‘Blood-tests’ for mental illnesses

Article originally written in July 2015

Brain scans may be used in biological tests of mental illnesses

Computer vision is looking into the brain images of unquiet minds in search for an objective assessment of mental illnesses.

Unlike other areas of medicine, psychiatry doesn’t yet have access to biological tests of mental disorders, such as depression, schizophrenia or bipolar disorder. There are no noticeable features in the brains scans of individual patients that could guide clinicians. To this day, they rely solely on the observation of behaviours and on the description of symptoms, which demand a good deal of clinical expertise.

An interdisciplinary line of research is applying computer algorithms and statistics to neuroimaging data. Scientists are paving the way for the development of software tools to support clinical psychiatrists.

Together with brain scanners, these tools could provide neurobiological tests that would highlight discriminative features in the brain of patients with mental health conditions. These tests would supplement psychiatrists’ expertise speeding up diagnosis and pointing to the treatments that are most effective on a subject by subject basis.

Subtle differences in the mind

Studies that analysed images of the brain of people with mental ill health have shown some evidence of differences in brain structure and functioning in relation to healthy control groups. However, these differences are very subtle and are found distributed across the whole brain.

Moreover, the evidence of the implicated brain regions only emerges from the combined brain scans of many patients with the same clinical profile. On the contrary, when individual subjects are analysed, their images don’t look remarkably different from brain scans of healthy people.

Peter Liddle, director of the Centre for Translational Neuroimaging in Mental Health in Nottingham, says that mental disorders, although very disabling, affect subtle things about the human mind.

“Even people with schizophrenia, most of what they say makes sense. It is only really a very small proportion of their thoughts or behaviours that is actually noticeably different, though also very disruptive,” he says.

When individual subjects are analysed, their images don’t look remarkably different from brain scans of healthy people.

This may explain why the differences in the brain are too subtle for the human eye to spot them. “You can’t actually look at the brain image of someone with schizophrenia and tell whether that image looks abnormal at all,” says Michael Brammer, head of the Brain Image Analysis Unit at the King’s College London.

“It’s very difficult to tell anything just by visual inspection. You need to do fairly subtle statistical analyses to pick up the differences.”

Although the aggregated differences in the brain scans of many patients show evidence of brain abnormalities, in the hospital or in the clinic, psychiatrists have to treat individuals, and sometimes they need to make hard guesses. This is why those studies that cover groups of people have had little impact on the clinical practice.

Around ten years ago, scientists started to realise that if they added more data about an individual, they should expect to gain a more detailed description of what was going on in his or her brain.

In neuroimaging, this meant gathering tridimensional digital images of the whole brain and finding correlations between brain regions, instead of analysing this organ just bit by bit.

A biological test that learns

Scientists have been applying methods in computational science and statistics to the wealth of data collected in the brain scanners. They have been investigating how much information is possible to infer about individuals with the help of computers.

Some of the statistical techniques and algorithms they have been employing are already popular in computer vision for the automatic classification of digital images, for example, in recognising human faces. These tools automatically extract regularities or distinguishable recurrent patterns in the image.

In the studies published in the course of the last seven years, computer programmes implementing those algorithms were presented with scans of the whole brain of many participants.

These participants were usually from two groups with clinical profiles that are often difficult to distinguish in the clinical setting. For instance, researchers used images from people who have schizophrenia and from people with severe bipolar disorder because both groups report psychotic episodes.

This process of training a computer programme to associate images to their respective clinical condition is called ‘machine learning’. After this training phase, which should have used a reasonably large sample, the programme tried to find a pattern in the images that discriminates between the clinical groups being investigated.

To test if the computer programme picked up the relevant differences and would be able to recognise them in new brain scans, the researchers presented the same programme with images not used in the training phase. For each testing image, it returned the clinical situation it believed that new image belonged to.

Probabilistic algorithms assume that there isn’t a hard boundary between traditionally defined diagnostic constructs and thus help to pick up the more ambiguous cases.

The most popular, and one of the first algorithms to be applied in the studies investigating the use of machine learning in psychiatry, aimed at finding the optimal boundary between groups of people and assigning each new, or not previously seen individual, to one of the groups.

Other algorithms give a probability of assignment to the groups. In quantifying the certainty of the prediction made by the classifier, probabilistic algorithms assume that there isn’t a hard boundary between traditionally defined diagnostic constructs and thus help to pick up the more ambiguous cases.

“We would say, for example, that somebody had an 85 per cent probability of having schizophrenia,” says Brammer. “A lot of the methods that we use recognise this dimensional continuum.

“There are also methods where you can correlate the brain imaging data with symptom scores, which would again explicitly recognise the fact that people could be ill to different extents,” he adds. Clinicians would be able to to make balanced decisions about treatment on the basis of how severe or mild the symptoms were.

Predicting the future

Most of the studies have been investigating the diagnostic value of these methods in order to prevent misdiagnoses. These can be caused by symptoms that are shared between illnesses, like between unipolar depression and bipolar disorder.

“If you could use the machine learning approach to predict diagnosis more accurately at the first consultation, then you could get patients on the right treatment sooner,” says Vince Calhoun, director of the Medical Imaging Lab, at the Mind Research Network in Albuquerque, New Mexico.

However, some researchers don’t expect that this will be the most promising application. Diagnosis is a difficult problem for most psychiatric conditions, and that may affect the training phase of the computer programme, says Andre Marquand at Donders Institute in Nijmegen, Netherlands.

“I think it’s reasonable to expect that even with all the care in the world there’s the possibility of having mislabeled training samples.”

The most likely use of machine learning techniques might be prognostic, which is something that cannot be done in any other way right now.

For example, some people might have a risk factor for a psychiatric illness, either because they have a history of mental illness in their family, or because they present early signs of a mental condition, referred to as prodromal symptoms. They might want to know what their real risk is of developing an illness later in life.

In prognostic applications, data from patients who were followed over time is analysed with the purpose of relating initial brain changes with how the illness actually evolved.

“This is the ability to see sensitively enough that there are already changes in the brain scan before somebody gets ill, and actually predict whether they were gonna get ill or not,” says Brammer. “That means you could direct the treatment to those people who might need it and not give it to people who don’t.”

The most likely use of machine learning techniques might be prognostic, which is something that cannot be done in any other way right now.

The prediction of future outcomes may influence people to make decisions not only about treatment but also about their lifestyle, adopting preventive measures in the same way as in physical health. How this knowledge is used, however, is likely to raise ethical issues.

Calhoun says that when there is some evidence of schizophrenia symptoms in children that don’t have schizophrenia yet, predicting that they will get it will have implications for them. “You might want them to initiate some treatment prior to illness, but treatments are far from perfect.”

Predicting an individual’s response to specific treatments is also a very useful application. Psychotherapy, for example, doesn’t work with everybody.

A classifier trained with images of the brain taken before the start of the therapy might be able to pick up differences in those people for whom the sessions later proved to be effective. This could guide clinical decisions about the right treatment for new cases.

“That makes a very clear case for using this kind of techniques,” says Marquand. “It could prevent the patient from having to undergo several alternatives which take a number of weeks in many cases to determine that they are effective.”

The reality out there

At the moment, these techniques are still in an experimental phase. Implemented clinical applications are unlikely to happen in the next five years, says Marquand. A number of challenges and open questions have to be addressed by future research.

The prognotsic of an individual’s response to specific treatments could prevent the patient from having to undergo several alternatives.

For instance, how useful in clinical practice is it to have a biological test that has 80 per cent confidence when it states that someone is not ill? In statistics, this percentage is called specificity. This test will fail to assess as healthy the other 20 per cent of the population who are not ill.

Let’s consider the disease in question affects about one per cent of the population, as is the case with schizophrenia. The test will misleadingly label as ill 20 per cent of the majority of the population who is not affected by that disease.

This rate is unacceptably high for what statisticians call a false positive test result. Thus, clinically useful tests need to reach a very high specificity when addressing diseases that are not common.

One way to do that is by adding more data to the mix. Each mental illness is characterised by many symptoms and scientists believe that each of them is linked to a different neural process.

Any neuroimaging modality gives only a partial view onto these processes. “Imaging technologies are good, powerful, but no way near to tell us everything about the brain,” says Calhoun, “so we really need to leverage them combining multiple kinds of data.

“How much data we need to learn about these complex mental disorders is a big open question,” he adds. Another open question is how well these techniques perform when presented with brain scans of people who may have more than one disorder.

“People are complicated in all sorts of ways,” says Liddle, “ranging from other pathological things in their life, different social circumstances that may change their mind and influence what we see in their brain.”

How well do these techniques perform when presented with brain scans of people who may have more than one disorder?

Finally, researchers still don’t know how generalisable the reported results are. “The reported accuracy was determined on the same scanning site within the same sample,” says Marquand. “How well is that generalised across different scanner manufacturers, across different study centres and so on, needs to be validated.”

Thus, these machine learning techniques will have to be translated from research settings to real clinical populations where the prevalence of the diseases varies widely.

One way forward will be enlarging the training data set, for example, combining data that was collected independently in different centres.

“There is starting to be more consortia forming, more open data, more sharing,” says Calhoun, “and that needs to continue so we can get numbers that actually give us something meaningful in the data.”

Brammer says that this kind of large scale validation requires thousands of people, funding and a great deal of organisation between the multiple imaging centres involved.

And it is a very multidisciplinary procedure, says Marquand. “We need people who know the algorithms well, we need access to good clinical samples, for which our clinical network is required, and we need good neuroimaging data acquisition procedures, which involve physical expertise.”

Despite the still high cost of doing brain scans, Brammer believes clinical applications of these methods will be cost-effective. We just have to think about the cost of keeping in hospital someone who is resistant to existing treatments, not to mention the human suffering.

This might become a rare scenario once psychiatrists have access to tools that will help them to arrive to better decisions and to better outcomes to their patients.

Have your say