High rates of unplanned readmission are often viewed as a sign of poor patient care and can adversely affect hospital revenue and prestige. The ability to identify patients who are likely to be readmitted may provide healthcare professionals the chance to arrange post-discharge follow-up which has the potential to reduce the risk of readmissions.

The Solutions for Patient Safety has identified 7-day unplanned readmission metric as an important safety metric in pediatrics. However, there

is paucity of general pediatric models in literature for 7-day unplanned pediatric readmission. The goal of this project is to develop a novel model for enhanced prediction and understanding of unplanned 7-day pediatric readmission using statistical and machine learning algorithms. We also aim to search for and identify novel risk factor predictors of 7-day

unplanned pediatric readmission based machine learning and artificial intelligence algorithms such as random undersampling with boosted regression trees, extreme gradient boosting trees, support vector machines, deep neural networks and other statistical learning algorithms.

Initial results indicate that our 7-day unplanned pediatric readmission models have the highest model performance as measured by the area under the receiver operator characteristic curve, AUC, in literature (AUC: 0.73 [0.71, 0.75]). We currently beat the LACE readmission index (when applied for predicting 7-day unplanned pediatric readmission) by 6 percentage points.

We have also identified several novel risk factors of readmission and expect further improvement in model performance as we tune our machine learning and AI algorithms. Lastly, we will compare the performance of selected statistical learning algorithms establishing which is best suited for predicting pediatric readmissions, and assess the statistical significance of any difference



Author: Louis Ehwerhemuepha

Coauthor(s): Louis Ehwerhemuepha William W. Feaster, MD, MBA, CHIO

Status: Work In Progress