Machine Learning in Predicting the Risk of Esophagogastric Variceal Bleeding Among Patients With Liver Cirrhosis: Systematic Review and Meta-Analysis.

Background: Liver cirrhosis (LC) can lead to several complications. Esophageal variceal bleeding (EVB) and esophagogastric variceal bleeding (EGVB) are particularly severe, leading to a high risk of mortality. Early identification of esophageal varices and esophagogastric varices is essential. Several studies have constructed prediction models for EVB and EGVB in patients with LC. However, robust systematic evidence to prove their performance is lacking. Objective: We included original studies that developed prediction models for esophageal or gastric variceal bleeding in patients with LC under different modeling variables. This study aimed to review the predictive performance of various models for EVB or EGVB in patients with LC, providing insights into the development or updating of simplified scoring tools in the future. Methods: PubMed, Web of Science, Embase, and the Cochrane Library were searched up to August 21, 2024, to collect original full-text studies on machine learning (ML) in the prediction of EVB and EGVB in patients with LC. The models were evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Subgroup analyses were carried out based on the modeling variables. Results: In total, 21 studies were included, with 7011 patients with LC, among whom 1412 (20.14%) developed EVB and 733 (10.45%) developed EGVB. The meta-analysis results suggested that the pooled c-index, sensitivity, and specificity of the prediction model for predicting EVB in the validation set were 0.85 (95% CI 0.77-0.92), 0.93 (95% CI 0.87-0.96), and 0.66 (95% CI 0.46-0.82), respectively. The pooled c-index, sensitivity, and specificity of the prediction model for predicting EGVB in the validation set were 0.89 (95% CI 0.85-0.94), 0.77 (95% CI 0.66-0.85), and 0.81 (95% CI 0.67-0.90), respectively. The subgroup analysis based on modeling variables revealed that, for predicting EVB, the c-index in the validation set was 0.84 (95% CI 0.80-0.88) for models based on clinical features, 0.82 (95% CI 0.69-0.96) for radiomics-based models, 0.78 (95% CI 0.67-0.89) for models based on radiomics and clinical features, and 0.97 (95% CI 0.95-1.00) for models based on endoscopic features. Subgroup analyses based on modeling variables revealed that, for predicting EGVB, the c-index in the validation set was 0.91 (95% CI 0.86-0.96) for models based on clinical features and 0.85 (95% CI 0.75-0.96) for models based on radiomics and clinical features. Conclusions: ML methods are feasible for predicting EVB and EGVB in patients with LC. Nevertheless, the number of included original studies is limited. In the future, more studies with larger sample sizes are needed to promote the application of ML in the early assessment of EVB and EGVB in patients with LC in clinical practice.