Machine learning based CAGIB score predicts in-hospital mortality of cirrhotic patients with acute gastrointestinal bleeding

Abstract

Acute gastrointestinal bleeding (AGIB) is a potentially lethal complication in cirrhosis. In this prospective international multi-center study, the performance of CAGIB score for predicting the risk of in-hospital death in 2467 cirrhotic patients with AGIB was validated. Machine learning (ML) models were established based on CAGIB components, and their area under curves (AUCs) were calculated and compared. Gray zone approach was employed to further stratify the risk of death. In training cohort, the AUC of CAGIB score was 0.789. Among the ML models, the least square support vector machine regression (LS-SVMR) model had the best predictive performance (AUC = 0.986). Patients were further divided into low- (LS-SVMR score <0.084), moderate- (LS-SVMR score 0.084-0.160), and high-risk (LS-SVMR score >0.160) groups with in-hospital mortality of 0.38%, 2.22%, and 64.37%, respectively. Statistical results were retained in validation cohort. LS-SVMR model has an excellent predictive performance for in-hospital death in cirrhotic patients with AGIB (ClinicalTrials.gov; NCT04662918).

Fig. 1. ROC curves of CAGIB score, Child-Pugh score, MELD-Na score, MELD 3.0 score, D’Amico model, and Augustin model for predicting the risk of in-hospital death in patients with cirrhosis and acute gastrointestinal bleeding in the training cohort.

a All patients; b Patients with variceal bleeding; c Patients who underwent endoscopic treatment; d Patients who received pharmacological treatment alone without endoscopic treatment. The performance of CAGIB score for predicting in-hospital death was statistically similar to that of Child-Pugh, MELD-Na, MELD 3.0, D’Amico, and Augustin scores both in overall and subgroup analyses.

Fig. 2. ROC curves of CAGIB score, Child-Pugh score, MELD-Na score, MELD 3.0 score, D’Amico model, and Augustin model for predicting the risk of in-hospital death in patients with cirrhosis and acute gastrointestinal bleeding in the validation cohort.

a All patients; b Patients with variceal bleeding; c Patients who underwent endoscopic treatment; d Patients who received pharmacological treatment alone without endoscopic treatment. The performance of CAGIB score for predicting in-hospital death was statistically similar to that of Child-Pugh, MELD-Na, MELD 3.0, D’Amico, and Augustin scores both in overall and subgroup analyses.

Fig. 3. ROC curves of ANN model, KNN model, decision tree model, RF model, XGBoost model, and LS-SVMR model for predicting the risk of in-hospital death in patients with cirrhosis and acute gastrointestinal bleeding in the training cohort.

a All patients; b Patients with variceal bleeding; c Patients who underwent endoscopic treatment; d Patients who received pharmacological treatment alone without endoscopic treatment. The performance of LS-SVMR model was significantly higher than ANN, KNN, and decision tree, and statistically similar to XGBoost model, but significantly lower than RF both in overall and subgroup analyses. ANN artificial neural network, KNN K-nearest neighbors, RF random forest, LS-SVMR least square support vector machine regression.

Fig. 4. Gray zone of LS-SVMR model.

a All patients; b Patients with variceal bleeding; c Patients who underwent endoscopic treatment; d Patients who received pharmacological treatment without endoscopic treatment. Patients were divided into low-, moderate-, and high-risk group of in-hospital death based on LS-SVMR score in overall and subgroup analyses.

Fig. 5. ROC curves of ANN model, KNN model, decision tree model, RF model, XGBoost model, and LS-SVMR model for predicting the risk of in-hospital death in patients with cirrhosis and acute gastrointestinal bleeding in the validation cohort.

a All patients; b Patients with variceal bleeding; c Patients who underwent endoscopic treatment; d Patients who received pharmacological treatment alone without endoscopic treatment. The performance of LS-SVMR model was significantly higher than ANN, KNN, decision tree, XGBoost, and RF models both in overall and subgroup analyses. ANN artificial neural network, KNN K-nearest neighbors, RF random forest, LS-SVMR least square support vector machine regression.