These authors use machine learning techniques to build sparse predictors based on grey/white matter volumes of specific regions. Correlations obtained are ~ 0.7 (see figure).
I predict that genomic estimators of this kind will be available once ~ 1 million genomes and cognitive scores are available for analysis. See also Myths, Sisyphus and g.
MRI-Based Intelligence Quotient (IQ) Estimation with Sparse Learning (PLOS)Training and testing of models was performed as described below. They had only 164 individuals in their sample, so IIUC the average correlation is computed on test samples of ~16 individuals. It would be good to see their predictors tested on larger data sets. I wonder how stable the predictor variables (feature coefficients) were across partitions.
In this paper, we propose a novel framework for IQ estimation using Magnetic Resonance Imaging (MRI) data. In particular, we devise a new feature selection method based on an extended dirty model for jointly considering both element-wise sparsity and group-wise sparsity. Meanwhile, due to the absence of large dataset with consistent scanning protocols for the IQ estimation, we integrate multiple datasets scanned from different sites with different scanning parameters and protocols. In this way, there is large variability in these different datasets. To address this issue, we design a two-step procedure for 1) first identifying the possible scanning site for each testing subject and 2) then estimating the testing subject’s IQ by using a specific estimator designed for that scanning site. We perform two experiments to test the performance of our method by using the MRI data collected from 164 typically developing children between 6 and 15 years old. In the first experiment, we use a multi-kernel Support Vector Regression (SVR) for estimating IQ values, and obtain an average correlation coefficient of 0.718 and also an average root mean square error of 8.695 between the true IQs and the estimated ones. In the second experiment, we use a single-kernel SVR for IQ estimation, and achieve an average correlation coefficient of 0.684 and an average root mean square error of 9.166. All these results show the effectiveness of using imaging data for IQ prediction, which is rarely done in the field according to our knowledge.
We performed experiments with 10-fold cross-validations. Specifically, we randomly partitioned each dataset into 10 subsets with no replacement, and used 9 out of the 10 subsets for training and the remaining one for testing. To further avoid a possible bias during partitioning, we repeated the experiments 10 times.Some background from the paper. Strangely, they don't cite the Thompson lab (UCLA) results on brain size and intelligence (21k individuals). IIRC from their results, brain size alone correlates 0.4 with IQ.
... Uncovering human intelligence has always been of major interest in cognitive neuroscience. With the advent of brain imaging, there have been efforts to investigate the relation between brain anatomy and intelligence [3,4], and substantial understanding has been achieved in the field. For example, Supekar et al. showed that the size and circuitry of certain parts of children’s brains could be a potential predictor for how well they would respond to intensive math tutoring [5]. Chen et al. [6] demonstrated that the volumetric analysis of gray matter (GM) from structural Magnetic Resonance Imaging (MRI) could be used to predict a subsequent decline in IQ in children with sickle cell disease. McDaniel et al. [3] found that the volume of the brain is positively correlated with IQ according to MRI-based experiments. Frangou et al. [7] reported positive correlations between IQ score and GM density of the orbitofrontal cortex, cingulate gyrus, cerebellum, and thalamus, but negative correlation between IQ score and the caudate nucleus. On the other hand, Navas-Sanchez et al. [8] investigated the relationship between IQ score and microstructure of white matter (WM) tracts using diffusion tensor imaging (DTI), and found that IQ score is positively correlated with fractional anisotropy (FA). Kim et al. [9] found that lower performance in verbal IQ score is correlated with the decrease of FA values. In another DTI-based study, Welcome et al. [10] discovered that the volume of WM fiber tracts is correlated with nonverbal IQ score. Inspired by these strong correlations between brain anatomy and IQ score, we propose, in this study, a novel framework to estimate IQ by using GM and WM features extracted from structural MRI. ...Their results might give some indication as to which regions of the brain are responsible for most of the population variation in IQ. Below are the brain regions most commonly identified as "features" by sparse learning methods.
From the comments (55% of variance means a correlation just larger than 0.7):
There are lots of recent studies that have tried to estimate IQ from MRI or EEG readings (sometimes called "neurometric" IQ); many of the teams are based in South Korea and Malaysia. The Malaysian group, based at the MARA University of Technology, has published about a dozen papers over the past two years, involving hundreds of subjects. They can now use EEG readings to sort subjects into one of seven IQ ranges (e.g. 90-100, 120-130) with 83% accuracy; this figure jumps to 98% when subjects are sorted into one of three IQ ranges (low, medium, or high). The South Korean researchers, at Seoul National University, have been combining MRI and fMRI scans to predict IQ scores, and in late 2012 they were granted a patent for their "neurobiological method for measuring human intelligence," which can explain up to 55% of the variance between individual IQ scores. An example (from Dec 2013) is at http://ieeexplore.ieee.org/...
Additional links:
http://dx.doi.org/10.1016/j.cmpb.2014.01.016
http://www.google.com/patents/WO2008018763A1?cl=en
25 comments:
There are lots of recent studies that have tried to estimate IQ from MRI or EEG readings (sometimes called "neurometric" IQ); many of the teams are based in South Korea and Malaysia. The Malaysian group, based at the MARA University of Technology, has
published about a dozen papers over the past two years, involving
hundreds of subjects. They can now use EEG readings to sort subjects
into one of seven IQ ranges (e.g. 90-100, 120-130) with 83% accuracy;
this figure jumps to 98% when subjects are sorted into one of three IQ
ranges (low, medium, or high). The South Korean researchers, at Seoul
National University, have been combining MRI and fMRI scans to predict
IQ scores, and in late 2012 they were granted a patent for their
"neurobiological method for measuring human intelligence," which can
explain up to 55% of the variance between individual IQ scores. An example (from Dec 2013) is at http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6735132&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D6735132
I see the three group EEG work at http://dx.doi.org/10.1016/j.cmpb.2014.01.016
Can you point me to the seven group work? In particular I am curious about the error pattern (do they give a classification matrix?). If they achieve 83% accuracy with most of the errors being off by one group that would be extremely impressive.
This looks like a link to that patent: http://www.google.com/patents/WO2008018763A1?cl=en
Is race a confounding variable in these studies?
One thing that seemed odd to me is that the paper discussed adding age to the model but said nothing about height (which I think would make more sense both in a developmental sense and because it might capture gender variation and/or general body size variation). It also seems strange that they did the research over a broad age range of children. Wouldn't adults (say 20 to 40?) give much more comparable numbers with less developmental variation?
Some remarkable results in various biological fields have been claimed in Korea which turned out to be the product of fraud.
The "seven" paper is called "IQ Classification Using EEG Spectrogram via GLCM Approach", from 2012, but I can't find it on Google scholar. It was presented at a conference called ICCIT 2012.
Explaining IQ with MRI appears to me a bit like explaining the horsepower of a car by looking at the motor from 1 m distance. Of course it works somehow as one can maybe guess the engine displacement but to really determine the horsepower one would have to disassemble the motor. In the analogy this would mean to look at the brain in microscopic dimensions.
Yes, it's a crude method. But, to rebut far left IQ skeptics it is important to show that these "socially constructed" test results are correlated to (indeed, can be predicted by) measurements of physical quantities such as brain region volumes. The Thompson lab results indicate that brain structure itself is heritable.
Thanks. I found the abstract on page 63 of http://faculty.mu.edu.sa/public/uploads/1331936104.7554Abstract_Book_V2_b.pdf
but I was unable to find the paper.
It was not on these pages for the lead author: http://www.researchgate.net/profile/Mahfuzah_Mustafa/publications
http://umpir.ump.edu.my/view/creators/Mahfuzah=3AMustafa=3A=3A.type.html
If I had time I would look into this more seriously. Thompson's collaboration has ~21k individuals so if these ML techniques work at all they should be able to build a pretty good predictor.
There is a problem with using IQ as a measure of intelligence in children age 6 - 15 (or any children): IQ is not a measure of intelligence, but of rarity of intelligence compared to a reference group of the same age. The authors of the study you discuss were of course limited by the prior MRI studies whose data they used only included IQ and age rather than Rasch measures that are a direct, equal-interval scale measure of intelligence. Even so, the results are remarkable - 0.7 correlation makes MRI potentially an adequate intelligence test.
A Rasch measure, the "change-sensitive score" or CSS is provided on the Woodcock-Johnson (WJ) or Stanford Binet (SB) (at least for the past decade or so). Getting good information on the distribution of CSS vs. age is difficult, though I recently found a presentation on constructing the WJ which gave an age vs. CSS plot with not only an average line but also +/- 1,2, & 3 standard deviations. The average line is a log-curve as expected, but there is a lot of variation in developmental rates at the top and bottom.
The interaction of age and s.d. (which is ~ IQ on the subtest) gives some shocking conclusions - an adult at (-1.8 s.d.) - not retarded, a score higher than at least 1 in 6 American Blacks - has the same absolute ability score on the subtest as a (+1 s.d.) 5 year-old.
A (+3 s.d.) 5 year old performs as well on this subtest as a (+1 s.d.) 18 year-old.
Another shocker (though it may be due to statistical noise) is that there is no increase in scores for the +3 s.d. line between ages 5 and 8, while all the other groups make substantial gains. If they gained the same number of CSS points as the average line, they would be over (+5.5 s.d.) compared to the current 8.33 year-old population - 1 in 64 million instead of 1 in 740. Hmm.... what happens at age 5? That kind of developmental depression would be considered evidence of severe abuse in other contexts. (Keep in mind that this is only on the block rotation subtest, the charts for other subtests and the test as a whole has not been released, perhaps due to Riverside Publishing's management's instinct to never release more data than they absolutely have to, but perhaps due to the policy implications of the data.)
Ref.: Kevin McGrew: "Art and Science of Applied Test
Development" / Part F--Psychometric/technical statistical analysis:
Internal", slide 19:
I wonder if anyone has tried incorporating reaction time measurements into these MRI predictors? It seems to me that size (assuming it correlates highly with interconnection density as well) and speed would be a good starting point for a predictor. Page 474 of Thompson's 2004 article mentions inspection time briefly and reference 61 looks interesting: http://www.sciencedirect.com/science/article/pii/S0160289601000770
These MRI studies were financed by the personal assistant recruitment firms referenced in Steve's previous post.
Is this a topic that East Asian-based teams are more likely than U.S.-based teams to be working on in 2015?
I have a PDF of the "IQ Classification Using EEG Spectrogram via GLCM Approach" paper if anyone's interested.
Sometimes it rains in Las Vegas but that doesn't mean that Las Vegas is a rainy city.
Many studies have also shown that brain volume (and other measurable physical quantities) can explain some of the variance in individual cognitive abilities in non-human primates, mice, rats, pigs, birds, etc. (example at http://rstb.royalsocietypublishing.org/content/366/1567/1017)
I am interested. Would it be possible to email it to my first initial last name at sonic.net ?
Thank you!
The dataset I am working on has the same problem. No data for body size and removing the effect of age is tricky as it has non-linear relationships to cognitive abilities and brain structures.
Really? Can you email me those studies?
There are quite a few. Rats, from back in 1935: "The association between brain size and maze ability in the white rat" showed that rats that were selected over generations to be either "maze-bright" or "maze-dull" differed by about 2.5 SD in brain weight. For birds, from 2010: "Brain size, head size and behaviour of a passerine bird"; head volume significantly predicted individual differences in "recapture probability", "suggesting that head volume is related to learning ability". From 2011, "Mapping behavioural evolution onto brain evolution: the strategic roles of conserved organization in individuals and species" looks at the correlates of individual brain-size differences in pigs, minks, and mice. Etc.
In the credit default models I use, we model nonlinear age effects using cubic splines. Saves degrees of freedom over using dummies for each age.
I struggled with modeling nonlinear effects while doing some analysis of NHANES data. Splines worked well for capturing fine variation and giving a nice graphical display. I found that the degrees of freedom needed were justified for age (there is often a large variation by age), but not so much for the other variables I was using. Here is a brief writeup I did then (extracted from a larger document just now so missing some context) http://rpubs.com/rseiter/72806 . I would be interested in any feedback. Did you try bucketed age categories? That's the usual approach in the medical literature. (As may be apparent by my comments elsewhere in Steve's blog I am a bit obsessed with nonlinear effects. This is because I think failing to account, or even test, for them is a serious omission in much of the medical literature. For example, cholesterol mortality is a U shaped curve but is seldom modeled that way.)
Thanks for the pointers (and the other emailed paper!). That 2011 paper is cited by a fair number of more recent papers: http://www.ncbi.nlm.nih.gov/pubmed?linkname=pubmed_pubmed_citedin&from_uid=21690129
Interesting...
Yes. I did not have much time recently to finish that project up.
Post a Comment