Using Graph Neural Networks to Investigate the Relationship Between the Socioeconomic Factors and Emergency Medical Service (EMS) Median Response Time in New York City

In this working-in-progress study, we plan to use Graph Neural Networks (GNNs) to investigate whether and how socio-demographic factors influence p redicted r esponse t imes of EMS services in New York City (NYC). Currently, the application of GNNs to predict the EMS response time is relatively novel. Operational models, which prioritize task-specific features (such as call priority level and day/time of incident) have been deployed in several contexts around the world. Leveraging unique capabilities of different emergent GNN architectures, we will evaluate whether the predictive accuracy of operational models of EMS response time are improved when neighborhood-level socio-economic factors are included. Focusing on different neighborhoods of NYC, we plan to use EMS Incident Dispatch Data obtained from the Fire Department of New York (FDNY) and accessed through the NYC Open Data Portal. We couple incident data with socio-economic factors at the neighborhood level, such as average income, population density, poverty level, and percentage of the population speaking a second language, obtained from the U.S Census Bureau and other sources. Our starting assumption is that response time is associated with different operational features (such as Call Priority Level) and that these features are consistent across neighborhoods. We then represent the call Priority Level Set and Neighborhood Set in a graph G = (U,V,E). The U and V are the node sets for Priority Levels (PLs) and Neighborhoods (NBs), respectively. The relationships or the links between the nodes U and V are represented by the edge set denoted as E, which is a subset of U × V. If relationships between call Priority Level and median response time fluctuate a cross n eighborhoods, i t s uggests that extra-operational factors influence EMS response times, and that there are inequities in response shaped by the socio-demographic features of where the call originated. We explore how we can leverage the novel techniques of Graph Neural Networks in this field of study compared to previous work done using traditional machine learning techniques by the research community.