Online Domain Adaptation of Glucose Forecasting Models Personalised Blood Sugar Prediction That Learns as It Goes

This patent describes an adaptive system for monitoring and forecasting blood glucose, the sugar level in a person’s blood, that continually tunes its predictions to the individual user as new data arrives. It comes from the University of Virginia, a centre of diabetes technology research.

The Problem

Managing diabetes well depends on anticipating where blood sugar is heading, not just knowing where it is now. Forecasting is hard because every person responds differently to food, activity, insulin and stress, and those responses drift over time. A model trained on a general population can be a poor fit for any one individual, and a model fixed to one person can go stale as their physiology and habits change. What is needed is a forecasting system that starts from population knowledge but keeps adapting itself to the specific user, on the fly, so its predictions stay accurate.

What This Invention Does

The patent sets out an adaptive glycemia monitoring and forecasting system with two cooperating parts. An event monitor receives a person’s blood glucose levels, or information about an activity they performed, and produces an event output. A control module then pulls observation data, predictor variables and a population-based set of weighting coefficients from a database, and uses the event output to generate an updated, personalised set of weighting coefficients for that individual. The update is computed using a cross-entropy loss objective function, a standard tool in machine learning for tuning a model to better match observed outcomes. This is a form of online domain adaptation: the model continually shifts from the general population toward the specific user, sharpening its glucose forecasts as it gathers more of their data.

Key Features

• Adaptive forecasting. The system updates its glucose predictions for each individual over time.
• Event monitoring. It ingests glucose readings and activity information as events.
• Population to personal. It starts from population weighting coefficients and personalises them.
• Cross-entropy tuning. Updated coefficients are derived using a cross-entropy loss objective function.
• Online operation. Adaptation happens continually as new data arrives rather than in a one-off training run.

Who Is Behind It

The applicant is the University of Virginia Patent Foundation, the technology commercialisation arm of the University of Virginia, whose researchers are prominent in artificial pancreas and diabetes modelling work. The named inventors are Marc D. Breton, Jonathan Hughes and Stacey Anderson.

Why It Matters

Accurate, personalised glucose forecasting is central to the next generation of diabetes care, including closed-loop systems that adjust insulin automatically. A model that keeps adapting to the individual can deliver better predictions than a static one, which translates into safer, more effective glucose management and fewer dangerous highs and lows. Protecting the method in Australia supports bringing advanced diabetes technology to the large local population living with the condition.

Related Concepts

• Blood glucose monitoring – the activity the system supports and forecasts.
• Artificial pancreas – the closed-loop application such forecasting enables.
• Domain adaptation – the machine learning idea of shifting a model to a new context.
• Cross-entropy – the loss function used to tune the personalised model.
• Diabetes – the condition this technology helps manage.

---

AU 2026201926 was published in the Australian Official Journal of Patents on 2 April 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.