A CNN-based modular classification scheme for motor imagery using a novel EEG sampling protocol suitable for IoT healthcare systems

• Published in: Neural computing & applications Vol. 35; no. 31; pp. 22865 - 22886
• Main Authors: Chacon-Murguia, Mario I., Rivas-Posada, Eduardo
• Format: Journal Article
• Language: English
• Published: London Springer London 01.11.2023; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Artificial neural networks; Classification schemes; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Datasets; Electroencephalography; Feature extraction; Frequencies; Human-computer interface; Image Processing and Computer Vision; Methodology; Probability and Statistics in Computer Science; Real time; S.I. : Neural Computing for IOT based Intelligent Healthcare Systems; Sampling; Signal classification; Special Issue on Neural Computing for IOT based Intelligent Healthcare Systems; Real-time brain computer interface; Motor imagery; Convolutional neural network; Continuous wavelet transform; Common spatial patterns
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-021-06716-x

The implementation of brain-computer interfaces (BCI) for real-time has become a paramount technology. Implementation of real-time BCI systems requires of methodologies that achieve high performance on classification over general brain signals of different subjects. Therefore, this work presents two simple and efficient methodologies to classify two and four motor imageries. The methodology to classify two motor imageries (MC-TM) includes an analysis of feature extraction methods based on spatial patterns and time–frequency transforms; and a convolutional neural network that preserves the information of the magnitude in the frequency bands of the sensorimotor rhythms (CNN-PIM). Besides, the methodology to classify four motor imageries (MC-FM) includes a modular classification scheme that instances 6 CNN-PIM; a new algorithm that uses the output of the softmax of each CNN-PIM to enhance the performance of the MC-FM methodology; an electroencephalogram sampling protocol that includes a specific procedure for 4 MIs classes; and a new dataset with the brain signals of 15 subjects. The MC-TM methodology achieved an accuracy of 94.44 ± 02.18% evaluated in BCI Competition IV dataset 2a (BCI-IV-2a), and accuracy of 97.67 ± 02.06% when evaluated in the EEGdataset. Meanwhile, the MC-FM achieved accuracies of 91.37 ± 3.29% and 86.48 ± 4.74% when evaluated in BCI-IV-2a and in the proposed dataset, respectively. These results situate our methodologies in a competitive position in comparison with the state-of-the-art methods. Moreover, the approximate processing time that the MC-FM methodology takes to classify EEG signals is 270 ms. Thus, it is suitable to be implemented in a real-time BCI system.