Short-term risk stratification using parallel admission and reassessment features in PICU patients with infection.

Objective: To develop and validate an early risk prediction model for short-term major adverse events (MAE) in pediatric intensive care unit (PICU) admissions with infection using admission (M0) and early reassessment (M1) features. Methods: Using the PhysioNet Paediatric Intensive Care database, 684 infection-related PICU admissions were included in the final landmark-defined analytic cohort. The primary outcome was 72-hour MAE, defined as new vasoactive drug use, invasive mechanical ventilation, or death within 72 h after PICU admission. For the primary M0 + M1 analysis, prediction targeted events occurring after completion of the reassessment window. Features were extracted from M0 (0-6 h) and M1 (12-36 h). LASSO, random forest (RF), XGBoost, and a stacked ensemble were compared. Performance was assessed in an internal test cohort and a de-identified timestamp ordered validation cohort using AUC, PR-AUC, calibration metrics, Brier score, and decision curve analysis. Results: Among the candidate models, RF showed the most favorable overall numerical performance. In the internal test set, RF achieved an AUC of 0.724 and a PR-AUC of 0.741. In the de-identified timestamp-ordered validation cohort, the corresponding values were 0.718 and 0.766. Calibration was good in the internal test set but attenuated in the later cohort, suggesting the need for recalibration. SHAP analysis indicated complementary contributions of M0 and M1 features. Conclusions: A parallel M0 + M1 framework showed moderate discrimination and potential utility for reassessment-oriented early risk stratification in infection-related PICU admissions. Further external validation and recalibration are needed before broader application.