Ahmed Arshad
University of Chicago Medical Center
Chicago, Illinois
Disclosure information not submitted.
Priti Jani, MD, MPH
Assistant Professor
University of Chicago Comer Children's Hospital, United States
Disclosure information not submitted.
Kyle Carey, MPH
Senior Clinical Research Data Manager
University of Chicago, United States
Disclosure information not submitted.
Anoop Mayampurath, PhD
Assistant Professor
University of Wisconsin Madison
Disclosure information not submitted.
Title: Predicting Intensive Care Readmission Among Hospitalized Children
Introduction: Readmission to the pediatric ICU (PICU) is associated with increased morbidity and mortality. The incidence of readmissions to the PICU has been reported to be up to 3.9% Children who are readmitted to the PICU are more likely to have chronic conditions, be younger, or have an unscheduled index admission. Methods to identify the risk of PICU readmission during hospitalization include early warning scores, clinical scoring systems, and the adaptation of adult scoring systems. Studies in adults demonstrate utilizing a machine learning approach is superior to current standards. However, a similar approach has not been developed for pediatric patients. The objective of this study was to derive a machine learning model using data from the electronic health record to predict PICU readmission prior to transfer to the general ward.
Methods: We conducted an observational cohort study on pediatric patients admitted to the PICU between 2012 and 2020 in an urban academic hospital. The primary outcome was defined as readmission to the PICU following transfer to the ward within the same hospital admission. Predictor variables included six commonly used vital signs, patient characteristics, and laboratory results. Data were split into independent derivation (years 2012 – 2017) and prospective validation (years 2018 – 2020) cohorts. We developed logistic regression, elastic net, and random forest models to predict PICU readmission.
Results: There were 433 (3.7%) readmissions to the PICU out of a total 11730 PICU admissions during the study period. The random forest model was the most accurate (AUC: 0.70, 95% CI: 0.65 – 0.74), followed by the logistic regression model (AUC: 0.69, 95% CI: 0.64 - 0.74, P = 0.846) and the elastic net model (AUC: 0.69, 95% CI: 0.64 – 0.74, P = 0.765). Variables most important to predicting PICU readmission within the random forest model include heart rate, respiratory rate, systolic blood pressure, temperature, and diastolic blood pressure. The most important laboratory value was the platelet count, followed by the white blood cell count.
Conclusion: We developed novel models to predict PICU readmission at the time of transfer to the ward. Implementation of this model into clinical practice may help identify patients at risk of readmission and improve outcomes in critically ill children.