“The idea is to develop a system that recommends a suitable transport destination for ambulances based on how seriously injured the patient is,” said Anna Bakidou, who has recently started working at the Department of Caring Science at the University of Borås and will work in the centre PreHospen, but is linked to Chalmers University of Technology, where she has just started her doctoral studies and where a large part of the research will be conducted.

The background to her research is that several studies show that patients who have been subjected to severe trauma have the greatest chance of surviving if they are transported to a trauma centre with high competence and readiness to treat serious injuries. In Sweden, these trauma centres are equivalent to university hospitals, and it is critical that patients with serious injuries are transported directly there. This is not always done today, due to several reasons.

No national guiding principles

For example, there are no national guidelines that describe where trauma patients should be transported based on their risk assessment, and some symptoms may be hidden, which means that the patient's injuries are considered less serious and the patient is then transported to the nearest hospital, where there is no access to advanced care. These problems have been addressed in a new study* from Stefan Candefjord at Chalmers University of Technology and Linn Asker and Eva-Corina Caragounis at Sahlgrenska University Hospital, which showed that there is a need for improved decision support regarding transport destinations, thus laying the foundation for this research.

“The hope with the new decision support that we want to develop is to improve the precision in the decision chain so that the patient is taken to the hospital with the right resources,” said Anna Bakidou, and further makes comparisons with today's decision support. “Today, national trauma emergency criteria are used in most of the prehospital healthcare systems in Sweden. These are inspired by RETTS (Rapid Emergency Triage and Treatment System) and serves as a checklist in which the ambulance staff fill in the patient's vitals and compare them with threshold values. Based on how many critical values exceed the predefined threshold values, the patient is assigned a process level, in which a higher level corresponds to an increased severity. This is then used by the ambulance staff to make a decision about what management the patient needs and where the patient should be transported,” she said.

AI differs

She believes that AI-based decision support differs from this, and describes the process in development: the first step is creating a mathematical model, where the various documented values for patients are entered into the system. The model analyses input data and identifies patterns for different patient groups. These patterns then form the basis of the recommended transport destination produced by the system.

“In the existing trauma criteria, there are no underlying mathematics in the assessment of a patient. This may lead to one criterion dominating the triage protocol in comparison with the other parameters. The difference with AI models is that they are not limited to the ‘simple’ patterns used in today's checklist, but instead have the opportunity to learn more complex patterns, and therefore have the potential to more reliably predict the right care needs and transport destination. By analysing the available data and finding patterns, the model can learn patterns that would be difficult for paramedics to detect. The more data that are available, the better the model can be developed, as the system becomes more confident in the patterns it identifies. In this way, we develop the model ,” she said.

The model is then tested on new data that the model has not analysed during the development stage to see how well it predicts future patient cases. After this step, the model is integrated into the ambulance. This means the ambulance staff fill in the patient's values as before, but enter them as input to the finished model. The model analyses the values according to the patterns it has learned and predicts which destination is most suitable based on, for example, care needs and distance to the nearest university hospital. The most likely destination will be the system's recommendation and this will be a support for the ambulance staff in deciding whether or not to transport the patient to a university hospital.

“We choose to develop AI models because they have the potential to find more complicated patterns and thereby increase precision in the decision chain,” she said.

What is the main benefit for ambulance personnel?

“The system will be a digital aid that fits into the ambulance staff's existing work procedure. There may be fewer fields to fill in when compared to checklists used today. But the absolute main advantage that we can see is that the staff can feel more confident in making tough decisions as the given recommendation is well-founded and based on applied mathematics. The staff will not be alone in deciding whether a critical patient should be transported to another hospital than the closest one, which is a support if  something happens during the longer transport distance. In addition, one can supplement difficult decisions with extra expert medical support via, for example, video conferencing directly from the ambulance as in the project ViPHS (video support in the prehospital stroke chain), another project with links to our interdisciplinary research group.”

What do you see as challenges in the project?

“There are many for this project! One of the biggest challenges in developing AI models revolves around how the data to be used is structured and how it is used. An example is if a certain group is overrepresented, as that can lead to the model learning stronger patterns for this group in comparison with the others, which in turn could lead to unequal care. It is therefore important to use a sufficiently large dataset that represents the included patient groups well.

Another challenge, she believes, may be that the data can be structured in different ways. For example, some parameters may be documented numerically while others are described with text. To find patterns in the different structures, several mathematical models are needed and it will be a challenge to decide how they should be combined.

“Technically, there are many different types of AI models and new ones continue to appear as it is an attractive area to research. It will therefore be a challenge to identify which model is best suited for this project.”

In addition to the technical issues, there are many other challenges, according to Anna Bakidou, who mentions several: the developed system must have a user-friendly interface, follow regulations, and be ethically approved as it uses patient data to build the model. For the same reason, it is important that the system is inaccessible to unauthorised persons.

What does the schedule look like?

“We hope to develop the model with prehospital data within a couple of years, which includes retrospective studies. Then we hope to test the model in a prehospital environment.”

What is your view on AI in the prehospital field?

“It is an exciting area that is starting to be more visible in research! The strength of AI is that the tool can analyse large amounts of data to find patterns that are not intuitive to us humans and thus can make predictions we cannot make ourselves. In the prehospital environment, the decisions that are made are important for the patient's survival, while at the same time, there is a limited amount of data to base these decisions on. It is therefore an exciting opportunity to include AI models that can use previous data to give a real-time recommendation and to support the ambulance staff in their decision-making process!”

How does it feel to do research in this area?

“It feels great to start up with this as a doctoral student. In this project, it will be important to put together many different pieces to develop a system that can be used in practice and be approved by the staff. It is a challenge that I am motivated by and it will be interesting to collaborate with competent and inspiring people from both the University of Borås and Chalmers University of Technology. In addition, I will now be part of the strong interdisciplinary regional research and innovation cluster in prehospital healthcare, which includes Chalmers, the University of Borås, healthcare in VGR, Sahlgrenska Academy, and PICTA (Prehospital ICT Arena) at Lindholmen Science Park,” said Anna Bakidou.

Anna Bakidou

Profession: Doctoral student at the Division of Signal Processing and Biomedical Engineering at the Department of Electrical Engineering at Chalmers University of Technology and employed at the Department of Caring Sciences at the University of Borås

Education: Master’s of Science in Electrical Engineering and has a Master's degree in Biomedical Engineering

Research area: AI as decision support for equal care

About the collaboration project

The research is conducted as a sub-project within the overall project "Artificial intelligence (AI) as decision support - for equal care" led by the University of Borås together with the University of Østfold in Norway and with Chalmers University of Technology as one of the parties responsible for the development of the mathematical models behind decision support. The project runs for three years and is financed through the ERDF (European Regional Development Fund) and the Interreg Sweden-Norway collaboration programme with just over The aim of the project is to design AI tools that put the patient's needs at the centre by giving the care provider the opportunity to design the care in such a way that individual variations between different care providers are minimised.

* Candefjord S, Asker L & Caragounis EC (2020). Mortality of trauma patients treated at trauma centers compared to non-trauma centers in Sweden: a retrospective study. European Journal of Trauma and Emergency Surgery.