Life-threatening blood glucose fluctuations are an overlooked problem in intensive care units (ICUs) – especially for people with diabetes. A new computer model can predict in real time who is at risk of developing low or high blood glucose so that doctors can intervene in time.
According to the WHO, more than 800 million people have type 2 diabetes – and many end up in the ICU, where blood glucose levels can become dangerously unstable. The danger is especially acute for ICU patients with type 2 diabetes, since both high and low blood glucose can be harmful.
High or low blood glucose levels are associated with longer hospitalisation, increased risk of developing other diseases and increased risk of death during hospitalisation.
To tackle this, researchers have developed a computer model that can assess in real time which patients risk dangerous blood glucose swings – so doctors can intervene in time.
According to a researcher behind the development of the model, it could save thousands of lives.
“In the United States alone, about 5 million people are admitted to ICUs each year. Our model could support hundreds of thousands of earlier interventions – potentially reducing hospital stays and preventing illness or death,” says Arijit Khan, Associate Professor, Department of Computer Science, Aalborg University, Denmark.
How the model detects dangerous blood glucose levels in time
Being able to predict dangerous swings in blood glucose during hospitalisation could make a real difference.
Blood glucose can be measured directly – but by the time it is clearly too high or too low, it may already be too late to intervene.
This is why researchers have long been working on developing models that can identify high-risk patients based on all the data already collected about them.
For example, when patients are admitted to an ICU, they have their blood pressure and pulse measured and provide a medical history.
Doctors also write extensive notes in patient records.
“Data on hospitalised patients are very diverse and collected at different time intervals. Some are recorded every minute, while others are taken once an hour. Patient data are measured at different times and in different systems – creating chaos that doctors may find difficult to comprehend. This is where the artificial intelligence model helps them find patterns they might otherwise miss,” explains Mohammad Hadi Mehdizavareh, a PhD student from the same department and another researcher behind the study.
The same artificial intelligence technique that powers ChatGPT
The team developed a model that can handle time-staggered data and uncover hidden patterns.
To build the model, they used the same techniques found in large language models such as ChatGPT.
The way language models work is that they try to predict which word is next in line when you ask a question based on the quantity of data available.
Similarly, the new model uses the available data to predict where the blood glucose is heading: up or down?
Just like ChatGPT guesses the next word in a sentence, the new model guesses whether blood glucose is going up or down – and chooses which data to use for that.
Traditional models required researchers to handpick which data to include – an approach that was time-consuming, resource-intensive, and often failed to produce the best prediction results.
Model can save thousands of lives
The researchers tested their model on data from more than 200,000 ICU admissions across 208 hospitals.
The model’s ability to predict who developed high or low blood glucose levels was 5–7% better than existing models.
In the dataset studied, this means that if the model had been used in real time, the analysis shows that more than 1,000 lives could have been saved based on previous interventions.
“The idea is that the model should have access to patients’ data in real time, so that it can issue a warning if, based on the data, it believes that a patient will develop too high or too low blood glucose within a period of time. This can enable doctors to intervene and thereby save lives, prevent the development of further disease and shorten hospitalisation,” explains a third researcher behind the study, Simon Lebech Cichosz, Associate Professor, Department of Health Science and Technology, Aalborg University.
Can also predict blood clots and death
According to Arijit Khan, the type of models the researchers have developed have great potential.
Therefore, the researchers plan to further develop it and make it capable of assessing more than just the risk of developing high or low blood glucose.
This could include the risk of blood clots, death or extended hospital stays.
The better the models become at extracting knowledge from data, the more lives can be saved – and the better hospitals can utilise their resources.
The researchers would also like to develop the model to make it more usable.
“We want to make it easier to be trained, because it currently takes a very long time, and training it again every time it has to predict something new would be expensive. Second, the model converts all data to zeros and ones, which may make sense to computers, but doctors and others also need to understand what is happening inside the model. This is why we are also working on solutions so that others can understand how the model works – and help ensure that it does not make critical errors,” concludes Arijit Khan.
A successful utilisation of AI is not just about performance.
“It is very important that AI models are not viewed as a replacement but rather as a complement to clinical care, which could improve patient outcomes by providing evidence-based recommendations and analysis at the bedside, reducing cognitive load, and enhancing confidence in decision-making. Implementing such models in the ICU needs to be done in close collaboration with clinicians and with consideration of the clinical workflow to realise their full potential,” comments Simon Cichosz.
