Research participants were recruited as a convenience sample. Volunteers were sought from within Google LLC via emails sent to relevant full-time employee listservs. There was no incentive offered for participation and no punitive consequences for lack of participation or completion of the study. Interested volunteers were asked to complete a screening questionnaire to determine eligibility for study enrollment, based on the inclusion/exclusion criteria (Table 1). Because the sleep schedule recommendation and coaching messages were based on sleep-related behaviors, individuals with sleep schedule irregularity—defined as a >1 h difference in bedtime and/or wake time, at least once a week, on a regular basis—were preferentially recruited.

Individuals participated in the study for a total of 14 weeks (Figure 1). After providing informed consent, participants set up the Fullpower Sleeptracker^® under-mattress sensor (“the sensor”) to start passively monitoring their sleep patterns (Ding et al., 2022). This data was abstracted to a secure spreadsheet for analysis on a daily basis. Throughout the study, participants did not have access to their data or the features of the Fullpower Sleeptracker^® app. Participants also received twice daily SMS messages to survey their perceived sleep quality—Overall, how was the quality of your sleep last night? “1” (very poor) to “5” (very good)—and daytime alertness levels—How alert were you today? “A” (not at all alert) to “E” (most alert); the quality SMS survey was sent within an hour of usual wake time, in order to get an accurate reflection on the previous night's sleep without the bias of sleep inertia, and the alertness SMS survey was sent about 6 h before usual bedtime, in order to follow the usual midday, circadian dip in energy levels. This data was also added to the secure spreadsheet for analysis and team decision-making.

At the beginning of the study, and again at the end of the 14 weeks, participants completed a brief survey about their age, self-identified sex, general sleep habits (e.g., usual bedtime, average sleep duration, etc.), perceived sleep issues, and sleep goals. The exit survey focused on feedback regarding the coaching experience (e.g., frequency of SMS, relevance of content, etc.). Neither survey was used in the Intervention phase of the study, but the data were gathered for qualitative feedback on the coaching experience and, in the case of age and sex, for model adjustment in quantitative analyses.

The first 2 weeks of the study included only passive sleep pattern monitoring and the twice daily SMS text messages.

Following 2 weeks of sleep pattern and quality/alertness data collection, participants entered the 4-week intervention phase. In addition to the ongoing sleep monitoring and twice-daily SMS surveys, individuals received a sleep schedule recommendation (recommended bedtime and wake time) and thrice-weekly sleep coaching messages. The sleep schedule recommendations were automatically calculated from the collected data from the previous 14 days. After the research team (blinded to participant identity) reviewed the sleep data and approved the sleep schedule recommendation, the recommended bedtime and wake time were shared with the participant via SMS and entered into a Google Calendar, which were associated only to the participant's de-identified study ID. The research team also reviewed the relevant coaching messages that were identified by a series of rules that automatically analyzed the collected sleep data (see “Coaching message selection” subsection). Once the coaching messages were confirmed to be appropriate by consensus opinion, they were sent via SMS, using the same platform as the twice-daily survey questions. In order to standardize and simplify the study process, the research team met on Monday, Wednesday, and Friday in order to review the participant data and schedule/coaching message recommendations.

Following the 4-week Intervention phase, participants were again passively monitored with the under-mattress sensor and asked to respond to twice-daily SMS survey questions about their sleep quality and daytime alertness levels.

Each week, the preceding 14 days of data were used to generate a sleep schedule recommendation. In order to focus on regularity of the sleep schedule, bedtime and wake time recommendations were generated from usual bedtimes and wake times, accounting for the relationship between the usual sleep duration and nightly sleep opportunity. Prior to providing recommendations, all schedule recommendations were reviewed by the research team to ensure adequate sleep opportunity, based on the participant's historical data. If the recommended schedule was deemed inappropriately short or long, the team would override the recommendation, instead suggesting a 7- or 9-h sleep period, respectively, in accordance with generally accepted healthy sleep guidelines for this demographic (Hirshkowitz et al., 2015; Watson et al., 2015).

As part of the study protocol, sleep coaching messages were generated that helped identify the area of potential sleep improvement, provide education on basic principles of sleep-wake physiology, and suggest an action that could be undertaken to make the intended improvement. This content was developed based on the extant body of peer-reviewed scientific literature and covered various concepts (e.g., the homeostatic sleep drive, the circadian system, etc.). Sleep improvement areas were identified, based on a number of evidence-based rules derived from peer-reviewed publications, professional society guidelines (e.g., American Academy of Sleep Medicine, National Sleep Foundation, etc.), and general health and wellness principles that are discussed in public outreach messages from reputable organizations (e.g., MayoClinic.org). The sleep improvement areas focused on multiple aspects of healthy sleep, generally structured around the RU_SATED paradigm (Buysse, 2014): sleep schedule regularity, sleep quality, reports of daytime alertness, timing of the main sleep period, patterns of wakefulness preceding or interrupting the main sleep period (sleep efficiency, based on the metrics provided by the under-mattress sensor), and sleep duration. Coaching messages followed three main templates: an insight template which describes what aspect of the participant's sleep needs improving; an education template which provides context on why a sleep issue is important; and a suggestion template that provides a tip on how they can address a given issue (see examples of each in Table 2). Participants were given the messages in a set, where the order of the set was insight, education, and suggestion. Messages were written and saved in a database and included in the IRB-reviewed participant-facing materials. During the intervention phase, message content for a participant on a given day was selected from the database based on automated calculation of metrics for each rule and a system that variably weighted each area, based on the number of activated rules and degree of deviation from what might be considered “optimal” for that rule. The research team assessed the appropriateness of the content selections at their tri-weekly meetings and scheduled relevant SMS based on discussion and consensus. Each set was selected to be the most relevant aspect of sleep for them that particular week, with the next week being potentially a different set.

Given that the main recommendation in the study was for a bedtime/wake-time routine, overall protocol adherence was estimated based on frequency of days in which participants had a wake time (WT) within 30 minutes of the recommended WT. As there was no WT recommendation during the baseline phase, the median WT from the entire baseline phase was used as a surrogate WT target to estimate “adherence” to a regular routine during that phase. Bedtime was not used because the estimate of bedtime provided by the sensor does not accurately reflect the time of getting into bed, but, in most cases, just reflects a time 5 min prior to estimated sleep onset. Along these lines, given that sleep efficiency was calculated from a denominator that does not reflect true time in bed, the normal range for sleep efficiency was set at 90–98% (the 10 and 90th percentile of sleep efficiencies for the entire cohort). Additionally, because of differences in the way sleep is estimated by the sensor relative to the validation analyses performed by Brindle and colleagues (Brindle et al., 2019), the normal range for sleep duration was set at the recommended 7–9 h for a population with our demographics (Hirshkowitz et al., 2015; Watson et al., 2015). The midpoint of the sleep period (MSP) was calculated as the midpoint from sleep onset to final wake time; the normal range for the timing and standard deviation of the MSP were used from the validation study performed by Brindle and colleagues (Brindle et al., 2019). Because daily sleep quality and alertness questions inverted responses, validated thresholds (Brindle et al., 2019) were similarly inverted to define normal vs. not. In addition to examining the absolute and relative number of individual metrics in/out of the normal range on a daily basis, an aggregated daily sleep-health (RU_SATED) (Brindle et al., 2019) score was calculated for each participant by summing the number of sleep health metrics that were in the “normal” range, with a minimum possible score of 0 and maximum of 6.

Categorical variables are presented as percentages, and continuous variables as mean±standard deviation or median and interquartile range for non-Gaussian variables, as confirmed by the Shapiro-Wilk test for normality. Similarly, when normal distributions of variables were evident, period-to-period comparisons of continuous variables were performed using ANOVA (multi-period comparisons), and post hoc pairwise comparisons were performed with the t-test only in instances where the ANOVA p-value was below the significance threshold; otherwise, the Kruskal-Wallis test and post hoc Wilcoxon rank sum testing were used. χ² or Fisher's exact test (when counts fell below 5 in any category) was used to compare categorical variables between groups. Pearson coefficients were calculated to demonstrate magnitude of correlation between relevant variables. A statistical threshold of α = 0.05 was set, and, when relevant, Bonferroni correction for multiple comparisons was performed for each major analysis.

Analyses were performed in the R programming language (v4.0.4) (R Core Team, 2019). In order to generate aggregated daily sleep health scores, data were imputed using the mice package for R (van Buuren and Groothuis-Oudshoorn, 2011), assuming that missingness was at random. Otherwise, linear mixed models (LMMs) were fit with the lme4 package (Bates et al., 2015), due to their relative robustness vis a vis modest missingness and auto-correlated longitudinal data. The LMMs were used to model general changes of metrics over the course of the study after accounting for various potential sources of confounding: fixed effects for age and sex and random effects (slope and intercept) for study start date and participant (ID). In order to roughly model changes in adherence over time, generalized additive models (GAMs) were used with a penalized cubic regression spline, which had its penalty modified to shrink toward zero at high enough smoothing parameters.

Of the 4,010 potentially eligible individuals who were emailed, 116 completed the screening questionnaire to determine eligibility for study enrollment, based on the inclusion/exclusion criteria (Table 1). Because the sleep schedule recommendation and coaching messages were based on sleep-related behaviors, individuals with sleep schedule irregularity—defined as > 1 h difference in bedtime and/or wake time, at least once a week, on a regular basis—were preferentially recruited. Of those who remained eligible following the screening questionnaire, 37 were consented for participation in the study. Of these, 31 individuals actually participated in the study, with 1 individual (#19) being withdrawn before completion of the study due to their lack of responses to SMS surveys and inability to sleep with the device for > 80% of nights. Four participants' data were excluded from final quantitative analysis due to extrinsic factors that altered their sleep patterns resulting in an inconsistent ability to adhere to the protocol throughout the study period: #4 & #7, due significantly erratic sleep schedules and disruptions potentially due to hardware issues or external factors; #9 and #18, due to extreme changes in sleep duration and schedule due to illness midway through the study. The remaining 26 participants were included in this analysis.

The participants in this study were younger (34.92 ± 10.08 years old) and predominantly male (20M:6F). Despite an effort to recruit volunteers with an irregular sleep schedule, the average standard deviation of the MSP wasn't very high (Tables 3, 4, Baseline column), particularly in relation to the 1:05 min threshold proposed by Brindle et al. (2019). Additionally, in relation to the MIDUS II and MIDUS refresher validation cohorts that Brindle and colleagues used to derive metric cutoffs, this study's participants reported being significantly less alert and had generally later sleep periods and longer sleep duration, with higher efficiency (all p-values less than the Bonferroni-corrected α of 0.008; see Supplementary Table 8). The longer sleep duration and higher efficiency may have been, in part, due to the different sleep tracking methods used in this study.

Adherence to the recommended wake time (WT) schedule was somewhat inconsistent in this cohort (Figure 2). A shrinkage-penalized, cubic spline, generalized additive model (GAM) was used to ascertain any trend in WT adherence over the course of the study. There appeared to be a tendency toward progressive delay in WT for the cohort as a whole, with participants “falling out” of the recommended WT±30 min window by about the 87th day of the study (45 days after the intervention phase). From another vantage point, looking at weekly rates of adherence (i.e., proportion of the week with actual WT within 30 minutes of the recommended WT), ANOVA suggested that adherence rates differed by study week [F_{(13, 338)} 2.05, p = 0.017], but, due to a large number of post hoc comparisons for the 14 weeks of the study, no pairwise significance was observed. Therefore, linear mixed models were used to model the changes in adherence rates over time, accounting for self-correlation on repeated measures, inter-individual differences, and occasional missing data points. In general, after adjusting for potential confounders (age and sex) as well as start date and inter-individual differences (with random effects), there was a significant main effect for the week of study, which suggested an approximately 1.5% reduction in weekly WT adherence rates over the course of the study (Table 5). As expected, weekend schedule adherence was the lowest and mid-week adherence was the highest [F_{(6, 175)} 5.05, p = 8.28^*10⁻⁵; Tue and Wed > Mon, Thu, Fri > Sat > Sun, by post hoc Tukey's honestly significant difference testing] (Figure 3).

Adherence was also explored based on how many consecutive days individuals were able to wake within 30 min of the recommended WT. There was a significant difference in the length of the longest duration of adherence per individual between phases of the study [F_{(2, 50)} 3.64, p = 0.03], with the intervention phase having longer streaks than the baseline phase on post hoc analysis (5.23 ± 3.73 vs. 3.73 ± 1.51 days, respectively; t₍₂₄₎ 2.40, p = 0.024) (Supplementary Figure 1). In order to explore the relationship between duration of adherence with subjective reports of alertness and sleep quality, LMMs that accounted for linear and quadratic relationships between length of each participant's longest continuous adherence (“streak”) revealed no statistically significant main effects of adherence length (Supplementary Figures 2, 3).

Consistent with the study phase comparisons in Tables 3, 4, none of the LMMs separately exploring changes in the 6 RU_SATED metrics demonstrated a main effect for the study day (all p-values >0.05), after adjustment for fixed effects of age and sex, and allowing for random slopes and intercepts based on start date and per participant (Table 6). The models' fixed effects explained, at most, a modest 22% of the variance in the metrics, with the only significant effect demonstrating that older age, expectedly, associated with an earlier sleep period (β = 5-min earlier MSP per year, p < 0.01). Moreover, the aggregate models' magnitudes of change over the intervention phase—from day 14 to day 42—weren't substantial: 16 min, 19 s increase in the standard deviation of the MSP; 0.18 point improvement in sleep quality perception; 0.30 point improvement in daytime alertness perception; 4 min, 18 s earlier timing of MSP; 0.10% reduction in sleep efficiency; and an 18 min reduction in sleep duration.

Because the sleep patterns of the participants were so diverse, it was felt that the individual metrics were possibly not an adequate reflection of changes in sleep health resulting from the intervention. As such, for the first 6 weeks of the study (2 baseline weeks and 4 intervention weeks), we performed an analysis of changes in a daily RU_SATED score that was calculated from imputed data (to ensure a full score could be calculated), in which 1 point was added for each metric in the prespecified ideal range, for a maximum possible score of 6 points. After adjusting for fixed effects of age and sex, and allowing for random slopes and intercepts based on start date and per participant, there was a significant, but modest daily improvement in the aggregate RU_SATED score over the course of the study (β = 0.02 point improvement per day, p = 1.90^*10⁻¹³), which translates to a 0.56-point improvement over the 4 weeks of the intervention phase, in general (Figure 4). Of note, the majority−15/28 (58%)—of the participants demonstrated an improvement in their overall sleep health over the course of the intervention.

Given the high degree of variability in both adherence rates and RU_SATED score changes, an analysis was performed to explore whether adherence to the WT recommendation over the 4-week intervention phase correlated with the change in RU_SATED score over this same time period. As demonstrated in Figure 5, there was a moderate positive correlation between adherence rate and RU_SATED score change (ρ = 0.54, p = 0.004).

Upon completion of the intervention phase, participants were given a survey on their experience. Of the 26 responses to the question “To what degree did you feel the insights provided were useful to you?” 8% selected very useful, 72% selected somewhat useful, and 20% selected not at all useful.