P239 A novel user friendly model to identify key predictors of corticosteroid utilization in newly diagnosed patients with ulcerative colitis

Abstract

Background: Corticosteroid (CS) use in ulcerative colitis (UC) has been identified as a poor prognostic marker. However, key predictors for CS utilization at the time of UC diagnosis are poorly defined. We aimed to develop and internally validated a predictive model to approximate the risk of CS utilization over the course of disease in newly diagnosed UC patients. Methods: Newly diagnosed UC patients from a US nationwide cohort treated in the VA health care system were followed over time to evaluate factors predictive of CS use. Multivariate logistic regression was performed. Model development was performed in a random 2/3 of the total cohort and then validated in the remaining 1/3 of the cohort. The primary outcome was to predict use of CS for the management of UC. Candidate predictors included routinely available data at the time of UC diagnosis, including demographics, laboratory results and index colonoscopy findings. Results: 699 patients met the inclusion criteria and followed for median duration of 8 years. 118 sites from 48 states and Puerto Rico were represented. Of 699 patients, 288 patients (41.2%) required CS use for the management of UC. Key predictors for CS utilization selected for the model were: younger age, non-African American ethnicity, presence of hypoalbuminemia as well as IDA at the time of UC diagnosis, pan-colitis/right sided colitis and increased severity of endoscopy disease at index colonoscopy. The AUC for the model based on extent of disease at UC diagnosis was 0.71 (95% CI: 0.66–0.76). The AUC for the model based on severity of disease at UC diagnosis was 0.71 [95% CI: 0.67–0.76]. Model calibration was consistently good in all models (Hosmer-Lemeshow goodness of fit p>0.05). The models performed similarly in the internal validation cohort. The predicted probability of the outcome event is: Probability = 1/[1 + exp(−(β0 + β1X1 + β2X2 + ... + βkXk))] where βj is the regression coefficient for predictor Xj, and β0 is the model intercept (Table 1). Conclusions: We developed and internally validated a novel prognostic model to predict CS use among newly diagnosed patients with UC. This is the first model to incorporate individual risk factors and develop a cumulative risk for identifying CS use. Once externally validated, this prediction model can be used by clinician to identify high risk patients who may require escalation of therapy and closer follow-up with serial biomarkers to potentially avoid the CS use in future.