Sleep Apnea + Wearables: Part 2: What can I learn about sleep apnea via wearables?

Never before in history have humans had the ability to collect so much data about ourselves. Billions of data points are stored about each of us in servers across the world. Now the opportunity and challenge is how we extract the potential of this information for good — how we leverage all of this technology to help us live longer, happier and healthier lives. We are at a crucial moment in history as we navigate how to best integrate rapidly advancing technology into society and use it for good.

Healthcare has been much slower to adopt personalized, big data approaches relative to many Silicon Valley internet companies. Healthcare systems are slowly embracing precision medicine, the idea that treatment and recommendations can be tailored specifically to individuals, as opposed to the average patient. However, there are many areas of medicine where this isn’t yet commonly practiced. What can we, as healthcare consumers, do to accelerate this? How can we take our healthcare into our own hands? How can we leverage real time measurements to gain insight into our own health?

Sleep apnea is an ideal case study for this exploration as it is a very common disorder, affecting 17–50% of people, and, at least in my experience, personalized approaches are not common. Furthermore, key technologies to monitor sleep apnea (continuous heart rate and blood oxygen (SpO2) monitoring) have recently become affordable and mainstream. Sleep apnea is a chronic, multifactorial disorder characterized by low oxygen events throughout the night, which can lead individuals to feel subpar during their waking hours and may lead to serious long-term conditions. Risk factors include obesity, diabetes, age, and being male. Some of these risk factors are modifiable, but like many medical conditions, many people have this disorder without presenting any of the standard modifiable risk factors. This blog post focuses on behaviors that may have an effect on sleep apnea severity.

In part 1 of the sleep apnea blog I wrote about a person with sleep apnea, two conflicting overnight sleep apnea tests, and broad, unpersonalized recommendations.

I sought to answer three initial questions:

• Why did the two sleep apnea tests have such different results?
• Is it important for the person to not sleep on their back? (One test indicated they were, and the other indicated they weren’t)
• Are there any other behaviors or signals tied to increased sleep apnea?

Data:

• Wellue O2 monitor, sleep position mobile app
• About one month of nightly real time measurements, including heart rate and SpO2.
• Sleep apnea can be characterized by low oxygen events, which I’m considering to be when oxygen drops below 92% for two 5 second intervals. The literature most often considers a low oxygen event to be a 4% decrease in SpO2 which, in general, is less conservative than this approach.

Question: Why did the two sleep apnea tests have such different results?

Answer: Sleep apnea severity differs dramatically by night, so it is unsurprising that two isolated nights would show different results

The two physician-prescribed tests were taken one year apart, and the second test indicated significantly worsening sleep apnea. The severity of sleep apnea is extremely variable by night with the number of low oxygen episodes per night varying widely. Overall, this represents mild sleep apnea, with most nights having < 20 ODI events. Given the high variability of sleep apnea presentation by night, it makes sense that two single night tests may have very different results.

Figure 1: Histogram of the number of low oxygen events per hour.

Question: Is it important to avoid back sleeping?

Answer: Yes

Yes. As one of the two physician-prescribed tests (but not the other) indicated, sleep apnea events are significantly more likely to start in a back position than a non-back position (chi squared test, p=5.3e-40). In fact, they are 5.3 times more likely to enter a low oxygen event when laying on their back than a non-back (side) position.

Figure 2: Histogram of the number of low oxygen events per hour.

Figure 3: This is a density plot of the total time over all nights in control (SpO2 >= 92%) and ODI event (SpO2 <92%). Orientation of 0 is back sleeping, larger numbers are the left and right sides.

Question: Besides back sleeping, are there other significant factors identified?

Answer: Yes.

It isn’t just back sleeping that results in more ODI events. There are also clear peaks around orientations (-)100 to 120 degrees, indicating additional nuance in how sleeping position affects low oxygen events. To understand what this corresponds to, I’ve added the images below. The ‘bad’ sleeping position (right) corresponds to that ~100–120 degrees, while the ‘good’ sleeping position (left) corresponds to ~50–90 degrees which have lower incidence of low oxygen events in this dataset. Published studies typically bin sleep position into four buckets, so it is unclear whether or not this has been observed previously. Another interesting finding is that ODI events are more likely to occur soon after sleeping position is switched. As of now low oxygen events cannot be well predicted from the current feature set.

Image Credit: https://lastbackpain.com/sleeping-position-side-sleeper/

Left: ‘Good’ sleep position, fewer low oxygen events

Right: ‘Bad’ sleep position, more low oxygen events

Sleep apnea is a complex disease that afflicts millions of people. Combining knowledge from n of 1 studies can further our understanding of symptoms and have the potential to crowdsource methods of reducing nightly symptoms. In part three, we will explore if behavior changes in sleep position decrease sleep apnea severity.

Appendix

Methods

Real time measurements of sleep position were taken via the Somnopose app in a shirt pocket every 2 seconds, and heart rate and SpO2 were recorded with the Wellue O2 monitor every 4 seconds. Data sets were merged and the first measurement from each device in every 5 second window was taken. This resulted in 26 nights with data and >156,000 rows. A ‘low oxygen event’ or ODI is when SpO2 drops below 92% for two 5-second intervals. Thresholds of 88% and 90% received similar results. In the literature, an ODI event is defined in various ways. Another common way it is defined is a 4% SpO2 lasting at least 10 seconds.

Disclaimers/Assumptions/Etc.

• Most importantly, I am not a physician and this is not medical advice. I have graduate level training in biomedical informatics and data science — the goal of this post is to explore how we can make our data work for us.
• Results above assume the monitoring equipment is accurate. Qualitative tests indicate accuracy.
• Code: https://github.com/kimberlymcm/digitalhealth_project/blob/master/src/Sleep_Position_O2_investigation_20200628_Sp02_92_for_blog.ipynb