Predicting Future Care Requirements Using Machine Learning for Pediatric Intensive and Routine Care Inpatients

Area of Science:

• Computational medicine and health informatics
• Pediatric critical care prediction modeling

Background:

• Accurate prediction of future healthcare needs is crucial for effective resource allocation in pediatric hospitals.
• Existing models may not adequately capture the dynamic nature of critical care requirements in children.

Purpose of the Study:

• To develop and compare distinct machine learning models for predicting intensive care unit (ICU) versus non-ICU care for hospitalized children.
• To assess model performance across four future time intervals: 6-12, 12-18, 18-24, and 24-30 hours.
• To validate models using an independent patient cohort and a simulated children's hospital environment.

Main Methods:

• Utilized the Health Facts database (Cerner Corporation) for predictive modeling, including patients with ICU admissions and routine care patients.
• Developed and validated models using training, validation, and testing sets derived from historical data (2009-2016).
• Employed four independent, sequential, fully connected neural networks calibrated to predict the risk of ICU care at specified future time points.

Main Results:

• Models demonstrated strong performance in the test sample, with high sensitivity, specificity, accuracy, and area under the receiver operating characteristic curves (AUROC).
• Performance remained robust in the independent 2017-2018 cohort, with sensitivity ≥ 0.545, specificity ≥ 0.972, accuracy ≥ 0.921, and AUROC ≥ 0.946.
• Comparable performance metrics were observed across simulated and real-world hospital settings, regardless of teaching status, bed capacity, or geographic location.

Conclusions:

• Machine learning models integrating physiology, therapy, and care intensity show promising performance in predicting future pediatric care needs.
• Model performance remained consistent and reliable across increasing prediction time horizons, from 6-12 hours to 24-30 hours.
• These predictive models can potentially enhance clinical decision-making and optimize resource allocation in pediatric healthcare settings.