Abstract
Background: Clinical criteria for Ultra-High Risk of psychosis (UHR) show moderate specificity for the prediction of the first psychotic episode (FEP), with an average true positive rate of approximately 30% at 3 years. We have recently developed a multimodal probabilistic model using the odds ratio form of Bayes’ rule that achieved sensitivity=73% and specificity=96% for FEP combining clinical data and biological data in a small sample of UHR patients. To validate this approach, we built a similar model using cognitive and clinical data from the Personal Assessment and Crisis Evaluation (“PACE 400”) study.
Methods: 430 UHR patients presenting to a specialist psychosis service in metropolitan Melbourne, Australia, were identified using the Comprehensive Assessment of At-Risk Mental State (CAARMS). Transition to FEP occurred in 114 within 13 years of presentation. Demographic and clinical data were available for the full cohort and cognitive data for 258 cases. Positive and negative likelihood ratios (LRs) for FEP were calculated for clinical and cognitive variables with statistically significant receiver–operating curves (ROCs). LRs were combined using the odds ratio form of Bayes’ Rule to calculate probability of transition for 258 cases with complete data. ROC curves, positive, and negative predictive values for this model were calculated at yearly intervals of follow-up. Model accuracy was calculated for brief limited intermittent psychotic symptoms (BLIPS), attenuated symptoms, and vulnerability subgroups of CAARMS risk.
Results: Significant predictors of transition included global assessment of function, duration of symptoms prior to presentation, quality-of-life scale, CAARMS items (Disorders of thought content, Conceptual disorganisation), performance IQ, and Full Scale IQ. An odds ratio form of Bayes’ rule model using these variables predicted transition with a sensitivity of 50%–64% and a specificity of 75%–92%(AUROC = 0.75–0.765) over 14 years post assessment. At 14 years, positive predictive value(PPV) = 50.7% and negative predictive value(NPV) = 86.4%. The model was more accurate in those presenting with BLIPS (PPV = 100%, NPV = 100%) in comparison to vulnerability criteria (PPV = 50%,NPV = 90.9%) or attenuated symptoms (PPV = 37.0%,NPV = 86.4%). Cases with overlapping symptoms were intermediate to these groups.
Conclusion: In a sample enriched by UHR criteria, multimodal modeling using simple Bayesian techniques can improve predictive power. Systematic use of clinical and cognitive assessment may facilitate personalized psychosis prevention strategies. The accuracy of UHR prediction may be improved by separate models for each UHR criteria subgroup. Further analysis should explore the value of additional modes of assessment such as imaging, electrophysiology, or other biomarkers, particularly in cases presenting with features other than BLIPS.