FRI0559 VALIDATION OF NORDIC JUVENILE IDIOPATHIC ARTHRITIS CLINICAL PREDICTION MODELS IN A CANADIAN COHORT

Abstract

Background Validation of clinical prediction tools for juvenile idiopathic arthritis (JIA) in populations different than those in which they were first developed is essential to understand their applicability across healthcare settings. Objectives To determine if clinical prediction tools to predict 1) non-achievement of remission off medication and 2) functional disability, developed in the Nordic cohort1 can be directly applied to JIA patients in the Canadian Research in Arthritis in Canadian Children emphasizing Outcomes (ReACCh-Out) cohort; and to assess performance of the prediction tools if model parameters are fine-tuned to the Canadian cohort. Methods Since the Nordic models were developed to predict outcomes 8 years after disease onset but the follow-up of the ReACCh-Out cohort was shorter, we chose to cross-validate the tools in a subpopulation of 513 subjects at the 3 years follow-up (3.75 years after onset). Attainment of remission off medications was determined by a panel of 3 pediatric rheumatologists as previously described2 and functional disability was defined as a Childhood Health Assessment Questionnaire Disability Index (CHAQ)>0. Missing data was handled with multiple imputation by chained equations and prediction ability was assessed with c-index and Receiver Operator Characteristic (ROC) curves. The Nordic models were first evaluated exactly as published on the entire Canadian cohort. Then we fine-tuned the model coefficients using repeated runs of cross-validation in the Canadian cohort. This way, fine-tuned models were tested in patients not included in the fine-tuning process while also minimizing the standard error of prediction. Results In total, 408 of 506 evaluable patients (81%) were not on remission and 137 of 361 evaluable patients (38%) had functional disability at the 3-year visit. The Nordic model for predicting non-achievement of remission had a c-index of 0.68 (95%CI 0.62-0.74) when directly applied, and a c-index of 0.74 (0.70-0.78) when it was fine-tuned for the Canadian population. The latter values are comparable to those reported in the Nordic cohort (median AUC 0.78, IQR 0.72-0.82). Table 1 shows fine-tuned coefficient values along-side the original values. The Nordic model for predicting functional disability had a c-index of 0.57 (0.50-0.63) when directly applied, and fine-tuning failed to improve its performance, a c-index of 0.53 (0.43-0.63). Figure 1 shows ROC curves for fine-tuned models in the multiple splits. Conclusion After fine-tuning of coefficients, the Nordic model for prediction of non-achievement of remission had similar prediction ability in Canadian patients with JIA. We could not confirm the validity of the Nordic model for prediction of functional disability in Canadian patients.