Improving Emergency Medical Services Using Simulation And Machine Learning

Abstract

Emergency Medical Services (EMS) are critical to public health, yet they face persistent challenges, particularly during periods of surge. This thesis advances the integration of machine learning (ML) and simulation methodologies to enhance EMS decision-making, resource allocation, and system resilience. First, a systematic literature review is conducted to investigate how simulation and ML can be combined in healthcare. This literature review provides a detailed categorization and research gaps. Second, an unsupervised ML clustering model based on genetic algorithms is developed to define EMS surge levels. This model objectively categorizes surges using EMS operational data without relying on subjective data. The model is adaptable, allowing for facilitation of regional customization and generalizability. A case study is conducted to test this model. The results show that the proposed model can correctly describe 92% of the busy periods. Third, the thesis questions the suitability of conventional ML evaluation metrics when ML models are used to generate inputs for simulation models. Using a discrete event simulation (DES) of EMS system, the study demonstrates inconsistencies between conventional metrics and simulation-based evaluations, highlighting the need for context-specific evaluation strategies. Fourth, this thesis proposes ML-based approaches for predicting ambulance travel time, comparing three artificial neural networks (ANN), Google Maps, and the traditional KWH model. Results show that ANN model aligns closely with historical data, making it more suitable for EMS simulation and decision-making, despite requiring more complex feature inputs. Finally, a comprehensive DES model is developed with the integration of ML-derived parameters for ambulance travel time and emergency department destination selection. The simulation model is validated against historical data of surges and used to test operational improvements, including ambulance upstaffing, alternative transportation, and surge-based low-priority call holding. Robustness analyses are conducted to test the potential of these strategies to enhance EMS performance under various demand scenarios. Collectively, this work contributes a novel framework for defining EMS surge levels and EMS simulation modelling. The findings underscore the importance of context-aware ML evaluation and offer scalable tools for improving EMS operational efficiency.