Frequency and Severity Prediction of Emergency Incidents from Socio-Demographic Data

Effective resource allocation and operational planning are significant challenges for emergency medical services (EMS) and fire and rescue services (FRS) operating in a dynamic and ever-changing environment. This thesis investigates the potential of socio-demographic, economic and ekistics data to predict the annual incident frequency and severity of emergency incidents in the Greater Reykjavík area. Using a 20-year dataset of EMS and FRS incidents and corresponding societal data, machine learning models were developed using the XGBoost algorithm. These models aimed to forecast incident rates by severity and type using societal data, as well as identifying the most influential societal factors driving emergency demand. The models were subject to large numbers of dataset transformations as well as rigorous hyperparameter tuning. The results demonstrate that the incident frequencies for EMS incidents can be predicted with notable accuracy, with FRS incident predictions fair but lacking interpretability. In the case of the EMS model, the results further showed clear patterns of societal factors being correlated with incident frequency, and furthermore those factors’ links to incident rate could be easily interpreted. In the case of the FRS model, despite its accuracy the model did not discern clear links between societal data and feature importance. The results demonstrate the potential of machine learning models to support data-driven decision-making in emergency services, particularly for EMS operations. They also suggest possibilities for further research, including more granular data collection and the incorporation to enhance the interpretability of predictive models for FRS planning.