. Ren. Eegbased brain-computer interfaces using motor-imagery: Techniques and challenges. Sensors, 19(66):1423, Jan. 2019. [28] S. Panwar, P. Rad, J. Quarles, E. Golob, and Y. Huang. A semisupervised wasserstein generative adversarial network for classifying driving fatigue from eeg signals. In 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC), page 3943–3948, Oct. 2019. [29] S. Paszkiel. Control based on brain-bomputer interface technology for video-gaming with virtual reality techniques. Journal of Automation, Mobile Robotics and Intelligent Systems, 10:3–7, Dec. 2016. [30] G. Pfurtscheller and C. Neuper. Motor imagery activates primary sensorimotor area in humans. Neuroscience Letters, 239(2):65–68, Dec. 1997. [31] G. Prapas, K. Glavas, A. T. Tzallas, K. D. Tzimourta, N. Giannakeas, and M. G. Tsipouras. Motor imagery approach for bci game development. In 2022 7th South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM), page 1–5, Sept. 2022. [32] Y. Renard, F. Lotte, G. Gibert, M. Congedo, E. Maby, V. Delannoy, O. Bertrand, and A. L´ ecuyer. Openvibe: An open-source software platform to design, test, and use brain–computer interfaces in real and virtual environments. Presence, 19(1):35–53, 2010. [33] A. Singh, S. Lal, and H. W. Guesgen. Small sample motor imagery classification using regularized riemannian features. IEEE Access, 7:46858–46869, 2019. [34] Y. Song, Q. Zheng, B. Liu, and X. Gao. Eeg conformer: Convolutional transformer for eeg decoding and visualization. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 31:710–719, 2022. [35] D. P. Subha, P. K. Joseph, R. Acharya U, and C. M. Lim. Eeg signal analysis: A survey. Journal of Medical Systems, 34(2):195–212, Apr. 2010. [36] Y. R. Tabar and U. Halici. A novel deep learning approach for classification of eeg motor imagery signals. Journal of Neural Engineering, 14(1):016003, Nov. 2016. [37] D. Wu, J.-T. King, C.-H. Chuang, C.-T. Lin, and T.-P. Jung. Spatial filtering for eeg-based regression problems in brain-computer interface (bci). IEEE Transactions on Fuzzy Systems, Feb. 2017. arXiv:1702.02914 [cs]. [38] Y. Wu. Applications of eeg-based brain-computer interface devices in rehabilitation. Highlights in Science, Engineering and Technology, 39:809–815, Apr. 2023. [39] D. Zhang, M. Ma, and L. Xia. A comprehensive review on gans for time-series signals. Neural Computing and Applications, 34(5):3551–3571, Mar. 2022.