Computerized Prediction of Mortality from Upper Gastrointestinal Bleeding

An artificial neural-network program that used pre-endoscopy data was superior to the Rockall scoring system.


Multiple scoring systems have been developed to identify patients with upper gastrointestinal bleeding (UGIB) who are at high risk for rebleeding and death. Most of these systems use evidence of comorbidities and large-volume bleeding with or without endoscopic information regarding high-risk stigmata.

Now, investigators in Italy have developed a computerized predictor of mortality using data from 2380 patients with nonvariceal UGIB collected from a network of 23 hospitals in a study consortium. The researchers used an artificial neural-network system called TWIST — an adaptive, computer program that analyzed pre-endoscopy data in nonlinear fashion to identify variables that correlated with 30-day mortality.

The result was a model based on 17 pre-endoscopy variables that had a sensitivity of 83.8% (95% confidence interval, 76.7–90.8) and a specificity of 97.5% (95% CI, 96.8–98.2). By comparison, the Rockall scoring system, which used endoscopic information, had a sensitivity of 71.4%; (95% CI, 62.8–80.0) and a specificity of 52.0% (95% CI, 49.8–54.2) in the same population. The area under the receiver operating characteristic curve was 0.95 for the neural-network program (95% CI, 0.92–0.98) and 0.67 for the Rockall scoring system (95% CI, 0.65–0.69).


Citation(s):

Rotondano G et al. Artificial neural networks accurately predict mortality in patients with nonvariceal upper GI bleeding. Gastrointest Endosc 2011 Feb; 73:218.