Dynamic prediction systems of transition to psychosis using joint modelling: extensions to the base system

Seeking risk factors and constructing prediction models for transition to psychosis in individuals at ultra-high risk (UHR) has been an important research area. Our previous work showed that dynamic prediction could perform better than the conventional approach of using only baseline predictors in predicting transition to a psychotic disorder in UHR individuals. Dynamic prediction is the prediction of the occurrence of an event outcome using longitudinal data and has been made possible using a statistical methodology called joint modelling. The application of joint modelling and dynamic prediction in our previous work was relatively simple. In this paper, we examined extensions to our previous work in three ways: how to use the estimated changes in transition probability at repeated assessments over time to perform prediction, how to model the trajectory of the longitudinal data and how to model the relationship between the longitudinal data and the risk of transition to psychosis. Data from the Pace400 study (n = 398 UHR individuals), a follow-up study with transition to psychosis as the primary outcome, were used to investigate these extensions. Our results indicated that these extensions can enhance improvement in terms of model fit and sensitivity and specificity values. We have shown that dynamic prediction through joint modelling not only can utilize the richness of longitudinal data but also offers versatility in how prediction can be conducted. Our results have again confirmed that dynamic prediction via joint modelling should be considered as a useful tool for predicting transition to psychosis.