2405: Machine Learning in SLE Diagnosis: Performance of the SLE Risk Probability Index Questionnaire in a Multicenter Cohort of Patients with Systemic Lupus Erythematosus

Background/Purpose: The SLE Risk Probability Index (SLERPI), a clinical prediction model for systemic lupus erythematosus (SLE), was developed using machine‑learning variable‑selection techniques (Random Forest, LASSO). Its score reflects diagnostic certainty, ranging from unlikely to definite SLE. A dichotomous model that uses a cut‑off > 7 points to distinguish SLE from non‑SLE previously showed high specificity and sensitivity and excellent diagnostic performance in early SLE (disease onset < 5 years). This study aimed to (i) evaluate the dichotomous SLERPI’s sensitivity, specificity, and likelihood ratios (LR) in a connective‑tissue‑disease cohort; (ii) correlate its score with SLE activity/damage indices; and (iii) assess its performance in early SLE, defined as disease duration < 5 years.

Methods: In this multicenter, cross‑sectional study, we included consecutive adults (≥18 years) with SLE (cases) and other systemic autoimmune diseases (controls). Exclusion criteria were active infections, active onco‑hematologic disease, drug‑induced lupus, and SLE overlap syndromes. We recorded sociodemographics, disease characteristics, activity/damage indices (SLEDAI, SLICC‑SDI), and treatments. Three blinded evaluators independently classified patients as SLE/No SLE; unanimous agreement served as the gold standard. Expert opinion was compared with ACR 1997, SLICC 2012, ACR/EULAR 2019, and SLERPI criteria. Analyses included descriptive statistics, sensitivity, specificity, LR ±, and Spearman correlations. All sites obtained institutional review‑board approval and complied with the Declaration of Helsinki.

Results: From seven centers, 365 patients were enrolled (183 cases, 182 controls). Table 1 for characteristics. SLERPI showed 98.3 % sensitivity (95 % CI 95.3–99.7), 88 % specificity (95 % CI 82.3–92.3), LR+ 8.1 (95 % CI 5.5–12.0) and LR‑ 0.02 (95 % CI 0.006–0.06). Comparative performance versus other criteria is presented in Table 2. SLERPI scores correlated with SLEDAI (r = 0.24; p = 0.001) and SLICC‑SDI (r = 0.23; p = 0.002). SLERPI was higher in patients with damage (20 [IQR 16.5–22]) than in those without (17 [IQR 13.5–21]; p = 0.02). In the early‑SLE subgroup (n = 125; duration < 5 years), sensitivity was 96.6 %, specificity 87.9 %, LR+ 8.0, and LR‑ 0.04. Frequently used treatments among SLE cases were hydroxychloroquine (86.4 %), glucocorticoids (38.8 %), and mycophenolate mofetil (17 %).

Conclusion: SLERPI demonstrated diagnostic accuracy in this multicenter cohort comparable to SLICC 2012 criteria. Although correlations with activity and damage indices were modest, the SLERPI score was significantly higher in patients with established damage. Diagnostic accuracy remained high in early SLE (disease duration < 5 years). SLERPI may serve as a rapid screening tool for SLE, particularly when serological testing is delayed.