JMIR Publications

ABSTRACT

Background:

Esophageal variceal bleeding (EVB) is one of the most frequent and life-threatening complications of patients with cirrhosis. Its high mortality risk and acute clinical burden substantially impair the quality of care near the end of life. Despite its clinical importance, predictive models for EVB development remain limited, particularly among patients with non-cancer-related cirrhosis.

Objective:

This study aimed to develop a machine learning (ML) framework utilizing a large-scale multicenter database to enhance the early prediction of EVB risk in non-cancer-related cirrhosis patients and support clinical decision-making.

Methods:

We conducted a multicenter retrospective cohort study utilizing the Chang Gung Research Database from January 2010 to December 2017. Multiple ML algorithms were developed to evaluate the predictability of EVB, and their performance was compared across training and testing datasets. The analytical pipeline involved 10-fold cross-validation with hyperparameter tuning and utilized the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance in the training cohort. Optimal sensitivity and specificity were determined using Youden’s index. The best- performing ML model was further interpreted using Shapley Additive exPlanations (SHAP) analysis.

Results:

Of 8,615 initially identified patients, 4,692 met the inclusion criteria after excluding individuals with hepatic malignancy, sepsis, or insufficient visits. The final cohort comprised 1,210 patients (25.8%) in the EVB group and 3,482 (74.2%) in the non-EVB group. Among the ML algorithms evaluated, the XGBoost demonstrated the best discriminative performance compared with other ML models, with an area under the receiver operating characteristic curve (AUROC) of 0.781 (95% CI 0.754–0.809), significantly outperforming both MELD (0.596, 95% CI 0.563–0.629) and HAS-BLED (0.551, p5% CI 0.515–0.588) scores in the testing dataset. Using an optimal probability threshold of 0.25, the model achieved a sensitivity of 79%, specificity of 62%, positive predictive value of 0.42, and a negative predictive value of 0.89. SHAP analysis identified platelet count as the paramount predictor, followed by hemoglobin, red blood cell count, C-reactive protein, and ammonia as the most influential predictors of EVB in non-cancer chronic liver disease patients.

Conclusions:

Our novel ML-based model suggests a feasible and interpretable approach to enhance the prognostication of EVB in patients with non-cancer-related cirrhosis through routinely collected clinical data. This model offers a practical tool to support clinicians and patients in shared decision-making regarding treatment strategies, with the ultimate aim of improving clinical care and reducing the burden of EVB-associated morbidity.