Born in 1976, Troels Martin Range originally graduated as a mathematical economist – MSc Mathematics-Economics – with a specialist in operations analysis and finance from the University of Southern Denmark before going on to complete a PhD in operations analysis. In short, operations analysis is an interdisciplinary field that uses mathematical methodologies and models to optimise decision-making and solve complex problems in different contexts, such as the optimisation of container logistics, train traffic and crew planning at airports. The discipline involves the use of tools such as linear programming, game theory, queuing theory and simulation models with the goal of analysing and improving processes, systems and operations. To many, this will probably all sound like Greek, but for Troels it’s his daily fare. “The methods are very abstract, but the solutions are very specific,” he says.
After completing his education and with a PhD under his belt, Troels was employed at the University of Southern Denmark where he worked as a researcher in his field for ten years. In 2016, he had reached a point where he needed change and wanted to find an outlet to apply his theory in a more practical way. And it just so happened that there was a relevant administrative position available at the Hospital of South Western Jutland in Esbjerg, which he applied for and got. He maintained his ties to the University of Southern Denmark with a part-time lectureship, which he still holds today.
In Esbjerg, Troels met Mikkel Brabrand who was the research lead in the hospital’s accident and emergency department. Mikkel challenged him with a task that he actually didn’t think he would be able to solve: “Can you predict how many patients will arrive at the A&E Department – several months into the future?”. Troels accepted the challenge, sat down at his keyboard and within a few weeks had created a rough sketch for a long-term forecast. “It was a simple model but it could actually do a lot,” Troels says.
The department had a lot of patient flow data. “But perhaps somewhat surprisingly, we could see that there were also some components that had a major influence on how many patients turned up at the department from one hour to the next. Among other things, the weather was not an insignificant factor,” Troels points out. “That’s why we contacted the Danish Meteorological Institute – to find out if they could provide the data we requested – but at the time it was extremely expensive, and we didn’t have funding for it. So, the project actually stalled there a bit.”
For a while. Meanwhile, Mikkel had been employed at the A&E Department at Odense University Hospital, where he noticed that there were sometimes extremely busy days when the staff had great difficulty keeping up. So, with Troels’ skills in mind he arranged for Troels to be transferred to Odense University Hospital and given the opportunity to look at the challenge more closely. This was not long after the brunt of the COVID-19 pandemic in 2021.
“To start with, we applied for funding from the innovation pool at Odense University Hospital and bagged a healthy grant. Since then, we have secured further funding from the University of Southern Denmark, the Region of Southern Denmark, the Innovation Fund and Beta.Health. That meant we could finally really get to work properly,” explains Troels.
In the meantime, the political decision had been made that the Danish Meteorological Institute should make a range of weather data available on an open and accessible basis, which combined with funds for development meant that the project could now really take off. Troels was now really able to get stuck into the system.
What followed were long, creative, methodical and labour-intensive months. But in August 2022, the day arrived when Troels was able to present a visual user interface to Mikkel for the first time with an overview of what predictions the system could calculate and display.
“It was bit of a wild day,” Troels remembers. “Because all of a sudden I was actually able to counter the scepticism Mikkel had shown all the way back at Southwest Jutland Hospital.”
Once Mikkel has seen it all he said: “That hits the mark really well!”
Troels and Mikkel were then given permission to put the system into operation in the Accident and Emergency Department at Odense University Hospital. “It was an anxiety-inducing day, to say the least. Suddenly we went from having our own small, kind-of secret system to releasing it into the world for real,” Troels says. How will it be received? Will it prove its worth? Would it meet expectations? Will it WORK?
And yes. It worked. In fact, it was so accurate that 95% of the time it delivers the right predictions – equivalent to +/- one patient an hour. And it’s actually quite difficult to get much closer.
“As a case study, hospitals are actually one of the most difficult to build forecasts for,” explains Troels. “For example, it is significantly easier to predict congestion in airports because airlines sell tickets, which is a relatively simple data set on which to base predictions. At hospitals, there is a tradition of rostering staff relatively evenly throughout the week – and so staff basically have to work harder if things are busy and will have more time on their hands if there are fewer patients.”
The motivation behind the Praemostro system all along has always been driven by the desire to be able to adapt the capacity offered to what is actually demanded. “We often find that staff burnout because there is too much divergence between capacity and actual incidents. And it’s exhausting for the staff on several levels. Wearing out shoe leather through being so busy. But also actually the opposite. “It can be enormously stressful to be constantly wondering if and when there will be a rush, even though there might not be one at all,” explains Troels.
There is a lot of historical data at hospitals in relation to patient arrivals but so far it has been difficult to use them for reliable forecasts. Everyday life now looks completely different in the departments using the Praemostro system, which includes knowledge from a number of additional data sets for example, about the weather, events, weekdays and holidays. “We’re getting great feedback on it and people get used to using it – and to doing calculations with it – very quickly. That’s because it’s really accurate in its predictions,” says Troels. “And that means that you can plan ahead and roster staff in a completely different, more precise and intelligent way than has been possible so far.”
If we scratch the surface of the more technical aspects of the Praemostro system, the concept of ‘machine learning’ quickly comes up.
“I’m not such a big fan of the term ‘artificial intelligence’ because people tend to confuse it with human intelligence. And that’s like using the term horsepower — it has nothing to do with actual horses,” laughs Troels. “Computers are basically not very intelligent. Where the magic happens is when we design models that can simulate something that resembles intelligence. When working with machine learning, we take some data and try to create a model that can help the computer with making choices. Quite simply, it’s done in two steps: 1) By continuously teaching it to become smarter from the real world; and 2) by fine-tuning the models when they make mistakes. When you train the network, which is what this exercise is known as, it becomes more and more accurate over time. And it can actually do that itself via algorithms. This ability isn’t particularly new or revolutionary – it was already possible in the 1970s. However, today we have significantly greater processing power that can handle it – and handle it much, much faster. “There’s a lot that’s happened on both the hardware and the data side since then,” Troels points out.
One question that Troels is often asked is : “Why don’t you just get statisticians to solve the task?” With all the types of data available, surely, they would be able to create a model and calculate it?
“In simple terms, you could say that statisticians will typically test one type of model at a time. And it can take a long time to set the model up and to test it in practice via the trial and error principle. With machine learning, it’s possible to test lots of models and types of model very, very quickly, which means we can go much further in a much shorter time. Your data is constantly being reviewed and the machine is training the system and continuously selecting which models work best. So in short, it scans a large number of models in parallel using a lot of computing power. That’s why I often refer to it as statistics on steroids,” laughs Troels.
Three fun facts about Troels:
