A Multi-label Classification Method for Detection of Combined Motor Imageries

Imaginary motor tasks cause brain oscillations that can be detected through the analysis of electroencephalographic (EEG) recordings. The imagination of hands movement allows inducing up to three different brain states by considering the activity that each hand produces separately and the one caused by the combination of both. This article presents a new method to extend the classic Common Spatial Pattern (CSP) algorithm to a multi-class approach which analyses both brain hemispheres separately to solve, together with a stepwise classification strategy, a multi-label Brain-Computer Interface (BCI) problem. The considered approach is based upon the assumption that the brain activity induced by the motor imagery (MI) of the combination of both hands corresponds to the superposition of the activity generated during simple hand MIs. In this way, based on the event-related desynchronization that is detected within each brain hemisphere, the multi-classification task can be reduced into two binary-classification problems, leading to a much simpler recognition scheme that overcomes the drawback of the classical CSP method of being suitable to discriminate only between two classes. After testing the proposed approach over the EEG signals of six healthy subjects performing a four-class multi-label task involving simple and combined hand MIs together with the rest condition, results show that this technique is plausible for BCI control. In terms of accuracy, it outperforms the classical one-vs-one approach by 20% and has the same performance as the one-vs-all method. Nevertheless, to solve a multi-label classification problem involving k classes, the proposed method requires only log2 (k) classifiers, whereas the one-vs-one method uses k (k-1)/2 classifiers and the one-vs-all k classifiers, thereby the new approach simplifies the classification task and seems promising for solving multi-label problems involving numerous classes.