%0 Journal Article
%T Noise-Assisted Instantaneous Coherence Analysis of Brain Connectivity
%A Meng Hu
%A Hualou Liang
%J Computational Intelligence and Neuroscience
%D 2012
%I Hindawi Publishing Corporation
%R 10.1155/2012/275073
%X Characterizing brain connectivity between neural signals is key to understanding brain function. Current measures such as coherence heavily rely on Fourier or wavelet transform, which inevitably assume the signal stationarity and place severe limits on its time-frequency resolution. Here we addressed these issues by introducing a noise-assisted instantaneous coherence (NAIC) measure based on multivariate mode empirical decomposition (MEMD) coupled with Hilbert transform to achieve high-resolution time frequency representation of neural coherence. In our method, fully data-driven MEMD, together with Hilbert transform, is first employed to provide time-frequency power spectra for neural data. Such power spectra are typically sparse and of high resolution, that is, there usually exist many zero values, which result in numerical problems for directly computing coherence. Hence, we propose to add random noise onto the spectra, making coherence calculation feasible. Furthermore, a statistical randomization procedure is designed to cancel out the effect of the added noise. Computer simulations are first performed to verify the effectiveness of NAIC. Local field potentials collected from visual cortex of macaque monkey while performing a generalized flash suppression task are then used to demonstrate the usefulness of our NAIC method to provide highresolution time-frequency coherence measure for connectivity analysis of neural data. 1. Introduction To understand how brain networks process information, it is crucial to accurately quantify their connectivity patterns. For analysis of brain connectivity between two signals, current measures such as coherence [1¨C3] rely upon spectral estimate of each signal, which is routinely computed based on Fourier or wavelet transform. Thus, the underlying nonstationary nature of neural data presents a significant challenge for the applications of current measures. Though short-time sliding window approaches, for example, short-time Fourier transform, have been used to alleviate this problem, this issue is not completely resolved for a number of reasons. First, the stationarity of neural data within each short-time window cannot be guaranteed. Second, even though the data are stationary within each time window, the resolution of time-frequency representation is limited by Heisenberg uncertainty principle [4]. Wavelet transform [4], albeit improved, is still subject to time-frequency resolution tradeoff, that is, frequency resolution is low at high frequencies and high at low frequencies. Moreover, wavelet analysis depends on
%U http://www.hindawi.com/journals/cin/2012/275073/