Does neuroimaging have a role in predicting outcomes in psychosis?

A key difficulty in the management of psychotic disorders is that clinical outcomes are difficult to predict on the basis of the patient's clinical features. As a result, patients with psychosis are generally treated in a similar way, even though there may be marked differences in their course of illness or response to medication. However, recent research using neuroimaging suggests that, within a sample of patients with psychosis, the pattern of abnormalities may vary in relation to different clinical outcomes. This raises the possibility that neuroimaging could be used to stratify patients according to clinical outcome; subgroups of patients could then be offered different forms of treatment.

Data from a number of structural magnetic resonance imaging (MRI) studies suggest that patients with relatively poor outcomes have, compared to those with good outcomes, more marked reductions in total and regional grey matter volume, and greater ventricular enlargement1. However, other studies have not found a relationship between alterations in brain structure and clinical outcomes2. This inconsistency may reflect the use of patient samples that were small, and heterogeneous for age, stage of illness, and pharmacological treatment, all of which can affect neuroimaging findings. Moreover, clinical outcomes have often been determined retrospectively, on the basis of clinical records.

Recent neurochemical imaging studies have suggested that the response to antipsychotic medication in patients with psychosis is related to both subcortical dopamine function, as measured using positron emission tomography, and regional brain glutamate levels, as assessed using magnetic resonance spectroscopy. A good therapeutic response has been associated with elevated dopamine function and relatively normal glutamate levels, whereas a poor response has been linked to normal dopamine function and elevated glutamate levels3. Independent work has also linked the response to antipsychotic medication to differences in cortical gyrification4, and to diffusion tensor imaging measures of white matter integrity5. However, again, these studies involved relatively small samples, and the patients were scanned after they had been treated with antipsychotic medication: it is thus unclear whether the neuroimaging findings predated treatment or were secondary to it.

Most studies to date have related clinical outcomes to a single cross‐sectional neuroimaging measure. Serial neuroimaging measurements provide data on how the brain changes over time within the same patient, and recent studies involving longitudinal scanning of patients suggest that measuring the progression of findings facilitates the prediction of outcome6. For example, longitudinal data from patients with first episode psychosis and from those with childhood‐onset schizophrenia suggest that reductions in hippocampal volume over the first few years of illness are associated with poorer functioning at follow‐up7.

All of the studies mentioned above reported differences between groups of patients. However, in order for neuroimaging to be useful in a clinical setting, it must be able to facilitate outcome prediction using data from an individual patient. Multivariate statistical approaches such as machine learning provide a means of addressing this issue. For example, application of machine learning analyses to MRI data from patients with first episode psychosis showed that baseline neuroimaging data could predict a non‐remitting course of illness over the subsequent six years with an accuracy of 72%8.

Ongoing studies in this field are seeking to address the methodological issues that may have limited earlier work. Sample sizes can be increased through the involvement of multiple research sites. Although multi‐centre studies are logistically challenging, and there are significant confounding factors associated with acquiring data on a variety of different scanners, these disadvantages are probably outweighed by the increased statistical power that results from having much larger samples. Similarly, serial neuroimaging studies are more difficult to carry out than those involving a single scan, but may provide more predictive power. Ongoing studies have also sought to enroll samples that are homogeneous with respect to stage of illness and previous treatment, and that are treated in a standardized way subsequent to scanning. A good example of this is OPTiMiSE (Optimization of Treatment and Management of Schizophrenia in Europe), a large multicenter study funded by the European Commission1. This involves a neuroimaging assessment of a large multi‐centre sample of medication‐naïve or minimally treated first episode patients, all of whom are then treated with amisulpride following a standardized protocol. Their clinical outcomes are evaluated prospectively.

Future studies may also benefit from using more than one modality of neuroimaging; there is some evidence that this may improve prediction of outcomes9, although other data do not support this10. Similarly, integrating neuroimaging data with non‐imaging measures that have independently been linked with altered outcomes in psychosis, such as polygenic risk score, substance use, inflammatory markers and central nervous system autoantibodies, may enhance predictive power. However, although this may be a reasonable expectation, it has yet to be tested.

Even if a neuroimaging measure is established as a robust statistical predictor of clinical outcomes, this does not necessarily mean that it can be translated into mainstream clinical practice. Financial and practical considerations will apply, such as the cost of scanning and the availability of the scanner. The development of tools that can be used in a clinical setting is likely to require neuroimaging measures that can be acquired without the need for highly specialized training or equipment. Some ongoing studies are explicitly focused on the development of such tools for psychosis (see, for instance, www.psyscan.eu).

Given that psychotic disorders are pathophysiologically heterogeneous, it is reasonable to expect that neuroimaging techniques which can identify pathophysiological differences within patient samples may be useful in predicting clinical outcomes. However, at present, it is unclear which particular neuroimaging measures will be the most useful, and whether combining these with non‐imaging biomarkers will enhance their ability to facilitate prediction of outcomes in psychosis.

Philip McGuire, Paola Dazzan
Department of Psychosis Studies, Institute of Psychiatry, Psychology and Neuroscience, King's College London; National Institute for Health Research (NIHR) Mental Health Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King's College London, London, UK