Maryam Fatemi; Mohammad Reza Daliri
Volume 13, Issue 4 , December 2019, , Pages 327-336
Abstract
Controlling of neuroprostheses to restore grasping ability in patients with paralyzed or amputated upper limbs is one of the important applications of BCI systems. The ability to get objects is necessary for daily works so, for a reliable function of the neuroprostheses, it is necessary for the user ...
Read More
Controlling of neuroprostheses to restore grasping ability in patients with paralyzed or amputated upper limbs is one of the important applications of BCI systems. The ability to get objects is necessary for daily works so, for a reliable function of the neuroprostheses, it is necessary for the user to control the amount of force needed for grasping. For this reason, increasing the accuracy of continuous force decoding is an important issue for the convenient function of these BCI systems. In most studies in the field of force decoding, linear models such as wiener filter, Kalman filter, PLS, etc. are used to decode force. So far, the effect of using nonlinear models is not investigated on force decoding. The goal of this study is to investigate the effect of using nonlinear regression models based on kernel functions on the accuracy of force decoding in Vistar rats using local field potential signals. To do this, we choose ridge regression, PCR and PLS methods and use the Gaussian kernel function to construct a generalized nonlinear model for the force decoding. Evaluating kernel ridge, kernel PCR and kernel PLS methods shows that considering nonlinear relations between brain signal’s features improves decoding accuracy. The mean coefficient of determination (R2) improves 12.7% in kernel ridge toward ridge regression, 25.5% in kernel PCR toward PCR and 19.1% in kernel PLS toward PLS method. The best decoding accuracy has been achieved by the kernel ridge regression method and the mean correlation coefficient between the estimated and measured force is 0.72 and R2 is 0.62.
