SAT0628 VALIDATION OF METHODS FOR PREDICTING LONG-TERM OUTCOME IN JUVENILE IDIOPATHIC ARTHRITIS: RESULTS FOR CANADIAN AND NORDIC PREDICTION MODELS IN THE NORDIC COHORT

Abstract

Background Models predicting outcome in juvenile idiopathic arthritis (JIA) have recently been proposed by Guzman et al.1 and Rypdal et al.2 Guzman et al. constructed a model for predicting severe disease course derived from the ReACCh-Out study, and Rypdal et al. constructed models for prediction of non-remission, functional disability and joint damage. Objectives To validate methods for prediction of long-term outcome in JIA by testing the ability of Guzman’s model and Rypdal’s model to predict severe disease course (the ReACCh-Out outcome) in the Nordic cohort. Methods The Nordic cohort is a prospective longitudinal multicenter cohort from defined geographical areas of 4 Nordic countries. Children with a baseline and an 8-year study visit were included. Missing data were imputed using low rank matrix factorization3, and a K-medoids algorithm4 was used to identify clusters corresponding to severe disease course in the ReACCh-Out study. With this outcome, the prediction model of Guzman et al. was tested with no re-estimation of parameters. A Receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC) were computed. For the same outcome, prediction models were built using the method of Rypdal et al. on randomly sampled training sets, and tested on disjoint validation sets. Results In the Nordic cohort 98/440 (22%) patients were identified with a severe disease course. This ratio is similar to the 125/610 (20%) found in the ReACCh-Out study. Characteristics of groups of patients with severe and non-severe disease course are similar in the two cohorts. The model of Guzman et al. had an AUC of 0.85 for prediction of severe disease course and an AUC of 0.66 for predicting remission off medication. In repeated cross-validations, the model of Rypdal et al. had a median AUC of 0.90 (IQR 0.86-0.92) for prediction of severe disease course, and a median AUC of 0.78 (IQR 0.72–0.82) for remission off medication. Conclusion Tests in the Nordic cohort validate the ability of the model of Guzman et al. to predict severe disease course. Repeated cross-validations of the model of Rypdal et al. indicate that validation results are highly dependent of the chosen outcome, and that prediction of long-term remission status is more challenging than prediction of a severe disease course.