A New Adaptive Sliding Window Method for fMRI Dynamic Functional Connectivity Analysis

Abstract

The fixed-window sliding time window method is widely used in exploring dynamics functional connectivity of functional magnetic resonance imaging data analysis, but it is difficult to select a suitable window to capture the dynamic changes in brain function. Therefore, a local polynomial regression (LPR) method is proposed to fit the region of interest (ROI) time series in this paper, in which observations are locally modeled by a least-squares polynomial with a kernel of a certain bandwidth that allows for better bias-variance tradeoff. It combines a data-driven variable bandwidth selection mechanism with intersection of confidence intervals (ICI) and a bandwidth optimization algorithm of particle swarm optimization (PSO). Among them, ICI is used to adaptively determine the locally optimal bandwidth that minimizes the mean square error (MSE), and then the bandwidth values at various time points within all ROIs are computed for each subject. Subsequently, the averaged bandwidth values at these time points is regarded as the bandwidth value for that subject at each time point, followed by generating a time-varying bandwidth sequence for each subject, which is used in the PSO-based bandwidth optimization algorithm. Finally, the results of experiments conducted on simulated data showed that the LPR–ICI–PSO method exhibited lower MSE values on time-varying correlation coefficient estimation for different noise scenarios. Furthermore, we applied the proposed method to the autism spectrum disorder (ASD) study, and obtained a classification accuracy of 74.1% from typical controls (TC) through support vector machine (SVM) with the 10-fold cross-validation strategy. These results demonstrated that our proposed method can effectively capture the dynamic changes in brain function, which is valid in clinical diagnosis and helps to reveal the differences in brain functional connectivity patterns.