Biomarkers to screen for autism (again)

Diagnosis of autism from biomarkers is a holy grail for biomedical researchers. The days when it was thought we would find “the autism gene” are long gone, and it’s clear that both the biology and the psychology of autism is highly complex and heterogeneous. One approach is to search for individual genes where mutations are more likely in those with autism. Another is to address the complexity head-on by looking for combinations of biomarkers that could predict who has autism.  The latter approach is adopted in a paper by Bao et al (2022) who claimed that an ensemble of gene expression measures taken from blood samples could accurately predict which toddlers were autistic (ASD) and which were typically-developing (TD). An anonymous commenter on PubPeer queried whether the method was as robust as the authors claimed, arguing that there was evidence for “overfitting”. I was asked for my thoughts by a journalist, and they were complicated enough to merit a blogpost.  The bottom line is that there are reasons to be cautious about the conclusion of the authors that they have developed “an innovative and accurate ASD gene expression classifier”.

 

Some of the points I raise here applied to a previous biomarker study that I blogged about in 2019. These are general issues about the mismatch between what is done in typical studies in this area and what is needed for a clinically useful screening test.

 

Base rates

Consider first how a screening test might be used. One possibility is that there might be a move towards universal screening, allowing early diagnosis that might help ensure intervention starts young.  But for effective screening in that context, you need extremely high diagnostic accuracy, and accuracy depends on the frequency of autism in the population.  I discussed this back in 2010. The levels of accurate classification reported by Bao et al would be of no use for population screening because there would be an extremely high rate of false positives, given that most children don’t have autism.

 

Diagnostic specificity

But, you may say, we aren’t talking about universal screening.  The test might be particularly useful for those who either (a) already have an older child with autism, or (b) are concerned about their child’s development.  Here the probability of a positive autism diagnosis is higher than in the general population.  However, if that’s what we are interested in, then we need a different comparison group – not typically-developing toddlers, but unaffected siblings of children with autism, and/or children with other neurodevelopmental disorders.   

When I had a look at the code that the authors deposited for data analysis, it implied that they did have data on children with more general developmental delays, and sibs of those with autism, but they are not reported in this paper. 

 

The analyses done by the researchers are extremely complex and time-consuming, and it is understandable that they may prefer to start out with the clearest case of comparing autism with typically-developing children. But the acid test of the suitability of the classifier for clinical use would be a demonstration that it could distinguish children with autism from unaffected siblings, and from nonautistic children with intellectual disability.

 

Reliability of measures

If you run a diagnostic test, an obvious question is whether you’d get the same result on a second test run.  With biological and psychological measures the answer is almost always no, but the key issue for a screener is just how much change there is. Gene expression levels could vary from occasion to occasion depending on time of day or what you’d eaten – I have no idea how important this might be, but it's not possible to evaluate in this paper, where measures come from a single blood sample. My personal view is that the whole field of biomedical research needs to wake up to the importance of reliability of measurement so that researchers don’t waste time exploring the predictive power of measures that may be too unreliable to be useful.  Information about stability of measures over time is a basic requirement for any diagnostic measure.

 

A related issue concerns comparability of procedures for autism and TD groups. Were blood samples collected by the same clinicians over the same period and processed in the same lab for these two groups? Were the blood analyses automated and/or done blind? It’s crucial to be confident that minor differences in clinical or lab procedures do not bias results in this kind of study.

 

Overfitting

Overfitting is really just a polite way of saying that the data may be noise. If you run enough analyses, something is bound to look significant, just by chance.  In the first step of the analysis, the researchers ran 42,840 models on “training” data from 93 autistic and 82 TD children and found 1,822 of them performed better than .80 on a measure that reflects diagnostic accuracy (AUC-ROC – which roughly corresponds to proportion correctly classified: .50 is chance, and 1.00 is perfect classification).  So we can see that just over 4% of the models (1822/42840) performed this well.

 

The researchers were aware of the possibility of overfitting, and they addressed it head-on, saying: “To test this, we permuted the sample labels (i.e., ASD and TD) for all subjects in our Training set and ran the pipeline to test all feature engineering and classification methods. Importantly, we tested all 42,840 candidate models and found the median AUC-ROC score was 0.5101 with the 95th CI (0.42–0.65) on the randomized samples. As expected, only rare chance instances of good 'classification' occurred.”  The distribution of scores is shown in Figure 2b. 

 

 

Figure 2b from Bao et al (2022)

 

They then ran a further analysis on a “test set” of 34 autistic and 31 TD children who had been held out of the original analysis, and found that 742 of the 1822 models performed better than .80 in classification. That’s 40% of the tested models.  Assuming I have understood the methods correctly, that does look meaningful and hard to explain just in terms of statistical noise.  In effect, they have run a replication study and found that a substantial subset of the identified models do continue to separate autism and TD groups when new children are considered. The claim is that there is substantial overlap in the models that fall in the right-hand area under the curve for the red and pink distributions.

 

The PubPeer commenter seems concerned that results look too good to be true. In particular, Figure 2b suggests the models perform a bit better in the test set than in the training set. But the figure shows the distribution of scores for all the models (not just the selected models) and, given the small sample sizes, the differences between distributions does not seem large to me. I was more surprised by the relatively tight distribution of AUC-ROC values obtained in the permutation analysis, as I would have anticipated some models would have given high classification accuracy just by chance in a sample of this size.

The researchers went on to present data for the set of models that achieved .8 classification in both training and test sets. This seemed a reasonable approach to me. The PubPeer commenter is correct in arguing that there will be some bias caused by selecting models this way, and that one would expect  less good performance in a completely new sample, but the 2-stage selection of models would seem to ensure there is not "massive overfitting". I think there would be a problem if only 4% of the 1822 selected models had given accurate classification, but the good rate of agreement between the models selected in the training and test samples, coupled with the lack of good models in the permuted data, suggests there is a genuine effect here. 

 

Conclusion

So, in sum, I think that the results can’t just be attributed to overfitting, but I nevertheless have reservations about whether they would be useful for screening for autism.  And one of the first things I’d check if I were the researchers would be the reliability of the diagnostic classification in repeated blood samples taken on different occasions, as that would need to be high for the test to be of clinical use.

 

Note: I'd welcome comments or corrections on this post. Please note, comments are moderated to avoid spam, and so may not appear immediately. If you post a comment and it has not appeared in 24 hr, please email me and I'll ensure it gets posted. 

 PS. See comment from original PubPeer poster attached. 

Also, 8th Dec 2022, I added a further PubPeer comment asking authors to comment on Figure 2B, which does seem odd. 

https://pubpeer.com/publications/B693366B2B51D143C713359F151F7B#4  

PPS. 10th Dec 2022. 

Author Eric Courchesne has responded to several of the points made in this blogpost on Pubpeer: https://pubpeer.com/publications/B693366B2B51D143C713359F151F7B#5  

 

 

 

 

 

 

 

 

 

Which neuroimaging measures are useful for individual differences research?

The tl;dr version

A neuroimaging measure is potentially useful for individual differences research if variation between people is substantially greater than variation within the same person tested on different occasions. This means that we need to know about the reliability of our measures, before launching into studies of individual differences.
High reliability is not sufficient to ensure a good measure, but it is necessary.

Individual differences research

Psychologists have used behavioural measures to study individual differences - in cognition and personality - for many years. The goal is complementary to psychological research that looks for universal principles that guide human behaviour: e.g. factors affecting learning or emotional reactions. Individual differences research also often focuses on underlying causes, looking for associations with genetic, experiential and/or neurobiological differences that could lead to individual differences.

Some basic psychometrics

Suppose I set up a study to assess individual differences in children’s vocabulary. I decide to look at three measures.

• Measure A involves asking children to define a predetermined set of words, ordered in difficulty, and scoring their responses by standard criteria.
• Measure B involves showing the child pictured objects that have to be named.
• Measure C involves recording the child talking with another child and measuring how many different words they use.

For each of these measures, we’d expect to see a distribution of scores, so we could potentially rank order children on their vocabulary ability. But are the three measures equally good indicators of individual differences?

We can see immediately one problem with Test B: the distribution of scores is bunched tightly, so it doesn’t capture individual variation very well. Test C, which has the greatest spread of scores, might seem the most suitable for detecting individual variation. But spread of scores, while important, is not the only test attribute to consider. We also need to consider whether the measure assesses a stable individual difference, or whether it is influenced by random or systematic factors that are not part of what we want to measure.

There is a huge literature addressing this issue, starting with Francis Galton in the 19th century, with major statistical advances in the 1950s and 1960s (see review by Wasserman & Bracken, 2003). The classical view treats test scores as a compound, with a ‘true score’ part, plus an ‘error’ part. We want a measure that minimises the impact of random or systematic error.

If there is a big influence of random error, then the test score is likely to change from one occasion to the next. Suppose we measure the same children on two occasions a month apart on three new three tests, and then plot scores on time 1 vs time 2. (To simplify this example, we assume that all three tests have the same normal distribution of scores - the same as for test A in Figure 1, and there is an average gain of 10 points from time 1 to time 2).

Figure 2

We can see that Test F is not very reliable: although there is a significant association between the scores on two test occasions, individual children can show remarkable changes from time to time. If our goal is to measure a reasonably stable attribute of the person, then Test F is clearly not suitable. aov
Just because a test is reliable, it does not mean it is valid. But if it is not reliable, then it won’t be valid. This is illustrated by this nice figure from https://explorable.com/research-methodology:
.

What about change scores?

Sometimes we explicitly want to measure change: for instance, we may be more interested in how quickly a child learns vocabulary, rather than how much they know at some specific point in time. Surely, then, we don’t want a stable measure, as it would not identify the change? Wouldn’t test F be better than D or E for this purpose?

Unfortunately, the logic here is flawed. It’s certainly possible that people may vary in how much they change from time to time, but if our interest is in change, then what we want is a reliable measure of change. There has been considerable debate in the psychological literature as to how best to establish the reliability of a change measure, but the key point is that you can find substantial change in test scores that is meaningless, and that the likelihood of it being meaningless is substantial if the underlying measure is unreliable. The data in Figure 2 were simulated by assuming that all children changed by the same amount from Time 1 to Time 2, but that tests varied in how much random error was incorporated in the test score. If you want to interpret a change score as meaningful, then the onus is on you to convince others that you are not just measuring random error.

What does this have to do with neuroimaging?

My concern with the neuroimaging literature, is that measures from functional or structural imaging are often used to measure individual differences, but it is rare to find any mention of reliability of those measures. In most cases, we simply don’t have any data on repeated testing using the same measures - or if we do, the sample size is too small, or too selected, to give a meaningful estimate of reliability. Such data as we have don’t inspire confidence that brain measurements achieve high level of reliability that is aimed for in psychometric tests. This does not mean that these measures are not useful, but it does make them unsuited for the study of individual differences.

I hesitated about blogging on this topic, because nothing I am saying here is new: the importance of reliability has been established in the literature on measurement theory since 1950. Yet, when different subject areas evolve independently, it seems that methodological practices that are seen as crucial in one discipline can be overlooked in another that is rediscovering the same issues but with different metrics.

There are signs that things are changing, and we are seeing a welcome trend for neuroscientists to start taking reliability seriously. I started thinking about blogging on this topic just a couple of weeks ago after seeing some high-profile papers that exemplified the problems in this area, but in that period, there have also been some nice studies that are starting to provide information on reliability of neuroscience measures. This might seem like relatively dull science to many, but to my mind it is a key step towards incorporating neuroscience in the study of individual differences. As I commented on Twitter recently, my view is that anyone who wants to using a neuroimaging measure as an endophenotype should first be required to establish that it has adequate reliability for that purpose.

Further reading

This review by Dubois and Adolphs (2016) covers the issue of reliability and much more, and is highly recommended.
Other recent papers of relevance:
Geerligs, L., Tsvetanov, K. A., Cam-CAN, Henson, R. N. 2017 Challenges in measuring individual differences in functional connectivity using fMRI: The case of healthy aging. Human Brain Mapping
Nord, C. L., Gray, A., Charpentier, C. J., Robinson, O. J., Roiser, J. P. 2017 Unreliability of putative fMRI biomarkers during emotional face processing.Neuroimage.

Note: Post updated on 17th June 2017 because figures from R Markdown html were not displaying correctly on all platforms.