475Machine learning approach: identifying the impact of heatwaves and air quality on children’s health

Background Heatwaves, air pollution and their effects on children’s health can vary temporally and spatially. With the emergence of advanced methods such as machine learning, there is an opportunity to improve prediction of children’s health events associated with those exposures. Methods Daily records on emergency department attendances (EDA) for children <15 years, heatwaves, landscape fire burns and air pollutants (CO, SO2, NO2, O3, PM10, PM2.5) were collected for Western Australia, 2006-2015. Decision tree, random forest (RF) and geographical RF (GRF) were compared to predict EDA, identify important risk factors and locations at elevated risk. Validation was performed by comparing actual and predicted EDA. Results RF was the best model with the lowest root mean squared error (MSE). The best RF validation model had an r-squared (R2) =0.95. The percentage increase in MSE indicated that PM10 and PM2.5 were important predictors of EDA for all children. Number of burns was more important in 5-9 year age group than other groups. GRF models (R2 0.90-0.98) showed that heatwave and PM2.5 were the important predictors in southern part of the study area for all age groups. Conclusions The importance of risk factors to predict EDA was varied by age groups and locations. Such differences are important when developing targeted health promotion strategies tailored to age groups and geographical locations. Key messages RF predicted EDA better than other models. Evaluation of spatial variation of heatwave and air quality effects on EDA for children by GRF modelling is useful to identify at risk geographical locations.