Using neuroimaging to dissect the heterogeneity of psychosis

Paola Dazzan, M.D, Ph.D.

 

Professor of Neurobiology of Psychosis, Institute of Psychiatry, Psychology and Neuroscience, King’s College London Director, Mind-Body Interface Research Centre, China Medical University Hospital

 

 

Difficulties in the prediction of outcome after the onset of psychosis are linked to the high clinical and biological heterogeneity of this disorder. Subtle and diffuse alterations in brain structure are already present at the onset of both affective and non-affective psychoses, although not in all patients. In fact, evidence mostly from single centre studies with relatively small sample sizes suggests that these alterations characterize a subgroup of patients with particularly poor outcome. To achieve a more detailed definition of the neurobiological subtypes associated with possible clinical outcomes, data from large multi-centre MRI studies should be used. These can provide the scale needed to identify the neuroimaging markers that specifically characterise treatment response and other clinical outcomes.

We used Magnetic Resonance Imaging in multiple datasets of patients scanned at their first episode of psychosis (n=410), and followed up clinically from 1 month to 6 years. We used machine learning approaches in both single and combined datasets, using measures of volume and morphology.

In the individual datasets, smaller volumes, particularly of frontal and temporal areas were predictive of illness episodes over 6 years with significant accuracy (70% correctly classified; p=0.005). Combining data from multiple centres, the accuracy in predicting long-term clinical outcome decreased to just above chance. Interestingly, accuracy in classification improved when biological (rather than clinical) heterogeneity was reduced, for example by restricting the analyses to male patients only. Furthermore, the prediction of treatment response at 1-month showed that accuracy was affected by imbalances in the number of subjects included in each outcome class, with better classification achieved when number of subjects with poor or good 1-month treatment response was almost equal.

Combining multi-center MRI data to create a well performing classification model for psychosis is possible, but each center should contribute a sample either large or homogeneous enough to first allow accurate classification within the single-center.