A Machine Learning Approach to Treatment Improvement in Type 2 Diabetes using Glucose Data

Type 2 diabetes (T2D) has become a major problem to global health and is now characterized as a pandemic. Despite the improved treatment options, diabetes technology, and international guidelines, a critical number of patients do not reach their recommended glycemic goals. To address this unfortunate trend, coupled with a worldwide shortage of healthcare professionals, there is a need for innovative approaches to improve treatment outcomes and maintain glycemic targets. In this context, continuous glucose monitoring (CGM) technology has revolutionized the field of diabetes by improving insights into temporal glycemic variability and dynamics of the glucose level (i.e., hypo- and hyperglycemia). Thus, it has enabled significant advances in diabetes management. Simultaneously, machine learning (ML) has shown promising performances in the fields of medicine and healthcare, particularly when used for decision support. CGM use in patients with T2D is increasing and has proven beneficial with respect to improved glycemic outcomes. In light of the need for innovative solutions to improve treatment outcomes, the main objective of this thesis was to investigate potential advantages of applying ML on CGM data to address different challenging aspects of T2D treatment. Thus, we attempted to leverage the potential of ML applied on CGM data for personalised and cost-effective treatment decision support. A total of four different treatment aspects using CGM data were investigated. First, we quantified the reliability of glycemic metrics for glycemic control assessment using less than two weeks of CGM data, where two weeks is the recommended length of CGM data. Using less than two weeks, we observed relatively low deviations for time in range (TIR) and average based metrics, whereas large deviations were observed for metrics describing infrequent and critical events, such as time below range (hypoglycemia). Furthermore, discrepancies were observed between two consecutive two week periods. Second, we examined if three days of CGM could improve the prediction of treatment outcomes for patients with T2D initiating basal insulin injection. Based on the CGM derived consensus metrics, a trend towards improved performance was observed when combined with baseline hemoglobin A1C (HbA1c). However, it appears difficult to predict a binarized HbA1c outcome based on the considered patient information to a level adequate for clinical use. Third, using simulated CGM data in which adherence could be controlled, we explored simple to a range of more complex approaches for adherence detection of oncedaily basal insulin injection. Providing that the results of this study are reproducible using real CGM data, an early adherence detection algorithm could be developed as part of a decision support system. Finally, we investigated the advantages of deploying complex nonlinear methods as opposed to conventional time-series analysis for glucose level prediction. According
to the study findings, advanced predictive models can be applied when designing decision support tools with glucose level forecasting functionality. However, a conventional linear approach may be more suitable in some circumstances, notably when little CGM data are available. Altogether, we were able to inspect possible benefits, challenges, and limitations with respect to the considered ML methods. Studies undertaken as part of this thesis provide valuable insights into the predictive capabilities of ML. However, the benefits and promising potential of ML remain unclear due to data constraints. This thesis can be considered as part of early attempts towards leveraging ML and CGM data for improved diabetes management and provides a starting point for future research efforts. Simultaneously, considering the improving technologies and substantial efforts made to remedy the current shortcomings in this field, a promising future lies ahead.