People with diabetes prick their fingers multiple times a day to check their blood sugar and inject insulin, the hormone that controls blood sugar, accordingly. Artificial intelligence (AI) technology that can automate this cumbersome manual management has been developed in Korea.
A research team led by Professor Park Sung-min of the departments of mechanical engineering and electrical engineering and the Graduate School of Convergence at Pohang University of Science and Technology POSTECH said on the 21st that it developed DA-CMTL (Domain-Agnostic Continual Multi-Task Learning), a blood sugar management AI that anyone can use. The findings were published on the 16th in npj Digital Medicine, a sister journal of Nature.
Blood sugar fluctuates constantly depending on meals, exercise, and stress. In healthy people, insulin secreted by the pancreas controls these changes, but people with type 1 diabetes secrete little to no insulin and must manage blood sugar on their own.
If blood sugar drops rapidly into hypoglycemia, it can lead to dizziness, fainting, and in severe cases, cardiac arrest. For this reason, people with diabetes must measure blood sugar multiple times a day and adjust insulin doses based on the readings. However, this manual approach is cumbersome and has limits in predictability.
Professor Park's team developed a general-purpose AI technology that can predict changes in blood sugar in advance and detect dangerous hypoglycemia. Research on AI-based blood sugar prediction has continued, but most systems were tailored to a specific patient's data and were difficult to apply to others. In addition, existing technologies required separate computations for blood sugar prediction and hypoglycemia detection, which was inconvenient.
The DA-CMTL model learns from blood sugar levels recorded every five minutes and insulin infusion data collected by a continuous glucose monitor (CGM) attached to a patient's arm. Based on these data, it predicts how blood sugar will change and simultaneously calculates the likelihood of hypoglycemia.
The team combined three AI techniques to improve accuracy. First, with "continual learning," the system maintained stable performance even when learning patient-specific data sequentially, and by applying "multi-task learning," it handled blood sugar prediction and hypoglycemia detection at the same time. On top of that, they added a "sim-to-real transfer" technique so that a model trained in a simulated environment would work well with real patient data.
In experiments, the model outperformed existing models on metrics that indicate blood sugar prediction accuracy, the team said. Performance gains were also demonstrated in tests with an actual artificial pancreas system, confirming clinical applicability.
Park said, "This study laid the foundation for a universal AI blood sugar management technology that is not limited to specific patients," adding, "It is expected to evolve into next-generation artificial pancreas technology and greatly improve diabetes treatment methods and quality of life."
References
npj Digit. Med (2025), DOI: https://doi.org/10.1038/s41746-025-01994-4