Systematic Mathematical Modeling for Agile Development of Model Based Medical Control Systems

Today, many people live with diseases that require constant treatment. An example is diabetes. Diabetes is a growing world-wide problem. In 2021, 537 million adults (20-79 years) were living with diabetes with a total cost of at least USD 966 billion. Studies have shown that automatic treatment by a closed-loop system (artificial pancreas) based on feedback control can both improve glycemic control and lessen the burden of living with diabetes. The first hybrid closed-loop system became commercially available in 2016. Hybrid closed-loop systems are not fully automatic and require the user to manually announce, e.g., meals or exercise. Furthermore, the currently available systems are only able to administer insulin which lowers the glucose concentration. Consequently, they are unable to actively prevent hypoglycemia by, e.g., administration of glucagon. Severe hypoglycemia can have acute consequences, such as loss of consciousness and seizures. Therefore, there is still a significant interest in developing artificial pancreases (APs). Clinical trials are crucial to ensuring a high level of safety and efficacy, but are also very expensive and time-consuming which makes the development of medical devices (including APs) long and cumbersome. Here, virtual clinical trials (in-silico studies) are beneficial to evaluate the performance and identify potential risks before a real clinical trial. In this thesis, we 1) develop a parallelized high-performance Monte Carlo simulation toolbox to perform large-scale long-term virtual clinical trials, 2) develop the DiaCon dual-hormone (insulin and glucagon) AP and test it in a clinical trial with 11 adolescents, 3) describe the mathematical models applied in the virtual clinical trials and in the DiaCon AP as well as the models that were developed during the thesis, and 4) develop a web application to visualize and analyze diabetes data from, e.g., an AP or a virtual clinical trial. The Monte Carlo simulation toolbox is connected to a PostgreSQL database of virtual participants (represented by mathematical models) and protocols. The database makes it straightforward to reuse or add more participants and protocols. We show examples of a virtual clinical trial where two different closed-loop algorithms are compared in 1 mio. virtual participants over 1 year. Using high-performance computing, the virtual clinical trial is conducted in 82 min. The DiaCon AP is based on nonlinear model predictive control where we use an extension of the Medtronic virtual patient model for predictions. We estimate the model parameters with a prediction error method based on the continuous-discrete extended Kalman filter that is also used for state estimation. The DiaCon AP consists of the control algorithm (implemented in an Android smartphone), a Dexcom G6 continuous glucose monitor, and two Dana-RS pumps. The clinical trial displayed that it is feasible to use NMPC for APs and the DiaCon AP improved the time in range compared to the baseline, but identifying a model individualized to each participant is a challenging and very time-consuming process. Furthermore, we experienced several technical difficulties during the trial, such as, pressure induced sensor attenuations and loss of connection to the pumps. The web application allows users to login and view representations of the data depending on the permissions of the users. Individuals can view different representations of their own data from selected periods whereas, e.g., doctors can select between all their patients. The web application is build using a Vue.js frontend application, a Java Spring Boot backend application, and a PostgreSQL database. The web application is a prototype hosted on the localhost and currently only shows virtual people and simulated data. This thesis consists of a summary report and a collection of thirteen research papers and three technical reports.