A machine learning framework for modeling the associations between various environmental features and health

Kurzfassung

Abstract: Introduction: Human health has been associated with the exposure to several environmental variables as well as socio-economic and neighborhood settings. Yet, their interplay is not adequately analyzed and consequently not well understood. We aimed to build a machine learning (ML) pipeline, able to sufficiently identify the driving environmental, socio-economic and individual factors for cardiovascular health. Methods: For our use case, we included midterm data from the baseline examination of the largest German population-based cohort study NAKO conducted between 2014-19 in 18 study centers. We assigned environmental exposures (e.g. air pollution, air temperature, greenness) and neighborhood factors (e.g. urbanization) based on the participants’ residential addresses. We compared traditional regression approaches (Linear Regression, Elastic Net) with multiple ML algorithms, e.g. neighbor-based methods (K-Nearest Neighbour), Statistical Learning (Support Vector Machine), Ensemble Learning (Random Forest, XGBoost) and Neural Networks to identify the main risk factors for hypertension. Results: Of 101,601 participants included in our analysis, 45% were classified as hypertensive. Most models performed well and comparable with an accuracy ranging from 0.68 (K-Nearest Neighbour) to 0.73 (Support Vector Machine) in our test set. The different approaches identified similar factors as the main drivers for hypertension with highest feature importance for individual characteristics (age, Body Mass Index, sex) followed by environmental (non-optimal temperature, air pollution) and individual socio-economic (income, education) factors. Further neighborhood socio-economic factors (e.g. deprivation, household income) are currently assigned and will be incorporated in the next runs. Discussion: The ML pipeline shall be openly accessible soon for use in epidemiological analysis, specifically with binary health outcomes, but we also plan to incorporate continuous outcomes.