09 September 2010
fMRI brain age method not ready for prime time
The image above is a scatter plot showing a plot of "maturation scores" based on processing of images from a five minute rest state fMRI scan of 115 female and 123 male subjects between the ages of 7 and 30. The center line is the average, and the dotted lines mark two standard deviations from the average.
There Is A High Noise to Signal Ratio
According to the article's abstract, age accounts for about 55% of the variability in "maturation scores" assigned to the subjects based upon their fMRI scores.
These "maturation scores" are the product of a black box formula generated statistically with a host of inputs from the images. According to the Science News summary of the article, the most important components of the "maturation scores" reflect the extent to which the "prefrontal cortex is important for sophisticated cognitive control, including regulating behavior, adapting to new tasks and planning for the future" has strong connections to physically distant parts of the brain, and the extent to which the "precuneus [which] is known to be a major hub for relationships between separate brain regions" has strong connections with the rest of the brain.
The trouble is that when you look at the plot, with this sized data set, the variability at any given age largely overshadows the mean prediction. Mean ranges from about 0.6 to 1.0 within the study group, and the standard deviation is about +/- 0.25 scale score points. The researchers are touting the fMRI index as a "brain growth chart" but the spread of the numbers seems too great to apply with any great confidence to individuals.
Also troubling is my sense from the scatter plot that the subjects are spread quite evenly within the two standard deviation from the mean curves, rather than clustering strongly towards the mean as one would expect if "maturation scores" were really distributed normally in a Bell Curve distribution, as a mean and standard deviation approach assumes.
The question then, is whether the mathematical models to which the researchers try to fit the data really make sense as ways to summarize the data.
An Alternative Toy Model
Eyeballing the data, a simpler toy model, in which there is a normal adult range (with scores from about 0.6 to 1.2), and a normal childhood range (with scores from about 0.3 to 1.0) could also explain the data pretty well, particularly if we assume that girls usually become mental adults for fMRI purposes by about age nine, while boys usually become mental adults for fMRI purposes by about age fourteen.
Not everyone is on schedule. But, interpreting the outliers in terms of children who experience puberty early or late (which could be confirmed by evidence of physical maturity), certainly seems like a plausible approach to the data, and the notion that girls usually mature earlier than boys both physically and mentally, should surprise no one.
The Data From The Female Subjects
Just four girls older than nine years old in the study, all teenagers, fall below 0.6 on the index, and eight younger girls in the study fell below 0.6 on the index. No girls under nine scored above 1.0 on the index. The outliers could be explained by the theory that the four low scoring teenagers were late bloomers in terms of maturity in general.
Under the model used, six female subjects, one a teenager who is an outlier by either method, and the other five in their twenties, fall below the two standard deviation error bar, and one girl under the age of nine scored above the two standard deviation error bar.
The Data From The Male Subject
No boys in the study older than fourteen fall below 0.6 on the index, and eighteen younger boys in the study fell below 0.6 on the index. Two boys, both about thirteen years old (each of whom were also outliers in the model used) scored above 1.0 on the index. The outliers could be explained as boys who matured a couple of years early.
Under the model used, five male subjects in the study fall below the two standard deviation error bar, and four male subjects in the study (three between thirteen and fifteen years old) scored above the two standard deviation error bar. The male subjects who were high outliers and low outliers were both spread over a wide age range.
One can tease other results out of the data, but I doubt that the statistical hypothesis test could distinguishing the models shown in the chart from a simpler adult-childhood model with varying ages of maturity at a statistically significant level.
Why Care Which Model Is Used To Fit The Data?
The reason to care about which of two plausible models to apply to the scatter plot is that with sex differentiated maturity data, a simple adult/child distinction supports the kind of lines that we draw between children and adults in the law, and even provides a way to decide where those lines ought to be drawn.
Blending data from male and female subjects, in contrast, makes it seem to support a more gradualist view of what is going on in brain maturity, in which there is an extended adolescence in terms of maturity, instead of a relatively short and well defined transition.
The variability is still as significant as the prediction, however, in either model. Half of the boys in the study were well within the range observed in adult men. More than two-thirds of the girls in the study were well within the range not infrequently observed in adult women. Only about a third of the adult men were above the range observed in the pre-maturity boys. Perhaps half of the adult women were above the range observed in pre-maturity girls. Either way, the data shows only trends and isn't very useful in going from a "maturity score" to a useful prediction about an individual subject.
Also, if this study, like so many academic studies, used many college students for adult subjects, the older part of the sample may be biased because particularly mentally immature adults may be less likely to go to college (or at least, are more likely to be unavailable because they are in the prison system) and hence may be less likely to end up as subjects in this kind of study.
Of course, as in most fields of science, the best way to resolve apparently conflicting interpretation of data that can be resolved conclusively due to a lack of statistical power is to gather more and better data and spend more time analyzing it in different ways.
The study discussed is N. Dosenbach et al. Prediction of individual brain maturity using fMRI. Science. Vol. 329, September 10, 2010, p. 1358. doi: 10.1126/science.1194144.