Machine Learning Models for Predicting Pediatric Hospitalizations Due to Air Pollution and Humidity: A Retrospective Study

This study aimed to better predict pediatric hospitalizations based on environmental exposure to nitric oxides (NOx), particulate matter (PM), and relative humidity (RH) and clinical features thereby supporting public health planning and awareness of populations at risk and exposures of which to be aware.This study included a retrospective analysis of 2500 children aged 1 year to 18 years presenting to the emergency department (ED) for respiratory symptomatology from 2016 to 2017 in Hadera, Israel, with prioritization for the data from 2016 due to missing data from 2017. Due to implications of viral infections and attempting to separate bronchiolitis from other pulmonary etiologies, age groups were created for those older than and younger than age 2 years. Of the ED presentations, 41.6% were younger than 2 years and 60.1% were male individuals, with 452 hospitalizations for all.Records were reviewed using International Classification of Diseases, Ninth Revision diagnosis for respiratory symptoms. Environmental data were obtained from 9 regional monitoring stations specifically for PM, NOx, RH, and temperature with separation for polluted days and clean days throughout all seasons. Demographics and environmental factors were entering into 2 modeling options (Random Forests [RF] and eXtreme Gradient Boosting [XGBoost]) to create decision trees and then training the models upon the parameters to improve performance and better provide insights for health care professions in their decision-making.Increased ED presentations (4.8 arrivals/d) were noted on highly polluted days with greater than the 90th percentile for PM2.5, NOx, and SO2 (53 days) compared with nonpolluted days with less than the 10th percentile (30 days). These observations were consistent across all seasons, except for summer, because there was insufficient polluted vs nonpolluted days to draw conclusions. Higher hospitalization rates were seen in the winter months; however, environmental PM was still an important predictor. RH was assessed 6 hours prior to presentation, with lower RH exposure seen in hospitalized patients. The authors then took these data and used them to train different machine learning models to enable predictive modeling with RF and XG. RF indicated NOx, RH, and SO2 levels 6 hours prior to arrival were the most contributing factors. XG indicated the same environmental factors as well as chief concerns, with cough/hemoptysis as important predictive features. XG did have lower reliability due to lower recall issues.This study shows the importance of environmental exposures on pediatric ED presentation and hospitalizations for respiratory complaints in a retrospective fashion, specifically analyzing for lower RH and higher PM including NOx and sulfur dioxides in the 6 hours prior to presentation. These details were able to train machine learning models to rapidly assimilate multiple key data points for prediction purposes.Artificial intelligence and machine learning are ongoing clinical decision-making tools. This retrospective study shows how key data points are identified and then predictive modeling must be trained to enable predicting hospitalization. Machine learning has the capability to simultaneously assimilate and assess multiple key data points to identify higher risk times and exposures and be able to drive data-supported resource allocation and clinician decision-making.