Hosted by the Interactive AI Centre for Doctoral Training
Abstract: In the wake of the 1952 Great Smog of London, which brought widespread awareness to the health impacts of air pollution exposure, the field of air quality research has evolved significantly. Today, applied machine learning (ML) techniques have showcased the ability to handle vast input data, offering a comprehensive description of health factors without the need for prior assumptions. In this work, we developed predictive regional ML models for daily mortality in Greater London. Input features consisted of nitrogen dioxide concentrations, temperature, and income. We compared linear and neural network regressor models, excluding input features in turn to assess their roles as mortality predictors. Further work aims to apply graph network architectures, in which nodes represent London boroughs and the model is trained to predict mortality. Using the trained graph, we can examine the interplay between air quality and socioeconomic inequity as mortality predictors at the borough level. To obtain consistent pollution estimates at these spatial resolutions, we applied graph networks to data recorded by the London Air Quality Network (LAQN). Leveraging edge connections between monitoring stations, we predicted missing data points in the LAQN time series data, using land use records to contextualise edges and explain predictions.
Contact <iai-cdt@bristol.ac.uk> to reserve a space (for catering purposes). Lunch at 13.30 will be followed by the talk at 14.00.