Background/Purpose: Childhood-onset systemic lupus erythematosus (cSLE) is characterized by changing disease activity: episodes of improvement are often followed by episodes of worsening, i.e. flares. Considering changes of previously defined core response variables (proteinuria [mostly from protein creatinine ratio, PCR], disease activity (SLEDAI, BILAG), physician global assessment of cSLE activity), and ESR, preliminary flare algorithms have been derived by Classification Tree Analysis (CART) and multinomial logistic regression (MLR), respectively. International consensus yielded 4 preferred criteria for detection of cSLE flares, based on initial validation in 2010-11. These algorithms yield flare-scores, with higher scores reflecting worse flares. Our objective was to validate the preliminary cSLE flare algorithms, and delineate flare scores that reflect mild/moderate/major flares.

Methods: A total of 1,860 Patient Profiles (PPs) providing relevant data at baseline and follow-up were generated from existing international prospective cSLE cohorts, with systematic imputation of missing algorithm variables. Pediatric rheumatologists (PP-raters) were asked to judge disease courses between baseline and follow-up as:  improved, no change, mild flare, moderate flare, or major flare. Based on majority opinion of PP-raters, each PP was assigned a true disease course, against which the flare algorithms were tested for accuracy. Using two complementary approaches (distribution-weighted, MLR-based) potential threshold flare-scores for flare severity (mild, moderate, major) were derived and tested for accuracy via area under the receiver operating characteristic curve (AUC).  Analyses were reviewed by an expert panel that used nominal group technique to achieve consensus (≥75% agreement).

Results:   Among 503 invited, feedback from 274 PP-raters (53% response rate) was available for analysis. Based on PP-rater feedback, true disease courses were as follows: 540 no change or improved and 510/483/325 PPs as mild flare/ moderate flare/ major flare. There was consensus to use MLR estimates for all flare thresholds. As shown (Table 1), CART-based models, though maintaining very good to excellent accuracy (AUCs for mild/moderate/major flares all >70%), were less suited to discriminate mild and moderate flares. Algorithms from MLR maintained outstanding accuracies (all AUC > 90.79%) and had sensitivities/specificities for mild/moderate/major flares all >82%/>82% in the 2017 validation analyses. This was true for both BILAG and SLEDAI-based algorithms (see Table 1).

Conclusion:  The provisional criteria for flare of global disease of cSLE have outstanding accuracy in identifying flare in cSLE with different degrees of severity. Based on our results and consensus, the BILAG and the SLEDAI-based algorithm from MLR are equally acceptable to measure flares in cSLE.