Background/Purpose: A key goal of treatment of SLE is the prevention of irreversible organ damage. The ability to identify patients at increased risk for damage could select patients for early intervention but is currently lacking. Conventional studies of damage outcomes in SLE utilize composite disease activity scores, and associations with damage are only modest. We evaluated whether using all available data in an unbiased fashion, including continuous laboratory data usually analysed dichotomously in disease activity scores, could give rise to superior predictive algorithms for organ damage in SLE.

Methods: Prospectively collected longitudinal data from a 13-centre multinational cohort were used. Each visit was assigned yes/no as being in a damage-transition period based on whether the nearest subsequent annual measurement of organ damage (SLICC damage index) increased. Candidate variables included demographic (3), baseline serology (4), medication classes (anti-malarial, immunosuppressant, and glucocorticoid), routine clinical laboratory parameters (14) and SLEDAI-2K. Logistic regression models were selected to predict the probability of damage transition using backward, forward, and hybrid stepwise selection methods, either including or excluding SLEDAI-2K. The primary metric evaluating performance was area under the receiver operating characteristic curve (AUC) for association with damage transition.

Results: Data from 15,625 visits of 1,621 patients were split randomly 80:20 into training and test datasets; a separate cohort (2,178 visits, 188 patients) was used for independent validation. As there were only 1,157 damage transition visits in the training dataset, oversampling was used to create a dataset with 23,120 visits with a ‘transition visit’ ratio of 50%.

Altogether 2²¹ (~ 2 million) models with different combinations of predictors were analysed. After excluding models with >20 variables or training AUC < 0.62 we ranked 43,332 models excluding SLEDAI, and 224,349 models including SLEDAI, according to the mean of the training, test, and validation set AUC.

The highest performing model (mean AUC=0.659) included age, sex, and ethnicity; baseline ANA, anti-Sm, and anti-phospholipid antibody positivity; prednisolone use; and routine clinical lab variables ESR, eGFR, haemoglobin, platelet & neutrophil count, and urine protein & red cells. All top 20 models included age, sex, ethnicity, some baseline serology, prednisolone use, and both ESR & eGFR, with anti-phospholipid status, haemoglobin, platelet and neutrophil count and urine protein and red cells in most. In this analysis, the inclusion in model selection of clinician measurement of disease activity using SLEDAI2K did not improve performance (highest AUC including SLEDAI2K=0.651).

Conclusion: We have derived algorithms predictive of organ damage, an important clinical outcome in SLE, using an unbiased ‘big data’ approach. Continuous laboratory variables contributed to damage risk, while clinician measured disease activity did not improve prediction. This suggests continuous laboratory measures should be included in future SLE clinical trial endpoints.

« Back to ACR Convergence 2020

ACR Meeting Abstracts - https://acrabstracts.org/abstract/prediction-of-damage-in-sle-using-unbiased-analysis-of-large-datasets/