Noninvasive prediction of large esophageal varices in liver cirrhosis patients Abstract Purpose: Esophageal varices are a dangerous complication of liver cirrhosis. e development of cost e ective, noninvasive means for prediction of large esophageal varices could reduce the use of upper gastrointestinal endoscopy in variceal screening and also provide an alternative way to con rm the results of conventional endoscopic diagnosis. Previously proposed predictive models are neither sensitive nor speci c.Methods: A retrospective study based on a group of 104 liver cirrhosis patients was performed. Multiple statistical approaches were used to evaluate the association of large esophageal varices with 20 individual and six compound clinical laboratory variables. A new predictive model was developed.Results: Univariate analysis suggested that eight out of 26 variables were signi cantly associated with large esophageal varices. Further stepwise logistic regression eventually identied three variables (hemoglobin level, portal vein diameter and the ratio of platelet count/ spleen diameter) that contributed signi cantly to the nal regression model. Receiver operating characteristic (ROC) curve analysis showed that this new regression model achieved 77.8% and 72% of diagnostic sensitivity and speci city, respectively, for the prediction of large esophageal varices. In our study group, its diagnostic accuracy (AUROC=0.814) was found to be signi cantly higher than six predictive models previously published.Conclusions: No single variable o ers self-su cient predictive function for large esophageal varices. A comprehensive model using multiple variables signi cantly improves the predictive accuracy in screening the most at risk patients with potential variceal hemorrhage.