Nonlinear analysis of motor imagery and motor execution to improve accuracy and efficiency of brain-computer interfaces

Abstract

Abstract Brain-computer interface (BCI) offers the possibility to mitigate disability arising from paralysis by using neural signals to control external devices. The capability of the BCI system to control the devices is essentially determined by decoding strategy for motor execution (ME) and kinesthetic motor imageries (KMIs) from neural signals. Previous studies have utilized various linear feature selection strategies to improve the decoding accuracy of KMIs. However, using a nonlinear feature selection strategy has not been examined fully yet. Here, we aim to propose a computationally efficient nonlinear feature selection strategy that improves decoding accuracy significantly. This aim is achieved by identifying common or specific features between ME and KMI that lead to the higher efficient accuracy using nonlinear dynamical measures of the electrocorticography (ECoG). The ECoG data across ME and KMI were collected from 9 patients and the approximate entropy and correlation dimension of the ECoG were computed in both states. The decoding accuracies of a combination of common or specific features were then compared using support vector machines (SVM). In this study, we demonstrate a highly efficient strategy for BCI using nonlinear dynamics of the cerebral and subcortical brains by providing better decoding accuracy or sufficient accuracy with less computation.The results of the current study are potentially helpful for real-time control for external devices using ECoG in paralyzed patients with high accuracy. Keywords: brain-computer interface, electrocorticography (ECoG), machine learning, decoding, motor execution (ME), kinesthetic motor imageries (KMIs)