Data augmentation and feature extraction using deep learning for motor imagery EEG-based brain–computer interface classification

• Published in: Neural computing & applications Vol. 37; no. 23; pp. 19339 - 19369
• Main Authors: Anjerani, Marzieh, Pedram, Mir Mohsen, Mirzarezaee, Mitra
• Format: Journal Article
• Language: English
• Published: London Springer London 01.08.2025; Springer Nature B.V
• Subjects: Accuracy; Artificial Intelligence; Brain research; Classification; Communication; Competition; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data augmentation; Data Mining and Knowledge Discovery; Datasets; Deep learning; Electroencephalography; Feature extraction; Human-computer interface; Image Processing and Computer Vision; Imagery; Interfaces; Machine learning; Original Article; Probability and Statistics in Computer Science; Recall; Signal to noise ratio; Spatial resolution; Deep learning; Motor imagery (MI); Electroencephalography (EEG); Data augmentation (DA); Probabilistic graphical models; Brain–computer interface (BCI)
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-025-11403-2

Motor imagery EEG-based brain–computer interfaces (BCIs) have recently been developed for communicating between the brain and external devices. One of the most difficult issues in BCI is classifying the brain activity of motor imagery. To achieve optimal results, the most suitable feature extraction and classification approach must be used because of the poor signal-to-noise ratio, low spatial resolution, and variation in brain activity among subjects in EEG. Gathering enough EEG data takes a lot of time and labor since EEG signals are nonstationary and variable. The main contribution of the paper is to propose models based on deep learning to augment EEG trials to improve EEG classification performance for multi-class motor imagery BCI even when small amounts of EEG data are provided for training. To augment the EEG data in this investigation, generative adversarial models and autoencoders were employed. Then, using various autoencoders, discriminative features were extracted from all the training data and augmented signals. The final step involved classifying multi-class motor imagery of the EEG BCI using deep learning and probabilistic graphical models. For the purpose of classifying EEG signals, the effectiveness of our various feature extraction and data augmentation models was examined using a standard BCI benchmark dataset in terms of classification accuracy, Cohen’s kappa, F1-score, and recall. The suggested model outperforms other cutting-edge models with an average classification accuracy of 90.20%, a kappa value of 0.98, and a recall value of 0.99 on the BCI Competition IV dataset 2a. The proposed model is subject-independent since it can be integrated into a single model by employing training on all subjects. Therefore, the proposed models may be of considerable interest for BCI applications in the real world.