Abstract
Automated closed-loop drug delivery systems, in which continuous adjustments of treatment parameters are based on frequent sampling of physiological processes, have been studied for years and are largely envisioned for use in clinical settings where the system can be under constant supervision by a clinician. Recently, there have been efforts in using these systems as advanced treatment for chronic illnesses, which is timely since chronic illness rate is on the rise, but also raises the possibility that the system user may now be the patient (not a clinician). Indeed, miniaturization and smarter algorithms promise that one day these technologies can be taken out of the hospital and brought to patient’s home. For example, recent clinical studies of artificial pancreas (AP) systems have demonstrated the feasibility of closed-loop treatment of Type 1 diabetes. While efforts to commercialize AP systems are underway, it is not yet clear that closed-loop treatment will appeal to a large segments of the patient population.
This work started with a desire to understand the factors that contribute to the success or failure of systems that (i) involve continuous measurement (as in the AP) but (ii) do not involve automated adjustment of all treatment parameters but rather present some treatment options to patients on demand where such options are model-based optimal recommendations. This approach accounts for human factors considerations and serves as a reasonable precursor to a fully closed-loop AP, but can be more readily adopted by the patients. The approach employs a model that addresses the risk asymmetry of the patient state space and takes advantage of a previously developed safety feature to create a semi-automated system.
One outcome of this work is a proposition of a controller design where individualization of the action of the system is achieved through the development of a mathematical model that is adapted to patient’s individual physiology. Previous model-based advisory systems have generally relied upon a “population average” model and achieved individualization through careful construction of optimization objective function. Such approach of defining state deviation penalties proved to be fragile because, for example, the patient’s pump therapy parameters that the objective function becomes sensitive to are often misestimated. The in silico preclinical trials using the new controller design suggested dramatic improvement over conventional therapy by better keeping the blood glucose in range and reducing the risk of implications such as hypoglycemia, without requiring ad hoc tuning of objective function parameters.
In vivo validation of the bolus advisory system confirmed safe and effective operation at meal times, but, due to model uncertainty, demonstrated that a different approach should be employed to retain the efficacy in the timeframe immediately after meals. Consequently, the work continued with the advice request limited to meal times and correction boluses decoupled from meal boluses. In addition, the system’s application was extended for multiple daily injection (MDI) therapy to serve a larger population. To address issues encountered earlier, the robustness of the system to uncertainty about the model’s pharmacokinetic parameters was tested. In addition, the system’s robustness to irregularities in the timing of long-acting insulin dose administration was tested, accounting for the reality of MDI therapy in practice. It was shown that the system can handle completely skipped long-acting insulin injections used in MDI therapy.
While the engineering design of automated systems for the management of chronic disease like Type 1 diabetes is a complex problem, it still does not encompass the larger challenge of designing human-machine systems of this sort. A formal framework is therefore proposed for holistic pre-design AP system analysis. In addition, a framework for risk identification is proposed that allows to locate and address the causes of suboptimal system performance.
