Artificial Neural Network (ANN) Is Superior to the Pre-Endoscopic Rockall Score in Predicting Which Patients Will Benefit From Urgent EGD in Acute Upper GI Bleeding (UGIB)

Abstract

Artificial Neural Network (ANN) Is Superior to the PreEndoscopic Rockall Score in Predicting Which Patients Will Benefit From Urgent EGD in Acute Upper GI Bleeding (UGIB) Ananya Das, Farees Farooq, Richard Wong, Tamir Ben-Menachem, Gregory Cooper, Amitabh Chak, Michael Sivak Introduction: In practice, not all patients with acute UGIB can receive urgent EGD (within 12 hours). A model for predicting need for urgent EGD based on nonendoscopic, clinical variables would be an important tool in patient triage. We have previously shown that ANN accurately predicts outcome in acute LGI bleeding (Lancet 2003;362:1261). Methods: We constructed an ANN to predict endoscopic stigmata of recent hemorrhage (SRH) and need for endoscopic therapy (ET) using prospectively collected data from 194 patients at an academic medical center who presented with acute UGIB over a 6-month period (training group). Confirmed variceal bleeding was excluded. The ANN model was a multi-layered perceptron network trained by back propagation using pre-endoscopic clinical input variables available at the time of triage. The trained ANN was applied to an internal validation (IV) group of 193 patients with UGIB during the same time period. The ANN was then applied to an external validation (EV) group of 200 patients at a tertiary care center in a different city over a 6-month period. The performance of the ANN was also compared to the widely used Rockall score. A pre-endoscopic Rockall score of 0 is considered low risk for adverse outcome. Results: The clinical characteristics of the EV group were different then the training and IV groups (increased men, CAD, cirrhosis, hematemesis, shock, ICU admit, rebleeding, and death). Using ANN, the areas under the ROC curve for predicting SRH and ETwere 0.94 & 0.84 in the IV group and 0.81 & 0.78 in the EV group, respectively. ANN had better specificity and predictive values than pre-endoscopic Rockall score in both validation groups and was comparable to total (post-endoscopic) Rockall score. Conclusions: ANN is an effective tool in the pre-endoscopic triage of patients with acute UGIB, even in a dissimilar external cohort. ANN is more specific than the preendoscopic Rockall score in identifying patients who may benefit from urgent EGD and may be used to exclude low risk patients.