Objective: to seek for predictors of inactive disease (ID) in juvenile idiopathic arthritis (JIA) with artificial intelligence. Methods: The clinical charts of patients seen within 6 months after disease onset between 2007 and 2019 and with follow-up visits at 6, 12, 18, and 24 months were reviewed retrospectively. Sixty-eight potential predictors were recorded at each visit. The primary endpoint was ID at 24 months by 2004 Wallace criteria. Data obtained from diverse combinations of visits were examined to identify the best forecasting model. After pre-processing, the cohort was divided into training (50%) and testing (50%) cohorts. Multivariate time series forecasting, coupled with the Random Forest method, was used to train the machine learning (ML) forecasting model. Predictive performance was assessed through the Matthews correlation coefficient (MCC). Results: A total of 414 patients were included. The best performance in predicting ID at 24 months in the training cohort was provided by the 0-12 months interval (MCC = 0.68). In the testing cohort, the same ML model confirmed a high forecasting performance (MCC = 0.65). Assessment of feature importance and impact analysis showed that the most relevant predictor of ID was the physician's global assessment (PhGA), followed by the count of active joints (AJC). Conclusions: PhGA and AJC values over the first 12 months were the strongest predictors of ID at 24 months. This finding highlights the importance of regular quantitative assessment of disease activity by the caring physician in monitoring the course of the patient toward achievement of complete disease quiescence.
Keywords: artificial intelligence; inactive disease; juvenile idiopathic arthritis; machine learning; outcome predictors; pediatric rheumatology.