Sleep is vital for human health, as it promotes physical and mental well-being at the individual and population levels [1]. Sleep affects brain function and the performance of other systems within the body such as digestive, cardiovascular, and endocrine [2,3]. The lack of proper sleep can cause fatigue, reduced concentration, and depression [4]. In addition, inefficient or disturbed sleep, due to behaviors such as technology use (eg, use of mobile device screens), can also lead to chronic stress and poor mental health [5]. Researchers have found that insufficient sleep is associated with a significant increase in mortality, diabetes, cardiovascular disease, coronary heart disease, and obesity [6]. Today, sleep health is understood not as an isolated portion of the day but as a significant part of a healthy 24-hour cycle [7].

The Public Health Agency of Canada recommends that individuals aged 18 to 64 years get between 7 and 9 hours of sleep per night and those aged ≥65 years between 7 and 8 hours per night [8]. In Canada, at least one in four adults is not getting enough sleep [8]. Even higher levels of sleep deprivation are reported for those aged 35 to 64 years. Similarly, in the United States, 25% of adults self-reported having <7 to 9 hours of sleep [9]. Insufficient sleep duration occurs despite the time spent indoors having increased over the last century [10].

Sleep research has historically relied on single-sleep questionnaires or sleep diaries [11]. However, self-reported sleep data are prone to recall bias and social desirability bias [11]. Sleep patterns have also been inferred by measuring human brain activity, breathing and blood oxygenation levels, muscle and eye movements, and heart rate [12]. These types of studies, although informative, rely on sleep data collected from artificial laboratory environments and do not reflect sleep patterns in a real-world setting, as individuals are often sleeping in a controlled environment with different, albeit fewer, disturbances. There is a need for the modernization of research methodologies to enable the unbiased collection of sleep pattern data in real-world settings.

The development of smartphones and wearable devices has enabled continuous behavioral monitoring [9]. The assessment of sleep behavior can now be performed using wearable devices such as smartwatches or actigraphs [11] or by interpreting the interactions of a user with the device (eg, smartphone screen) [4].

Advances in smart home technology and the Internet of Things (IoT) have the potential to take behavioral monitoring even further [13]. Smart devices collect data objectively; they are unobtrusive and require zero effort from study participants [14]. This technology has the additional advantage that it can assess behavior in individuals with physical or mental impairments who may be unable to interact with smartphones or wearable devices [15]. These advances offer a previously unprecedented opportunity to monitor sleep behaviors in a real-world setting using methods that can reduce participant bias [16]. Previous studies have successfully used smart thermostats to monitor indoor behavior including sleep [17].

The objective of this study is to evaluate the potential use of smart home thermostats to help us understand population-wide sleep patterns, as well as the time the population spends indoors during the year. To assess the impact of different seasonal patterns (eg, days of the week, weekdays vs weekends, and seasons of the year) on sleep health as well as indoor activity, we developed the IoT-based population-level indicators for sleep duration, sleep time, wake-up time, and time spent at home. These indicators are compared across multiple seasonal patterns as weekdays, seasons of the years, and a combined cross-analysis between weekdays and seasons of the year. Ultimately, this study will provide evidence of the potential use of large-scale IoT data to help public health researchers understand the sleep patterns of their population by using the nonobtrusive data collection methods, which will lead to the future use of these data to understand the effects of large-scale global health events (eg, pandemics and climate change).

In this study, our team explored the use of the Donate Your Data (DYD) data from the ecobee smart home thermostat. The data are composed of the anonymized indoor activity of households captured every 5 minutes through the embedded motion sensors [18]. Approximately 98% of participating households in the DYD program are in North America. Taking the specification of ecobee motion sensors into account, for a family size of up to 4, the distribution of floor area has been identified. On the basis of previous exploratory work, we identified the optimum number of sensors based on the average floor area and minimum distance coverage of the sensors. Our household selection criteria included a minimum of 300 days of data available on the data set, the presence of at least six motion sensors in the home, and a maximum of 4 residents. The data management and analysis have been performed on Microsoft Azure Databricks and the scikit learn library [19] in Python (version 3.6).

The original DYD data were reported every 5 minutes. To avoid unnecessary uncertainty, the data were aggregated in 30-minute intervals [18,20,21]. The sum of activation of all sensors, in every 30-minute interval, corresponds to the activity level for that period.

The activation of 1 sensor for 5 minutes corresponds to a score of 1. In a 30-minute interval, the active interval was defined as a score ≥4 (eg, 1 sensor active for 4×5-minute interval [20 minutes], 4 sensors each active for 5 minutes, or any combination of the aforementioned parameters). Intervals with activation sums below this threshold were considered noise. The data were compounded into a binary vector representing a daily record with 48 time slots [18].

This activation pattern was identified to ensure that, while avoiding unnecessary noise, two types of behavior can be picked up by the system: (1) individuals staying in the same room for extended periods, hence activating one sensor sequentially, or (2) individuals moving around the house, hence activating multiple sensors in a shorter time frame.

To develop the different sleep indicators discussed above, we divided each day into two parts, namely, (1) midnight until noon and (2) 8 PM to midnight, and disregarded the time interval from noon to 8 PM. In every 2 consecutive days (eg, days 0 and 1), a sleep cycle was defined as the second part of day 0 combined with the first part of day 1 (Figure 1).

To identify the sleep indicators of each household, the following steps were performed:

Ultimately, the weighted average of each indicator demonstrates the result of sleep parameters for each household at the selected time scale, where the weighted average is defined by the following:

In addition to sleep indicators, we also explored the amount of time spent at home, where the daily cycle records are used to identify the different patterns in each cluster. From each pattern, the amount of time spent at home is defined by the sum of activation of sensors at each interval. The weighted average of each cluster demonstrates the ultimate result.

To explore some of the seasonal effects on sleep and indoor activity patterns, we stratified the data based on different time frames: weekdays and seasons of the year. Next, we compared the different indicators in each of the different time frames (ie, sleep and time spent at home). For each stratified group, the indicators were calculated and the statistical significance between groups was evaluated using analysis of variance (ANOVA) [23]. The statistically significant differences of two indicators (ie, sleep and time at home) were further explored using Tukey post hoc tests [24].

For each of the stratified groups, we present complete descriptive statistics: sample size, mean (SD), SE, and 99% CI of the mean. We assumed that the subsets are independent and distributed normally and the variances are homogeneous [25]. The homogeneity of variance was evaluated using the Levene test [25].

Data were stratified by seasons of the year, dividing the annual record of each household into 4 seasons [26]. The start and end dates of each season are as follows: winter (January 1 to March 21), spring (March 22 to June 21), summer (June 22 to September 21), and fall (September 22 to December 21).

The distribution of all the subsets as well as the homogeneity of variances was checked using the Levene test [23,27]. The 1-way ANOVA and Tukey post hoc tests were used to evaluate the statistically significant relationship between sleep time, wake-up time, sleep duration, and time spent at home with respect to different seasons.

A similar approach was used for stratifying the data on different weekdays. The annual records of each household were stratified into 7 subsets, representing data from each weekday. The distribution of all the subsets, as well as the homogeneity of variances, was investigated using the Levene test. The 1-way ANOVA and Tukey post hoc tests were used to evaluate the statistically significant relationship among sleep time, wake-up time, sleep duration, and time spent at home when comparing different weekdays.

In the combined analysis exploring seasons of the year and weekdays, the annual record of each household was divided into two independent variables: each of 4 seasons and each of 7 days of the week.

The same approach has been replicated for the analysis of different sleep parameters and time spent at home. The 2-way ANOVA [23] and Tukey post hoc test were used to compare the differences of each parameter for different seasons and weekdays simultaneously.

To examine sleep patterns across large populations, we explored the DYD ecobee smart home thermostat data set. The DYD program is hosted by ecobee and provides researchers with access to anonymized data from 110,000 households. After we assessed these households for eligibility, a total of 481 households met the inclusion criteria and were included in this study: they had at least 300 days of data available in the DYD data set, at least six sensors, and a maximum of 4 residents. Of the 481 households included in the study, 390 (81.1%) were in the United States, 63 (13.1%) were in Canada, and 28 (5.8%) had undeclared locations. The largest proportion (40/390, 10%) of households in the United States were in the state of California. In Canada, most households were from the province of Ontario (40/63, 65%).

To examine the effect of different seasons on sleep parameters, household data were stratified into 4 seasons. For each of the four indicators (sleep time, wake-up time, sleep duration, and time spent at home), the statistical distributions are presented in Table 1.

The sleep duration and time spent at home, is presented as the total number of minutes. While, the sleep time and wake-up time are presented using the standard hh:mm format.

Knowing that the null hypothesis of the Levene test is that the groups we are comparing all have equal population variances, the results would confirm the homogeneity of variances in each of the stratified groups is (Table 2).

We can declare all groups homogeneous with a significance threshold of (P<.01). A 1-way ANOVA test was used to compare the seasonal differences for different sleep parameters and time spent at home (Table 3).

Assuming that all other variables are constant, there is a statistically significant difference in time spent at home in different seasons of the year. However, season alone has no statistically significant impact on sleep time, wake-up time, and sleep duration.

The Tukey post hoc test was performed to identify the significant pairs. The difference in means, CIs, and adjusted P values per pair are presented in Table 4.

The results indicate a significant difference in the time spent at home between winter and summer (P<.001; odds ratio [OR] 1.6, 95% CI 1.3-2.3).

The results of the Tukey post hoc test demonstrate that the time spent at home among these households during the winter is statistically significant from that during the summer. On average, individuals in these households spend, in the summer, 1 extra hour outside when compared with that in the winter. These results demonstrate the potential of this IoT data set to inform public health practice by providing insights on population-level behaviors in different conditions.

The ability to identify sleep indicators and time spent at home on different weekdays, using an IoT data set, is further evidence of the potential use of these data to understand and monitor the behaviors of a population [28]. To examine the influence of different weekdays on the 4 indicators, household data were divided between 7 weekdays.

For each indicator, the descriptive statistics for each weekday as well as the homogeneity of variances within different weekdays are presented in Tables 5 and 6, respectively.

We found no significant difference of variances among the 7 weekdays for wake-up time, sleep time, and time spent at home. However, the sleep duration is not fulfilling the variance homogeneity assumption, and the results need to be generalized with precaution. As explained in Figure 1, a typical sleep cycle is spread across 2 days, beginning and ending on different dates. Therefore, the sleep time and sleep duration occur on one day and wake-up time on the next day (Figure 1), which is likely one of the reasons for the nonhomogeneous distribution of variances. The sleep time, wake-up time, sleep duration, and time spent at home on weekdays are illustrated in Figure 2.

The average sleep time for the entire sample was 10:40 PM (Figure 2A). Most households had a sleep time earlier than 10:40 PM on Mondays, Tuesdays, Wednesdays, Thursdays, and Sundays. However, 50.1% (232/463) and 52.8% (245/464) of the households had sleep time of equal to or later than 10:40 PM on Fridays and Saturdays, respectively.

The average wake-up time of all the households in the entire sample was 6:20 AM (Figure 2B). Most households have a wake-up time earlier than 6:20 AM during weekdays, and the average wake-up time on weekends was greater than 6:20 AM (Figure 2B).

The average sleep duration for the entire sample was 8 hours (Figure 2C). The average sleep duration of the households was 7½ hours on Mondays, Tuesdays, Wednesdays, Thursdays, and Sundays. The sleep duration was longer than 8 hours on Fridays and Saturdays.

The average time spent at home for the entire sample was 9 hours (Figure 2D). The time spent at home on weekends was >9 hours.

The 1-way ANOVA showed statistically significant differences in sleep indicators and time spent at home for the stratified weekdays, assuming that all other variables were constant (Table 7).

The Tukey post hoc test comparing sleep time, sleep duration, wake-up time, and time spent at home for different weekday pairs are presented in Table 8. The results of the overall comparison and the significant pairs are presented in Multimedia Appendix 1 and Table 8, respectively. Owing to a large number of comparisons, we only present the statistically significant results in Table 8. Out of all the 21 possible pair combinations for each indicator, we had 8 (38%) statistically significant pairs when comparing sleep time, 10 (47%) when comparing wake-up time, 10 (47%) when comparing sleep duration, and 6 (29%) when comparing time spent at home.

The Tukey post hoc test provides evidence that the sleep time on Fridays and Saturdays was statistically different from that on Mondays, Wednesdays, and Thursdays (OR 1.8, 95% CI 1.5-3; P<.001). The most significant difference in sleep time was between Mondays and Saturdays, with an average of 40 minutes earlier on Mondays than on Saturdays (Table 8).

There was a statistically significant difference in wake-up time on Saturdays and Sundays compared with that on the remaining weekdays (OR 5.6, 95% CI 4.3-6.3; P<.001). The maximum wake-up time difference was between Tuesdays and Sundays, with households waking an average of 76 minutes later on Sundays (Table 8).

The sleep duration on Fridays and Saturdays was statistically significant from the other days of the week (OR 1.8, 95% CI 1.4-2; P<.001), with households sleeping longer on Fridays and Saturdays than on the other days of the week. The highest sleep duration difference was between Tuesdays and Saturdays, with an extra 36 minutes of sleep on Saturdays (Table 8).

There was a statistically significant difference in the time spent at home on Sundays with respect to other days of the week (OR 2.06, 95% CI 1.64-2.5; P<.001). The highest difference in time spent at home is between Thursdays and Sundays, with an average of 96 minutes more time spent at home on Sundays (Table 8).

These results indicate that the data collected by IoT smart home sensors can provide evidence of expected differences between sleep time, wake-up time, sleep duration, and time spent at home. These results provide further evidence for using these data to monitor population-level changes caused by global events.

To further investigate the ability of the smart thermostat IoT data to differentiate patterns in the data and days with unique behavioral patterns, we investigated the combined effect of weekdays and seasons of the year. This proposed analysis will focus on comparing the 4 indicators on different seasons of the year but blocking out analysis by weekday. We first divided the data into 7 weekdays and then, within each weekday, into 4 seasons. Descriptive statistics of the season-weekday groups, as well as the validated results of homogeneity of variances among all subsets for wake-up time, sleep duration, and time spent at home have been provided in Multimedia Appendix 1.

We found no significant difference of variance among subsets, except for the Saturday sleep time, which requires caution when generalizing the results.

The 2-way ANOVA [29] demonstrates the statistically significant differences in sleep indicators and time spent at home with respect to the variation of seasons and weekdays (Table 9).

The greatest impact of seasonality on weekday-specific variations was seen with respect to time spent at home. Figure 3 illustrates the differences through box plots for Thursdays (Figure 3A), Fridays (Figure 3B), Saturdays (Figure 3C), and Sundays (Figure 3D) with respect to the different seasons. Through ANOVA, followed by the Tukey post hoc test, there was a statistically significant difference in time spent at home on Thursdays in the summer in contrast with other seasons. The average time households spend at home on Thursdays is 8½ hours. The time spent at home during the summer was significantly less than that in all other seasons (Figure 3A–D). The time spent at home in the summer on Fridays (Figure 3B), Saturdays (Figure 3C), and Sundays (Figure 3D) is significantly less than in the winter. The time spent at home on Saturdays in the summer is statistically different from that in winter and fall. Households spend less time at home on Saturdays during the summer than during the fall (Figure 3C).

The results demonstrate that the sleep indicators and the time spent at home are significantly associated with variation of seasons of the year and weekdays. To identify the season’s impact on each weekday and compare the different parameters of sleep and time spent at home, we used the Tukey post hoc test to compare the variation of seasons with respect to the specific weekday. The overall comparison and the significant pair results are presented in the Multimedia Appendix 1 and Table 10, respectively.

Different seasons did not have a statistically significant effect on sleep time and sleep duration. There was a statistically significant difference between the wake-up time on Fridays between summer and winter, with an average of 27 (SD 5) minutes (Table 10).

This study is the first of its kind, leveraging smart home thermostat data to monitor sleep indicators and time spent at home on different temporal scales (weekdays, seasonal, and seasonal weekdays) at the population level. This approach not only uses nonadhesive and zero-effort technology to collect the data but can also improve the decision-making of public health officials through large-scale, near–real-time IoT data. Our results demonstrate the variations in sleep indicators and time spent at home for different days of the week and different seasons of the year, which provides evidence of the discriminability of the data and will potentially lead to wider-scale use of IoT data in public health, which has the potential to monitor the effects of global events such as climate change and pandemics.

Some of the limitations of this study include the absence of household socioeconomic and demographic data. The DYD data set is collected and compiled by ecobee, which unfortunately prevents us from collecting important additional data from our participants. As a result, no stratified analysis is possible for age, sex, race, and health indicators. In addition, the self-reported number of occupants and the size of the house may be inaccurate. As this study uses data at the household level, an individual-level analysis is not possible. Similarly, the data provided by ecobee only include information from indoor movement using sensors, limiting the analysis to indoor behaviors. Integration of these data with other IoT devices such as smartwatches, cell phones, and fitness trackers would provide more comprehensive insights into population-level behaviors.