Abstract / Summary
The etiology of late recurrent intussusception (LRI) in children remains unclear, and the recurrence rate is high. However, systematic studies on its prediction are currently lacking. This study aimed to develop a machine learning–based prediction model for LRI and a web-based risk calculator to support clinical decision-making for the early identification of high-risk children. We retrospectively analyzed 1,671 children from two centers who experienced a first episode of intussusception and underwent successful pneumatic reduction. A total of 24 clinical and laboratory variables were collected. Key predictors were selected using the Boruta and recursive feature elimination (RFE) algorithms. Ten machine learning algorithms were trained in the internal training set to identify the optimal model, and model performance was evaluated in both the internal validation set and an independent external validation set. SHapley Additive exPlanations (SHAP) analysis was performed to interpret the model predictions. A web-based risk prediction calculator was developed and deployed using the Shiny platform. The internal dataset was randomly divided into a training set and an internal validation set at a ratio of 8:2. Nine predictive features were ultimately identified. LightGBM demonstrated the best performance and was selected as the final model. In the internal validation set, the model achieved a ROC-AUC of 0.972 (95% CI, 0.947–0.996) and a Brier score of 0.045 (95% CI, 0.025–0.066). In the independent external validation set, the ROC-AUC was 0.972 (95% CI, 0.961–0.983), with a Brier score of 0.047 (95% CI, 0.035–0.060). Decision curve analysis (DCA) demonstrated that the model provided potential clinical net benefit across a range of threshold probabilities. SHAP analysis elucidated the contributions of the key features to model predictions and the directions of their associations with the predicted risk of LRI. The web-based calculator developed from the LightGBM model enables rapid estimation of the probability of LRI for individual patients. The LightGBM model developed in this study demonstrated good discriminative ability for predicting the risk of LRI in children, and the results of external validation support its potential clinical utility. The web-based risk calculator developed from this model enables individualized risk assessment and may serve as a supportive tool for the early identification of children at high risk for LRI.