Understanding Reflection Needs for Personal Health Data in Diabetes

GlucoMine

“…For example, in diabetes care, it is common and recommended that clinicians to review short-term CGM data (i.e., the most recent 14-days) to assess diabetes management since the last clinical visit which is usually 3 or even 6 months prior for an adherent patient [12]. This recommendation is based on prior work that suggests the sufficiency of short-term CGM data based on correlation analysis, more specifically correlations in the range of 0.66 -0.86 between 14-days and 3-months of CGM data [55,58,70]. In this study, we conducted similar analysis as shown in [58,70] to quantify the correlation between the most recent 14-days and 3-months (i.e., 90-days) of CGM data.…”

“…Exploratory analysis in section 4.2 did not show many general patterns of management across the full dataset, hence this section explores the presence of individualized patterns and need for automated algorithms that can find such patterns [8,41,55,72]. This evaluation is motivated by prior work such as Iyengar et al [41] that emphasizes the need for "decision-support algorithms" that "help providers identify trends that may otherwise go unnoticed or be hard to find (e.g., the patient is consistently low [having low blood glucose] in the afternoon).…”

“…Following discretization of blood glucose values, contrasting colors were chosen to represent each category. Although there is no consensus on colors for encoding blood glucose categories, our choice of colors was informed by prior literature [55,74] and informal conversations with patients and clinicians. Based on these, the colors orange, green, and blue were used to represent the 'low', 'normal', and 'high' blood glucose categories while the neutral color, white, was used to represent the 'missing' category.…”

“…Hence, day of the week and time of day are promising features to leverage for identifying temporal patterns. In addition, tools for analyzing blood glucose data including conventional tools highlight the importance of facilitating comparison by date, time of day, and day of week [15,40,55,59]. Building on this knowledge, in this step we transform an extensive and longitudinal 1-D time-series dataset (𝐵𝐺 1 , 𝐵𝐺 2 , ...𝐵𝐺 𝑡 ) to a 3-D grid structure (x,y,z), where the x-axis represents days of week (Monday, ..., Sunday), the y-axis represents time of day (12 AM, ... 11:59 PM), and the z-axis represents the number of weeks of data (𝑊 𝑒𝑒𝑘 1 ,𝑊 𝑒𝑒𝑘 2 , ...,𝑊 𝑒𝑒𝑘 𝑛 ).…”

“…Specific to diabetes care, Iyengar et al [41] highlight the need for automated algorithms that can "identify trends that may otherwise go unnoticed or be hard to find (e.g., the patient is consistently low [has low blood glucose] in the afternoon)," and thus "reduce the burden of deciphering glucose data". Similarly, Prioleau et al [55] highlight the need for tools that "extract trends/patterns in blood glucose data" and data interpretation solutions that minimize cognitive load. Toward a solution to the aforementioned challenge, the objectives of this paper are to:…”