Dynamic Emergency Medical Service Dispatch: Role of Spatiotemporal Machine Learning

Abstract

Previous research has suggested that providing prompt access to emergency medical services (EMS) may greatly improve the health outcomes of patients with urgent conditions. However, there has not been enough research on ways in which planning resources for ambulance dispatch may enhance the response time of EMS. GIS has been used to manage and visualize the spatial distribution of EMS demand, but there is still a need for more empirical evidence from spatiotemporal demand-based prediction techniques, such as machine learning. We applied the long short-term memory (LSTM) method to forecast EMS demands based on past records and reallocated service locations using a dynamic maximal covering location model. The training of the prediction models and validation were conducted with 323,993 emergency calls in the Gyeongnam Province in Korea in 2014. We found that conventional hotspot-based emergency dispatch systems, ignoring temporal variations of service demands, could fail to fulfill a desired coverage standard. This study shows an evidence that demand-based spatiotemporal demand prediction and dynamic dispatch protocol based on machine learning algorithm have the potential to support more efficient allocation of resources, especially when resources are limited.