Motor imagery-based neural networks for assisting tetraplegic patients.

Nowadays, deep network-based classification algorithms are used in a myriad of applications for brain-computer interfaces (BCIs). These interfaces can enhance the daily lives of quadriplegic patients. Electroencephalography (EEG) based motor imagery (MI) is an integral part of BCI, and the performance of the available deep classifiers is still limited. This paper presents a novel convolutional neural network (CNN) architecture designed to enhance the multiclass classification accuracy of motor imagery (MI) signals acquired through EEG-based sensing. We have selected the electrodes over the sensorimotor cortex region of the brain in the 8-30 Hz EEG frequency band. Further, we have computed the classification accuracy and kappa scores in an end-to-end deep classification network. Our framework surpasses the contemporary literature algorithms in classifying BCI competition IV-2a, a four-class MI dataset of nine subjects (left hand, right hand, both feet, tongue). The proposed network architecture has achieved an average and maximum accuracy of 95.19% and 99.28%, respectively. We have outperformed state-of-the-art accuracies of the individual subjects S1, S2, S3, S4, S5, S6, S8, and the average accuracy of the dataset by 8.28%, 40.97%, 5.54%, 14.83%, 19.09%, 25.5%, 10.43%, and 12.82% respectively.