Long before type 2 diabetes becomes clinically visible, the body already signals that something is wrong — and an algorithm can now detect those signals. Instead of focusing on aggregated measurements, the model learns the patterns of day-to-day blood sugar fluctuation. It is like a ChatGPT for blood sugar measurements, says the researcher behind the algorithm.
The body often reveals that something is wrong long before we notice it ourselves. One of the clearest clues is hidden in the small day-to-day fluctuations in blood sugar.
Based on these patterns, the model can identify with high precision whether a person is on a trajectory towards type 2 diabetes, cardiovascular disease or other serious health problems – even when traditional blood sugar measurements suggest that the person is metabolically healthy.
Previously it was only possible to react once disease had already become clinically apparent, but the model indicates risks that build up quietly over many years.
The research therefore opens up new opportunities for gaining insight into health – in research, in disease and at a more personal level.
“Even when blood sugar measurements would immediately suggest that a person is healthy, our model can identify that this may not be the case,” explains Jordi Merino, Associate Professor at the Novo Nordisk Foundation Centre for Basic Metabolic Research at the University of Copenhagen in Denmark.
“Our model is better than conventional risk markers because we can predict the development of disease up to 12 years before it manifests clinically.”
The research was carried out by an international consortium supported by the European Innovation Council called GLUCOTYPES and published in Nature.
When blood sugar fluctuations matter more than the number
Continuous glucose monitors are becoming increasingly widespread. They were originally developed for people with type 1 diabetes, enabling them to keep track of their blood sugar levels – especially at night – and avoid dangerously low levels.
Later, the devices were allowed to be used among people with type 2 diabetes, and today blood glucose monitors are broadly available and are being used outside the healthcare in people without diabetes who want more detailed insight into their health.
Experience has shown that blood sugar is not only important in terms of its absolute level but also in how it fluctuates over the course of the day. Two people can have the same average blood sugar level – yet display very different patterns, depending on how pronounced and how frequent their fluctuations are.
“It is precisely the fluctuations in blood sugar over time that contain information about how disease develops,” explains Jordi Merino.
The aim was to find out whether these everyday fluctuations quietly foreshadow disease — years before diagnosis.
To determine whether such patterns actually predict disease required data on a completely different scale than traditional blood sugar measurements.
Learning these patterns required data on a completely different scale
In developing the algorithm, the researchers used more than 10 million blood sugar measurements from 10,812 individuals, most of whom did not have diabetes.
The algorithm works on the same principle as ChatGPT: it learns from large volumes of data to predict what comes next. ChatGPT predicts the next word in a sentence, whereas this model predicts the next blood sugar measurement based on previous measurements and patterns in the data.
By systematically comparing the predicted and actual measurements, the model gradually learns which patterns characterise healthy development – and which point to early signs of disease.
“The model learns patterns that cannot be seen in individual measurements but only become visible when blood sugar is monitored continuously,” says Merino.
The model identifies high-risk groups long before disease develops
After training on the dataset, the model was validated in different studies including individuals with type 1 diabetes, type 2 diabetes, gestational diabetes, prediabetes or severe obesity – and performed similarly across eight different continuous glucose monitoring devices.
The model, GluFormer, predicted which individuals in a group of 580 people would develop type 2 diabetes or die from cardiovascular disease within up to 11 years – even though the participants had only used a continuous glucose monitor for a short period of time.
The quarter of participants whom the model assessed as being at highest risk accounted for 66% of all cases of type 2 diabetes, and only 7% in the lowest-risk group went on to develop the disease.
Similarly, 69% of cardiovascular deaths occurred in the high-risk group and none in the low-risk group.
“Diabetes or cardiovascular disease does not occur overnight but develops slowly over time. The earliest signs cannot be detected using traditional blood sugar measurements – but we can detect them with this model,” says Merino.
Same food – widely different blood sugar responses
The researchers further developed the model by incorporating data on food intake as well. This enabled the researchers to predict how an individual’s blood sugar responds to specific foods and thus identify which dietary choices are associated with the unhealthiest blood sugar patterns for that person.
“We can see that the same food can trigger very different blood sugar responses from one person to another,” explains Merino.
In a research context, the algorithm can provide new insight into how early unhealthy blood sugar patterns over time lead to disease – and how it may be possible to intervene before the disease develops.
“It allows us to understand how disease develops long before it can be diagnosed clinically,” he says.
At the same time, the algorithm could play a role in the growing market for personal health monitoring, in which many people already track blood pressure, heart rate, sleep and physical activity. Blood sugar could become another important piece of that picture.
This would, however, require new measurement methods that can record blood sugar without a needle under the skin – for example, using light or laser-based techniques.
“The world is moving towards ever better health measurements. With this algorithm, we now have a tool that can begin to pick up the body’s signals long before disease shows itself,” concludes Jordi Merino.
