To further identify clinical features, phenotypes and inflammatory patterns of refractory systemic juvenile idiopathic arthritis (sJIA) through machine learning approaches.

 In this retrospective cohort study, total 151 patients with sJIA were categorized as refractory (R-sJIA) or non-refractory  (nR-sJIA). Baseline clinical and laboratory features, treatment patterns, and outcomes were analyzed. Unsupervised clustering using factor analysis of mixed data with partitioning around medoids was applied to define baseline phenotypes. Group-based trajectory modeling was applied to characterize 6-month longitudinal inflammatory trajectories. Cross-model correspondence analysis evaluated the relationships between phenotypes and trajectories.

Among 151 sJIA patients, 31.1% developed R-sJIA. Baseline demographics, joint involvement patterns, and conventional inflammatory markers were similar between R-sJIA and nR-sJIA, with no significant difference in overall MAS prevalence. However, recurrent MAS (≥2 episodes) was strongly associated with R-sJIA (OR = 8.54, 95% CI: 2.33–33.54, P=0.0012). Despite frequent treatment escalation, over half of R-sJIA patients (61.5%) ultimately achieved disease stabilization with biologic therapy. Four distinct baseline phenotypes and four inflammatory trajectories were identified. The cross-model mapping analysis enhanced the clinical interpretability of both models. High initial systemic inflammation did not uniformly predict poor outcomes. Instead, refractory disease was characterized by more frequent polyarticular involvement and delayed inflammatory resolution.Among 151 sJIA patients, 31.1% developed R-sJIA. Baseline demographics, joint involvement patterns, and conventional inflammatory markers were similar between R-sJIA and nR-sJIA, with no significant difference in overall MAS prevalence. However, recurrent MAS (≥2 episodes) was strongly associated with R-sJIA (OR = 8.54, 95% CI: 2.33–33.54, P=0.0012). Despite frequent treatment escalation, over half of R-sJIA patients (61.5%) ultimately achieved disease stabilization with biologic therapy. Four distinct baseline phenotypes and four inflammatory trajectories were identified. The cross-model mapping analysis enhanced the clinical interpretability of both models. High initial systemic inflammation did not uniformly predict poor outcomes. Instead, refractory disease was characterized by more frequent polyarticular involvement and delayed inflammatory resolution.

sJIA exhibits marked clinical and inflammatory heterogeneity, and refractory disease cannot be reliably predicted by baseline severity alone. Recurrent MAS, polyarticular involvement, and delayed inflammatory resolution may predict R-sJIA. Early relieve of inflammation may make change the trajectories to develop R-sJIA.