Investigating the interplay of environmental, neighborhood and socio-economic features to predict cardiovascular health: a nationwide machine learning framework

Abstract

Although there is strong evidence that human health and well-being are strongly associated with the exposure to several environmental variables as well as our socio-economic and neighborhood settings, yet their interplay is not adequately analyzed and consequently not well understood. We thus aimed to build a Machine Learning (ML) framework, able to sufficiently identify the driving contextual factors for various health outcomes. We first compared traditional regression approaches such as linear regression with multiple ML approaches, from neighbor-based algorithms to ensemble and deep learning methodologies, to predict cardiovascular disease (CVD) mortality across Germany in 5 km grid cells for 2017. The performance of all models was good for the training stage, resulting in R2 higher or equal to 85% and mean squared error (MSE) smaller or equal to 0.01. In the testing stage, R² dropped in the range of 27% to 66%, depending on the model, while the errors remained quite low, i.e., MSE smaller or equal to 0.02. The predicted CVD mortality rates of the different models were highly correlated, and the models identified similar main predictors (e.g., deprivation index, proportion of foreigners, unemployed, median income and air pollution). The predictions captured the North-East to South-West CVD mortality trend, i.e., highest to lowest gradient, but showed countrywide spatial heterogeneity when mapping. In our ongoing work, we aim to extend these prediction models by adding environmental maps of higher resolution and individual information from participants of the German National Cohort (NAKO), to investigate the additional influence of individual risk factors for hypertension, an important risk factor for CVD morbidity and mortality. First results will be presented at the conference.