Generic filters

Artificial Intelligence Could Help Early Discharge of Patients With UGI Bleeding

Prateek Sharma, MD, FASGE

Prateek Sharma, MD, FASGE reviewing Shung DL, et al. Gastroenterology 2019 Sep 25.

A number of clinical scoring systems (eg, the Glasgow-Blatchford, Rockall, and AIMS65) have been developed for identifying patients with upper GI bleeding (UGIB) who are at low risk for blood transfusions, hemostatic interventions (endoscopic, surgical, or interventional radiological), and mortality. Investigators in this study used machine learning to develop a model to calculate the risk of intervention or death in patients with UGIB and compared its performance with clinical scoring systems. 

Data from consecutive, unselected patients presenting with UGIB were collected between March 2014 and March 2015 from 6 large hospitals: Yale New Haven Hospital (United States), Glasgow Royal Infirmary (Scotland), Royal Cornwall Hospital (England), Odense University Hospital (Denmark), Singapore General Hospital (Singapore), and Dunedin Hospital (New Zealand). Only patients presenting with overt bleeding, defined as hematemesis or melena, were included, whereas patients who were already inpatients with UGIB were excluded from the analysis. The endpoint used for the machine learning models was a composite endpoint: hospital-based intervention with blood transfusion or hemostatic intervention, or 30-day all-cause mortality. 

The study included 2357 patients: 1958 for the training/internal validation group (U.S. and European centers) and 399 for the external validation group (Asian-Pacific centers). For the composite outcome, the machine learning model with an area under receiver operating characteristic curve (AUC) value of 0.91 (95% confidence interval [CI], 0.90-0.93) performed better than the Glasgow-Blatchford score (AUC=0.88; 95% CI, 0.86-0.90; P=0.001), Rockall score (AUC=0.69; 95% CI, 0.66-0.71; P<0.001), and AIMS65 score (AUC=0.72; 95% CI, 0.69-0.74; P<0.001). Similarly, for the external validation cohort, the model performed better than all clinical risk scores: AUC=0.90 (95% CI, 0.87-0.93) versus the Glasgow-Blatchford score (AUC=0.87; 95% CI, 0.84-0.91; P=0.004); Rockall score (AUC=0.65; 95% CI, 0.60-0.71; P<0.001); and the AIMS65 score (AUC=0.64; 95% CI, 0.59-0.69; P<0.001).


In acute UGIB, machine learning can predict the composite outcome of transfusion, hemostatic intervention, or death better than the current, commonly used clinical risk scores, such as the Glasgow-Blatchford, Rockall, and AIMS65. From a clinical perspective, the results of a machine learning model could provide recommendations for outpatient management of UGIB patients who are at very low risk for blood transfusion, hemostatic intervention, or death.

Note to readers: At the time we reviewed this paper, its publisher noted that it was not in final form and that subsequent changes might be made.

Prateek Sharma, MD, FASGE


Shung DL, Au B, Taylor RA, et al. Validation of a machine learning model that outperforms clinical risk scoring systems for upper gastrointestinal bleeding. Gastroenterology 2019 Sep 25. (Epub ahead of print) (

Related Posts

Scroll to Top