Identifying Children at Readmission Risk: At-Admission versus Traditional At-Discharge Readmission Prediction Model.

Area of Science:

• Pediatric healthcare outcomes research
• Clinical informatics
• Machine learning in medicine

Background:

• Pediatric readmissions are a significant concern, with many occurring within a week of discharge.
• Current prediction models are typically developed post-discharge, limiting intervention windows.

Purpose of the Study:

• To develop and evaluate a novel pediatric readmission prediction model at the time of hospital admission.
• To compare the performance of at-admission models against standard at-discharge models.

Main Methods:

• Utilized the Hospital Cost and Utilization Project database (Florida, 2016-2017).
• Developed four machine learning algorithms (logistic regression, decision tree, SVM, Gradient Boosting) for both at-admission and at-discharge predictions.
• Employed recursive feature elimination and cross-validation; performance measured by Area Under the Curve (AUC).

Main Results:

• At-admission models demonstrated performance comparable to at-discharge models across all algorithms.
• Support Vector Machines (SVM) with a Polynomial Kernel achieved superior performance for both prediction timings.
• Key risk factors included patient demographics, social determinants, clinical factors, and hospital characteristics.

Conclusions:

• An at-admission pediatric readmission prediction model provides a valuable decision support tool.
• Early risk identification allows for proactive intervention planning, especially addressing social determinants of health.