Psy30 Life Expectancy and Lifetime Cost of Rheumatoid Arthritis in the Era of Biologics in Taiwan

Abstract

two-phase study used data from Reliant Medical Group electronic medical records linked to healthcare claims for 1,647 patients with SICCA syndrome (estimated prevalence=15%). Primary outcomes included presence and early indicators of SjS. Patient, clinical, and treatment characteristics were included as predictors. Random forest (RF) and optimal classification tree (OCT) machinelearning (ML) models were used. In phase 1, the RF model was used to 1) identify patients with high risk of SjS; 2) assist with a targeted chart review; 3) identify SNOMED codes to use as proxy for SjS diagnosis. In phase 2, RF and OCT models were used to identify early indicators of SjS up to 3 years pre-SjS diagnosis. Area under the curve (AUC), sensitivity, and specificity were used to estimate model performance. Results: Phase 1: Research nurses reviewed 200 (stratified) randomly-selected patient charts to confirm SjS. The charts were used to train the RF model, subsequently applied on the 1,447 patients to select additional 200 patient charts for the ML-guided chart review. ML-guided-chart-review results were used to refine model prediction and identify and confirm the use of SNOMED codes as an SjS-diagnosis proxy (AUC=85%, sensitivity=75%, speci-ficity=70%). Phase 2: RF and OCT models were estimated on all 1,647 patients to identify early indicators of SjS. Notable early predictors from the RF model (AUC=86%, sensitivity=90%, specificity=65%) included: patient's age and sex, specialist visits (e.g., rheumatology, ophthalmology), tests (e.g., anti-SSA/SSB, RNP antibody/SCL-70, ANA and RF titer, complement, immunoglobulin blood), and vaccinations (e.g., DPT). OCT results were consistent with RF results though model performance was slightly worse (e.g., AUC=64%). Conclusions: Advanced ML methods can be used to inform clinicians about early indicators to facilitate timely diagnosis of SjS.