A Risk Scoring System to Predict In-hospital Mortality in Patients With Cirrhosis Presenting With Upper Gastrointestinal Bleeding

Abstract

We aimed to develop a simple and practical risk scoring system to predict in-hospital mortality in cirrhotics presenting with upper gastrointestinal (GI) bleeding. Extensive clinical data were captured in patients with documented cirrhosis who underwent endoscopic evaluation for upper GI bleeding between January 1, 2003 and June 30, 2011 at Parkland Memorial Hospital. Predictors of mortality were identified by multivariate regression analysis. A total of 884 patients with cirrhosis admitted for upper GI bleeding were identified; 809 patients survived and 75 died (8.4%). The etiology of bleeding was similar in both groups, with bleeding attributed to esophageal varices in 59% of survivors and 60% of non-survivors (ulcer disease and other etiologies of bleeding accounted for the other causes of bleeding). Mortality was 8.6% and 8.3% in patients with variceal bleeding and nonvariceal bleeding, respectively. While survivors and those who died were similarly matched with regard to gender, age, ethnicity and etiology of cirrhosis, patients who died had lower systolic blood pressures, higher pulse rates and lower mean arterial pressures at admission than patients who survived. Non-survivors were more likely to be Childs C (61% vs. 19%, P<0.001). Multivariate regression analysis identified the following 4 predictors of in-hospital mortality: use of vasoactive pressors, number of packed red blood cells transfused, model for end-stage liver disease (MELD) score, and serum albumin. A receiver operating characteristic curve including these 4 variables yielded an area under the receiver operating characteristic (AUROC) curve of 0.94 (95% confidence interval, 0.91-0.98). Classification and Regression Tree analysis yielded similar results, identifying vasoactive pressors and then MELD>21 as the most important decision nodes for predicting death. By comparison, using the Rockall scoring system in the same patients, the AUROC curve was 0.70 (95% confidence interval, 0.64-0.76 and the comparison of the University of Texas Southwestern model to the Rockall model revealed P<0.0001). A validation set comprised of 150 unique admissions between July 1, 2011 and July 31, 2012, had an AUROC of 0.92, and the outcomes of 97% of the subjects in this set were accurately predicted by the risk score model. Use of vasoactive agents, packed red blood cell transfusion, albumin, and MELD score were highly predictive of in-hospital mortality in cirrhotics presenting with upper GI bleeding. These variables were used to formulate a clinical risk scoring system for in-hospital mortality, which is available at: http://medweb.musc.edu/LogisticModelPredictor.