DOP61 Development and validation of novel models in prediction of intravenous corticosteroids resistance in patients with Acute Severe Ulcerative Colitis

Abstract

Abstract Background Early prediction of intravenous corticosteroid (IVCS) resistance in Acute severe ulcerative colitis (ASUC) patients could reduce costs and delay in rescue therapy. However, most prediction models for ASUC were at high risk of bias with a lack of external validation. This study aims to construct and validate a model that accurately predicts IVCS resistance using various statistical methods. Methods A retrospective cohort of patients who were diagnosed with ASUC and had undergone IVCS treatment between March 2012 to January 2020 was established. Predictors evaluated included age, gender, race, medications before admission, infections, and laboratory data at baseline and during IVCS treatment, and endoscopic outcomes relied on blinded centralized endoscopy reading. The LASSO regression was used in feature selection for multivariate logistic regression model. Models based on machine learning methods (decision tree and random forest [RF]) were also constructed. Internal validity was confirmed and model performances were compared. External validation was conducted using data using an independent cohort from a tertiary referral centre. Results A total of 129 patients were included in the derivation cohort. During index hospitalization, 102 (79.1%) responded to IVCS, and 27 (20.9%) failed; 16 patients underwent colectomy, 6 received cyclosporin, and 5 succeeded with IFX as rescue therapy. Ulcerative Colitis Endoscopic Index of Severity (UCEIS; odds ratio [OR] 5.39, 95% confidence interval [CI] 2.52–14.0, p<0.001) and C-reactive protein (CRP) level on the third day (OR 1.05, 95% CI 1.03–1.08, p<0.001) were selected by LASSO regression and identified as the only two independent predictors of IVCS resistance in logistic regression. The decision tree model identified a UCEIS higher than 6.5 points and CRP level at day 3 higher than 33.57 mg/dL as the proxy for IVCS resistance. UCEIS and CRP level at day 3 were also the most important predictors in the RF model. Areas under the curve receiver operating characteristic (AUC) of logistic model, decision tree model, and RF model were 0.64 (95% CI 0.49–0.80), 0.81 (95% CI 0.71–0.90), and 0.88 (95% CI 0.82–0.95), respectively. A validation cohort of 65 ASUC patients were established, and the AUC of the models in external validation were 0.57 (95% CI 0.45–0.70), 0.70 (95% CI 0.61–0.80), and 0.71 (95% CI 0.48–0.94), respectively. Conclusion In patients with ASUC, UCEIS and CRP level at day 3 of IVCS treatment appeared to allow the prompt prediction of likely IVCS nonresponders. Machine learning-based models outperformed the traditional statistical model in the prediction. The models may aid therapeutic decision-making in ASUC patients.