Comparison of machine learning methods for the prediction of cardiovascular mortality from environmental and socio-economic neighbourhood factors

Kurzfassung

BACKGROUND AND AIM: Several environmental and socio-economic neighbourhood factors have been reported to majorly impact human health, but their interplay is not yet well understood as simultaneous analyses are sparse. We aimed to identify the driving contextual factors and assess their predictive ability for cardiovascular disease (CVD) mortality in a neighbourhood setting by comparing several machine learning methods. METHODS: We obtained the number of CVD deaths per county for entire Germany for 2017. Most socio-economic factors (e.g. proportion of unemployed, foreigners, household income, deprivation index) were also only available on a county level, whereas most environmental factors (e.g. imperviousness, greenness, air pollution, noise, building density) could be gained on a higher spatial resolution (10m-2.4km). All data was aggregated to a 5km*5km grid to compile compliant prediction maps. In addition to traditional linear and additive regression models, we applied one neighbour-based method and several statistical, ensemble and deep learning approaches using a random search strategy for hyper-parameter tuning. Variable importance was assessed by SHapely Additive exPlanations (SHAP) values. RESULTS: CVD mortality for the 5km grid ranged from 2.5 to 8.1 per 10,000 residents with a mean of 4.7 (standard deviation 1.0). The models performed well in the training phase with R² between 0.85-1.00, mean squared error (MSE) between 0.001-0.005, and moderate to well in our test data (R²: 0.27-0.66; MSE: 0.011-0.024). Most models identified the deprivation index as the most important predictor followed by the proportion of foreigners, unemployed, median income and air pollution. Comparison of our predicted and observed CVD mortality rates showed high correlations for all models (Spearman r: 0.69-0.82), though prediction maps indicated heterogeneity in spatial patterns. CONCLUSIONS: The selected approaches differed in their predictive ability but identified similar predictors as the main drivers for CVD mortality.