Saturday, August 18, 2012

Predicting Brain Age from 231 Neuroanatomical Measures

Is your child's brain on track to reach normal developmental milestones? A paper in Current Biology reports on a new, composite neuroanatomical metric of maturity that predicts 92% of the variance in brain age (Brown et al., 2012). Structural MRI scans were obtained from 885 healthy children and young adults ranging from 3 to 20 years of age. A set of 231 different measurements, or biomarkers, were used to determine the age that provided the best "fit" for each subject. The model made the most accurate predictions at the youngest ages, and the margin of error was 1.03 years across all ages.



Figure 3 (Brown et al., 2012). Multimodal Quantitative Anatomical Prediction of Age.
For 885 individuals, estimated brain age is plotted as a function of actual chronological age. Colors correspond to different sites and scanners. Symbol size represents subject sex (larger = female, smaller = male). A spline-fit curve (solid line) with 5% and 95% prediction intervals (dashed lines) is also shown.



The 231 measures were chosen because they are "known or suspected to change over the ages" (Brown et al., 2012):
This collection of variables was derived from T1-, T2-, and diffusion-weighted imaging and included quantitative measures of brain morphology, signal intensity, and water diffusivity within different tissue types, reflecting anatomical structural organization. Specifically, we measured cortical thickness and area, volumes of segmented subcortical structures, normalized signal intensities, and measures of diffusion magnitude and directionality within cerebral, cerebellar, and white matter fiber tract regions of interest.

The data from each of these imaging modalities alone could explain 81-83% of the variance, and that number rose to 92% when the T1-, T2-, and diffusion-weighted images were combined. The relative contribution from each type of measure changed with age, as shown below.

Figure 4 (adapted from Brown et al., 2012). Age-Varying Contributions of Different Imaging Measures to the Prediction of Age. The relative contributions of separate morphological, diffusivity, and signal intensity measures within different brain structures are plotted as a function of age. Colors correspond to measure and structure type (T1 cortical area; T1 cortical thickness; T1 subcortical volumes; diffusion (FA/ADC) within white matter tracts; diffusion (FA/ADC) within subcortical ROIs; T2 signal intensity within white matter tracts; T2 signal intensity within subcortical ROIs). Contributions are computed as units of the proportion of total explained variance.


The data were from the Pediatric Imaging, Neurocognition, and Genetics (PING) Study database (http://ping.chd.ucsd.edu), which is open access.
The primary goal of PING has been to create a data resource of highly standardized and carefully curated MRI data, whole genome SNP genotyping data, and developmental and neuropsychological assessments for a large cohort of developing children aged 3 to 20 years. The scientific aim of the project is, by openly sharing these data, to amplify the power and productivity of investigations of healthy and disordered development in children and to increase understanding of the origins of variation in neurobehavioral phenotypes.

Does it sound like the PING investigators are creating a normative database for possible diagnostic purposes in the future?
Perhaps further development of techniques to quantify the complex multidimensional nature of typical brain maturation can also help to improve the early identification of individuals with abnormal developmental trajectories. Our findings suggest that a multimodal neuroanatomical imaging assessment may hold promise for making an objective, quantitative contribution to our clinical evaluations of brain development.

Why yes it does. We already know this promise is not right around the corner (Where Are the Clinical Tests for Psychiatric Disorders?), and we know about the possible hazards of premature commercial ventures that make bold claims not supported by solid scientific evidence (The Dark Side of Diagnosis by Brain Scan). Returning to the first sentence of this post ["Is your child's brain on track to reach normal developmental milestones?"], you can see I was already anticipating franchised scanning facilities in strip malls ready to give worried parents the verdict on their child's neurodevelopment. Kind of like genetic testing outfits that make silly claims:
...unscrupulous businesses like My Gene Profile (which offers the "Inborn Talent Genetic Test" for the low low price of $1,397) have capitalized on the public's desire for simple explanations. Now you can find out whether your child has the Split Personality Gene! The Propensity for Teenage Romance Gene! The Self Detoxifying Gene!
The article in Biopolitical Times is highly recommended, especially since the mygeneprofile.com url seems to be defunct.1





To conclude with some important points about the Current Biology paper, the predictive accuracy of 92% is very impressive. The authors suggest there is a "latent brain phenotype that is tightly linked to chronological age." But they also issue a caveat about psychological maturity, which cannot be inferred from their measurements:
Brain scans, though informative about anatomical and physiological states, cannot be used to make inferences about an individual’s psychological maturity. Rather, these results speak only to the degree to which typically developing children differ among each other in their fundamental structural brain properties.


Footnote

1 Biopolitical Times also mentioned the "sprawling website" that sells the "Inborn Talent Genetic Test", which you can still view via the Wayback Machine. [NOTE: EVERYONE wants their child to be the next Tiger Woods!]


Reference

Brown, T., and 21 others. (2012). Neuroanatomical Assessment of Biological Maturity. Current Biology. DOI: 10.1016/j.cub.2012.07.002

Subscribe to Post Comments [Atom]

3 Comments:

At August 29, 2012 12:06 AM, Anonymous Anonymous said...

You were too easy on this article... You should note that the same data was used for both training and testing of the model, rendering the results a bit suspicious. I also did not buy the authors contention that they rigorously guarded against over-fitting by choosin (200+!!!) regions supported by the literature. I, quite frankly, would be shocked if this was replicated to anywhere near the same degree of accuracy in an independent sample.

 
At March 25, 2020 3:23 PM, Anonymous Anonymous said...

This comment hasn't aged well, and the reason is that it was inaccurate. Leave-one-out cross-validation inherently does not use the same observations for training and testing. So there's that. Further, eight years after the publication of this paper, several groups have independently replicated the main findings using both similar machine learning methods and anatomical brain measures and somewhat different methods, with similar coefficients of determination (COD) being found. The "brain-age" prediction concept is so well replicated, in fact, that it has been found to be useful in aging and dementia research. You can see in the original paper that the COD actually increases with increasing sample size, it doesn't decrease. This is the sine qua non test of whether any model is overfitting. If you're overfitting, the COD will never increase, and it will decrease. Read up on regularization, pre-whitening, and the use of Mahalanobis distance in multidimensional fitting. I think that will help.

 
At August 31, 2022 1:08 PM, Anonymous Anonymous said...

Thanks for correcting that misleading original comment. He clearly didn't read the paper and doesn't know enough about the cross-validation methods to weigh in on the results. Yes, leave-one-out CV never uses the same data for training and testing. Hence, the whole leave-one-out concept. SMH.

 

Post a Comment

<< Home

eXTReMe Tracker