An ensemble boosting model for predicting transfer to the pediatric intensive care unit

Jonathan Rubin, Cristhian Potes, Minnan Xu-Wilson, Junzi Dong, Asif Rahman, Hiep Nguyen, David MoromisatoInt. J. Medical Informatics 2018 · DOI

Introduction

In hospital settings, approximately 1% to 3% of pediatric patients in general wards deteriorate to the point of requiring transfer to a Pediatric Intensive Care Unit (PICU). Early detection of clinical decline is critical for improving outcomes, yet existing Early Warning Score (EWS) systems face significant challenges in pediatric care.

Traditional systems, such as the Pediatric Early Warning Score (PEWS), often rely on subjective manual observations—like respiratory effort or capillary refill—which are prone to inter-observer variability and difficult to automate. Furthermore, pediatric vital signs are highly age-dependent; a heart rate that is normal for an infant may indicate severe distress in a teenager. To address these hurdles, researchers developed an automated, objective, and age-aware ensemble machine learning model using Electronic Health Record (EHR) data.

Methodology

The study utilized 5.5 years of EHR data from two medical centers: Banner Desert Medical Center (for training and internal testing) and Banner Thunderbird Medical Center (for independent validation). The researchers focused on 10 objective features: six direct vital signs (HR, O2 Sat, RR, Temp, dBP, sBP), patient age, and three derived measures (Pulse Pressure, Mean Arterial Pressure, and Shock Index).

The proposed ensemble approach combines two distinct boosting variants:

  1. Modified AdaBoost-abstain: This algorithm uses simple decision stumps modified to handle pediatric physiology by learning age-dependent risk thresholds within predefined age brackets. It also handles missing data by "abstaining" from voting when specific features are absent.
  2. Extreme Gradient Boosting (XGBoost): A high-performance implementation that treats age as a standard feature, allowing the model to learn complex, non-linear interactions between age and vital signs.
  3. The Ensemble: The final prediction is a simple average of the probabilities generated by the two models.

Key Results

The ensemble model consistently outperformed the modified PEWS baseline across all primary metrics. On the Desert facility test set, the Ensemble achieved an Area Under the Receiver Operating Characteristic (AUROC) curve of 0.84, compared to 0.75 for PEWS. Notably, the Ensemble demonstrated significantly higher sensitivity (0.83 vs. 0.73), making it more effective at identifying patients at risk of deterioration.

Performance metrics for the Desert facility test set, showing the Ensemble's superior sensitivity and AUROC.

Performance remained robust when applied to the independent Thunderbird facility dataset, where the Ensemble achieved an AUROC of 0.85. This suggests that the model captures fundamental physiological signals of deterioration rather than hospital-specific charting habits.

ROC curves for both facilities, illustrating the Ensemble (green) consistently outperforming individual models and the baseline.

Discussion and Limitations

By removing subjective inputs and utilizing age-specific thresholding, the model provides a reliable, automated background process for clinical monitoring. However, several areas for future work remain. The current model utilizes a "snapshot" of the most recent vitals; incorporating temporal trends over 6–12 hours could further enhance accuracy. Additionally, while the model generalized well between two regional hospitals, further validation on geographically diverse and international datasets is necessary to ensure broader clinical utility.

Summary updated

BibTeX
@article{rubin2018ensemble,
  author = {Jonathan Rubin and Cristhian Potes and Minnan Xu-Wilson and Junzi Dong and Asif Rahman and Hiep Nguyen and David Moromisato},
  title = {{An ensemble boosting model for predicting transfer to the pediatric intensive care unit}},
  journal = {Int. J. Medical Informatics},
  year = {2018},
  doi = {10.1016/j.ijmedinf.2018.01.001},
}
Figure