Risk models for rebleeding and postoperative mortality in bleeding gastric ulcer

In order to better define management policies we attempted to construct risk models for rebleeding on initial conservative management and mortality after emergency surgery for failure of medical therapy in 387 patients with bleeding gastric ulcer. Several different models were constructed using logistic regression analysis with validation by the 'leaving-one-out' method. However, despite large patient numbers, modelling in this way is difficult because of inherent wide variation between patients. Suitable models for rebleeding were regarded as rather unsatisfactory, for although overall accuracy was 86%, sensitivity was only 54%. More promising was a model for mortality after emergency surgery which had an accuracy of 93% and a sensitivity of 80%. Such mortality models incorporating age, history of previous malignant disease or dyspepsia, the presence or absence of ascites and total transfusion requirements may well prove to be of value in surgical practice. This paper seeks to examine the process of modelling rebleeding and mortality and of interpreting the models produced.