The development of human functional brain networks
Jonathan Power, Damien Fair, Brad Schlaggar, Steve Petersen
Neuron. 2010 Sep 9; 67(7):735-48.
This paper describes the field I intended to study as a graduate student: human development in resting state fMRI. In 2010 there were relatively few papers on this topic and many were from our group, the papers by Damien Fair primarily. The story being found, across groups, was that school-age children seemed to have strong signal correlations between nearby regions, but that these strong short-distance correlations weakened as a more distributed, "adult" pattern emerged in the second and third decades of life. The effect sizes were large and compelling. Fleshing this story out with high subject numbers, whole-brain coverage, and attention to physiologic changes that accompany development was going to be my thesis.
I never got to do that thesis. About one year after this article was published a series of studies reported that small motions caused spurious patterns in correlations, and that many techniques for fMRI image processing didn't remove these spurious patterns. Some of the previous developmental results were frank artifact due to children moving more than adults. How much of the effects were artifact (some/most/all) remains an open question. It's difficult to convince people one way or another since the methods for detecting and removing motion (and other) artifact are still under development and debate.
Within the last few years, after spending much time trying to control motion and other artifact, I found that in my most pristine developmental data, I had no more resting state effects to speak of when contrasting 7-9 year old against 21-23 year old subjects. No distance dependent effects, no obvious story overall. But my subject numbers were reduced by the quality standards I had imposed. I was now studying about 30 children, a number which could be underpowered (Ted Satterthwaite recently suggested that hundreds of subjects are needed to find developmental effects; if this is true then resting state developmental effects are remarkably subtle). I can't say much with confidence about my data in terms of telling a developmental story (other than that classifiers still work well in separating children from adults). But what is clear is that in a single dataset I used to have compelling effects when the data were processed one way, but now I have marginal and unclear results when the data are processed another way. And the second way better controls artifact than the first way.
Many papers have emerged on development of resting state signal correlations in the past 5 years. To me, the various findings in these papers have to be viewed first through the lens of data processing and artifact control. It is easy to produce a "developmental effect", such as distance dependence, by letting artifact remain in the data. This review, written in 2010 before the heightened attention to issues of artifact in resting state studies, closes with these lines: "Unfortunately, over childhood and adolescence, changes in functional connectivity will likely correlate with many behavioral or physiologic measures, such as height or hormonal levels, as well as various cognitive measures. Developmental studies of correlates between rs-fcMRI measures and cognitive or behavioral measures will therefore need tightly constrained hypotheses, or large datasets that can achieve the power needed to tease apart such covariance." This statement is the only mention of confounding factors and the word artifact is not even in the paper. Times have changed.