Temporal Characterization of Glycemia Risk Index Sequences through Random Walk Bayesian Clustering

Pattern detection in Continuous Glucose Monitoring (CGM) signals could support long-term diabetic patient monitoring. The Glycemia Risk Index (GRI) has been recently proposed to summarize segments of the CGM signals focusing on the identification of glycemic profiles associated with hyperglycemic and hypoglycemic events. In this study, we show how GRI time sequences can be used to identify groups of glycemic risk patterns that can effectively describe the temporal sequence of glycemic events. We computed the daily GRI value in 880 Monday-Sunday aligned weekly signals extracted from the CGM measurements of 21 type 1 diabetic patients. The Random Walk Bayesian Clustering (RWBC) method was applied to the obtained GRI weekly profiles, resulting in 16 clusters with clearly different temporal behaviours. The Area Under the Median Curve (AUMC) was chosen as an indicator of the pattern severity to characterize each cluster. In this way, we defined a risk-related sub-grouping of weekly GRI patterns, which provides an informative inter- and intra-patient description of the glycemic risk profile over time. We used Kruskal-Wallis and Dunn’s post hoc tests to compare the distributions of state-of-the-art glycemic variability metrics among clusters, achieving significant results (p-value < 0.001) for most of the comparisons. Specifically, six clusters with elevated AUMC showed significantly high-risk profiles in terms of glucose variability. In addition, we conducted a preliminary qualitative validation analysis to associate the clusters with lifestyle-related information extracted from medical reports of patients’ regular visits using the open-access LLM numind/NuExtract-1.5. This qualitative analysis revealed meaningful insights into the relationship between quality of life and glucose profiles.