Risk Factors for Pediatric Acute Chest Syndrome Utilizing Machine Learning Techniques

Abstract

Introduction: Acute chest syndrome (ACS) causes significant morbidity and mortality in both adult and pediatric sickle cell disease (SCD). Fifty to thirty percent of all pediatric SCD admissions develop ACS, with each episode potentially causing irreversible lung damage. Although there are some known risk factors for developing ACS (e.g., history of asthma), often this complication occurs unexpectedly. Our hypothesis was that a machine learning model could be developed to predict progression to ACS. Methods: A retrospective cohort study was designed to develop this prediction model. Data was abstracted from the electronic health record (EHR) from a single institution for all pediatric SCD admissions from 0-21 years old between June 1, 2016, and December 31, 2022. Inputs into the model included demographics (e.g., age, genotype), vital signs (e.g., heart rate, oxygen saturation), laboratory values (e.g., hemoglobin, neutrophil count), medication orders (e.g., albuterol, oxycodone) and additional co-morbidities (e.g., asthma, obstructive sleep apnea). The inputs had maximum and minimum values based on normal ranges by age. The output was a diagnosis of ACS as a continuous dependent variable to develop a continuous likelihood estimation score from 0 to 1. A positive output of ACS was defined as 2 doses of the antibiotic azithromycin and a chest x-ray ordered in the same encounter. The negative output included all patients who did not meet criteria for a positive output as defined above. Both positive and negative cohorts for ACS were included in the model. The time of ACS diagnosis was when the first dose of azithromycin was ordered. All ACS encounters were verified by chart review. A random forest model was implemented using Sklearn Python library. Different depth of trees was experimented between 1 to 24. A prediction for ACS was calculated every 6 hours after admission to the inpatient setting. Eighty percent of the data was used for training and twenty percent for testing. Results: Two thousand two hundred and sixty encounters were included in the negative cohort and 512 encounters were included in the positive cohort. One thousand four hundred and fourteen (51%) were male patients and 1358 were female patients (49%). The maximum depth of the tree was determined to be 10. Sixty-five percent of the positive cohort developed ACS 24 hours after the admission time. The median time that the random forest model predicted ACS was 18 hours before the diagnosis was determined in the chart. The model had a sensitivity of 80% and specificity of 64% at the cutoff point of 0.17. The area under the receiver operating characteristic (ROC) curve was 71%. The model's negative predictive value was 92% and the positive predictive value was 36% of diagnosing ACS. The features that were important predictors in the model were age, height, weight, pulse, respiratory rate, presence of hypoxia, temperature, neutrophil count, hemoglobin, absolute reticulocyte count and reticulocyte percent. Conclusion: A machine learning model demonstrated markers of hemolysis and vital sign changes were important to predict ACS in pediatric patients. This is the first model utilizing machine learning techniques in pediatric SCD to predict ACS. Detecting patients who are a higher risk for ACS could impact treatment options and approach to inpatient management.