Artificial Neural Network as a Predictive Instrument in Patients With Acute Nonvariceal Upper Gastrointestinal Hemorrhage

Background & Aims: Triage of patients with acute upper gastrointestinal hemorrhage (UGIH) has traditionally required urgent upper endoscopy. The aim of this study is to evaluate the use of artificial neural network for nonendoscopic triage. Methods: A cohort of 387 patients was used to train (n = 194) and internally validate (n = 193) the neural network, which was then externally validated in 200 patients and compared with the clinical and complete Rockall score. Two outcome variables were assessed: major stigmata of recent hemorrhage and need for endoscopic therapy. Patient cohort data from 2 independent tertiary-care medical centers were prospectively collected. Adult patients hospitalized at both sites during the same time period with a primary diagnosis of acute nonvariceal UGIH. Results: In predicting the 2 measured outcomes, sensitivity of neural network was >80%, with high negative predictive values (92–96%) in both cohorts but with lower specificity in the external cohort. Both Rockall scores had adequate sensitivity (>80%) but poor specificity (<40%) at outcome prediction. Comparing areas under receiver operating characteristic curves, the clinical Rockall score was significantly inferior to neural network in both cohorts (≤0.65 vs. ≥ 0.78), while in the external cohort, neural network performed similarly to the complete Rockall score (≥ 0.78). Conclusions: In acute nonvariceal UGIH, artificial neural network (nonendoscopic triage) performed as well as the complete Rockall score (endoscopic triage) at predicting stigmata of recent hemorrhage and need for endoscopic therapy, even when tested in an external patient population.