AI-driven environmental epidemiology: a generalisable framework for predicting blood pressure using fair and explainable machine learning

Kurzfassung

Understanding the relationship between environmental exposures, demographics, individual characteristics and health outcomes is crucial in the context of a changing climate. ML methods provide promise to model these complex relationships, but their application in environmental epidemiology is under-utilised. This study presents a generalisable supervised learning regression framework to investigate the joint impacts of environmental exposures, contextual and individual socioeconomic factors and demographics, and clinical data on key continuous health outcomes. The approach is illustrated using blood pressure, a key risk factor for cardiovascular disease (the world’s leading cause of mortality and morbidity). Blood pressure has also been shown to be impacted by environmental exposures. We leverage the NAKO data (health, demographic, and socio-economic data from 205,000 participants collected through extensive health examinations and questionnaires at nationwide study centres) linked to high-resolution environmental data. We evaluate and compare a variety of traditional statistical regression methods with ML approaches, including ElasticNet, XGBoost, Neural Networks, Random Forest, Generalized Linear Models, and Support Vector Regression. The predictive performance, explainability, and fairness of the models is compared, providing insight into their ability to capture complex interactions. This is done using performance metrics such as MSE, as well as SHAP values, permutation feature importance, and fairness metrics across demographic groups. Cross-validation is used to ensure robustness. This framework builds on the Helmholtz AI funded Noise2NAKOAI project, in which a similar framework was developed for binary health outcomes. Expected Outcomes - ML framework for predicting continuous cardiovascular metrics from environmental and health data. - Comparative analysis of predictive performance, explainability, and fairness across statistical and ML-based models. - Insights into relationships between environmental exposures and blood pressure. Analysis is ongoing; results will be reported at the conference. This research contributes to the advancement of AI in environmental epidemiology, deepening the understanding of associations between environmental exposures, socioeconomic factors, and blood pressure. This generalisable methodology may be used to investigate a wide range of health outcomes, resulting in downstream impacts on public health policy.