A new study from Kaunas University of Technology reveals advanced AI models that could revolutionise glucose monitoring by predicting swings up to an hour in advance, offering hope for improved diabetes care in Europe’s rising diabetes landscape.
Type 1 diabetes management is becoming more data-rich, but not necessarily easier. A new study from researchers at Kaunas University of Technology suggests that artificial intelligence could help clinicians anticipate glucose swings before they happen, using continuous glucose monitor readings alongside insulin doses, carbohydrate intake and physical activity to forecast changes 30 and 60 minutes ahead. The work comes against a wider European backdrop in which diabetes remains a major public health burden, with the World Health Organisation’s European office estimating that at least 64 million adults and about 300,000 children and adolescents live with the condition across the region.
The KTU team’s approach tries to go beyond simple time-series forecasting. In place of a model that only reads glucose data in order, the researchers built a graph-based system that links similar moments in a patient’s history and then uses an attention mechanism to decide which earlier readings matter most. Rytis Maskeliūnas, a professor at KTU, compared that process to a clinician weighing the most relevant signs rather than treating every datapoint as equally important. The same design is intended to improve explainability, which remains a central concern in medical AI.
The model was tested on international type 1 diabetes datasets and, according to the researchers, performed strongly both in glucose forecasting and in identifying hypoglycaemia risk. They also reported that the system worked better when it was trained to handle both tasks at once rather than separately. Muhammad Abdullah Sarwar, a KTU PhD student involved in the project, said the model retained high forecasting accuracy even for patients whose data it had not seen before. The team also examined personalised insulin adjustments, although only as a decision-support tool for clinical review rather than something patients should use to change treatment on their own.
The researchers caution that the work is still at an early stage. Historical data can show promise, but it does not prove how well a model will perform in day-to-day care, and the next step will be clinical testing in real-world settings. They also point to unresolved questions around data security, privacy and how such software would fit into doctors’ workflows. Still, the study adds to a growing body of research suggesting that AI may help clinicians make better use of continuous glucose monitoring data, particularly at a time when the WHO says Europe faces the world’s heaviest burden of type 1 diabetes and projections point to continued growth in the overall diabetes population.
Disclaimer: This content is for informational purposes only and is not intended to be a substitute for professional medical judgment, advice, diagnosis, or treatment.
