Predicting late-onset infections in the pediatric intensive care unit: a multimodel machine learning framework

Late-onset healthcare-associated infections in pediatric intensive care units (PICUs) are associated with increased mortality and healthcare costs. Early prediction remains difficult because of overlapping systemic inflammation and data leakage flaws in the existing models. This study developed a robust, interpretable machine learning framework for early risk stratification. We retrospectively analyzed 1,012 patients admitted to the PICU. To reduce temporal leakage, predictors were extracted within a prespecified − 24 to + 48 h periadmission window, a strict dynamic baseline observation window (-24 to + 48 h) was established to extract high-frequency physiological, laboratory, and interventional data. Missing values were handled using nonparametric random forest imputation. Eight machine-learning algorithms were developed using the training set and evaluated using a held-out internal test set. The algorithm comparison was used for performance benchmarking and SHAP-based feature prioritization, whereas the final clinical tool was developed separately using multivariable logistic regression. SHapley Additive exPlanations (SHAPs) were employed to interpret the optimal models. Finally, nested multivariate logistic regression was used to construct a parsimonious clinical tool. Tree-based ensemble algorithms outperformed traditional models. eXtreme Gradient Boosting (XGBoost) achieved the highest test-set area under the receiver operating characteristic curve [AUC] (0.839) and accuracy (0.752), whereas RSVM achieved the highest sensitivity (0.891) and MCC (0.496). Following SHAP-guided feature selection and backward elimination, a parsimonious five-variable predictive framework was finalized, incorporating actual base excess, pH, platelets, hemoglobin, and hematocrit. The final logistic-regression model showed acceptable apparent discrimination in the full development cohort (Concordance index = 0.725), while reducing the number of required variables. The framework was deployed as an interactive, web-based risk calculator. By addressing data leakage through a dynamic baseline window and mitigating the artificial intelligence (AI) “black-box” dilemma via SHAP, we developed a highly interpretable, five-variable early warning tool. This practical framework facilitates real-time bedside risk stratification for late-onset PICU infections and offers the potential to guide timely therapeutic interventions and improve clinical outcomes.