Individualized prediction of 2-year risk of relapse as indexed by psychiatric hospitalization following psychosis onset: Model development in two first episode samples

Abstract

Background: Although most patients with psychotic disorders experience relapse, it is not possible to predict whether or when an individual patient is going to relapse. We aimed to develop a multifactorial risk prediction algorithm for predicting risk of relapse in first episode psychosis (FEP).

Methods: Data from two prospectively collected cohorts of FEP patients (N = 1803) were used to develop three multiple logistic prediction models to predict risk of relapse (defined as hospitalization) within the first 2 years of onset of psychosis. Model 1 (M1_S1) used data obtained from clinical notes (Sample 1) while model 2 (M2_S2) applied the same set of predictors using data obtained from research interviews (Sample 2). The final model (Sample 2: M3_S2) used the same predictors plus additional detailed information on predictors. Model performance was evaluated employing measures of overall accuracy, calibration, discrimination and internal validation.

Results: In both samples, the 2-year probability of psychiatric hospitalization was 37%. Of all the models, discrimination accuracy was lowest when limited information (such as socio-demographic and clinical parameters) was included in the prediction model. Model M3_S2 using additional information (descriptors of pattern of cannabis, nicotine, alcohol and other illicit drug use) obtained from research interview had the best discrimination accuracy (Harrell's C index 0.749).

Conclusions: The measures that contributed most to predicting hospitalization are readily accessible in routine clinical practice, suggesting that a risk prediction tool based on these models would be clinically practicable following validation in independent samples and permit a personalized approach to relapse prevention in psychosis.

Keywords: Hospitalization; Model development; Psychosis; Relapse; Risk prediction.