Improving Ambulance Dispatching with Machine Learning and Simulation

As an industry where performance improvements can save lives, but resources are often scarce, emergency medical services (EMS) providers continuously look for ways to deploy available resources more efficiently. In this paper, we report a case study executed at a Dutch EMS region to improve ambulance dispatching. We first capture the way in which dispatch human agents currently make decisions on which ambulance to dispatch to a request. We build a decision tree based on historical data to learn human agents’ dispatch decisions. Then, insights from the fitted decision tree are used to enrich the commonly assumed closest-idle dispatch policy. Subsequently, we use the captured dispatch policy as input to a discrete event simulation to investigate two enhancements to current practices and evaluate their performance relative to the current policy. Our results show that complementing the current dispatch policy with redispatching and reevaluation policies yields an improvement of the on-time performance of highly urgent ambulance requests of 0.77% points. The performance gain is significant, which is equivalent to adding additional seven weekly ambulance shifts.