Using generalized random forests to characterize vulnerability to adverse health outcomes following wildfire smoke...

As the health burden attributable to wildfire activity increases under climate change, it is crucial to determine which subgroups face heightened vulnerability to wildfire smoke. Marginalized communities may experience disproportionate risk from overlapping individual and community vulnerability factors. We leverage recent developments in machine learning methods for high-dimensional settings to construct detailed profiles of California communities disproportionately impacted by wildfire smoke across 27 potential effect modifiers.

We used daily 2006-2019 data on hospital admissions and emergency department (ED) visits for cardio-respiratory diseases in California. We applied a time-stratified case-crossover study design to analyze the effect of wildfire-specific fine particulate matter (PM_2.5) on cardio-respiratory diseases. Then, we investigated heterogeneous effects using a generalized random forest approach, which can handle a large set of individual-level (age, sex, race/ethnicity) and area-level (e.g., poverty level, racial/ethnic segregation) factors to construct vulnerability profiles for each Air Basin, representing areas with similar meteorological and geographic conditions.

A 10 µg/m³ increase in wildfire PM_2.5 concentration (2-day moving average) was associated with an increased risk of hospital admissions and ED visits related to respiratory diseases (OR = 1.014, 95 % confidence interval = 1.012-1.016). No association was found for cardiovascular diseases. Associations between exposure to wildfire PM_2.5 and respiratory diseases varied strongly by individual- (age, sex, race/ethnicity) and area-level factors (such as A/C prevalence, Black/White dissimilarity index). The importance of these effect modifiers, and vulnerability profiles, changed across Air Basins.

Machine learning can characterize the complex heterogeneity in wildfire smoke-related health impacts.