Abstract
<title>Abstract</title> <p>Background Although some non-invasive indicators have already been applied in diagnosing pediatric portal hypertension (PHTN), predictive tools for assessing the risk of gastrointestinal (GI) bleeding remain scarce. This study aimed to develop a predictive model for PHTN combined with GI bleeding in pediatric patients. Methods We retrospectively reviewed clinical data of 308 patients diagnosed with PHTN between January 2008 and December 2023 in Children's Hospital of Chongqing Medical University. We calculated the importance of every indicator by XGBoost and CatBoost and ranked the importance using Entropy-Weight Topsis. The set of essential indices was screened from 28 characteristic variables as the independent variable, and whether there was a combination of GI bleeding or not was taken as the dependent variable. The patient data were randomly divided into a training set (n = 216) and an internal validation set (n = 92) at a ratio of 7:3. The logistic regression (LR), support vector machines (SVM), extreme-gradient boosting (XGB) and random forest (RF) were constructed. The area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and F1 value were analyzed to evaluate the predictive performance of models. Furthermore, the predictive ability of models was further validated on the validation set. Results Six predictors, including hemoglobin, AST, bilirubin, platelets, albumin, and aetiological classification (prehepatic) were finally identified. The AUC of LR, SVM, XGB, and RF in the validation cohort were 0. 867 (95% CI 0.790–0.943),0.901 (95% CI 0.835–0.967), 0.915 (95% CI 0.853–0.976) and 0.927 (95% CI 0.871–0.984) respectively. The RF model achieved the highest accuracy (88.8%) and outperformed the other models. Conclusions These models consisting of hemoglobin, AST, serum total bilirubin, etiology classification (prehepatic type), serum albumin, and PLT are valuable in predicting the risk of GI bleeding in children with PHTN. Among them, the random forest model demonstrates the best performance and can be crucial in disease assessment and clinical decision-making.</p>