Generating Ten BCI Commands Using Four Simple Motor Imageries

TL;DR

This paper proposes a method to generate up to ten BCI commands using combined motor imageries of different body parts, employing artificial EEG signals and a small deep neural network to improve classification performance.

Contribution

It introduces a novel approach to create more BCI commands by combining simple motor imageries and validates the effectiveness with high classification accuracy.

Findings

Achieved 81.8% average classification accuracy for ten classes.

Successfully generated and classified combined MI EEG signals in real-time.

Validated the approach on multiple datasets with promising results.

Abstract

The brain computer interface (BCI) systems are utilized for transferring information among humans and computers by analyzing electroencephalogram (EEG) recordings.The process of mentally previewing a motor movement without generating the corporal output can be described as motor imagery (MI).In this emerging research field, the number of commands is also limited in relation to the number of MI tasks; in the current literature, mostly two or four commands (classes) are studied. As a solution to this problem, it is recommended to use mental tasks as well as MI tasks. Unfortunately, the use of this approach reduces the classification performance of MI EEG signals. The fMRI analyses show that the resources in the brain associated with the motor imagery can be activated independently. It is assumed that the brain activity induced by the MI of the combination of body parts corresponds to the superposition of the activities generated during each body parts's simple MI. In this study, in order to create more than four BCI commands, we suggest to generate combined MI EEG signals artificially by using left hand, right hand, tongue, and feet motor imageries in pairs. A maximum of ten different BCI commands can be generated by using four motor imageries in pairs.This study aims to achieve high classification performances for BCI commands produced from four motor imageries by implementing a small-sized deep neural network (DNN).The presented method is evaluated on the four-class datasets of BCI Competitions III and IV, and an average classification performance of 81.8% is achieved for ten classes. The above assumption is also validated on a different dataset which consists of simple and combined MI EEG signals acquired in real time. Trained with the artificially generated combined MI EEG signals, DivFE resulted in an average of 76.5% success rate for the combined MI EEG signals acquired in real-time.