Machine learning predicts late-onset PICU infection risk in retrospective cohort

A retrospective study of 1,012 children admitted to a pediatric intensive care unit found that machine-learning models using data collected around admission identified risk of late-onset healthcare-associated infections with varying performance. In the held-out internal test set, XGBoost produced the highest area under the receiver operating characteristic curve, or AUC, at 0.839, and the highest accuracy, at 0.752. A radial basis function support vector machine achieved the highest sensitivity, at 0.891, and the highest Matthews correlation coefficient, at 0.496. The researchers then distilled the analysis into a five-variable logistic-regression framework that was deployed as an interactive web-based risk calculator.

The investigators analyzed high-frequency physiological, laboratory and interventional data from the 1,012-patient cohort. To reduce temporal data leakage, they restricted predictors to a prespecified dynamic baseline window extending from 24 hours before admission through 48 hours after admission. The window was intended to capture information available around the admission period while limiting reliance on observations accumulated later in the hospital stay. Missing values were addressed with nonparametric random-forest imputation. Eight algorithms were developed in the training set and evaluated on a held-out internal test set, while the comparisons served as performance benchmarking and SHAP-based feature prioritization.

The strongest results came from tree-based ensemble methods. XGBoost reached an AUC of 0.839 and accuracy of 0.752 on the internal test set, while the radial basis function support vector machine reached a sensitivity of 0.891 and a Matthews correlation coefficient of 0.496. These are internal performance estimates from the study cohort rather than measurements of effectiveness in routine care. The retrospective design, single development population and reliance on an internal test set limit the scope of the findings. The supplied study abstract does not report external validation, confidence intervals for the principal performance measures, or patient outcomes associated with use of the models.

To make the prediction framework less dependent on a large set of variables, the research team used SHapley Additive exPlanations, or SHAP, to prioritize features and then applied backward elimination. The resulting framework retained five variables: actual base excess, pH, platelets, hemoglobin and hematocrit. Unlike the tree-based algorithms used for comparison, this final clinical tool was developed separately through multivariable logistic regression, with nested regression used to construct the parsimonious predictor. In the full development cohort, the final model showed an apparent concordance index of 0.725. That estimate describes discrimination in the development data; the abstract does not establish how the five-variable model would perform in an independent patient population or in a different PICU workflow.

The technical approach addresses two challenges identified by the authors: limiting temporal leakage and interpreting the contribution of variables in complex machine-learning models. SHAP analysis made it possible to examine which features influenced the better-performing algorithms, and the streamlined logistic-regression tool reduced the number of variables needed for risk calculation. The authors implemented the framework as an interactive web-based calculator, providing a software demonstration rather than evidence of clinical benefit. The study’s results support the predictive performance reported within its retrospective development and internal-testing framework, but they do not show that risk-guided intervention reduces infections, mortality, treatment burden or costs. Prospective assessment, independent validation and evaluation of the calculator in clinical workflow would be needed to establish whether its predictions can support effective bedside decisions.

Subject of Research: Machine-learning prediction of late-onset healthcare-associated infections in children admitted to intensive care

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

Article References: Li, J., Xia, Y., & Zhang, L. (2026). Predicting late-onset infections in the pediatric intensive care unit: a multimodel machine learning framework. BMC Infectious Diseases. https://doi.org/10.1186/s12879-026-14361-9

Image Credits: AI Generated

DOI: 10.1186/s12879-026-14361-9

Keywords: pediatric intensive care, late-onset infection, healthcare-associated infection, machine learning, risk prediction, XGBoost, SHAP, logistic regression, electronic health records, clinical calculator, Predicting, late-onset