Files
Abstract
Out-of-Hospital Cardiac Arrest (OHCA) affects over 350,000 Americans annually, yet survival rates remain below 6%. Identifying causal factors and developing comprehensive interventions are critical for improving outcomes. Existing research typically falls into two categories: (1) traditional geospatial analysis, which correlates risk factors and survival outcomes but often provides limited insight into underlying mechanisms, and (2) machine learning (ML), which identifies various risk factors but often overlooks geospatial disparities. Emerging Geospatial Artificial Intelligence (GeoAI) shows that incorporating geospatial variables can significantly enhance ML performance. However, in the domain of geospatially explainable AI (GeoXAI), there is still no comprehensive framework for estimating causal effects in health geography—an essential step for devising effective regional interventions. In this dissertation, we propose a comprehensive framework designed to improve survival outcomes for OHCA patients. This framework comprises three key components, each corresponding to a chapter. First, we introduce the Overlayed Spatio-Temporal Optimization (OSTO), a spatio-temporal placement optimization method that maximizes coverage for potential OHCA patients by accounting for spatiotemporal heterogeneity, we apply this into Washington D.C. and demonstrated an improved OHCA coverage. Second, we present Spatial Counterfactual Explainable Deep Learning (SpaCE), which goes beyond AED placement optimization by exploring additional risk factors that affect survival outcomes. Using a Georgia OHCA outcome prediction task as an example, SpaCE uncovers how various risk factors—and the extent to which they correlate with OHCA survival—vary across different locations. Finally, we introduce Spatially-Aware Causal Inference (SpatialCausal) to estimate causal effects for each treatment variable across space. This method achieves an improved performance over baseline approaches in the Georgia case study. Two standing out variables—AED usage prior to EMS arrival and the identity of the OHCA witness—show a positive treatment effect on survival outcomes, indicating that expanding AED training programs and ensuring more professional individuals can significantly enhance survival chance. Through this framework, we improved OHCA survival by AED placement optimization and the causality estimation of risk factors. The flexibility of this framework demonstrates adaptability for broader geospatial health applications. The methodological advances also serve as a valuable reference for health geographers integrating GeoXAI.