Leveraging Artificial Pancreas Technology for Treatment Optimization in T2D

In type 2 diabetes (T2D), injections with long-acting insulin can become necessary to normalize blood glucose and avoid long-term complications. However, finding a safe and effective insulin dose, a process known as titration, is both challenging and time demanding. In this thesis, we propose a new titration method for swift and safe identification of a personalized insulin dose with long-acting insulin through short-term use of rapid-acting insulin in an artificial pancreas (AP).
We augment a published T2D model to simulate an AP driving the blood glucose into the clinical target range followed by a switch to injections with long-acting insulin. In simulation, the new titration method can reduce the titration period to a single week, compared to five weeks on standard-of-care titration. To explore how to best switch between rapid- and long-acting insulin, we use clinical trial data to assess the correlation between the insulin response to rapid- and long-acting insulin injections in the same individual. In an in silico cohort of a hundred people with T2D, we investigate how differences in bioavailability may influence the conversion from rapid-acting insulin delivered in a pump to an equivalent injection dose of long-acting insulin. The cohort simulation reveals that many individuals need more than one week of AP treatment to reach the clinical target range.
As an alternative to letting an AP drive the blood glucose into the target range, we explore how to predict a safe and effective long-acting insulin dose from 24 to 48 hours of AP data. With simulated AP data, we estimate parameters in dose-response models using maximum likelihood estimation (MLE). We apply the continuous-discrete extended Kalman filter (CDEKF) to approximate the likelihood function which is maximized in MLE. To improve the model-based dose predictions, we apply model-based design of experiment (MBDoE) and determine how to best run an AP system to collect data for parameter estimation. Finally, we obtain personalized dose-response models from the experimental data and evaluate their ability to predict a safe and effective insulin dose for each simulated individual.
In simulation, the proposed method is feasible. However, the efficacy and safety of the dose estimates heavily depend on the level of system excitation. The results indicate that MBDoE holds a potential to improve the performance of model-based dose-guidance solutions. Still, without clinical data, it is not possible to conclude on the clinical feasibility of a translating between pump- and pen-based treatment in T2D. In the future, commercial AP systems may enable clinical evaluation of the new titration method.