P0512 Predict the risk of acute severe ulcerative colitis based on machine learning algorithm: A multicenter cohort study

Abstract

Listen

Translate article

Abstract Background Severe ulcerative colitis (UC) is associated with with increasing risk of admission and colectomy. Till now, there are limited number of studies on the predictive models for severe UC especially in Asian population. In this study, we aimed to develop and validate a predictive model for severe UC Chinese patients using machine learning algorithm. Methods Patients diagnosed with mild-to-moderate UC were included retrospectively at four tertiary hospitals in China from 2013 to 2016. The primary outcome was severe UC requring admissions within 3 years after diagnosis. The machine learning algorithm were applied to develop and validate a predictive model for severe UC. The model performance was then compared to the well-established one for europeans patients - the Oxford’s model. Results Altogether 437 patients with mild-to-moderate UC were included. A total of 67(15.3%) of patients with mild to moderate UC developed severe UC requring hospitalization or surgery within 3 years. The predictive model selected 11 features, including age, gender, disease extent, CRP, ESR, hemoglobin, WBC, platelet, ALB and history of intravenous corticosteroids, immunosuppressor and/or biological agent. The area under receiver operating characteristic curve (AUC) of multilayer perception (MLP) in the internal validation dataset was 0.883. The AUC of MLP model in external validation dataset was 0.897, which was better than both the Oxford’s model (AUC=0.661, P＜0.012) and the model based on Cox’s regression analysis (AUC=0.764, P＜0.001). Conclusion The MLP predictive model had excellent discrimination for severe UC patients, which outperformed both the Oxford’s model and the model based on logistic regression analysis.