Title:A robust multivariate, non-parametric outlier identification method for scrubbing in fMRI

Abstract:Functional magnetic resonance imaging (fMRI) data contain high levels of noise and artifacts. To avoid contamination of downstream analyses, fMRI-based studies must identify and remove these noise sources prior to statistical analysis. One common approach is the "scrubbing" of fMRI volumes that are thought to contain high levels of noise. However, existing scrubbing techniques are based on ad hoc measures of signal change. We consider scrubbing via outlier detection, where volumes containing artifacts are considered multidimensional outliers. Robust multivariate outlier detection methods are proposed using robust distances (RDs), which are related to the Mahalanobis distance. These RDs have a known distribution when the data are i.i.d. normal, and that distribution can be used to determine a threshold for outliers where fMRI data violate these assumptions. Here, we develop a robust multivariate outlier detection method that is applicable to non-normal data. The objective is to obtain threshold values to flag outlying volumes based on their RDs. We propose two threshold candidates that embark on the same two steps, but the choice of which depends on a researcher's purpose. Our main steps are dimension reduction and selection, robust univariate outlier imputation to get rid of the effect of outliers on the distribution, and estimating an outlier threshold based on the upper quantile of the RD distribution without outliers. The first threshold candidate is an upper quantile of the empirical distribution of RDs obtained from the imputed data. The second threshold candidate calculates the upper quantile of the RD distribution that a nonparametric bootstrap uses to account for uncertainty in the empirical quantile. We compare our proposed fMRI scrubbing method to motion scrubbing, data-driven scrubbing, and restrictive parametric multivariate outlier detection methods.