In the future, psychiatrists may thus be able to use quantitatively analysed brain scans and other biomarkers as well as their own clinical judgement to stratify patients in terms of clinically meaningful outcomes. This would allow psychiatrists to tailor the form of treatment to the needs of each individual patient.
Almost a decade ago, around the time I graduated from medical school, I wrote the above in an editorial for the British Journal of Psychiatry, called Translating neuroimaging findings into psychiatric practice.
I argued that there were clear group differences in brain scans of people with mental disorders and healthy controls. I wrote about the potential of brain scans to be used to guide management in patients developing psychosis and the main barrier was the requirement of psychiatrists to operate outwith their comfort zone in embracing research developments.
Around the same I won a Royal College of Psychiatrists essay competition - The science of psychiatry in uncertain times: What does the future hold?
Here, I declared:
It is my opinion that clinical application of neuroimaging will be one of the most important advances in medicine.
The power of creating images of the human brain in vivo should not be underplayed. It may seem obvious now, but neuroimaging conclusively proved that psychiatric disease is brain disease. The opportunity to see inside the human brain is a boon which previous generations of psychiatrists could only dream of. It now falls to tomorrow’s generation to take the leap from theory to clinical practice.
As a graduating doctor it seemed clear that we were on the cusp of a revolution in psychiatry through the application of neuroscientific tools, like brain imaging. All it needed was a little effort to bring psychiatric diagnosis and treatment into the 21st century.
In the years that have passed there have been no real therapeutic advances in schizophrenia or related conditions. Clearly I was, at best, hopelessly naive. Today, I feel less optimistic that brain scans will ever play a role in the clinical management of disorders like schizophrenia than I did ten years ago.
So what went wrong, and why don’t we have a brain scan for schizophrenia?
Don’t we already do MRI scans in a work-up for first psychotic episodes?
In current clinical practice (in the UK) it’s relatively common to order an MRI scan for a patient experiencing a first psychotic episode. The purpose of this is usually to exclude organic cause, i.e. to make sure there isn’t a structural brain lesion causing the patient’s symptoms.
On the face of it, this seems sensible. MRI scans are relatively harmless (they don’t irradiate the brain like CT scans). Don’t we want to be absolutely sure there isn’t a treatable cause of the psychosis before starting a journey of antipsychotic medications which have significant side-effects?
Indeed, ordering an MRI, based on clinical indication (see here and here for discussion) is recommended by the American Psychiatric Association. Examples of clinical features that might push a clinician toward this includes new onset of seizures, focal neurological signs, late age of onset, fluctuating consciousness or severe headaches.
However, the yield from MRI, particularly if there are no features suggestive of an underlying secondary cause (think of a typical patient presenting in early adulthood with a long prodrome, cannabis use and positive family history of schizophrenia), is practically zero.
In one study of two samples with first episode psychosis (n=108 and n=241), there were more MRI abnormalities compared with healthy controls (12.3% vs 5.5%), but crucially these were all ‘incidental’ - meaning none required a change in clinical management.
Matt Butler and colleagues argue that ordering low-yield screening tests in psychiatry can have negative consequences: through wasting limited resources, giving false reassurance and harmful false positives. I agree that blanket screening of first episode psychosis with MRI is not justified - demonstrated in a formal cost-benefit analysis.
While there is no current role for MRI scans in typical cases of schizophrenia, developing neuroimaging into a relevant clinical tool has been a major focus of psychiatric research over the past 40 years.
The era of neuroimaging in schizophrenia research
The dawn of the modern neuroimaging era came in the 1970s. A Scottish psychiatrist Eve Johnstone published a study in the Lancet comparing CT scans of patients with chronic schizophrenia with aged-matched controls.
All of the patients (n=18) had spent at least 20 years of their lives in hospital, four of them had undergone leucotomy, half had undergone insulin coma therapy. The researchers took photographs of the CT scans and measured the lateral ventricles (fluid filled spaces in each side of brain, which can be taken as a measure of brain atrophy) using a planimeter. The ventricles of the patients were significantly bigger than the controls.
If brain abnormalities are so obvious in schizophrenia that they can be detected in a group of less than 20 patients, using unsophisticated technology, why does this not inform clinical practice? Well, as pointed out at the time, there are confounding factors that could be driving these changes, such as antipsychotic medication, institutionalisation, smoking, and deprivation. A later study demonstrated that antipsychotic medication, in monkeys, could result in brain shrinkage of ~20%.
In order to mitigate against the confounding effects that accompany a chronic mental illness, research participants with psychosis are often scanned at their first presentation, ideally before antipsychotic medication is started.
In terms of clinical application, it also isn’t particularly helpful having a brain scan to distinguish between patients with chronic schizophrenia and healthy controls. It would be much more clinically relevant to know:
Who will go on the develop psychosis in those at risk
Who will respond to antipsychotics during their first psychotic episode
What is the likely illness trajectory after the first episode
The potential of neuroimaging was enhanced with the development of scans that do not use ionising radiation, particularly those related to magnetic resonance imaging. These have made CT scans obsolete in research. MRI provides an array of tools that can investigate different aspects of the brain such as structure and functional connectivity. These have resulted in a large number of publications showing differences between groups but no applications have been clinically translated.
I assumed advances in machine learning would tie together disparate biological markers into something with clinical utility for individual patients. A lack of progress in real world medical applications (even in radiology where differences between disease and non-disease are much clearer) does not fill me with confidence though.
My first hint that translating neuroimaging into clinical practice might not be easy came in 2016, with a paper by Fusar-Poli and Meyer-Lindenberg. They highlighted that the field was characterised by small effect sizes, lack of replication and had failed to produce any clinically useful biomarkers.
One passage in particular stood out to me:
…it may be time to consider the alternative possibility that brain structure, while certainly altered in the majority of patients with psychosis, may never be—in isolation—a useful biomarker for psychosis just as high blood pressure, while certainly altered in the majority of patients with type II diabetes, is not a useful biomarker for this condition.
In other words, if structural brain alterations are too far downstream or only distantly related to causal pathology of psychotic illness, they may not index informative biology related to the clinical critical parameters of differential diagnosis and prediction of course and therapeutic stratification.
Perhaps our current investigations are just not at the right scale for psychiatric disorders, after all there can be millions of neurons in a single MRI voxel.
The end of an era?
Last month, a paper was published in Nature, claiming that reproducible brain-wide association studies require thousands of individuals (for a digestible commentary see Wiring the Brain). The study used large neuroimaging datasets (total n>50,000) to quantify effect sizes in structural and functional connectivity measured with MRI scans- examining correlations across the whole brain with common measures of psychopathology.
They took a range of subsamples from small (n=25) to large (n=30,000), and showed that small samples were associated with large spurious effect sizes. Somewhat paradoxically, as small samples require stringent statistical thresholds, only large effects become statistically significant. Aided by publication bias, high degrees of researcher freedom and chance findings, this leads to the publication of false results in small studies which do not replicate.
The authors suggest that typical effect sizes found in brain-wide association studies are small (r<0.2). Although this study used non-clinical samples and assessed for correlations, complex diseases like schizophrenia are likely to have similarly small effect sizes in brain wide associations.
Due to previous concerns about small sample sizes in neuroimaging, large consortia have been formed to pool large numbers of participants, such as ENIGMA. Comparing n>4000 patients with schizophrenia with healthy controls, they found Cohen’s d effect sizes between 0.2-0.3 for regional differences in cortical thickness.
My assumption is if we reach sample sizes of 40,000 patients, the effect size will reduce. I think effect sizes will decrease further if we compare groups that are more similar than schizophrenia versus healthy controls - such as high risk participants who later go on to develop psychosis versus those who don’t, or patients with first episode psychosis who respond to conventional antipsychotics versus those who don’t.
Small robust effects in brain structure and function when thousands of brain scans are aggregated could give clues as to the aetiology of conditions like schizophrenia, but they are unlikely to provide useful information for individual patients.
Any good news?
The good news is that lots of people in the field recognised this was a problem before I did and have taken appropriate steps (this is why we have consortia like ENIGMA).
As clearly stated in the Nature paper, this doesn’t apply to other neuroimaging studies that for example use task-based paradigms, typically having small sample sizes but running tasks many times for each individual. However, these sorts of experiments aren’t usually translated for clinical benefit - they tend to focus more on understanding how brain processes are affected by psychiatric disorders.
Meanwhile, other types of neuroimaging operate at different scales to MRI. Magnetoencephalography and electroencephalography are able to detect electrical changes at the level of milliseconds, much closer to the timescale that individual neurons operate. Positron Emission Tomography has less spatial resolution but better molecular specificity; using radioactive tracers to quantify targets like neurotransmitter receptors.
Even if it doesn’t provide immediate clinical applications, neuroimaging could eventually lead to therapeutic advances through improved understanding of the human brain and what can go wrong in mental disorders.
Conclusion
For me, this has been a lesson in being wrong. I under-estimated the complexity of schizophrenia and over-estimated how close neuroimaging was to clinical application.
To answer my own question, we don’t have a brain scan for schizophrenia because it’s a complex mental disorder with only small differences in structure and function detectable by MRI scans.
Hopefully neuroimaging along with other neuroscientific methods will lead to a better understanding in the decades to come. But it’s sobering to think that if the abnormalities associated with schizophrenia are simple not on the macroscale of neuroimaging we may never have a brain scan for schizophrenia.
On this point, I’d be happy to be proven wrong twice.
Deepak Sarpal has wonderful work from AJP .... https://ajp.psychiatryonline.org/doi/10.1176/appi.ajp.2015.14121571
I assume that you are excluding functional MRI from your discussion? As a chemistry who used multinuclear NMR to characterize individual Molecules, it would seem that it would get you a lot closer to the biochemistry- physiology of individual brain structures and neurons. Is it be ause you. Can't do multinuclear fMRI on brains of living people at sufficient resolution (geometric and chemical shift)?