Exploring the role of air quality in machine learning predictions of mortality

Wan, Michelle

Exploring the role of air quality in machine learning predictions of mortality

Authors

Wan, Michelle

Publisher

University of Cambridge
Department of Chemistry

Description

Air pollution has well-established risks to human health, ranging from acute to chronic disease, cancers, and mortality. This thesis focuses on the application of machine learning (ML) approaches to predict daily mortality rate in Greater London, UK, where comprehensive datasets are available describing air quality, meteorology, and socioeconomic factors. First, data missingness in high-resolution air quality monitoring data is addressed using graph propagation. Sporadically incomplete air quality datasets must be imputed for downstream use in ML mortality models which require complete data features. The graph propagation algorithm applied here balances imputation power with computational efficiency. Monitoring instruments in the London Air Quality Network (LAQN) are represented by nodes in a graph, connected by edges which propagate information to estimate missing values. Single- and multi-species approaches are introduced and evaluated for robustness under real-world emissions scenarios. Completed LAQN datasets are produced, which are used in the research subsequently presented in this thesis. Neural network models are applied to predict regional daily mortality rates in Greater London, from environmental and socioeconomic input features. Temporal and spatial structure are each explored in the relationships between input features and mortality rate. Temporal dependencies are investigated in mortality prediction models which take regionally aggregated input features. These features describe air quality, meteorology, and socioeconomic factors for the region of Greater London, as multi-day input windows of varying lengths. Comparison between multi-layer perceptron (MLP) and long short-term memory (LSTM) models finds that the capability of LSTMs, to process sequential input windows, confers superior predictive performance (MLP: 9.627 ± 0.965% error; LSTM: 8.802 ± 0.379% error). Varying window lengths reveals insights into time lags which may be epidemiologically meaningful. Feature importance analysis finds that socioeconomic factors are the strongest mortality predictors at both 1-week and 8-week timescales, while the importance of PM10 pollution increases at the 8-week timescale. Spatial dependencies are explored by predicting regional daily mortality rate using borough-level input features describing air quality, meteorology, and socioeconomic factors. Comparison between prediction models finds that borough-resolution inputs (9.353 ± 0.355% error) provide more useful information than the equivalent regionally aggregated data (10.229 ± 1.376% error). However, the application of graph neural network (GNN) models and comparison of GNN edge connection methods suggests that spatial structure does not play an important role in prediction with this input feature set. Spatially varying feature importance analysis finds that socioeconomic factors in Inner London boroughs are the strongest mortality predictors, while pollution is the most significant predictor in Outer London boroughs. Overall, this thesis demonstrates machine learning frameworks which combine environmental and socioeconomic data to improve understanding and modelling of mortality in Greater London. Novel applications to explore temporal and spatial dependencies highlight the importance of both pollution and socioeconomic factors, providing insights for public health management and further environmental health research.
Applications of AI to the study of environmental risk (AI4ER) UKRI Centre for Doctoral Training