Online Adaptive Personalization of Supervised Learning Models in Mobile Health Treatment Systems

There is often significant heterogeneity present in the context of systems engineering problems. This heterogeneity can limit the effectiveness of policies and models that are designed to operate at a coarse, population level when the actual point of intervention is at the level of the specific and varying subgroups or individuals constituting the population. Thus, methods of model personalization may be required to achieve desired outcomes. In this dissertation, we propose a means of rapid, online model personalization of decision rules based on statistical learning models, GMAdapt, which is informed by the context of decision support systems for the management of type 1 diabetes. To evaluate the effectiveness of this procedure, we performed experiments using both numerical simulations and retrospective data analysis based on real-world clinical trials conducted at the UVa Center for Diabetes Technology. In addition to the adaptation procedure itself, we present a simulation based methodology for deconfounding data to address the issue of intervention generated label noise. This method is evaluated in silico using the UVa/Padova type 1 diabetes simulator and compared against some alternative methodologies for creating end-to-end systems capable of adaptively learning personalized decision rules in spite of system generated interventions and resulting label noise.