Background/Purpose: Response to therapy in SLE varies significantly and presents a considerable challenge in terms of predictability. Development of predictive models with ability to accurately foresee attainment of remission or low disease activity would be of major clinical value.

Methods: We extracted 288 baseline features from five phase III trials of belimumab, comprising 3645 patients. Using a recursive regression and partitioning tree, the top 20 most important factors were identified. Next, clinical expertise was integrated to narrow down the features qualified for the final model to only three. To investigate a potentially differential predictive ability of the model with different treatment regimens, we stratified the study population into placebo and belimumab recipients. Outcomes were attainment of sustained remission as defined by the Definitions of Remission in SLE (DORIS) criteria and sustained lupus low disease activity state (LLDAS), maintained through week 52. Remission or LLDAS was considered sustained if present at two or more consecutive visits. Different predictive models i.e., least absolute shrinkage and selection operator (LASSO), neural network (NNet), support vector machine (SVM), and extreme gradient boosting (XGBoost) were trained 10 times in a 10-fold cross-validation. Performance was assessed using 20% of the data not used in training. Area under the curve (AUC), confusion matrix, and calibration plots were generated.

Results: In total, 11% met the criteria for sustained LLDAS and 5% those for sustained remission. Lower urinary protein/creatinine ratio (UPCR), lower prednisone dose, and increasing C3 levels were the features yielding the greatest predictive ability. The largest AUC was achieved for predicting sustained LLDAS in the placebo group (LASSO with AUC: 0.80, sensitivity: 0.73, specificity: 0.67, accuracy 0.70) while the belimumab subgroup yielded inferior performance (LASSO with AUC: 0.71, sensitivity: 0.77, specificity: 0.57, accuracy: 0.67). Although calibration seemed satisfactory, class imbalance and the small positive sample size overall might be a reason underlying the superior performance of the first model. Sustained remission was more challenging to predict among belimumab-treated patients (NNet with AUC: 0.63, sensitivity: 0.62, specificity: 0.60, accuracy: 0.61) compared with the placebo group (LASSO with AUC: 0.76, sensitivity: 0.63, specificity: 0.63, accuracy: 0.63). Prediction of sustained LLDAS yielded superior results (LASSO with AUC: 0.76, sensitivity: 0.68, specificity: 0.65, accuracy: 0.67) compared with remission (SVM with AUC: 0.71, sensitivity: 0.61, specificity: 0.71, accuracy: 0.66).

Conclusion: The models developed to predict attainment of sustained remission or LLDAS yielded moderate sensitivity and specificity. Overall, sustained LLDAS was more precisely predicted than sustained remission and accuracy in the placebo population was superior to that in belimumab-treated patients. By reducing the complexity to only 3 factors (UPCR, prednisone dose, C3 levels), our approach enabled a quick estimation of the probability to attain sustained remission or LLDAS in this SLE population.

« Back to ACR Convergence 2023

ACR Meeting Abstracts - https://acrabstracts.org/abstract/predicting-remission-and-low-disease-activity-attainment-in-patients-with-systemic-lupus-erythematosus/