Objective:To construct a risk prediction model for the inter-hospital transfer of critically ill children using machine learning methods, identify key medical features affecting transfer outcomes, and improve the success rate of transfers.Methods:A prospective study was conducted on critically ill children admitted to the pediatric transfer center of Hunan Children's Hospital from January 2020 to January 2021. Medical data on critical care features and relevant data from the Pediatric Risk of Mortality (PRISMⅢ) scoring system were collected and processed. Three machine learning models, including logistic regression, decision tree, and Relief algorithm, were used to construct the risk prediction model. A back propagation neural network was employed to build a referral outcome prediction model to verify and analyze the selected medical features from the risk prediction model, exploring the key medical features influencing inter-hospital transfer risk.Results:Among the 549 transferred children included in the study, 222 were neonates (40.44%) and 327 were non-neonates (59.56%). There were 50 children in-hospital deaths, resulting in a mortality rate of 9.11%. After processing 151 critical care medical feature data points, each model selected the top 15 important features influencing transfer outcomes, with a total of 34 selected features. The decision tree model had an overlap of 72.7% with PRISMⅢ indicators, higher than logistic regression (36.4%) and Relief algorithm (27.3%). The training prediction accuracy of the decision tree model was 0.94, higher than the accuracy of 0.90 when including all features, indicating its clinical utility. Among the top 15 important features selected by the decision tree model, the impact on transfer outcomes was ranked as follows based on quantitative feature violin plots: base excess, total bilirubin, ionized calcium, total time, arterial oxygen pressure, blood parameters (including white blood cells, platelets, prothrombin time/activated partial thromboplastin time), carbon dioxide pressure, blood glucose, systolic blood pressure, heart rate, organ failure, lactate, capillary refill time, temperature, and cyanosis. Eight of these important features overlapped with PRISMⅢ indicators, including systolic blood pressure, heart rate, temperature, pupillary reflex, consciousness, acidosis, arterial oxygen pressure, carbon dioxide pressure, blood parameters, and blood glucose. The decision tree was used to select the top 15 medical features with high impact on the neonatal and non-neonatal datasets, respectively. A total of 19 features were selected, among which there were 8 differences and 11 overlap terms between the important features of the neonatal and non-neonatal.Conclusions:Machine learning models could serve as reliable tools for predicting the risk of inter-hospital transfer of critically ill children. The decision tree model exhibits superior performance and helps identify key medical features affecting inter-hospital transfer risk, thereby improving the success rate of inter-hospital transfers for critically ill children.