Machine learning models identify a distinct urinary metabolomic signature in children with non-systemic juvenile idiopathic arthritis.

OBJECTIVES: Despite advances in understanding the molecular mechanisms behind juvenile idiopathic arthritis (JIA), reliable non-invasive biomarkers for diagnosis and monitoring remain limited. Urinary metabolomics offers a promising approach to investigate disease-specific metabolic alterations, especially in the paediatric setting. This exploratory study aimed to characterise the urinary metabolomic profile of children with non-systemic JIA with a machine learning-based approach. METHODS: We conducted a cross-sectional study including 98 participants under 18 years of age (58 non-systemic JIA patients and 40 healthy controls). Urine samples were collected and analysed by targeted metabolomics. Feature selection was conducted using Boruta and recursive feature elimination (RFE). Supervised machine learning (ML) models, including logistic regression (LR), k-nearest neighbors (KNN), and random forest (RF), were developed and evaluated using stratified 10-fold cross-validation. RESULTS: Univariate analysis did not reveal any statistically significant metabolites between groups after false discovery rate correction. However, a trend towards upregulation was observed in JIA patients for urinary glycine and long-chain acylcarnitines. Boruta and RFE identified a shared subset of informative features, which were used for classification modelling. The RF model achieved the best performance (AUC 0.989; 95% CI: 0.78-1.00), followed by LR (AUC 0.898; 95% CI: 0.78-1.00) and KNN (AUC 0.795; 95% CI: 0.71-0.99). CONCLUSIONS: A multivariate approach and supervised ML models revealed a subtle but biologically relevant urinary metabolomic signature in children with non-systemic JIA, leading to highly accurate patient classification. Given the exploratory nature of this research, further studies with larger and more stratified cohorts are needed to confirm these findings.