Can we use a machine learning approach to predict the impact of heatwaves on emergency department attendance?

Global warming has contributed to more frequent and severe extreme weather events, which has led to increased research on the health impacts of extreme heat. However, research on heatwaves, air quality, and their spatial impact on health service demand is limited. This study used machine learning (ML) approaches to obtain the optimised model to predict health service demand associated with those risk factors for an all-age model and compared it with young children (0-4 years) model in Perth. Ten years' data (2006-2015) on emergency department attendances (EDA), socioeconomic status (SES), heatwaves, landscape fires, and gaseous and particulate air pollutants were collected. ML approaches, including decision tree, random forest (RF), and geographical random forest (GRF) models, were used to compare and select the best model for predicting EDA and identify important risk factors. Five-hundred cross validations were performed using the testing data, and a construct validation was performed by comparing actual and predicted EDA data. The results showed that the RF model outperformed other models, and SES, air quality, and heatwaves were among the important risk factors to predict EDA. The GRF model was fitted well to the data (R-2 = 0.975) and further showed that heatwaves had significant geographic variations and a joint effect with PM2.5 in the southern suburbs of the study area for young children. The RF and GRF models have satisfactory performance in predicting the impact of heatwaves, air quality, and SES on EDA. Heatwaves and air quality have great spatial heterogeneity. Spatial interactions between heatwaves, SES, and air quality measures were the most important predictive risk factors of EDA for young children in the Perth southern suburbs. Future studies are warranted to confirm the findings from this study on a wider scale.