Using CGM data to answer your questions about diabetes
At Steady Health, data and care is our focus. To understand our members’ data, we started asking some fundamental questions and thought we’d share our early findings. In coming blog posts, we’ll cover the process of how we use technology to empower our clinical staff to provide world-class care.
If you haven’t heard of Continuous Glucose Monitoring (CGM), it’s a wearable sensor that measures blood glucose every 1–5 minutes. It has revolutionized the ability to make data-driven diabetes care decisions. To read more about the impact of CGM check out The wearable that changed my life. The author, Henrik Berggren, has also gracefully provided 90-days of his Dexcom data that we’re about to explore in this post.
Okay, let’s jump in! We have a bunch of questions on our data, let’s start visualizing the data to reveal some interesting insights.
These sections will be laid out by asking questions and try to let the data and plots give us the answers.
What are the daily trends?
Let’s have a look at an average day for a Patient, meaning all glucose values aggregated and shown over 24 hours. Our goal is to uncover any trends and cycles in the Blood Glucose data.
Here we can see the variability and trends of blood sugar throughout the day. There’s a general trend of night-time highs — where we can see higher than average glucose levels after bedtime.
Further, we see higher variability around the same time. Something to revisit.
Time In Range, a well-known success-metric for people living with diabetes, however, is not perfect. It tracks how much time a Patient had spent neither with a too High not too Low blood glucose. However, we want to account for variance too. As we observe, most of the member data fall within the standard definition of In Range. Broadly suggesting good glycemic control. Yet looking at the valleys in the plot, we can see the lowest blood glucose reading occurs pre-breakfast and early in the morning. Showing a cyclic nature shown, which Time In Range fails to surface.
Do Weekends differ from Weekdays?
I always look for an excuse to use violin plots, they look so nice! In this case, they actually make a lot of sense too. Let’s group our data by day-of-week and observe how the distributions differ. Et violá.
Immediately we notice that Thursdays and Friday display a larger variance than other days. (Which can be fine, if the member applies the appropriate correction Insulin dosages.) Definitely, something to revisit with this member.
Other days look pretty solid, with a well-bounded distribution. Monday, Tuesday, Wednesday and Saturday all look to be In Range.
From these insights, we can uncover behavioral triggers, like going for runs on Wednesday, or Sunday brunches. We want to make actionable items with the member based on these insights.
We can even take it one step further.
Let’s add in data after a coaching session with these behavior goals. The left-side (green color) of the plot below is pre-coaching and right-side (salmon color) is post-coaching.
We can then observe that Fridays are getting a lot less variable, but on Sunday’s there’s been more time in high recently compared to before. Cool!
Are there any daily or periodic trends in the blood glucose data?
As we saw in the first plot, recurring patterns can be seen during the day. Let’s dig into this idea and observe the Blood Glucose samples in a heatmap.
A heatmap gives us a bird’s eye view on the data and enables easy detection of patterns. For example, scanning across the nighttime portion of the plot, we can see infrequent streaks of red (highs) broken up by chunks of dark blue (lows).
Perhaps tighter glycemic control during nighttime/morning might be worth experimenting with.
Looking at the big picture, apart from the short episodes mentioned above, no major periodic trends are evident for the 90-day period. In cases where the member is moving cities, starting a new job, or training for a marathon, these birds-eye view plots is a useful tool to uncover larger trends.
This was just a brief introduction of how member data is our ground truth and how we let our focus narrow until we can draw conclusions. With this process, we can build amazing things like meal detectors, predictive systems, queryable data and so on.
Thanks to Sid Ghodke and Henrik Berggren.