Clinical-metabolic machine learning model differentiating rheumatoid arthritis from high-inflammatory Sjögren...

Although rheumatoid arthritis (RA) and the high-inflammatory phenotype of Sjögren disease (SjD-Phenotype 1) share clinical features including RF positivity and articular involvement, their differentiation remains challenging, particularly in anti-CCP-negative or indeterminate cases. Here, we developed and validated a machine learning model based on routine metabolic biomarkers to distinguish these two conditions. In the discovery cohort (Tianjin, n = 5,330: 4,736 RA and 594 SjD-Phenotype 1), XGBoost modeling was performed after excluding underlying liver diseases, with glucocorticoids (GCs) and hydroxychloroquine (HCQ) incorporated as key covariates. The model achieved an AUC of 0.88 (95% CI: 0.84-0.91) on the held-out test set, with 5-fold cross-validation confirming robust calibration (mean Brier score = 0.141, calibration slope = 1.03). Exploratory analysis in the ACPA-negative subgroup yielded an AUC of 0.910, though this estimate includes training samples and should be interpreted with caution. External validation in an independent Nanchang cohort (n = 319: 236 RA, 83 SjD) confirmed generalizability, yielding a pooled AUC of 0.840 (95% CI: 0.791-0.890; Rubin's rules) with consistent SHAP feature importance rankings. Model ablation analysis confirmed that while glucocorticoid use was the strongest individual discriminating feature, metabolic features provided significant incremental value beyond treatment variables alone (ΔAUC = +0.086, DeLong p < 0.001). Multivariable analysis demonstrated that elevated glucose (OR = 1.20, p = 0.028) and CRP (OR = 2.54, p < 0.001) remained independent discriminating features for RA after dual adjustment for GC and HCQ use. To explore the molecular basis of this metabolic divergence, we performed parallel transcriptomic analyses of publicly available PBMC datasets (GSE51092 for pSS, GSE93272 for RA), revealing fundamentally distinct pathway signatures: RA exhibited dominant mitochondrial oxidative phosphorylation and ribosomal gene programs, while SjD was characterized by robust type I interferon activation-providing molecular-level corroboration for the clinically observed metabolic differences. These findings demonstrate that routine biochemical parameters can provide auxiliary diagnostic value in the serological gray zone where anti-CCP fails, supported by both multi-center clinical validation and transcriptomic biological evidence.