Use of a Machine Learning Algorithm to Predict Rebleeding and Mortality for Oesophageal Variceal Bleeding in Cirrhotic Patients - European Medical Journal


Use of a Machine Learning Algorithm to Predict Rebleeding and Mortality for Oesophageal Variceal Bleeding in Cirrhotic Patients

| Gastroenterology
3 Mins
EMJ Gastroenterology 9.1 2020 Feature Image
*Jonathan Soldera,1 Fernanda Tomé,2 Leandro Luís Corso,3 Matheus Machado Rech,1 Andressa Daiane Ferrazza,1 Alana Zulian Terres,1 Bruna Teston Cini,1 Louise Zanotto Eberhardt,1 Juline Isabel Leichtweis Balensiefer,1 Rafael Sartori Balbinot,1 Ana Laura Facco Muscope,1 Morgana Luisa Longen,1 Bruna Schena,1 Gilberto Luis Rost Jr.,1 Rafaelle Galiotto Furlan,1 Raul Angelo Balbinot,1 Silvana Sartori Balbinot1

The authors have declared no conflicts of interest.

EMJ Gastroenterol. ;9[1]:46-48. Abstract Review No. AR2.

Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.


Oesophageal variceal bleeding (OVB) is one of the most common complications of cirrhosis. Mortality rates range from 15% to 20% in the first episode.1,2 Therefore, identifying patients with high chances of survival is paramount to allocate resources into treatment with accuracy.3 The purpose of this study was to use a machine learning algorithm to predict rebleeding and mortality for OVB in patients with cirrhosis and to analyse its accuracy.4-6


A historical cohort study was conducted, analysing data from hospital charts from January 2010 to December 2016. Patients were found by searching every use of terlipressin during the time period. Medical charts were hand-analysed. Patients over 18 years old with laboratory and imaging data supporting the diagnosis of cirrhosis and with a definitive diagnosis of OVB were included.

This analysis used data from 74 patients with cirrhosis, taking into account 36 variables, which had OVB as a complication. The preliminary analysis of the study was Pearson correlation, which compared the 36 variables in the study with outcomes of death and rebleeding, aiming to verify the linear correlation strength, positive or negative.

When artificial intelligence was applied, an artificial neural network (ANN) was utilised to recognise patterns in outcomes through supervised learning. The results were analysed on a confusion matrix, which presented the probabilities of the positive predictive value, negative predictive value, sensitivity, specificity, and network accuracy. A receiver operating characteristic (ROC) curve analysis was then performed.


Electronic search retrieved 177 hospital admissions with use of terlipressin, 101 of which were due to OVB. All-cause mortality was 36.0%, 41.5%, and 50.4% for 30-, 90-, and 365-day, respectively. Mean age was 56 years and 79% were male. The most frequent cause of cirrhosis was alcohol abuse, followed by hepatitis C.

The Pearson correlation analysis showed that the variables had values of linear correlation ranging from -0.34 to 0.30 for mortality and -0.31 to 0.21 for rebleeding. Both values represent weak correlations with the outcomes. Thus, it is notably difficult to define which variables are the ones with major leverage on the outcomes. Therefore, the use of artificial intelligence could be a key tool to identify the patterns in such a complex data-evolved situation.

For patients who had a rebleeding outcome, the specificity value showed that the ANN was able to identify 66.7% of cases. The predictive value showed when the ANN predicted rebleeding, 100% of the patients did indeed rebleed. The overall accuracy was 97.4% and the area under the ROC curve (AUROC) was 0.942.

For patients who had a mortality outcome, the specificity value showed that the ANN was able to identify 95.0% of cases. The predictive value shows when the ANN predicted mortality, 95.0% of the patients did indeed die. The overall accuracy was 97.4% and the AUROC was 0.993, which demonstrates a high performance of the network.


The ANN could more accurately predict mortality by OVB when compared with two other assessment tools, Chronic Liver Failure-Sequential Organ Failure Assessment (CLIF-SOFA) and Model for End-stage Liver Disease (MELD) Score.7,8,9 The AUROC of CLIF-SOFA found in the literature for the outcome death was 0.943 and the AUROC of the MELD score was 0.80,10 whereas the AUROC of the ANN was 0.993. Therefore, machine learning could be a useful tool to improve clinical practice, with the possibility of outperforming the current tools.

Berzigotti A et al. Assessing portal hypertension in liver diseases. Expert Rev Gastroenterol Hepatol. 2013;7(2): 141-55. Boregowda U et al. Update on the management of gastrointestinal varices. World J Gastrointest Pharmacol Ther. 2019;10(1):1-21. Lyles T et al. A risk scoring system to predict in-hospital mortality in patients with cirrhosis presenting with upper gastrointestinal bleeding. J Clin Gastroenterol. 2014;48(8):712-20. Le Berre C et al. Application of artificial intelligence to gastroenterology and hepatology. Gastroenterology. 2020;158(1):76-94.e2. Rotondano G et al. Artificial neural networks accurately predict mortality in patients with nonvariceal upper GI bleeding. Gastrointest Endosc. 2011;73:218-26.e2. Das A et al. Artificial neural network as a predictive instrument in patients with acute nonvariceal upper gastrointestinal hemorrhage. Gastroenterology. 2008;134:65-74. Cerqueira RM et al. Risk factors for in-hospital mortality in cirrhotic patients with esophageal variceal bleeding. Eur J Gastroenterol Hepatol. 2012;24(5):551-7. Chalasani N et al. Model For End-stage Liver Disease (MELD) for predicting mortality in patients with acute variceal bleeding. Hepatology. 2002;35:1282-4. Amitrano L et al. MELD score and hepatocellular carcinoma identify patients at different risk of short-term mortality among cirrhotics bleeding from esophageal varices. J Hepatol. 2005;42(6):820-5. Lee M et al. CLIF-SOFA scoring system accurately predicts short-term mortality in acutely decompensated patients with alcoholic cirrhosis: a retrospective analysis. Liver Int. 2015;35(1):46-57.