An algorithm to decode movement and laterality from Deep Brain Local Field Potentials utilizing time and frequency domain features

Abstract

Deep Brain Local Field Potentials (LFP) are one of the most promising techniques in the process of understanding the inner workings of our the central nervous system (CNS). The information gathered on neural activity is a key component of interfaces between the human brain and artificial devices. The LFPs recorded from subthalamic nucleus (STN) are important sources of information related to the preparation, execution and imaging of movements. In this paper, time and frequency domain features are used to decode movements and its laterality, left or right sided visually cued movements from STN LFPs. First the frequency components of LFPs are separated using wavelet packet transform (WPT). The feature selection process involves combining three approaches using time window based power features, causality features computed using granger causality and cross correlation, and frequency domain features computed using discrete cosine transform (DCT). A weighted sequential feature selection process is then used to prune the feature vector. A Bayesian classifier along with cross validation procedure is applied in the classification process. The classification results show a marked improvement in performance over previous reports.