Treatment of baseline drifts in fMRI time series analysis.

Abstract

Gradual drifting of baseline signal intensity is common in functional MRI (fMRI) time course data. Methods for dealing with this effect are studied. Simulations and fMRI data are used to study three statistical models that account for baseline drift. A method is proposed in which the time course data are linear least-squares fit to a reference function that includes the slope of the baseline drift as a free parameter. It is shown that the least-squares method is equivalent to cross-correlation with Gram-Schmidt orthogonalization. Additionally, it is shown that certain paradigm designs improve the sensitivity of statistical tests when using any of the drift correction methods commonly employed. The least-squares method results in a variety of useful parameters such as activation amplitude, with a well characterized error. Very simple techniques can effectively account for observed drifts. It is important to design paradigms that are symmetric about the midpoint of the time series. In calculating confidence levels, a proper statistical model that accounts for baseline drifts is necessary to ensure accurate confidence level assessment.