[Comparative estimation of currently available approaches to early identification of severe acute pancreatitis]

It is difficult to precisely predict the severity of each specific patient with acute pancreatitis (AP), by using conventional statistical methods or single clinical and laboratory criteria. By using the patient database, the authors developed an artificial neuronal network (ANN) to predict the severity of AP and compared it with the Ranson, Imrie/Glasgow, APACHE II, and Physiological Condition Severity scoring systems, ultrasound/radiological criteria, and linear regression analysis. ANN was found to be significantly better than all the scoring systems, but not better than a linear regression model (the area under the receiver operating characteristic curves were equal to 0.83). With 81% sensitivity, ANN showed a 70% specificity, and positive/negative predictive values of 73 and 79%, respectively. ANN is an effective tool in developing prediction models for poor outcomes in patients with AP that is precisely similar with linear regression models.