Non-component clinical feature-based machine learning for disease activity risk stratification in juvenile idiopathic arthritis: a multi-center retrospective validation study.

BACKGROUND

JADAS27 is widely used to assess juvenile idiopathic arthritis (JIA) disease activity, but complete scoring is impractical in many clinical settings because it requires simultaneous physician global assessment (PhGA), patient/parent global assessment (PtGA), active joint count (AJC), and erythrocyte sedimentation rate (ESR). We asked whether non-component clinical variables alone-strictly excluding all four JADAS27 constituents-can classify JADAS27-defined activity strata.

METHODS

In this retrospective multi-center study, 800 patients with JIA were enrolled from CARRA (n = 400), PRCSG (n = 240), and LHTCM (n = 160). Disease activity was categorized as inactive, low, moderate, or high by established JADAS27 cutoffs. PhGA, PtGA, AJC, and ESR were excluded from all model inputs to prevent circular prediction. Thirteen remaining non-component variables (18 encoded features) were used to train four machine learning algorithms on CARRA + PRCSG (n = 640), with independent external validation on LHTCM (n = 160). Pre-specified ablation analyses quantified contributions of proxy variables (CHAQ, pain score, limited joint count).

RESULTS

SVM achieved the best external validation performance (accuracy 0.731, 95% CI 0.657-0.797; macro AUC 0.918, 95% CI 0.887-0.945). Class-wise recall was highest for inactive (82.4%) and high activity (77.4%), with most errors between adjacent classes. SHAP and permutation importance analyses consistently identified CHAQ score, JIA subtype, pain score, limited joint count, and C-reactive protein as the most influential predictors. Removing all three proxy variables reduced accuracy to 0.619 and macro AUC to 0.843, indicating that both proxy and non-proxy features contribute independently. Sensitivity analysis with subtype-specific cutoffs yielded comparable performance (accuracy 0.713).

CONCLUSION

Strictly non-component clinical features can stratify JADAS27-defined disease activity with clinically meaningful external performance. This approach may support early risk stratification when formal JADAS27 scoring is unavailable, and complements rather than replaces physician assessment.