Short-term Risk Stratification for Pediatric ICU Patients with Infection

684 infection-related pediatric ICU admissions from the PhysioNet Paediatric Intensive Care database were used to develop a risk prediction model. The study compared LASSO, random forest, XGBoost, and a stacked ensemble to predict major adverse events within 72 hours. A random forest model achieved an AUC of 0.724 in an internal test set, using features extracted at admission and a 12-36 hour reassessment window.

Use Cases

• Training machine learning models for early risk stratification based on admission and reassessment features.
• Benchmarking predictive performance of different ML algorithms (LASSO, RF, XGBoost) on clinical time-series data.
• Investigating the complementary value of early reassessment data for improving prediction of adverse events.

Strengths

• Model performance was validated in a de-identified timestamp-ordered validation cohort, achieving an AUC of 0.718.
• The dataset includes 684 infection-related PICU admissions, providing a focused cohort for analysis.
• Features were extracted from two distinct time windows (0-6h and 12-36h), enabling a parallel assessment framework.

Limitations

• Row count and column-level documentation are absent, which limits suitability assessment and requires manual inspection after download.
• The model's calibration was noted to attenuate in the later validation cohort, suggesting a need for recalibration for broader application.

Provenance

Source

PhysioNet Paediatric Intensive Care database

Collection Method

Retrospective analysis of clinical records for model development and validation.

Freshness

Last updated 2026-06-04 05:44:16