Predicting Wildfire Smoke-Associated Acute Care Utilization in Washington State Using Observed PM2.5 Data

Abstract

Background: Wildfires are increasing in frequency and severity across the Pacific Northwest, and wildfire smoke (WFS) is the primary driver of rising PM2.5 concentrations in Washington State. WFS exposure is associated with increased cardiorespiratory acute care utilization, yet Washington remains understudied compared to other heavily impacted regions. Climate adaptation requires health systems to develop practical, real-time forecasting tools to anticipate shifting acute care utilization patterns during WFS events.Objectives: This study has two aims: (1) to assess whether a simple WFS exposure classifier derived from real-time public air monitor data adequately approximates a validated satellite- and ensemble model-based smoke-day classifier (Childs et al., 2022); (2) to determine whether this classifier can support prediction of cardiorespiratory acute care encounters in the week following WFS exposure among commercially insured Washington State residents. Methods: For Aim 1, we used daily PM2.5 data from air monitors across Washington State (2010–2018) to construct a binary WFS exposure classifier. We validated classifier performance against Childs et al. estimates using standard classification metrics. For Aim 2, we used longitudinal commercial claims data (2015–2018) from the Merative MarketScan® Research Databases to identify a study population of 83,350 enrollees with active asthma or COPD diagnoses. We trained LightGBM gradient boosting frameworks to predict cardiorespiratory ED visits and inpatient admissions within one week of smoke exposure. Results: The simple exposure classifier achieved recall of 0.84, precision of 0.62, F1 of 0.71, specificity of 0.98, and MCC of 0.71 against Childs et al. estimates, demonstrating strong alignment with the validated ensemble method. Conversely, both the ED and inpatient LightGBM models failed to discriminate between positive and negative cases (AUPRC = 0.001 for both). Conclusions: PM2.5 monitoring data can adequately approximate more resource-intensive WFS exposure methods in Washington State, offering a practical foundation for climate-informed health systems planning. However, these results indicated individual-level prediction of acute care utilization following WFS exposure may not be supported using commercially insured claims data. Structural limitations including low outcome prevalence, absence of environmental confounders, and underrepresentation of high-risk groups constrained our approach. Future climate-resilient health systems research in Washington State should use syndromic surveillance data and aggregated count-based outcomes to maximize predictive power.