4E-01 - Severe infections risk assessment score for patients with rheumatoid arthritis receiving targeted therapies: development and validation in a Nationwide cohort

Background: RA patients receiving targeted therapies face an increased risk of severe infections due to immunosuppression, comorbidities, and prior infections. Existing risk models, such as the RABBIT Risk Score, have limitations in predicting infection risk across newer therapies.

Objectives: This study aimed to estimate severe infection risk in RA patients initiating targeted therapies and develop a predictive tool for clinical decision-making.

Methods: Using the French national healthcare database (SNDS), we conducted a cohort study of RA patients initiating targeted therapy from 2010 to 2023. Severe infections requiring hospitalization were identified, and risk factors were assessed using a Cox proportional hazards model, adjusting for key covariates at baseline (age, comorbidities, prior infections, corticosteroid use, and treatment history). The dataset was split into development (66%) and validation (34%) cohorts. Model performance was evaluated via the C-index and calibration plots. We developed RAISE (RA Infection Score Estimator), a calculator providing real-time infection risk predictions.

Results: We included 97,888 RA patients (73.3% women, mean age 57.0 ± 14.1 years). Most received TNFi (67.8%), followed by CD20i (10.6%), CTLA-4i (8.6%), IL-6i (8.8%), and JAKi (3.8%). Severe infections occurred in 6,997 patients (7.1%), with an incidence rate of 41.1 per 1,000 patient-years. In multivariate analysis, older age, corticosteroid use, prior infections (HR 1.57, 95% CI 1.40–1.75), diabetes (HR 1.38, 95% CI 1.27–1.50), cancer (HR 1.33, 95% CI 1.21–1.46), and specific targeted therapies were significant predictors. Compared to Etanercept, Infliximab (HR 1.76, 95% CI 1.57–1.98) and Rituximab (HR 2.12, 95% CI 1.91–2.35) were associated with higher infection risk. The model demonstrated good discrimination (C-index: 0.71) and acceptable calibration (mean error at 24 months: 0.019). Patients were stratified into low (84.3%), moderate (11.3%), and high-risk (4.4%) groups. Time-dependent AUCs were 0.73 (12 months) and 0.72 (24 months). RAISE was developed to estimate real-time infection probability based on patient characteristics.

Conclusions: Our model identified key predictors of severe infections in RA patients on targeted therapy and demonstrated good discrimination and calibration, supporting its clinical applicability. We developed RAISE, a tool for real-time infection probability estimation to assist physicians in treatment decision-making.