Tuo Shi1, D Rangaprakash1, and Gopikrishna Deshpande1,2,3
1AU MRI Research Center, Department of Electrical and Computer Engineering, Auburn University, Auburn, AL, United States, 2Department of Psychology, Auburn University, Auburn, AL, United States, 3Alabama Advanced Imaging Consortium, Auburn University and University of Alabama Birmingham, Auburn, AL, United States
Synopsis
In this work, we propose a novel
strategy for selecting the minimum window length required to capture maximum
dynamics as well as reliably estimate correlation during dynamic functional
connectivity analysis. Using the error in estimated correlation compared to simulated
ground-truth correlation as the metric, we compared our method with (i) the fixed
window length approach, and (ii) the DCC method. We show that our method can
provide minimum window lengths which give more reliable correlation estimates
than those obtained from DCC and fixed-window methods. Further, we show that
our method can accurately track fast variations in connectivity.Introduction
Most
functional Magnetic Resonance Imaging (fMRI) studies implicitly assume that the
functional connectivity (FC) between brain regions is constant over time.
However, FC has been shown to be non-stationary. Studies indicate that dynamic
FC (DFC) contains neurobiologically meaningful information not available in
conventional static connectivity [1]. Evaluation of DFC involves successive sliding-windows,
with correlation being evaluated in each of them. One of the critical issues
concerning sliding-window analysis is the choice of window size, with arbitrary
choice of fixed window size leading to arbitrary results [2]. Techniques based
on Dynamic Conditional Correlation (DCC) [3] and timeseries stationarity [1]
have been developed lately to address this issue. They search for the window
length within a predefined range which satisfies certain mathematical criteria.
However the choice of minimum window length (MWL) from which to start the
search is an arbitrary small number. While we would like the window to be as
small as possible so that maximum dynamics is captured, the main issue is that
the correlation estimated for such small windows might not be reliable, that
is, the estimated correlation might deviate largely from the ground-truth
correlation. It is thus necessary to objectively assess the impact of window
length on the reliability of estimated correlation. In this work, we propose a novel
strategy for selecting MWL. Since the ground-truth can be precisely controlled
in artificial data, we performed simulations. Using the error in evaluated
correlation compared to ground-truth correlation as the metric, we compared our
method with (i) the popular arbitrary fixed window length approach, and (ii) the
DCC method [3], which is model-based and requires no ad-hoc choices. Note that
while ours is a method to obtain MWL required to give reliable correlation, DCC
and fixed-window approaches are DFC methods which do not consider reliability
of correlations in their formulation. Hence we hypothesize that our method
would result in minimum window lengths which give more reliable correlation
estimates than those obtained from DCC and fixed-window methods.
Methods
The
MWL required to give reliable correlations can be obtained only when the
ground-truth correlation is known, which is possible through simulations. We
simulated timeseries pairs with 1000 time points each, with predefined
correlation between them. As in our pervious study [1], we used a multivariate
vector autoregressive (MVAR) model to generate pairs of timeseries as follows:
$$Y(t)=\sum_{i=1}^pA_{i}*Y(t-i)+\epsilon$$ where $$$\epsilon$$$ represents noise
vector with covariance matrix $$$C=\begin{bmatrix}1 & {v(t)} \\{v(t)} & 1 \end{bmatrix}$$$ and $$$Y(t)=\begin{bmatrix}{y_{1}(t)} & {y_{2}(t)}\end{bmatrix}$$$ denotes simulated
time series. $$$A_{i}$$$ is regression
coefficient matrix for delay $$$i$$$ , chosen to be zero
matrices such that only zero-lag correlation (no time-lagged relationships)
were considered. We simulated timeseries pairs with constant predefined
correlation between them($$$v(t)=constant$$$),
with correlation being varied from 0 to 1 with step-size of 0.05. Window
length was varied from 5 to 100 with step-size of 1, and DFC was evaluated over
the range of window lengths. We used sliding-windowed Pearson’s correlation to
evaluate DFC. This procedure was iterated 100 times. Employing the widely used
Dickey-Fuller (DF) test [4], we obtained MWL required to ensure reliable
correlation, that is, to ensure stationarity of correlation values with 95% confidence.
Having calculated the
MWL required to estimate correlation reliably, we performed simulations to
validate our minimum stationary window length (MSWL) method and compared it
with DCC and fixed-window methods. We considered two scenarios as follows. Case-1:$$$v(t)=0.5+0.3*\sin(0.01t)$$$
, which represents a
slow and less-varying periodic change in correlation, and case-2:$$$v(t)=0.5+5*\sin(0.5t)$$$,which
represents a fast and largely varying periodic change in correlation. We
iterated 100 times and compared the error between estimated correlations and
ground-truth correlation for all methods.
Results and
Discussion
The
reliability with which the estimated correlation from simulated data matched
the ground-truth correlation value (in terms of estimation error in Fig.1 and
correlation stationarity in Fig.2) depended on the window length and the
simulated correlation value itself, with shorter windows and smaller simulated
correlations leading to larger error and less reliability/stationarity. We then
obtained the relationship between MWL and correlation (Fig.3), which showed
that the highest value of MWL (=45 time points) was obtained with a correlation
around 0. Results for case-1 (Fig.4) showed
that all methods gave comparable performances when correlation was nearly
constant over time. However with case-2 (Fig.5), wherein correlation changes were faster like in real data, our MSWL method performed the best with lowest error and error
variability. This suggests that MSWL can automatically determine MLWs during
the evaluation of DFC successfully, thus giving highly reliable connectivity
estimates. These findings should be taken into consideration in the evaluation
of sliding-window based dynamic connectivity.
Acknowledgements
The authors gratefully acknowledge Dr. Martin Lindquist for sharing the code to implement the DCC methodReferences
1. Jia, et al, Brain
Connect., 4(9):741-58, 2014.
2. Hutchison, et al, NeuroImage 80:360-378,
2013. 3. Lindquist, et al,
NeuroImage 101:531-546, 2014. 4. Said,
et al, Biometrika 71:599–607, 1984.