Predicting Changes in Pediatric Medical Complexity using Large Longitudinal Health Records

Area of Science:

• Pediatric Healthcare Analytics
• Clinical Informatics
• Predictive Modeling in Medicine

Background:

• Medically complex pediatric patients utilize significant healthcare resources.
• Sub-optimal clinical outcomes persist despite high resource consumption.
• Predicting complexity changes can optimize care and resource utilization.

Purpose of the Study:

• To model and predict changes in medical complexity status for pediatric patients.
• To identify factors influencing shifts in patient complexity.
• To improve care quality and reduce hospital resource utilization.

Main Methods:

• Utilized a dataset of 226,000 pediatric patients over five years from Children's Healthcare of Atlanta.
• Compared logistic regression, random forest, gradient boosting trees, and multilayer perceptron models.
• Predicted complexity status changes in the final year based on prior data.

Main Results:

• Achieved an 88% area under the ROC curve (AUC) for predicting non-complex patients becoming complex.
• Attained a 74% AUC for predicting complex patients remaining complex.
• Identified key factors associated with patient complexity changes.

Conclusions:

• Predictive models show promise in forecasting pediatric patient complexity shifts.
• Accurate prediction can guide interventions to improve outcomes and resource management.
• Understanding influencing factors is crucial for targeted healthcare strategies.