Sleep Stage Comparisons: Apple Watch vs. Polar Unite

TL;DR

1. The Apple Watch and Polar Unite are found to occasionally match sleep stage data and detect similar times for when changes in sleep stage occur, but the estimates frequently disagree. This is especially true at shorter timescales (which is an ongoing theme of our sleep stage comparisons).
2. The estimated total time spent in each sleep stage is quite similar between the Polar Unite and Apple Watch, even when the shorter timescale sleep stage data disagrees.
3. The Apple Watch shows higher total time spent in light sleep. Polar also always estimates sleep latency values that are far too low. On some nights, both devices estimate acceptable deep sleep and REM sleep durations, while on other nights the estimated time for REM and deep sleep are too low.
4. Polar records far too many interruptions during sleep and the algorithm Polar uses seems to favor the awake state (likely as it is easier to determine this using only accelerometer data). As a result, Polar actually picks up on true interruptions really well, but determining which of the recorded interruptions from Polar are real and which are erroneous is not possible without comparing to other devices or EEG readings. In cases where the Apple Watch also records interruptions, Polar detects an interruption at the same time as well (these are likely true interruptions to sleep).
5. The Apple Watch sleep stage data is again found to estimate long continuous periods spent in a single sleep stage, as well as estimating longer time spent in REM sleep as the night goes on (expected behavior). Polar's sleep stage data has greater variability throughout the night.

We continue our sleep stage comparison series by looking at Apple Watch vs. Polar Unite. We last looked at both Apple Watch and Polar's sleep stage data in our Apple Watch vs. Oura and Polar Unite vs. Oura sleep stage comparison articles.

Sleep stages are determined using accelerometer data to detect movement combined with heart rate, HRV, and respiration rate readings during sleep. These additional metrics are known to depend on different sleep stages, so utilizing all of these metrics together allows wearable devices to more accurately detect sleep stages and the moment when changes in sleep stages occur. For more info on the different sleep stages and how wearables detect them, refer to the intro in our Polar Unite vs. Oura sleep stage article linked above.

When we last looked at Polar, we found that the sleep stage readings occasionally matched with Oura's data, however, a large number of interruptions were registered by Polar throughout the sleep session. We also saw that Polar's sleep latency estimates were far too low.

For the Apple Watch, we previously saw that Apple Watch detected changes in sleep stages quite well (matching Oura's sleep stage data), but the Apple Watch was seen to overestimate the time spent in light sleep and did not record interruptions very often (at times interruptions were also mislabeled as light sleep instead of a switch to the awake state). The Apple Watch was also found to have less variability in the stages for a given night of sleep and remained in a single sleep stage for longer periods.

Let's take a look at Apple Watch and Polar's sleep stage data.

We can see in the graph above that there are times when the data matches, but at other times, there is obvious disagreement. Firstly, we see that between 2AM–4AM the devices record similar sleep stages, but Polar records a larger number of interruptions which smear the data (more on this later). We can see that Apple Watch and Polar both detect deep sleep occurring near the beginning of sleep. Both devices detect a change from deep sleep → REM → light sleep (if the interruptions for Polar are not considered). Both devices also detect a switch to REM sleep and back to light sleep shortly after 4AM, but at slightly different times. A similar pattern occurs between 5AM–6AM from deep to light sleep, but again the change is detected at different times for the wearables. Between 6AM–9AM, we see differences in the detected sleep stages between the wearables and a large number of interruptions recorded by Polar during this time period. They do however agree that the sleep stage moves from light to REM at the same time around 7:30AM. As we discussed in our previous articles, two or more devices agreeing on changes in the sleep stage (especially if they detect a switch to the same sleep stage at the same time), indicate that the change in the sleep stage actually did occur. Closer to wake up time, Apple also records a long period of REM (which makes sense as REM is known to increase in duration with each successive REM stage during sleep). On the whole, we also see that Apple estimates longer times spent continuously in a single sleep stage and has less variability/fluctuations in the sleep stage, which was also seen when comparing Apple Watch with Oura.

Let's now discuss Polar's interruptions. We can clearly see a large number of interruptions (switching back to the awake state) recorded by Polar, which we also observed to be the case when comparing Polar with Oura. These are likely erroneous readings considering that the Apple Watch does not record many interruptions for this night of sleep (and neither did Oura last time). An interesting point is that nearly every time that the Apple Watch detects an interruption, Polar does too (this is showcased better in the second night of sleep discussed below). As with the case for matching sleep stages, two wearables detecting the same interruption makes it more likely that there actually was an interruption during sleep at that point in time. We also saw when we last looked at the Apple Watch data that it will occasionally mislabel an interruption as light sleep, which may be occurring here. It's difficult to say for certain if Apple Watch is mislabeling interruptions here or if Polar erroneously records too many interruptions, but the evidence lends itself to the latter case. Let's take a look at the total sleep stage durations for this night in the bar graph below.

The sleep stage durations for Apple Watch and Polar match quite well for this night. We saw in our last comparison that the Apple Watch typically overestimates light sleep duration. In this case, the light sleep stage duration for both devices is quite high, with approximately 70% of time spent in light sleep. This night was the only time where Polar's light sleep stage duration was found to be greater than the Apple Watch. It is interesting to note that the total durations of the different stages are quite close even though the sleep stage data on shorter timescales disagree (for both the detected sleep stage and the frequency of interruptions). Polar was seen to fluctuate between stages more often, but it occasionally corrected itself to match the Apple Watch sleep stages for certain time periods. This correction from the Polar Unite may have been enough to provide comparable duration estimates on the whole, even when the minute-by-minute sleep stage data did not agree well. The Apple Watch records a fairly long amount of time in REM sleep for this night. While Apple Watch sleep stage data does not fluctuate much within a given night, it was found that between different nights there is a large variation in the estimated duration of time spent in REM or deep sleep. It's difficult to say if this variability is due to expected variations in the time spent in sleep stages for different nights, or if the variability is a result of the Apple Watch struggling to differentiate between REM and Deep Sleep in general.

As we saw last time, Polar estimates far too low values for the sleep latency (15 seconds here), which doesn't even appear on the graph above. The Apple Watch sleep latency estimates are more logical at around 20 minutes. Let's take a look at another night of data.

For this night, an interesting trend is that while the Polar Unite records far more interruptions than the Apple Watch, we find that nearly every interruption detected by the Apple Watch is also detected by Polar (e.g. the interruptions between 4AM–5AM or the two successive interruptions starting at 5:15AM). It's clear from our analysis of Polar Unite's sleep stages that it is sensitive to interruptions and records far too many, but the fact that it matches the Apple Watch in certain places is likely indicative of places where true interruptions occur. In these cases where Apple Watch also detects an interruption, we still see that Polar records more interruptions for that time frame. For example, Apple Watch detects two interruptions just after 4AM, while Polar detects 3 interruptions. Or for the single interruption from Apple at 6:45AM, Polar again detects a greater number of interruptions and has greater variability in the data. These examples further showcase Polar's sensitivity to recording sleep interruptions, even for times when a single, true interruption may have occurred. It is as though the Polar Unite has a hard time "settling" and going back to detecting sleep stages again after picking up a potential interruption.

In terms of the sleep stages in the graph above, there is a greater amount of deviation here than the previous night. The wearables still manage to detect changes in sleep stage at similar times, but they disagree on what the sleep stage changed to more often. For instance, at 4:30AM, both devices detect a change in sleep stage, but the Polar Unite fluctuates between deep sleep and light sleep. When comparing this to Apple Watch, we can see that deep sleep was likely the correct stage for this time frame and that the Polar Unite was struggling to determine this. Moving along, aside from the matching double interruptions, the devices disagree often on the sleep stage between 5AM–6AM. Due to the novelty of sleep stage data from wearables, it's difficult to say which device is detecting the correct sleep stage for this time period without a comparison to EEG readings.

Just after 6AM, we can also see a sequence of interruptions not detected by the Apple Watch followed by a change to Deep Sleep detected by both devices. However, Polar records this change to deep sleep earlier than the Apple Watch. Both devices then detect a change from deep sleep → light sleep → REM before deviating from each other again. Apple records a long REM period around 7:45AM. which polar only picks up as a shorter duration of REM sleep at 8:15AM. Considering that REM sleep tends to last longer with each successive REM stage during sleep, the Apple Watch is likely correct here. A jump to the light sleep stage is then detected by both devices, but again at different times. Finally, just before waking up both wearables detect switches to the awake state. The Polar Unite seems to have mistakenly recorded sleep stage data after 9AM even though the individual had already woken up by this time. Let's now look at the sleep stage durations for this night in the bar graph below.

Even though the sleep stages at shorter timescales disagreed quite frequently, we again see from the above graph how the total sleep stage durations for this night are still fairly close. For this night, the Apple Watch overestimates light sleep as we have seen in other cases, and estimates a short amount of total time spent in deep and REM sleep. Finally, as we have every night so far, the Polar Unite estimates sleep latency that is far too low. The Apple Watch estimates a lower sleep latency than the previous night, which simply could be due to the individual actually falling asleep quicker. For the data enthusiasts out there, check out another night of sleep stage data and durations in the graphs below.

The graphs above provide more evidence for the trends we have already seen for these wearables. We can see places where sleep stages agree (e.g. after 4AM and 5AM) as well as changes in sleep stages being picked up on at the same time. On the flip side, there is clear disagreement at many times throughout the night. We again see Polar's high variability at shorter time scales and a large number of interruptions. We again find Polar matches nearly all interruptions that the Apple Watch detects. In terms of sleep stage duration, there is a greater difference between the devices for this night. We again see the Apple Watch overestimates light sleep duration as well as the variability in estimated REM and deep sleep duration compared with other nights (this time Apple Watch estimates a very short total time in REM). Polar's REM and deep sleep durations are closer to the expected amounts, but Polar estimates a short time spent in light sleep compared with other nights as well as far too low sleep latency as usual.

Wearables use similar techniques and algorithms to determine sleep stages. As a result, it makes sense that the trends match in some places. However, this means that for both agreements or deviations in the wearables' estimates, it's difficult to conclude if the devices (or which device) are showcasing the correct sleep stage without being able to compare the data against EEG/brain activity readings (especially for cases when the sleep stage data from both devices is not volatile and neither device records many interruptions for the time period in question). It should be noted that the different estimates of sleep stages between the devices may also be due to prioritizing different metrics in the wearable's sleep stage algorithms.

The accuracy of sleep stage algorithms that wearables use will improve, especially after utilizing research that compares wearable sleep stage data with medical-grade EEG readings. This would then allow for: accurately detecting sleep stages at shorter timescales with less variability, detecting exactly which sleep stage you woke up from, as well as differentiating between mislabeled or true interruptions. This would then allow you to confidently add sleep stage data from wearables to the tools you can use to better understand your sleep quality, why you feel refreshed vs. groggy when waking up, and any potential factors impacting your sleep.