This study measured sleep patterns and wrist temperature patterns of 3 groups of participants: (1) healthy younger adults (aged 20-30 years), (2) healthy older adults (aged≥65 years), and (3) OAWD. Inclusion criteria were any gender, any race, age in the targeted populations, and ability to understand English. An additional criterion for OAWD was early- to middle-stage dementia. Individuals diagnosed with severe sleep disorders were excluded from the study.

We recruited 10 participants for each group based on the convenience sampling method. Healthy adults were recruited through university-wide poster advertisements and word of mouth. Healthy older adults were recruited through phone calls and email invitations. OAWD participants were recruited from a local long-term care (LTC) residence. The staff at the LTC setting requested that they coordinate the initial identification of potential participants and support the recruitment process. Each potential participant was then contacted and introduced to the study by the researcher. Assent to continue participating in the study was obtained from each OAWD participant every time they were contacted by the researchers. This study was approved by the Human Research Ethics Committee of the University of Waterloo (ORE #31860 and ORE #40459).

Each participant was asked to wear a commercially available smart wristband (Mi Band 2, XiaoMi; Figure 2) and a custom-built wristband developed by the authors (Figure 3) on their left arm for 14 days. The Mi Band 2 was chosen because of its popularity, low cost, long battery life, and accessibility of actigraphy data. The custom-built band consisted of a temperature sensor (iButton DS1922L, Maxim) that maintained contact with the anterior side of the wrist and a 3-axis accelerometer sensor (AX3, Axivity). Both sensors operated offline by storing data locally; data were extracted after the 2-week study period using USB adaptors. All participants were offered to keep the Mi Band 2 if they wished on completion of the study.

The study protocol was the same for healthy younger adults and healthy older adults, with an altered OAWD protocol. Before the start of the study, all healthy participants were asked to complete 4 questionnaires: (1) a demographic form; (2) morningness-eveningness questionnaire, which is a measure of circadian rhythm type [44]; (3) Pittsburgh Sleep Quality Index (PSQI), which is a measure of subjective sleep quality [45]; and (4) Epworth Sleepiness Scale (ESS), which is a measure of daytime sleepiness [46]. Participants were then given the 2 wristbands; instructed on how to wear them; and asked to wear both on their left wrist for 14 days, except when showering or bathing. Participants were also asked to fill out a daily sleep journal that was adapted from a study by Monk et al [47] to capture sleep onset, offset, and subjective sleep quality.

As OAWD were not able to fill out the questionnaires, only the demographic form was completed and was done with the help of staff from their LTC residence. In addition, the Mini-Mental State Examination (MMSE) [48] was conducted to approximate the level of cognition for each OAWD. After giving each OAWD the wristbands, their personal support workers (PSWs) were asked to help with getting wristbands off and on when they assisted OAWD shower or bathe as well as to routinely check that the person was not adversely affected by wearing the wristbands. In addition, PSWs were asked to observe participants’ sleep status and check off either awake or sleeping on a sheet every half an hour throughout the day and night as part of their routine in with the OAWD. Sleep status from this sheet was then used to generate an observed sleep journal.

Sleep onset and offset, the number of nighttime wake-ups, and sleep quality scores were obtained from the sleep journals and were used to help interpret actigraphy and temperature data. Although we thought it was unlikely that people could estimate their sleep at a minute level, we considered sleep journals of younger adults and healthy older adults reliable at the half-hour interval level. Occasionally, participants would miss reporting sleep onset and offset. In these cases, the sleep onset and offset were manually identified from the data from the custom-built wristband AX3 actigraphy data.

For OAWD, the intention was to use the sleep status sheet filled out by PSWs to extract sleep parameters and use these as ground truths for sleep. However, at the end of the study, it was found that, on average, 40% of the sheet was filled (ie, the PSWs did not report sleep status of OAWD 60% of the time). Owing to the large amount of data missing from the OAWD sleep journals, we chose to manually identify sleep periods from the AX3 data and combined those identified sleep periods with PSW sheets to estimate sleep for all OAWD.

One-tailed t test was used to compare the sleep parameters and rhythm indices between different groups (young vs old and old vs OAWD). The Pearson correlation coefficient was used to assess correlations between different circadian rhythms and sleep parameters calculated; P<.05 was considered to be significant.

When analyzing the collected data, it was found that the sleep of some participants was not correctly reported by the Mi Band 2. For example, most PSWs reported that OAWD had 8 hours of sleep at night, but the Mi Band 2 detected as few as 40 minutes of sleep. The AX3 data from the custom-built band were inspected for the periods Mi Band 2 had classified as wake periods during the night as well as wrist temperature measurements for the same time. A healthy younger participant who had good sleep quality; a healthy older participant with poor sleep quality; and 2 OAWD who had poor, interrupted sleep were selected as case studies to illustrate this phenomenon and are presented later in this paper.

In total, we recruited 10 healthy younger adults, 10 healthy older adults, and 8 OAWD. A total of 7 OAWD were able to give their own consent and sign the consent form; for the 1 participant who could not self-consent, his power of attorney signed the consent form, and assent was obtained from the participant. Participants’ demographic data are presented in Table 2. The average MMSE score of the 8 OAWD participants was 20. Moreover, 7 OAWD were in the mild cognitive impairment category (ie, a score between 18 and 23), and 1 OAWD had a score of 16, which indicates severe cognitive impairment [50].

The sleep parameters obtained from the sleep journals for all participants are presented in Table 1. Compared with older adult groups, healthy younger participants tended to sleep later and wake up later, with the shortest average sleep duration among the 3 groups. Between the 2 older adult groups, OAWD slept earlier and had a longer sleep duration. Although we did not require participants to record daytime naps, we asked them to report whether they had the habit of napping. Only 1 younger adult and 1 healthy older adult reported that they took regular noontime naps. All OAWD napped during the day, sometimes more than once. As the recreated sleep journals of OAWD only summarized nighttime sleep and did not capture daytime naps, the total sleep duration of OAWD might be longer than reported. Healthy older adults had better compliance in reporting and provided more detailed sleep journals (eg, a few participants consistently recorded their nap time during the day).

The absolute differences in sleep parameters between the Mi Band 2 and sleep journals are summarized in Table 3, and significant differences were found in sleep duration between the younger adult group and the healthy older adult group. Between the 2 older adult groups, the absolute differences in sleep onset, offset, and duration of OAWD were significantly higher than those in the healthy older adult group. As shown in Table 4, of the 106 days of data measured from the 8 OAWD, for 28 of the days (28/106, 26% of days monitored), the Mi Band 2 did not detect any sleep, and we confirmed that all OAWDs were wearing the Mi Band 2 correctly throughout the study. Discrepancies between sleep parameters reported in the recreated sleep journal and detected by Mi Band 2 for OAWD were large, namely, sleep onsets detected by Mi Band 2 were often noticeably delayed (ie, on the order of hours), and the offset was earlier than what was reported. The average absolute difference in sleep duration between the Mi Band 2 data and the recreated sleep journal for OAWD was more than 6 hours. A discussion of this detection error is presented later in the Case Studies section.

The average wrist temperature data of the 3 groups are shown in Figure 4. Although all curves have higher daytime (awake) temperatures, the OAWD group has the flattest temperature curve with the lowest nighttime wrist temperature and the highest daytime wrist temperature compared with the other groups. Although the average nighttime wrist temperature was similar for healthy older and younger adults, older adults tended to have a higher daytime wrist temperature, causing a flatter average curve for healthy older adults compared with younger adults. The wrist temperature indices are presented in Table 5.

Data from 4 participants are presented below as illustrative case studies. One of the case studies is of a younger healthy sleeper, and the other 3 cases are representative data from irregular sleepers, 1 of which is a healthy older adult and the other 2 are OAWD.

In line with other recent studies [17,19], this study found that wrist temperature increased when people were asleep, even when their sleep patterns were irregular. As illustrated in the case studies presented earlier, wrist temperature appears to correspond with sleep status changes regardless of whether body movements occurred; wrist temperature increased when people were asleep and decreased when they woke up. This adds evidence that wrist temperature is correlated with circadian rhythms and associated changes in thermal regulation of the body during sleep. This appears to hold true for OAWD as well, whose circadian rhythms are impacted by both aging and dementia (and often other morbidities); their wrist temperature appears to reflect when they are asleep and awake. A larger sample size is needed to draw more definitive conclusions regarding the impact dementia may have on CBT and wrist temperature.

From the results presented in Table 5 and Figure 4, there seems to be a trend of increasing wake temperature and decreasing sleep and wake temperature difference as people age and people become more pronounced dementia in this study. It is not clear from our data whether and how much this increased discrepancy is due to dementia or other factors, such as increased average age of OAWD, greater prevalence of comorbidities, and/or how living in an (often busy) LTC environment may disrupt sleep. The less significant day versus night wrist temperature contrast suggests changes in the circadian rhythms of older adults and OAWD compared with younger adults. This finding is consistent with other studies (eg, [51]).

The IS values suggest that the 2 healthy groups had similar IS for their wrist temperature rhythm, and the dementia group had lower wrist temperature rhythm stability. The lower IV values of the older adult groups indicate that changes in wrist temperature rhythms tend to be less variable intraday as people get older, especially with people who have dementia. As IS reflects the stability and the IV reflects the fragmentation of one’s wrist temperature rhythm, the significantly lower IS and IV of OAWD could be associated with irregular sleep patterns and increased daytime sleepiness. Most OAWD in this study took frequent naps and experienced insomnia at night, which almost certainly increased the average daytime wrist temperature, lowered the average nighttime wrist temperature, and accounted for a greater amount of flux in nighttime wrist temperature. Rhythm indices, such as IS and IV, could be further explored with these and other populations to determine how they might be used to infer circadian rhythm health.

Some studies [30,52-54] have reported excessive body or wrist movements of OAWD during sleep. This research supports the hypothesis that advanced sleep monitoring systems should detect nighttime body movements. Although the exact cause of abnormal body movements by the older adults in this study is unknown, body movement during sleep seems to become increasingly common with age and was present in all our OAWD participants. A system that can detect involuntary movements could be valuable in the detection of previously unknown conditions (such as the case with our OA1 participant) and possibly support the diagnosis and management of conditions that impact sleep.

There are 2 noteworthy findings regarding sleep detection using the Mi Band 2. First, by comparing the results of the Mi Band 2 with AX3 data from the custom-built wristband and reported sleep from sleep journals, it appears that the Mi Band 2 does not accurately detect people’s sleep when they have pronounced or irregular body movements during sleep. Second, the Mi Band 2 sleep detection algorithm does not appear to work well if the user has fragmented sleep (ie, waking up multiple times during the night). For example, OAWD2 slept in short and sporadic episodes; most of her sleep was undetected by Mi Band 2, which reported sleep for only 2 of her 13 recorded days. From the example data in Figure 9, we can observe that each sleep episode of OAWD2 was short and each wake-up accompanied body movements and that the Mi Band 2 did not detect any sleep. As the Mi Band 2 detects sleep based on accemetry data, it is plausible that periodic body movements during sleep, such as those seen in the older adults presented in the case studies, may be the cause of the misclassification of these periods as awake instead of asleep. Moreover, the Mi Band 2 may be trained on data from people with 1 or 2 long nighttime sleeping periods than multiple short ones. The hypothesis that movements and short sleeping episodes cause the Mi Band to miss episodes of sleep is supported by the data from our other participants.

Although the sleep detection algorithms for most commercial smart wristbands are proprietary information, as they rely on accelerometers and gyroscopes, their sleep detection is based on movement. For a regular sleeper, frequent and substantial body movements at night usually indicate the person is awake, and fewer and infrequent movements indicate sleep. Although an activity-based rule for sleep detection may be robust for people who are mostly still when they sleep, this study shows that sleep detection based on activity is not accurate for some older adults, especially OAWD. As younger adults are the primary consumers of wearable wristbands, the Mi Band 2 sleep detection algorithms may have been trained predominantly on data from younger adults with healthy sleep patterns, thus biasing the algorithm to detect sleep for people who fit that profile. This aligns with other studies that have found that commercial wristbands performed poorly on populations with sleep disorders [55]. Therefore, other smart wristbands that track sleep based on movement alone may have similar performance as the Mi Band 2 for the same reasons; however, this was not investigated in this study; therefore, it remains a question for future research.

This research highlights the need for the older adult population to be included in the development of sleep detection algorithms, including the possibility of developing smarter algorithms that autonomously identify the type of sleeper a person is and self-adjust appropriately to that person’s sleep habits. This requires research into how to categorize what good sleep is for that particular person in a way that helps people, their care partners, and/or caregivers understand. This information could then be used to better manage sleep without missing information that may indicate a change in sleep or sleep patterns that are of possible concern. As wrist temperature appears to be highly correlated with sleep regardless of age or dementia, the inclusion of wrist temperature as a complementary sensor to movement-based ones could enable their strengths to be leveraged and fused to support more accurate sleep detection.

More accurate sleep detection for older adults using wearable systems could support better short-term and long-term monitoring of sleep. Through daily sleep tracking, irregular sleep quality and patterns can be recognized by machine learning algorithms and then be reported to older adults or their health care providers. Such monitoring systems are especially useful for OAWD because the incidence of poor sleep is quite high in this population, and self-reporting of sleep is often difficult or impossible. Wearable technologies for sleep tracking and sleep problem identification could help to better understand older adults’ sleep using objective measures, but this approach is only possible if the results are accurate. Accurate sleep data would enable the creation of tailored sleep management plans, enabling better sleep support for each individual, thus supporting better health outcomes and quality of life.

This study had several limitations. First, only the Mi Band 2 was the only commercially available wearable device that was examined. As other commercially available smart wristbands collect other types of data (eg, PPG) and use different sleep detection algorithms, their performance will likely be different. These alternatives should be examined in healthy older adults and OAWD in future studies. Second, the sample size of this study was small. Future research with larger sample sizes is required to provide a deeper, comprehensive understanding that needs to support greater generalization of the results. Third, all our OAWD participants lived in the same LTC home. Data from other LTC homes as well as OAWD living in the community need to be investigated to determine how the environment plays a role in sleep detection and dementia. Finally, sleep onset and offset were estimated from sleep journals and/or actigraphy data. Using another method, such as computer vision, could estimate onset and offset more accurately as well as provide information about what was happening during sleep (eg, giving insight into repetitive movements).

This research adds evidence that wrist temperature can be used as an indicator of sleep status, including for OAWD and people with irregular sleep patterns. As illustrated by case studies from our data, suboptimal sleep detection performance by a commercial wristband was likely because of broken sleep patterns and body movements. As wearable technologies are increasingly being used by LTC homes and healthy older adults to track sleep and inform sleep management and support, this research suggests that caution should be used when interpreting sleep data when monitoring the sleep of older adults, particularly those living with dementia. This highlights the need for future research and development to create systems that better complement older adult populations. Future sleep monitoring wearable systems could consider adding a temperature sensor to capture the waist temperature as an extra indicator of sleep combined with motion-based data and machine learning algorithms of typical sleep patterns.