Predicting short-term emergency medical service demand on different levels of temporal and spatial aggregation: A machine learning approach
- Zusatzfeld:
Emergency medical services are essential for today’s modern societies, because they provide quick help for people in urgent medical need. The planning process in most middle European countries orients itself around the response time metric, namely how fast a medical professional can arrive at an incident site. There exist many ideas on how to further optimize the planning process according to the response time metric, like the dynamic relocation of ambulances in real time. What most of these models have in common is that they will need a reasonable accurate forecast of the emergency medical service demand, on a short-term temporal interval and as granular as possible concerning spatial dimensions, as an input factor. In this thesis three machine learning models were computed next to a baseline model on a data set from Ludwigsburg, Baden-Württemberg in Germany. This was done on eight different temporal and spatial aggregation levels, in order to have a better understanding on how granular the aggregation interval can be to yield a reasonably good forecasting performance. It is shown that the boosting technique LightGBM does overall perform best on most aggregation horizons and that all of the chosen machine learning techniques outperform the baseline model. Models computed on the more highly aggregated data sets, in terms of temporal and spatial dimensions, can account for much of the variation in the emergency medical service demand and therefore have the potential of being used as a input factor for operations research models in this context. However, the smaller the aggregation interval, the poorer is the forecasting performance. On the most granular aggregation level it is shown that the factor of randomness can not be reproduced by any of the forecasting models. It does nevertheless show the potential of continuous development and optimization of the application of machine learning techniques in this context, because by the inclusion of further explanatory variables and more computational power even better results can most likely be achieved.