Abstract
Ureteropelvic junction obstruction (UPJO) is a common pediatric condition often treated with pyeloplasty. Despite the surgical intervention, postoperative urinary tract infections (UTIs) occur in over 30% of cases within six months, adversely affecting recovery and increasing both clinical and economic burdens. Current prediction methods for postoperative UTIs rely on empirical judgment and limited clinical parameters, underscoring the need for a robust, multifactorial predictive model. We retrospectively analyzed data from 764 pediatric patients who underwent unilateral pyeloplasty at the Children's Hospital affiliated with the Capital Institute of Pediatrics between January 2012 and January 2023. A total of 25 clinical features were extracted, including patient demographics, medical history, surgical details, and various postoperative indicators. Feature engineering was initially performed, followed by a comparative analysis of five machine learning algorithms (Logistic Regression, SVM, Random Forest, XGBoost, and LightGBM) and the deep learning TabNet model. This comparison highlighted the respective strengths and limitations of traditional machine learning versus deep learning approaches. Building on these findings, we developed an ensemble learning model, meta-learner, that effectively integrates both methodologies, and utilized SHAP(Shapley Additive Explanation, SHAP) to complete the visualization of the integrated black-box model. Among the 764 pediatric pyeloplasty cases analyzed, 265 (34.7%) developed postoperative UTIs, predominantly within the first three months. Early UTIs significantly increased the likelihood of re-obstruction (P < 0.01), underscoring the critical impact of infection on surgical outcomes. In evaluating the performance of six algorithms, TabNet outperformed traditional models, with the order from lowest to highest as follows: Logistic Regression, SVM, Random Forest, XGBoost, LightGBM, and TabNet. Feature engineering markedly improved the predictive accuracy of traditional models, as evidenced by the enhanced performance of LightGBM (Accuracy: 0.71, AUC: 0.78 post-engineering). The proposed ensemble approach, combining LightGBM and TabNet with a Logistic Regression meta-learner, achieved superior predictive accuracy (Accuracy: 0.80, AUC: 0.80) while reducing dependence on feature engineering. SHAP analysis further revealed eGFR and ALB as significant predictors of UTIs post-pyeloplasty, providing new clinical insights into risk factors. In summary, we have introduced the first ensemble prediction model, incorporating both machine learning and deep learning (meta-learner), to predict urinary tract infections following pediatric pyeloplasty. This ensemble approach mitigates the dependency of machine learning models on feature engineering while addressing the issue of overfitting in deep learning-based models like TabNet, particularly in the context of small medical datasets. By improving prediction accuracy, this model supports proactive interventions, reduces postoperative infections and re-obstruction rates, enhances pyeloplasty outcomes, and alleviates health and economic burdens.Level of evidence IV Case series with no comparison group.