Uncertainty-aware modelling of insulin sensitivity for automated insulin delivery in Type 1 Diabetes

Background

In type 1 diabetes the immune system destroys the insulin-producing cells of the pancreas, and insulin must be replaced. Getting the dose right is difficult, and the consequences of getting it wrong are not symmetric: too little insulin leaves glucose high, driving long-term blood vessel, kidney and eye complications, while too much causes hypoglycaemia, which is acutely dangerous. Safe dosing means steering between two unequal risks.

This is made harder because sensitivity to insulin is not constant. It shifts acutely with illness, exercise and stress, and varies with physiological status including circadian rhythm and menstrual cycle. Sensitivity also falls during periods of high glucose and recovers only slowly, so the larger doses needed to bring glucose down, risk crashing into hypoglycaemia hours later.

Automated insulin delivery (AID) systems have transformed diabetes care, with both commercial systems and open-source algorithms validated in randomised trials (1). Every five minutes such a system forecasts where glucose is heading and doses insulin accordingly. The efficacy and safety of each insulin injection depend on an individual's changing insulin sensitivity, and automated dosing decisions are often suboptimal. Open-source algorithms adapt to changing sensitivity only through limited adjustments, and commercial systems may adapt further but are proprietary and cannot be inspected or improved by the research community. In both cases the system produces a single predicted curve with no measure of uncertainty, so it cannot become more cautious when it is less confident. Algorithms are also typically evaluated on average accuracy across cohorts that over-represent particular groups, so it is largely unknown whether they serve everyone equally well (3).

This project will develop a model in which insulin sensitivity is a latent state that evolves over time, inferred within a Bayesian framework, yielding both adaptation and calibrated uncertainty at once. You will analyse large, harmonised datasets of real-world CGM and insulin records (2) and target the improvement of open-source AID algorithms.

Aims

1. Characterise where prediction is weakest, and for whom. Using harmonised open datasets, benchmark existing open-source prediction algorithms conditional on physiological state, characterise cyclical variation, and stratify by sex, age, and ethnicity where recorded.

2. Build models that track changing insulin sensitivity. Develop Bayesian state-space models of glucose-insulin dynamics in which insulin sensitivity is a time-varying latent state, capturing both acute shifts and slow cyclic variation.

3. Turn uncertainty into safer decisions. Because the risks of high and low glucose are not equal, evaluate dosing rules that act on the predictive distribution rather than its mean, and test whether uncertainty-aware dosing improves outcomes and narrows differences found in Aim 1 between groups.

Training outcomes

The student will be trained in Bayesian inference and dynamical systems modelling, machine learning for time series data, and the design of evaluation frameworks for predictive models in healthcare. They will join a team of mathematical and computational biologists in the lead supervisor’s lab.

They will gain biomedical grounding in type 1 diabetes through the clinical co-supervisor, learning how insulin therapy decisions are made in practice and what makes a model clinically credible. They will also develop expertise in health data equity: assessing who is represented in the datasets, and testing whether algorithms perform equally across groups.

This fully computational project suits a student with a quantitative background and skills who wants to apply these to a major chronic disease, and prepares them for careers in computational health research, clinical AI, or diabetes technology.

References

1. Burnside MJ et al. Open-source automated insulin delivery in type 1 diabetes. N Engl J Med 387:869–881, 2022. doi: 10.1056/NEJMoa2203913
2. Wolff M et al. MetaboNet: The Largest Publicly Available Consolidated Data Set for Type 1 Diabetes Management. Journal of Diabetes Science and Technology. 2026. doi:10.1177/19322968261441637 
3. Wolff M et al. Blood glucose prediction algorithms require clinically relevant performance criteria beyond accuracy. Diabetes Technol Ther, 2025. doi: doi.org/10.1089/dia.2025.0074