Mining associations between glycemic variability in awake-time and in-sleep among non-diabetic adults

Abstract

It is often assumed that healthy people have the genuine ability to maintain tight blood glucose regulation. However, a few recent studies revealed that glucose dysregulation such as hyperglycemia may occur even in people who are considered normoglycemic by standard measures and were more prevalent than initially thought, suggesting that more investigations are needed to fully understand the within-day glucose dynamics of healthy people. In this paper, we conducted an analysis on a multi-modal dataset to examine the relationships between glycemic variability when people were awake and that when they were sleeping. The interstitial glucose levels were measured with a wearable continuous glucose monitoring (CGM) technology FreeStyle Libre 2 at every 15 min interval. In contrast to the traditional single-time-point measurements, the CGM data allow the investigation into the temporal patterns of glucose dynamics at high granularity. Sleep onset and offset timestamps were recorded daily with a Fitbit Charge 3 wristband. Our analysis leveraged the sleep data to split the glucose readings into segments of awake-time and in-sleep, instead of using fixed cut-off time points as has been done in existing literature. We combined repeated measure correlation analysis and quantitative association rules mining, together with an original post-filtering method, to identify significant and most relevant associations. Our results showed that low overall glucose in awake time was strongly correlated to low glucose in subsequent sleep, which in turn correlated to overall low glucose in the next day. Moreover, both analysis techniques identified significant associations between the minimal glucose reading in sleep and the low blood glucose index the next day. In addition, the association rules discovered in this study achieved high confidence (0.75-0.88) and lift (4.1-11.5), which implies that the proposed post-filtering method was effective in selecting quality rules.

Keywords: association rules mining; continuous glucose monitoring; data mining; glycemic variability; repeated measure correlation.

Figure 1

Boxplots of glycemic variability metrics (ns: p > 0.5; *: 0.01 < p ≤ 0.05; **: 0.001 < p ≤ 0.01; ***: 0.0001 < p ≤ 0.001; ****: p ≤ 0.0001; HBGI was scaled up 100 times).

Figure 2

Heatmap of significant correlations between in sleep GV metrics of day N and awake time GV metrics of day N−1. Only moderate (0.4 < rmcor ≤ 0.6) to strong correlations (rmcor > 0.6) are shown.

Figure 3

Heatmap of significant correlations between in sleep GV metrics of day N and awake-time GV metrics of day N. Only moderate (0.4 < rmcor ≤ 0.6) to strong correlations (rmcor > 0.6) are shown.