Using autoregressive-dynamic conditional correlation model with residual analysis to extract dynamic functional connectivity

Statistical methods that simultaneously model temporal and spatial variations of fMRI data are promising tools for dynamic functional connectivity (FC) estimation. Although different approaches are available, they need to manually set the parameters, or may disregard some important fMRI features such as the autocorrelation. In addition, no reliable method exists for the validation of dynamic FC analysis models.

In the present study, we have proposed an autoregressive dynamic conditional correlation model to deal with the temporal autocorrelation and non-stationarity in fMRI time-series. This model assumes that the brain time courses follow a multivariate Gaussian distribution, and that the conditional mean, variance and covariances change in an autoregressive form. Also, we proposed a new measurement index for the evaluation of the statistical consistency between the inferred dynamic functional connectivity and the real fMRI data. The performance of our model was tested in both simulated and real fMRI data.

The model was associated with independent Gaussian residuals, and identified the dynamic connectivity patterns with high precision. Applying the model to the fMRI data from typically developing and attention deficit hyperactivity disorder subjects, brain connectivities were significantly different between the two groups.

Prominent features of our model were the consideration of the fMRI autocorrelation, no need to adjust the window length, and also elimination of the variance changes in each brain time-course from its connectivity changes.