To our knowledge, this is the first study to investigate whether spatially distributed information in brain tissue data, obtained at the first presentation of psychosis, can potentially be used to predict subsequent illness course type at the individual level. The study provides proof of concept that MRI can be used as a tool for the prediction of illness course. Our findings suggest that at the time of first presentation, patients who will subsequently develop a continuous illness course may be already distinguished, above chance level, from those who will develop an episodic course, using a classifier trained only on patient groups. By contrast, patients with a future episodic course cannot be distinguished from healthy individuals. Finally, our findings provide preliminary evidence that, even for patients with a ‘less extreme’ (or intermediate) type of outcome, it is possible to predict, to a certain extent, who will develop further episodes and who will be less likely to do so.
Although significant, the accuracy with which we have correctly classified patients is lower than that reported in other studies that have used SVM (Fan et al.
; Sun et al.
). Of note, these studies have applied SVM to patients with an established illness, rather than at first episode, and have used SVM to define subject status (patient versus
control) rather than future illness course. It is possible that, as more marked and distributed brain changes occur with illness progression (Ellison-Wright et al.
), the classification and distinction between patients and healthy individuals becomes more accurate in later illness stages. Interestingly, a study that applied SVM to the early detection of different at-risk mental states for subthreshold psychotic symptoms reported accuracies lower than those reported in patients with established schizophrenia (Koutsouleris et al.
). Consistent with our finding, a study that used a different classification approach to predict 1-year outcome in first-episode psychosis patients also reported an accuracy similar to the one we observed (Kasparek et al.
). Nevertheless, the better classification we found in individuals destined to develop a continuous illness course suggests that significant brain changes have already occurred by the time of first contact in this subgroup.
Although, with these methods, it is not possible to make local inferences on the discriminating regions, it is noteworthy that the anatomical regions with the highest contribution to the discrimination between groups include the cingulate and parahippocampal gyri, the basal ganglia and the thalami, which have been consistently reported as important in the pathophysiology of schizophrenia (Ellison-Wright et al.
Our study should be taken as preliminary, but it proposes a promising approach for the future translation of imaging into patient benefit. Although our approach requires replication and validation in a larger sample, it provides initial evidence of an easy and accessible methodology that could potentially enable rapid clinical decisions. We believe that future development should consider that, although a better classification accuracy could potentially be achieved with feature selection approaches preceding the SVM classifier, this would involve applying cross-validation pipelines that are time-consuming, calculation intensive and/or methodologically challenging. Such approaches can be useful in the research setting but might not enable a rapid and user-friendly assessment of patients' prognosis. Furthermore, there is at present no conclusive evidence that feature selection increases the accuracy for MRI-based diagnosis, while substantially increasing computation time without improving classification performance (Cuingnet et al.
). Therefore, we have preferred here a whole brain approach, which has stronger potential for future applicability in the clinical setting. Once the SVM classifier is trained and a decision function is generated, a new example could be classified in minutes. The pipeline we used includes pre-processing of the structural images using standard neuroimaging software (SPM), extraction of a feature vector containing gray-matter values, and application of the SVM decision function to the feature vector. Moreover, the greater the complexity of the method (i.e. the number of steps in the pipeline), the greater the danger of overfitting the data. By contrast, maintaining a simple pipeline improves the translational potential of the method, as such a technique will be more likely to be incorporated into routine clinical investigations, because a structural MRI can be obtained in 10 minutes.
In psychiatry, clinicians do not have algorithms to quantify the risk of, for example, occurrence of further episodes, which could help in guiding clinical decisions. Such algorithms are of common use in other areas of medicine, such as heart medicine and oncology. Our results offer preliminary evidence that, once further developed and validated, an SVM approach could in the future be used by a psychiatrist to say, for example, that on the basis of the MRI scan, a patient may have 88% probability of having more illness episodes (positive predictive value: in this example, it is calculated as the proportion of intermediate subjects who went on to experience repeated episodes and who were classified as continuous by the SVM classifier), thus providing more information on which to base therapeutic decisions. This would imply applying risk quantification in psychiatry, something that has not yet been achieved. Furthermore, the quantification could improve even further and be refined if other factors related to outcome, such as duration of untreated psychosis, were taken into account in a decision tree.
As mentioned earlier, our study is preliminary and the findings should be considered with caution. An important limitation is the relatively small sample size. In addition, future work should consider validating the accuracy of our classifier with an independent larger sample of patients classified according to the same illness course criteria. Finally, as this study used data from a single site, it remains unclear to what extent differences in acquisition protocol or scanners affect the accuracy of the classifier. Nevertheless, it is encouraging that the classifier trained in using the two extremes groups of patients (i.e. episodic versus continuous) could be successfully applied to classify the intermediate group of patients. Future studies with multi-site images will allow an investigation of the extendibility of the approach to images from different sites.
Ultimately, once developed, these methods could allow a patient to be assigned to targeted assertive case-management at first presentation to services, with optimized pharmacological treatment, cognitive and family interventions, which have been shown to improve treatment adherence and reduce relapse rates (Malla et al.
; Eack et al.
), eventually improving outcome. At the same time, those patients most likely to have a good remitting illness after their first episode could avoid long-term exposure to antipsychotic medication. We propose that, with further development and validation on larger datasets, a simple and automated SVM pipeline offers a promising approach to help rapid and early clinical decisions based on imaging data, which will eventually benefit patient care and reduce health-care costs.