A machine learning framework for cardiovascular health prediction modeling the interplay between various environmental, neighborhood and socio-economic features: a German-wide application

Abstract

Environmental exposures and socio-economic neighborhood characteristics have a major impact on human health and well-being. However, little is known about their interplay. Machine Learning (ML) methodologies go beyond the conventional statistical approaches and help us towards identifying the driving contextual factors and assessing their predictive ability for various health outcomes even under high complexity. In this study, we first compared multiple ML techniques, from neighbor-based to deep learning approaches for the prediction of cardiovascular disease (CVD) mortality in 5×5 km grid cells across Germany during 2017. The models performed well in the training phase [R² ≥ 0.85, mean squared error (MSE) ≤ 0.005], and moderate to well in the testing set (0.27 ≤ R² ≤ 0.66, 0.011 ≤ MSE ≤ 0.024). All models were highly correlated (0.69 ≤ Spearman r ≤ 0.82) and identified similar predictors as the main drivers for CVD mortality (e.g., the deprivation index, proportion of foreigners and air pollution), though prediction maps indicated spatial heterogeneity across the country. Currently, we aim to extend this analysis on the prediction of hypertension, an important risk factor for CVD morbidity and mortality, by using advanced and highly resolved environmental maps and recent health data from the largest German cohort, the NAKO study. The work is still in progress and the results will be presented at the conference.