Constructing Community- and Outcome-specific Weather and Air Health Risk Indices Using Explainable Machine Learning: A Multicentre Study in 66 Australian Communities.

Abstract

Existing studies rarely assess and capture both the mortality and morbidity risks from the joint exposures of meteorological factors and air pollution, while also accounting for between-communities variations. We developed a Weather and Air Health Risk Index (WAHRI) to quantify and communicate short-term combined and community-specific mortality and morbidity risks in relation to multiple weather and air stressors. Daily counts of deaths, hospital admissions, and emergency department (ED) visits were collected from 2014 to 2019 across 66 Statistical Area Level 3 (SA3) communities in Victoria, Australia. Health outcomes were SA3-specific daily counts of all-cause, cardiovascular disease (CVD)-related, and respiratory disease-related mortality, hospital admissions, and ED visits. SA3-specific daily average weather and air pollution variables included temperature, relative humidity, air pressure, ultraviolet B radiation, wind speed, rainfall, fine particulate matter and daily maximum 8-hour ozone. Outcome-specific random forest models were trained to estimate health risks associated with combined environmental exposures, and SHAP values were used to construct interpretable WAHRI indices. We found that higher WAHRI was consistently associated with increased risks across all-cause, CVD and respiratory disease-related mortality, hospital admissions and ED visits. Among the included environmental factors, temperature and ultraviolet B radiation were the leading contributors across most outcomes. Although the geographical distributions of WAHRI varied across outcomes, high-risk days were more frequent in socioeconomically disadvantaged, lower-GDP, and communities with older populations. WAHRI provides an interpretable, outcome-specific framework for assessing combined weather-air pollution health risks and will support short-term environmental health surveillance and alerts, public risk communication, and health-service preparedness.