Objectives: Sjögren’s syndrome (SjS) systematic disease activity cannot be ascertained in administrative claims and/or electronic medical records (EMR). Prediction algorithms to classify patients with moderate-to-severe systemic activity using claims/EMR were evaluated.

Methods: A retrospective chart review was used to identify adult patients with SjS in a US integrated delivery network and assess disease severity using EULAR’s Sjögren's Syndrome Disease Activity Index (ESSDAI). Patients with ESSDAI scores ≥5 and <5 were considered to have moderate-to-severe and low systemic SjS severity, respectively. Demographic and 12-month baseline clinical predictors of moderate-to-severe SjS were obtained from linked claims/EMR prior to first ESSDAI assessment. Predictors included healthcare resource utilization (HRU; i.e., inpatient/outpatient/emergency room days), EMR (i.e., laboratory testing), comorbidities, number of specialist visits (e.g., rheumatologist), and medications. Logistic regression modeled the predicted probability of having moderate-to-severe SjS. Discrimination was assessed using the receiver operating curves (ROC) and positive/negative predictive values (PPV/NPV). Calibration was assessed using Hosmer-Lemeshow (HL) goodness-of-fit.

Results: Data from 213 patients were collected (mean age 59.6 years, 92% female). Based on chart review, 19 patients (8.9%) had moderate-to-severe SjS (i.e., gold standard). Three models were identified with high discrimination and were well-calibrated: Model 1: Comorbidities, specialists, and medications (ROC=0.886; HL p-value=0.93); Model 2: HRU, specialists, and medications (ROC=0.907; HL p-value=0.96); Model 3: HRU, EMR, specialists, and medications (ROC=0.939; HL p-value=0.88). Using 0.50 predicted probability cut-off, for Models 1 to 3, sensitivity ranged from 11% to 58%, specificity 97% to 99%, PPV 33% to 72% and NPV 92% to 96%.

Conclusions: Overall, these prediction models, particularly including EMR data, were successful in ruling out patients with moderate-to-severe systemic disease activity. Given the heterogeneity of SjS, low sample size and sensitivity, and increased likelihood of model overfitting, random forests will be implemented to rank-order importance of predictors and develop a more parsimonious model.

Sponsorship: Novartis