AI Model Created for Personalized Blood Glucose Monitoring

Type 1 diabetes (T1D) patients must consistently (and accurately) monitor their blood glucose levels. Scientists have looked to artificial intelligence (AI) to assume this task, but inter-patient variability and large data volumes remain key implementation challenges.

In a new study, however, researchers have developed BiT-MAML, a model-agnostic algorithm designed to provide personalized blood glucose predictions for T1D patients. The approach overcomes the limitations of existing models and enables precise predictions in real clinical settings, according to analysts.

Type 1 diabetes (T1D) is an autoimmune condition in which the body’s own immune system attacks insulin-producing cells. Thus, T1D patients must closely monitor their blood glucose (BG) levels and rely on insulin injections or pumps. Even small miscalculations or oversights can lead to unregulated blood sugar levels and potentially life-threatening complications.

Continuous glucose monitoring (CGM) systems have become a promising tool for predicting and forecasting BG levels. Over the past decade, researchers have attempted to use AI models to improve the prediction accuracy of CGM systems but differences in patient physiology and poor adaptation for new users have prevented the widespread adoption of this technology in real-world settings. In addition, traditional models often focus on either short-term or long-term glucose patterns, but not both.

In an attempt to address these issues, a research team led by Professor Jaehyuk Cho from the Department of Software Engineering at Jeonbuk National University in South Korea, has developed a model named BiT-MAML to tackle inter-patient variability in BG prediction. “BG dynamics are not uniform across all patients,” Cho said. “The physiological patterns of an elderly patient are vastly different from those of a young adult. Our model demonstrates how this variability can be accounted for by developing more personalized models.”

The team’s findings were published in Scientific Reports last summer in a paper titled, “Personalized blood glucose prediction in type 1 diabetes using meta-learning with bidirectional long short-term memory-transformer hybrid model.”

BiT-MAML (with BiT standing for Bidirectional LSTM-Transformer and MAML short for Model-Agnostic Meta-Learning) uses a hybrid architecture combining two deep learning models: bidirectional long-short-term memory (Bi-LSTM) and Transformer. Bi-LSTM processes time-series BG data bidirectionally, precisely capturing short-term patterns. Simultaneously, the transformer, using a multi-head attention approach, efficiently models long-term patterns, capturing complex day-to-day and lifestyle-based cyclical variations. During training, researchers applied a meta-learning approach known as Model-Agnostic Meta-Learning (MAML) that helps the model quickly adapt to new and diverse patients using only a small amount of training data by learning from various patient examples.

To test model performance, the researchers adopted a Leave-One-Patient-Out Cross-Validation (LOPO-CV) scheme. “In simple terms, we train the AI on five patients, then test it on the sixth patient it has never seen before,” Cho explained. “This is effective for assessing the model’s ability to generalize to unseen patients.”

The model demonstrated significantly reduced prediction error compared to conventional models. Notably, the prediction error varied from 19.64 milligram/decilitre (mg/dL) for one patient to 30.57 mg/dL for another. While these results represent a clear improvement over the standard LSTM models, they also highlight the persistent difficulty of managing inter-patient variability in real-world settings. “Our study shows how AI-based BG prediction models should be evaluated to improve both trust and model performance,” Cho concluded. “Addressing this challenge will contribute to the development of effective CGM models that can serve diverse patients with T1D, from children to the elderly.”

The team’s findings suggest that developing effective personalized BG prediction needs advanced AI models to incorporate robust evaluation methods that can transparently report the full performance spectrum.

Founded in 1947, Jeonbuk National University (JBNU) is a Korean flagship university based in Jeonju. The campus embodies an open academic community that harmonizes Korean heritage with a spirit of innovation. JBNU leads the Physical AI Demonstration Project valued at around $1 billion and spearheads national innovation initiatives such as RISE (Regional Innovation for Startup and Education) and the Global University 30, advancing as a global hub of AI innovation.

Cho is a professor in the Department of Software Engineering at Jeonbuk National University and director of the Adaptive AI Laboratory. He also serves as CEO of Human AI Plus, a university startup translating research into practical healthcare tools. His research focuses on medical AI, healthcare data analytics, and physical AI convergence, with particular emphasis on developing AI systems that are both accurate and usable in real clinical environments.