Discussion

To our knowledge, this is the first study to investigate the prospect of exploiting multivariate neuroanatomical information to predict clinical severity of schizophrenia at the individual level. Using a classifier based on the features of cortical folding, we can identify the degree of illness severity in medicated, community-living patients with clinically stable schizophrenia. This predictive ability appears to be a unique feature of folding patterns, as the classifiers based on thickness and volume do not perform significantly above chance when separating high and low illness severity groups. Furthermore, patients with greater illness severity had reduced cortical folding in most brain regions, suggesting that a distributed defect in cortical morphology influences prognosis. Although the accuracy achieved by the gyrification-based classifier is statistically significant, the performance of this classifier is considerably weaker when compared with the multivariate neuroanatomical classifiers tested in the separation of healthy controls from patients with schizophrenia. ^{Reference Orrù, Pettersson-Yeo, Marquand, Sartori and Mechelli16} There may be several reasons for this disparity. The use of median split to divide the sample into high and low severity could have contributed to the lack of strong between-groups discriminative features With larger samples, extreme prognostic groups (lying on either end of the severity continuum) could be used for training the classifier and improve the accuracy. It is worth noting that in clinical practice, it is rarely necessary to apply a test to differentiate a patient with schizophrenia from a healthy control. The classification of a patient with schizophrenia from a healthy control can be done clinically with a high degree of confidence, thus even a high-performance neuroimaging test will have limited clinical utility in this context. On the other hand, at present there are no reliable means of predicting prognostic group membership; even a test that increases the likelihood of identifying prognostic grouping to a moderate extent, could be of significant benefit to patients and clinicians.

We quantified illness severity on the basis of a number of variables; this approach offered a multidomain metric that reflected symptom burden across the three syndromes of schizophrenia, a cognitive function that is most prominently affected in schizophrenia i.e. processing speed, social and functional performance and persistence of illness. Nevertheless, various other metrics relevant for the assessment of severity (such as Clinical Global Impression, quality of life scales, self-rated recovery measures or assessments of daily living) were not collected in this study. Furthermore, from this cross- sectional study it is not possible to extrapolate whether a gyrification-based classifier applied at illness onset could prospectively predict later severity. Nevertheless, when compared with cortical thickness and volume, gyrification has been shown to be relatively stable during adult life. ^{Reference Raznahan, Shaw, Lalonde, Stockman, Wallace and Greenstein17} In addition, a large degree of variance in the cortical folding patterns relates to neurodevelopmental integrity during the fetal or early neonatal period. ^{Reference Dubois, Benders, Borradori-Tolsa, Cachia, Lazeyras and Ha-Vinh Leuchter18} Taken together, these observations suggest that the burden of neurodevelopmental abnormalities in a patient with schizophrenia could be a potential influence on illness severity.

Our results provide preliminary evidence for the utility of cortical folding in single participant-level prognostic imaging in schizophrenia. With larger validation studies that combine high-yield clinical prognostic indicators with gyrification metrics, the predictive value can be further improved, enabling an objective grading of outcome in the management of schizophrenia. This has the promise of assisting targeted service delivery and making personalised recommendations with regard to the required duration of antipsychotic treatment. Most importantly, this approach can be refined to provide accurate information on the chances of a satisfactory clinical recovery and thus potentially empower patients by addressing the uncertainty that surrounds prognosis in psychotic disorders.