There is wide interest in identifying EEG based biomarkers of mental health disorders.  However, there are numerous pitfalls that necessitate careful study design and a big data, multi disorder based approach.

There is increasing interest in identifying biomarkers in the EEG.  What are they and how should we go about it?

The general methodology is to look for differences in EEG metrics between a particular disease group compared to ‘healthy controls’ or correlations between some EEG metric and severity of mental health symptoms.  However, this approach is fraught with pitfalls.  Here’s why:

Mental Health diagnostic scores are wildly imperfect measures

Last week we published a post on the challenges of mental health diagnosis. Diagnosis of disorders in this domain rests in large part on highly subjective questionnaires.  ‘Disorders’ are therefore essentially groupings of symptoms based on subjective and somewhat arbitrary decisions of question inclusion and exclusion and do not directly reflect internal biological processes.  Definitionally, biomarkers are quantifiable characteristics of a biological process. Therefore simply finding a metric that correlates to an arbitrarily defined group of symptoms does not mean that it is a biomarker.  Nonethless, even if we accept that we are looking for diagnostic markers rather than biomarkers per se, we are still trying to predict biased and arbitrary groupings and scores.

The challenge of symptom-based diagnosis is easy to understand in the context of the following analogy.  Let’s say we categorized all diseases based on symptoms.  Maybe we would have one questionnaire asking about all kinds of possible pains in the head which we called ‘headache disorder’ and others that asked about muscle weakness and fatigue called ‘muscle weakness disorder/fatigue disorder’ or WDFD and another with stomach aches, diarrhea and nausea called ‘stomach spectrum disorder’. Imagine ‘stomach spectrum disorder’ or SSD consisted of questions like ‘how often have you felt stomach cramps in the last week?’ with answer choices such as Frequently, Rarely and Never.  Then there may be some other symptoms like ‘do you feel overheated or hot?’ that we wouldn’t know where to put so we include them in a variety of questionnaires, and maybe omit them in others.  Administering just one of these symptom group questionnaires could miss a whole lot of things. And it is plain to see that there is not a clear relationship between these symptoms and the underlying biological process.

Then there is the question of scoring.  Say we came up with a stomach spectrum disorder score based on some ten different questions covering different kinds of stomach complaints from feelings of nausea to cramps to sharp pain.  A stomach disorder severity score might combine scores of nausea and cramps such that the final score reflects only the total amount of stomach ailment you have rather than the specific symptoms. You can imagine that you could easily get the same score if you were vomiting all the time as if you were experiencing sharp pain but the two probably mean very different things biologically speaking.  Yet this is exactly how it works for mental health.  For instance, in the DSM depression questionnaire there are 128 combinations of symptoms that can produce the same score (see The Difficulty of Diagnosing Depression). Relating a metric to a score therefore does not even specifically target a specific constellation of symptoms but an entire range of permutations and combinations.

But let’s assume, just to keep the argument simple, that our symptom-based disorders are useful and meaningful categorizations.  Continuing the analogy of our physical symptom categories, let’s say we want to find a biomarker for MDFD our ‘muscle weakness and fatigue disorder’ defined above based on a bunch of questions about various aspects of muscle weakness and fatigue (for example: How often do you feel tired or drained of energy? How often do your legs feel heavy? How often do you drop an object when gripping it?). Now our goal is to find a surface measure – a biomarker – of MDFD.  The measure we have come up with is surface body temperature measured in the armpit. To do this we take a group of people with high MDFD and those who do not have MDFD related complaints (healthy controls) and we find a 0.65 correlation between MDFD and body temperature. We can now declare body temperature to be a biomarker of MDFD (and maybe even get FDA approval to market it as an MDFD diagnostic).  Of course, we may get a similarly high correlation between body temperature stomach spectrum disorder (SSD).  However, since we were studying MDFD we did not include SSD patients in the original study – this finding came later.  You can see the pitfalls of this approach…

This is exactly the challenge in the search for EEG biomarkers of mental health disorders.  Since studies only look at one disorder grouping in a study, it is very dangerous to interpret any arising correlations as specific to this ‘disorder’.  The Theta/Beta ratio for instance has been touted as a ‘biomarker’ for ADHD and has even received FDA approval.  However, assessment of other disorders such as Schizophrenia show similar patterns in the power spectrum relative to healthy controls and therefore are likely to have very similar theta/beta ratios.

Large intra-person variability

Another major challenge is that the characteristics of the EEG fluctuate over time within individuals in response to a variety of factors including consumption of various common stimulants such as caffeine and alcohol, sleep, state of mind at the time of the recording and so on.  This makes it difficult to obtain a reliable biometric or diagnostic marker. Indeed, correlations observed between a disorder and healthy group could well arise on account of tertiary aspects. For example, people diagnosed with depression may consume more alcohol as a group and therefore the results may reflect the effects of alcohol rather than depressive symptoms.  This calls for a wide range of tight controls (see related post Human Brains and the Control Issue)

Further, the measured EEG outcome will depend very much on the particular state they are in when they arrive for the test.  Like using body temperature as a metric for our hypothetical MDFD, you could arrive for testing when your fever has broken.

See related posts Intra-person variability in the EEG and How Consistent are you? Stimulus Quenching in the EEG.

Where to then?

Given that mental health disorders are symptom constellations at best, the search for EEG biomarkers requires a big data approach that explores symptom clusters and EEG metrics across a very large number of disorders and subjects along with clear metadata about the patients and their circumstance and state of mind at the time of recording. Of course, there is also tremendous potential to improve mental health symptom assessment with better designed question groups that are more quantitatively specific. And in the end, even with really big data with extensive metadata, success will only come from clever design of meaningful EEG metrics.