Beat the Curve: Designing Adaptive Blood Glucose Management Strategies for Non-Pump Patients with Type 1 Diabetes

Abstract

There are approximately 1.25 million people in the US living with Type 1 Diabetes (“T1D”), an autoimmune disease that triggers the body’s defense system to attack insulin-producing beta cells in the pancreas. Usually diagnosed between the ages of four and fourteen, T1D is a condition that dramatically alters a patient’s lifestyle, requiring multiple daily insulin injections and finger-stick tests to control and monitor blood glucose (“BG”) levels.

Recent technological breakthroughs have led to the development of closed-loop systems that mimic the supervisory action of a healthy, non-diabetic pancreas. Most artificial pancreas (“AP”) designs consist of a continuous glucose monitor, an insulin pump, and a control algorithm to direct the timing and quantity of insulin released into the body. In some ways, excitement for AP technology is outsized given the current care preferences of most T1D patients: around half administer insulin without a pump (choosing alternatives like injections or inhalations) and only 10% use a CGM (opting instead for traditional finger-stick measurements).

In the following thesis, we formalize the sequential decision-making process of measuring BG levels and dosing insulin for the non-pump, non-CGM T1D patient. We begin by implementing a mathematical model of the T1D patient system, which dynamically-changes over time given meal and insulin inputs and returns noisy BG measurements. Combining industry-standard dosing recommendations with inspiration from AP control algorithms, we design various dosing policy functions that incorporate beliefs about underlying patient-specific parameters to optimize BG outcomes.

By implementing our model in Python, we are able to evaluate the performance of policies over a large number of simulated trials, and find that policies involving deterministic “lookahead” features to forecast the uncertain future of the patient system outperform policies that dose in a myopic fashion. Finally, through modeling extensions and parameter tuning exercises, we demonstrate ways in which non-pump, non-CGM T1D patients can leverage insights from the artificial pancreas research community to power a proactive, rules-based approach to BG management.