Subject-Independent Classification of Motor Imagery Tasks in EEG Using Multisubject Ensemble CNN

• Published in: IEEE access Vol. 10; pp. 81355 - 81363
• Main Authors: Dolzhikova, Irina, Abibullaev, Berdakh, Sameni, Reza, Zollanvari, Amin
• Format: Journal Article
• Language: English
• Published: Piscataway The Institute of Electrical and Electronics Engineers, Inc. (IEEE) 2022; IEEE
• Subjects: Accuracy; Artificial neural networks; Brain-computer interface; Classification; Classifiers; convolutional neural network; Datasets; deep learning; Electroencephalography; Human-computer interface; Image classification; Median (statistics); multi-subject ensemble; Optimization; Performance evaluation; Training
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2022.3195513

Subject-independent (SI) classification is a major area of investigation in Brain-Computer Interface (BCI) that aims to construct classifiers of users’ mental states based on collected electroencephalogram (EEG) of independent subjects. Significant inter-subject variabilities in the EEG are among the most challenging issues in designing SI BCI systems. In this work, we propose and examine the utility of Multi-Subject Ensemble Convolutional Neural Network (MS-En-CNN) for SI classification of motor imagery (MI) tasks. The base classifiers used in MS-En-CNN have a fixed CNN architecture (referred to as DeepConvNet) that are trained using data collected from multiple subjects during the training process. In this regard, training subjects are divided into [Formula Omitted]-folds using which [Formula Omitted] base DeepConvNets are trained based on data from [Formula Omitted] folds, whereas the hyperparameter optimization is performed using the held-out fold. We evaluate the performance of the MS-En-CNN on the large open-access MI dataset from the literature, which includes 54 participants and a total number of 21,600 trials. The result shows that the MS-En-CNN achieves the highest single-trial SI classification performance reported on this dataset. In particular, we obtained SI classification performances with average and median accuracies of 85.42% and 86.50% (± 10.16%), respectively. This result exhibits a statistically significant improvement ([Formula Omitted]) over the best previously reported result with an average and a median accuracy of 84.19% and 84.50% (±10.08%), respectively.