Participants were identified by referral from either local studies that adjudicated MCI diagnoses or a local research recruitment registry that is led by the University of Pittsburgh Clinical and Translational Science Institute (Pitt+Me). The inclusion criteria were (1) being 50 years of age or older; (2) passing the San Diego Brief Assessment of Capacity to Consent [7] with scores ≥14.5; (3) have a score of >27 on the Telephone Interview for Cognitive Status (TICS) [8]; (4) having a TICS score of either ≤34 (high-risk group) or ≥39 (no cognitive disorder or NCD); and (5) have a prior adjudicated diagnosis of MCI (high-risk group) or reporting no concerns regarding cognitive decline (ie, NCD). The exclusion criteria included (1) self-reporting active behavioral health treatments for insomnia or depression; (2) self-reporting of performing the prescribed exercises; and (3) self-reported use of sleep medications every night or nearly every night. After explaining the study purpose and procedures, we obtained verbal consent to screen interested individuals for eligibility. Of 82 potentially eligible participants who completed the verbal eligibility screening, 35 were ineligible, 6 refused, and 41 were enrolled. One participant withdrew after enrollment, resulting in an analytic sample of 40 individuals (19 high-risk (ie, having a low TICS score and prior MCI diagnosis) and 21 low-risk (ie, having a high TICS score and no prior MCI diagnosis), as shown in Figure S1 in Multimedia Appendix 1 (ie, a flow diagram illustrating how we arrived at our analytic sample).

After completing web-based written informed consent forms, participants completed baseline procedures and assessments via Health Insurance Portability and Accountability Act (HIPPA)-compliant video-conferencing. We sent participants the following study devices: an Apple Watch 8, iPhone SE 2nd generation with active data plans, and nonstock 40-Watt charger. After the participants received the devices, we conducted an additional study visit to instruct them on logging in and using the myRW application. We specifically instructed participants that about 20‐30 minutes of charging a day is sufficient, and asked them to wear the watch whenever it was not charging for a week.

All study procedures were approved by the University of Pittsburgh Institutional Review Board (identifier: STUDY22080033). All the participants completed web-based written informed consent forms

The myRW software system, developed by the corresponding author, extracts triaxial accelerometer data that are recorded at 50 hertz on the Apple Watch. After bandpass filtering, we aggregate across axes and into 30-second counts, and transmit data for storage and further computations on the Amazon Web Services platform. To encourage data collection, we depict the amount of data collected per day graphically by proportionally filling and color coding a star (Figure 1). We explicitly told participants that sufficient data were required to “get a star,” that the goal was to get 7 stars in a row, and we also displayed information on how many days in a row participants obtained stars (to encourage data collection “streaks”). Participants received reminders three times daily to check in to the application. Other features, including personal data graphs and sleep/wake metrics, were disabled for this study to prevent the feedback from potentially altering the users’ behavior/typical sleep/wake patterns.

Consistent with prior publications using researcher-focused accelerometers [9-11], we required participants to have at least 3 valid continuous days of accelerometer data to consider the recording adequate for processing. Valid days were defined, consistent with National Sleep Research Resource data processing standards [12], as those with no more than 4 hours of nonwear time or any nonwear time during the main sleep period. Invalid days and remaining data missing within days were censored (not imputed). Although we applied this aforementioned 20 hours per day criterion, to maximize data collection, note that the app was programmed to reward participants with a green star only if they collected at least 22 hours of data per day.

Following previously published technical definitions, we calculated extended cosine-based [13] and nonparametric [14-17] sleep/wake measures using the R package ‘RAR’ and custom code. From the extended-cosine models, we extracted measures of 24-hour robustness (pseudo-F statistic, indicating how well the observed data fits the 24-hour curve); activity onset time (up-mesor, the time which the modeled activity level passes the middle modeled rhythm height prior to the peak); and activity offset time (down-mesor or the time which the modeled activity level passes the middle modeled rhythm height prior to the nadir). From the nonparametric method, we calculated the cross-daily stability (inter-daily stability, measuring the consistency of circadian sleep/wake activity rhythms across days); rhythm strength (relative amplitude, measuring the standardized peak-trough difference of 24-hour activity rhythms); and sleep/wake rhythm fragmentation (intradaily variability or IV, measuring the frequency and extent of transitions in activity levels). Although sleep/wake rhythm fragmentation is typically measured using differences in activity levels every hour, it can also be computed on a range of timescales [16,17]. For all nonparametric metrics, we used all-time series data and did not subsample to hourly activity levels, as described previously [14]. To explore the relevance of fragmentation timescale, we computed IV values using numerator timescales ranging from 5 to 60 minutes, but holding the denominator constant as done previously [16].

Cognitive performance was measured using the Hopkins Verbal Learning Test (HVLT; PAR, Lutz, FL) and the Oral Trail Making Test (O-TMT).

The HVLT is a verbal list learning measure with three trials of 12 words from multiple representative semantic categories, and includes delayed recall and delayed recognition memory components [18,19]. For the delayed recall, participants were asked to repeat as many terms as they could recall from the word listed presented to them thrice, 30 minutes earlier. For delayed memory recognition, participants were asked to recognize those same terms appearing on a broader list containing semantically related terms, ie, “horse” must be differentiated from “dog,” which was not part of the original term set. The HVLT assessed encoding, storage, and retrieval of noncontextual verbal information.

Processing speed was measured as the time taken on Part A of the O-TMT, which requires the subject to verbally count upwards from 1 to 25 as quickly as they are able [20]. Set-shifting, an aspect of cognitive function, was measured as the time taken using Part B of the O-TMT, which requires the subject to alternate between listing numbers and letters in sequence (“1, A, 2, B,” etc). Note that longer times on both parts of the O-TMT indicates worse performance [21].

To compare recording lengths between the MCI and NCD groups, we used an independent sample t test and density plots illustrating recording length distributions by groups. To examine relationships of sleep/wake rhythm measures with cognitive function, we first adjusted for key confounders by taking the residuals from linear regression models (one for each cognitive outcome variable) that had the confounders age, sex, education (college degree vs less than college degree), and accelerometer recording length as predictor variables. We used the residualized values as age, sex, education, and recording length-adjusted cognitive outcome variables in a series of linear regressions (one per sleep/wake predictor variable). Since there were many conceptually similar or highly inter-correlated intradaily variability metrics, for related analyses, we only report Benjamini-Hochberg corrected false discovery rates (which are here referred to as q values instead of P values) [22] to account for the 12 statistical tests relating intradaily variability metrics within each cognitive outcome.

The sample included older adults, with a mean (SD) age of 67.2 (8.4) years; the majority of the participants were females with college degrees (Table 1).

In both the groups, all participants achieved the minimum data requirement for computing 24-hour sleep/wake rhythm assessments. There was, however, a statistically significant difference in the mean (SD) recording lengths of about a half a day when comparing the NCD (7.2 [0.6] days) and MCI groups (6.6 [1.2] days), with the Satterthwaite test assuming unequal group variances (df 25.033; t value −2.15; P=.04). This difference was due to 3 individuals in the MCI group who collected 3‐6 days of data each (Figure 1).

As shown in Table 2, there were small effect size associations of later activity onset time with lower delayed memory performance (β=−.28, 95% CI −0.55 to ‐0.02; t=−2.17, df=38, P=.04) and more stable rhythms with better processing speed and performance (β=−0.27, 95% CI: −0.54 to 0.00; t=−2.00, df=38, P=.05). None of the other sleep/wake measures listed in Table 2 were associated with the cognitive outcomes. Regarding sleep/wake rhythm fragmentation, greater fragmentation levels in the 40‐60-minute timeframe was significantly associated with worse processing speed and performance (Table 3; β point estimate range: 0.36, 0.40; q=.022). When repeating analyses in the subgroup with at least 6 days of data, results were not substantively altered (n=37; see Tables S1 and S2 in Multimedia Appendix 1).

To illustrate the accelerometer data collected in this study and visualize 24-hour sleep/wake fragmentation, Figure 2 shows data from example participants with lower and higher degrees of 24-hour sleep/wake rhythm fragmentation.