Estimation of Imaginary Movements Quality Based on Machine Learning for Brain Computer Interface Applications

Abstract

Brain-computer interfaces based on motor activity aim at restoring motor skills for people with paralysis or providing devices to help such people (interfaces with computers, robotic hands, wheelchairs). In such interfaces, imaginary movements are widely used. Imaginary movement it is the mental process, during which a person imagines a certain movement, as a result of which the neural activity in the motor regions of the cerebral cortex is modulated. These oscillations can be observed in electroencephalograms. Imaginary movements are used to train neurocomputer interfaces, and it is often impossible to evaluate the quality of the subject's performance of the task. The process of performing imaginary movement is not an easy task and requires prior training. Therefore, the accuracy of classification for some subjects can be higher than for other. All methods for imagine movement classification could be improved if there would be a methodology for assessing the quality of subject performance of imaginary movements. Then the data of the experiments could be filtered, and only valid data for training could be used. The paper presents the application of machine learning to detect anomalies (outliers) in data and improve the accuracy of the imaginary movement classification for brain-computer interfaces. The database with electroencephalogram signals of 29 subjects that performed right and left hand imaginary movements (NIRx GmbH, Berlin) was used. The signals from the eight channels corresponding to the motor zones of the cerebral cortex are filtered in alpha and beta frequency ranges. For feature extraction was used parameter event related synchronization and desynchronization which was obtained from different frequency bands after filter the raw EEG. Using this index, the decreasing and increasing of oscillations in alpha and beta rhythms that occur in the localized areas of the motor cortex of the brain can be seen. Support Vector Machine with linear kernel used for classification. An approach is presented on the basis of unsupervised machine learning methods for anomaly detection. As a result, the outlier fraction is obtained. This parameter shows the percentage of saturation data with anomalies. This parameter is actually an indicator of the quality of the input data and is used to clear data before training the brain computer interface. As a result of proposed approach application, the accuracy of the imaginary movement classification increased by 14.9% for 8 subjects. For other subjects, the accuracy remained unchanged, or decreased. This approach is subject-specific and requires customization for a specific user and improvement before further wide usage, but now it is already possible to significantly improve the accuracy of imaginary movement classification for individual subjects. Perspective use of anomaly detection for estimation of imaginary movement quality established. Ref. 28, Fig. 2, Tabl. 2.