Data-driven optimization of prehospital care via statistics and machine learning
Research project
An aging population, urbanization, and medical advancements require a flexible ambulance service so that resources can be used sustainably, efficiently, and fairly.
Our vision is to use statistics and machine learning to transform healthcare data into a more efficient and equitable healthcare system. In this project, we are developing data-driven tools to visualize, simulate, analyze, and optimize ambulance services. The goal is to create the conditions for systematic quality work, where decisions are based on objective and transparent decision-making foundations.
Through collaborations with several regions, we have access to unique alarm data and expertise in ambulance care. One of our goals is to describe ambulance care and changes in ambulance care. The work includes developing tools to visualize large and complex alarm data, for example by studying different levels of geographical resolution, from regional level down to district level, studying alarm occurrence over time and space, studying different process times, and studying different patient groups. We use the tools, together with statistical methods, to study, for example, how response times differ between men and women.
ImageJörgen Lundälv
Simulate the impact of operational changes
Ambulance operations consist of a series of processes. 112 operators answers incoming calls, the alarm is then dispatched to the appropriate ambulance, the ambulance goes to the pick-up position, the patient is treated, and finally the patient is driven to a hospital. We use statistics and machine learning to model the alarm intensity, process times and the dispatchment of resources. The result of the modelling is a digital twin of the reality, which can be used to study the consequences of various operational changes, such as changing the scheduling of resources or moving ambulance stations.
Estimating a weighted target function
The digital twin generates simulated data that has the same structure as historical data, which makes it possible to calculate various key metrics that describe the ambulance care. For example, median response times, economic parameters, and measures describing differences between patient groups. One problem is how different key measures should be weighed together in decision-making. How do we choose between two alternatives, where the first results in shorter response times, but also greater regional and gender differences? The goal here is to identify a weighted target function that can be estimated via historical and simulated alarm data.
Optimize ambulance care
The combination of a digital twin, which can simulate the consequences of various operational changes, and an target function that can be estimated via data creates the conditions for optimizing the ambulance care. An allocation is determined by how the resources are scheduled, where the resources are located, how resources are dispatched, and various process directives. An allocation can therefore be considered as a solution in a high-dimensional room. We have a challenging optimization problem where we can simulate data for each allocation and estimate the target function. We develop methods that can be used to identify an optimal solution through robust optimization.