How can machine learning extend the investigation of relationships between air pollution, meteorological conditions and Covid-19 health data in Baden-Württemberg (Germany)?

Kurzfassung

Background: Our main objective is to explore the associations between environmental stressors and Covid-19 health data. In a first step, we analyzed the connections between surface-level meteorological and air pollution variables, such as particulate matter (PM10, PM2.5), nitrogen dioxide (NO2), ozone (O3), precipitation, air temperature, vapor pressure, and ultraviolet radiation. Secondly, we will use the initial analysis results to establish a link between mortality and severity of COVID-19 infections and environmental stressors. Using earth observation data, environmental stressors are available on a daily level. Spatially, we consider postal code areas in Germany's federal state, Baden-Württemberg, which includes urban and rural areas with mountains and rivers. Methods: We utilized data from the Copernicus Climate Change Service and Copernicus Atmosphere Monitoring Service to model environmental exposure in Baden-Württemberg. We performed a temporal analysis using cross-correlations and a spatial analysis using the Local Indicators of Spatial Association (LISA) method. The next goal is to establish a generalized additive model (GAM) that relates COVID-19 infections to environmental stressors based on the first analysis. Machine learning methods can optimize prediction models and assess the impact of individual environmental stressors on the number of COVID-19 infections in terms of prediction accuracy and reliability. Results: The first analysis shows clear spatial and seasonal patterns between meteorological conditions and air pollution. NO2 and O3 show a strong interdependent relationship with varying Pearson correlation coefficients over time. A negative correlation was observed in January (-0.84), April (-0.47), and October (-0.54), while a positive correlation was observed in July (0.45). The correlation direction for O3 and NO2 exhibited a noticeable change in the cross-correlation plot. Urban areas had higher concentrations of NO2, PM2.5, and PM10 than rural areas, while O3 showed the opposite trend. The LISA method confirmed the presence of distinct cold and hot spots for the environmental stressors. Discussion: We analyzed the relationship between air pollution and meteorological conditions and identified a set of variables to prioritize for future health impact analyses. Our data basis covers both spatial and temporal information. Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks are particularly suitable for capturing temporal dependencies (in our case, quarters). On the other hand, Convolutional Neural Networks (CNNs) can recognize spatial patterns (in our case, within the postal code areas). Is it appropriate to use an ensemble model incorporating both RNN/LSTM for the temporal information and CNN for the spatial patterns or a random forest model for our project? Conclusion: Machine learning methods are a powerful tool for extending classical statistical analysis and prediction.