Systematic review of machine learning models in predicting the risk of bleed/grade of esophageal varices in patients with liver cirrhosis: A comprehensive methodological analysis.

Esophageal varices (EV) in liver cirrhosis carry high mortality risks. Traditional endoscopy, which is costly and subjective, prompts a shift towards machine learning (ML). This review critically evaluates ML applications in predicting bleeding risks and grading EV in patients with liver cirrhosis. Following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, we conducted a systematic review of studies using ML to predict the risk of variceal bleeding and/or grade EV in liver disease patients. Data extraction and bias assessment followed the CHARMS (CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modeling Studies) checklist and PROBAST (Prediction model Risk Of Bias Assessment Tool) tool, respectively. Due to the heterogeneity of the study, a meta-analysis was not feasible; instead, descriptive statistics summarized the findings. Twelve studies were included, highlighting the use of various ML models such as extreme gradient boosting, artificial neural networks, and convolutional neural networks. These studies demonstrated high predictive accuracy, with some models achieving area under the curve values above 99%. However, significant heterogeneity was noted in input variables, methodologies, and outcome measures. Moreover, a substantial portion of the studies exhibited unclear or high risk of bias, mainly due to insufficient participant numbers, unclear handling of missing data, and a lack of detailed reporting on endoscopic procedures. ML models show significant promise in predicting the risk of variceal bleeding and grading EV in patients with cirrhosis, potentially reducing the need for invasive procedures. Nonetheless, the current literature reveals considerable heterogeneity and methodological limitations, including high or unclear risks of bias. Future research should focus on larger, prospective trials and the standardization of ML assessment criteria to confirm these models' practical utility in clinical settings.