Defining heatwave thresholds using an inductive machine learning approach

Abstract

Establishing appropriate heatwave thresholds is important in reducing adverse human health consequences as it enables a more effective heatwave warning system and response plan. This paper defined such thresholds by focusing on the non-linear relationship between heatwave outcomes and meteorological variables as part of an inductive approach. Daily data on emergency department visitors who were diagnosed with heat illnesses and information on 19 meteorological variables were obtained for the years 2011 to 2016 from relevant government agencies. A Multivariate Adaptive Regression Splines (MARS) analysis was performed to explore points (referred to as "knots") where the behaviour of the variables rapidly changed. For all emergency department visitors, two thresholds (a maximum daily temperature ≥ 32.58°C for 2 consecutive days and a heat index ≥ 79.64) were selected based on the dramatic rise of morbidity at these points. Nonetheless, visitors, who included children and outside workers diagnosed in the early summer season, were reported as being sensitive to heatwaves at lower thresholds. The average daytime temperature (from noon to 6 PM) was determined to represent an alternative threshold for heatwaves. The findings have implications for exploring complex heatwave-morbidity relationships and for developing appropriate intervention strategies to prevent and mitigate the health impact of heatwaves.

Fig 1. Study area and dataset [Data Source: [–23]].

Fig 2. Graphical representation of the MARS model.

Fig 3. Graphical representation of interaction terms from the MARS.