Investigation of the long-term effects of environmental factors on human health using AI/ML methods.
Helmholtz AI project call showcase: AI Methods linking Environment and Health - a Large-scale Cohort Application
The Helmholtz AI project Noise2NAKOAI, a collaboration between the German Aerospace Center (DLR) and Helmholtz Munich (HMGU), investigates how noise in our environment affects human health. Read more about this Helmholtz AI project in today’s showcase.
Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?
My name is Kathrin Wolf, I work at the Institute of Epidemiology at Helmholtz Munich in the field of environmental epidemiology, environmental exposure modelling and geographic information systems and I am the PI of the project “Noise2NAKOAI: AI Methods linking Environment and Health - a Large-scale Cohort Application”.
In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?
We apply AI/ML methods to investigate the long-term impact of environmental factors on human health. In our case study, we develop extensive noise maps for entire Germany and link them with neighbourhood and health data of participants of the German National Cohort (NAKO is one of the largest longitudinal studies with overall 200,000 participants) to identify vulnerable clusters for the risk of hypertension and cardiovascular mortality. In the next step, we enhance this network by individual socio-economic and health data to investigate the interplay of different risk factors on hypertension. That is, we explore and expand different ML techniques considering two approaches: i) applying inherently interpretable algorithms (e.g., XGBoost, random forests) and ii) employing explainable AI (XAI) techniques (e.g., SHAP, Lime) to interpret results of accurate but not interpretable algorithms (e.g., CNNs, DRNs). Finally, we will compare our findings to traditional models (e.g. linear/logistic regression, Generalised Additive Models (GAM)).
The exploration of several ML approaches and adaptation to our highly heterogeneous data structures is very challenging and time-consuming and we currently do not know if the new and computation-intensive methods really outperform our traditional approaches in modelling the interactions between environmental exposures, neighbourhood and individual characteristics but also explaining their respective effects on health.
How important has the Helmholtz AI funding and platform been to carry out this project?
Without the Helmholtz AI funding, this project and especially the collaboration with our partners at DLR and HMGU/ICT would not exist, since we basically started the collaboration with this project. We are really happy to work in this interdisciplinary project together since it not only broadens our horizons, but already enables further smaller joint projects. We are also planning to submit a joint follow-up proposal. Moreover, we highly appreciated the help from Helmholtz AI consultants via a realisation voucher with regard to explainable AI and Semi-Structured Deep Distributional Regression.
Figure: Complete logo of the Noise2NAKOAI project.