MULTIVARIATE HURST EXPONENT ESTIMATION IN FMRI. APPLICATION TO BRAIN DECODING OF PERCEPTUAL LEARNING

So far considered as noise in neuroscience, irregular arrhythmic field potential activity accounts for the majority of the signal power recorded in EEG or MEG [1, 2]. This brain activity follows a power law spectrum P(f) similar to 1/f(beta) in the limit of low frequencies, which is a hallmark of scale invariance. Recently, several studies [1, 3-6] have shown that the slope beta (or equivalently Hurst exponent H) tends to be modulated by task performance or cognitive state (eg, sleep vs awake). These observations were confirmed in fMRI [7-9] although the short length of fMRI time series makes these findings less reliable. In this paper, to compensate for the slower sampling rate in fMRI, we extend univariate wavelet-based Hurst exponent estimator to a multivariate setting using spatial regularization. Next, we demonstrate the relevance of the proposed tools on resting-state fMRI data recorded in three groups of individuals once they were specifically trained to a visual discrimination task during a MEG experiment [10]. In a supervised classification framework, our multivariate approach permits to better predict the type of training the participants received as compared to their univariate counterpart.