Decoding of finger movement using kinematic model classification and regression model switching

Abstract

Brain Computer Interface (BCI) is one of the clinical applications that may help to restore communication to people with motor disabilities. Electrocorticography (ECoG) is a semi invasive record to brain signals from electrode grids on the cortex surface. ECoG signal makes possible localization of the source of neural signals due to its high spatial resolution. This study is a step towards exploring the usability of ECoG signal as BCI input technique and a multidimensional BCI control. The objective of this deterministic approach is to predict individual finger movement from ECoG signal by combining both classification and regression problems in machine learning of signal responses (regression via classification), on the other hand addressing the signal responses variability within a single subject. The dataset used in this work is the one presented in the fourth dataset from BCI competition IV. The difficulty is that; there is no simple and direct relation between ECoG signals and finger movements. This research work starts in two directions. The first direction is related to the decoding of the finger position signal to obtain a finger movement state signal. The second direction is related to the ECoG recorded signal, in order to obtain the corresponding brain signal of each finger movement. The work consists of five main phases (decoding finger state, pre-processing, features acquisition, classification, and regression). This approach suggests kinematic finger model which is applied on the finger muscle signal to generate the finger kinematic state signal. For feature extraction we used shift invariant wavelet decomposition and multi-taper frequency spectrum, followed by Gram-Schmidt test for selection. Linear support vector machine (SVM) is used for classification. Regression models are established by using the finger position training signal and the acquired ECoG features. To predict the finger movement signal under test; switching between these regression models is made. Finally the predicted finger movement signal is correlated with the measured one for evaluation. Results show that the average correlation measure between real and predicted movement is 0.82. This result is higher than the one obtained by the competition winner (0.46).