New AI Model Could Help Type 1 Diabetes Patients Avoid Dangerous Glucose Swings

Researchers from Jeonbuk National University in South Korea have developed a new artificial intelligence model designed to improve personalized blood glucose prediction for people living with Type 1 diabetes. The innovation aims to address longstanding challenges in glucose monitoring, including differences in physiology between patients and the difficulty of adapting models to new users.

Patients with type 1 diabetes must consistently monitor their blood glucose levels and depend on insulin injections or pumps to regulate their condition. Even small miscalculations can lead to unstable blood sugar levels, which may result in serious or life-threatening complications.

Researchers have explored AI-powered approaches for more than a decade to improve predictions in Continuous glucose monitoring systems. However, many existing models struggle to account for differences among patients or to balance short-term and long-term glucose patterns.

Hybrid AI model aims to improve prediction accuracy

The research team, led by Jaehyuk Cho of the university’s Department of Software Engineering, developed a hybrid algorithm called BiT-MAML to tackle these limitations. According to Cho, blood glucose dynamics can vary widely depending on factors such as age, lifestyle, and physiology, making personalized prediction essential.

“BG dynamics are not uniform across all patients. The physiological patterns of an elderly patient are vastly different from those of a young adult,” Cho explained. “Our model demonstrates how this variability can be accounted for by developing more personalized models.”

The BiT-MAML model combines two deep learning architectures: bidirectional long-short-term memory and transformer technology. The bidirectional long-short-term memory component captures short-term time-series glucose patterns, while the transformer analyzes long-term trends and complex lifestyle-related variations in glucose levels.

Meta-learning approach enables adaptation for new patients

During training, researchers applied a meta-learning technique known as Model-Agnostic Meta-Learning to help the system adapt quickly to new patients using limited data. The study was published in Scientific Reports on August 20, 2025.

To evaluate the system’s performance, the researchers used a testing method called Leave-One-Patient-Out Cross-Validation, in which the AI model was trained on five patients and then tested on a sixth patient it had never previously encountered.

The results showed that the model significantly reduced prediction errors compared with conventional long-short-term memory models. Prediction errors ranged from 19.64 milligrams per deciliter for one patient to 30.57 milligrams per deciliter for another, highlighting both improvements in accuracy and the continuing challenge of managing inter-patient variability.

Cho noted that improved evaluation methods are also essential for building trust in AI-based glucose prediction systems. Researchers believe the approach could support the development of more effective continuous glucose monitoring tools capable of assisting diverse groups of patients with type 1 diabetes.