Learning Data-Driven Policies for Insulin Dosing Adjustments
Currently, endocrinologists make adjustments to their patient's insulin dosing by reviewing logbooks of glucose measurements. Clinicians look for patterns in these measurements (e.g., consistent elevation) and suggest changes to the carb ratio settings, the basal dose and/or the correction factor. However, it's not clear that these recommendations are made often enough. Some patients may not be getting the kind of attention or support they require. To increase the efficiency and bandwidth of interactions between clinicians and patient, we propose the development of decision-support tools that can help clinicians more rapidly make decisions regarding adjustments. Using observational data from patients with Type I diabetes and reinforcement learning techniques, we aim to learn data-driven policies for proposing insulin dosing adjustments. We will first learn to model the current clinical approach to making such adjustments, and then learn to improve on this policy with the goal of increasing the time a patient spends at the target blood sugar. Once implemented, such a system could help clinicians target their efforts, impacting a large number of people with Type 1 diabetics seen at UM hospitals.
$200,000 grant from the Juvenile Diabetes Research Foundation International (JDRF)
$9,000,000 grant from the JDRF and the M-Diabetes Center of Excellence at the University of Michigan
Presented at the International Conference on Machine Learning Workshop