The FHS was established in 1948, when 5,209 residents of Framingham, Massachusetts, USA, aged 28 to 62 years, were enrolled in a prospective epidemiologic cohort study. In 1971, an additional 5,124 subjects (offspring of the original cohort subjects and the offspring spouses) were enrolled in the Framingham Offspring Study. The Offspring cohort had cycles of periodic health examination on average every 4 years. The 9^th Offspring health exam began in April 2011 and concluded in April of 2014.

In the early 1990s, the need to establish a new group of participants reflecting the increasing diversity of the community was recognized. In 1994, the Omni Cohort 1 of the FHS was initiated. The original Omni cohort consisted of 507 men and women of African-American, Hispanic, Asian, Indian, Pacific Islander and Native American origins, who at the time of enrollment were residents of Framingham and the surrounding towns. The Omni Cohort 1 continues to be examined and followed at the same health examination cycle as the Offspring Cohort. Health exam 4 for Omni Cohort 1 coincided with the dates of the Offspring 9^th health exam.

The current study samples are comprised of Offspring/Omni-1 participants in the Wearables Sleep Study who also participated in the health exam 9/4 respectively. A subset of these subjects also underwent both structural brain magnetic resonance imaging (MRI) and DTI between 2009 and 2014 as part of a separate larger study of cognition and brain imaging [24]. Participants with prevalent stroke or dementia at the MRI evaluation were excluded from the Wearables Sleep Study. Demographic characteristics are summarized in Table 1. The Institutional Review Boards at all participating institutions approved this study, and subjects gave written informed consent.

The entire sleep physiology assessment was conducted by mail. Participants who agreed where mailed two sleep quality and sleep oxygenation devices, asked to use both for 2 consecutive nights, and then mailed the devices back. The main measures used were the oxygen saturation time under 90%, and the oxygen 3% desaturation index (ODI, number of 3% desaturations/hour of the recording).

The SleepImage^® System (https://sleepimage.com) was used to collect measures used to compute sleep quality. It is a Health Insurance Portability and Accountability Act (HIPAA)-compatible Cloud Computing system currently hosted in the Amazon Cloud, using small single-sensor recording devices. The analysis is also FDA approved as a stand-alone Software as a Medical Device (SaMD) and can be run on any signal recordings that include an ECG or similar information-content signal including plethysmography (PLETH) and is comparable to a full or limited polysomnography [25]. The SleepImage system allows a review of raw data, to the resolution of individual ECG complexes, snoring bursts, and activity-driven sensor displacements.

This small wearable recorder (https://sleepimage.com) measures continuous ECG, sampled at 600 Hz, expressed in millivolts, 12-bit quantization, with one adhesive pad under the device and a thin wire across the chest to a second pad. Activity and body position are measured by internal accelerometers and gyroscopes, and snoring is detected by induced vibration. The data is uploaded to the SleepImage website, and automatic analysis generating CPC variables, the sleep spectrogram graphs total sleep time (actigraphic), actigraphic wake (after sleep onset), and transient awakenings (by movement).

The device itself has the following dimensions: height: 79.6 mm, width 48.7 mm, thickness 11.7 mm, weight 20 g, and a storage capacity of 500 Mb. The accelerometer within the device has the following specifications: 12-bit quantization, units are in gravitational acceleration “G” units. The Z channel is sampled at 300 Hz, the Y at 37.5 Hz, and the X channel at 37.5 Hz. Collectively, the X, Y, and Z channels are referred to as the “gravity channels”, and used to compute actigraphy, body position and snore vibrations. The M1 starts recording when the ECG is sensed and stops when the ECG is no longer sensed. Participants were instructed to apply the device when ready to sleep. The movement sensor detects non-respiratory trunk movements to estimate the sleep period (start and stop of sleep) and actigraphic wake episodes when very low frequency coupling (VLFC) periods are associated with actigraphic motion.

CPC analysis of the ECG signal was performed as previously described [26–28]. The ECG-derived sleep spectrogram [27] uses a time series analysis of the ECG, to generate a measure of “CPC”. Both autonomic drive (through heart rate variability) and respiration (ECG R-wave amplitude fluctuations with individual breaths) are extracted from the continuous ECG and used for analysis. The product of the coherence and cross-spectral power is used to calculate the ratio of coherent cross power in the low frequency (0.01–0.1 Hz) band to that in the high frequency (0.1–0.4 Hz) band. The logarithm of the high to low frequency CPC ratio (log [high frequency coupling (HFC)/low frequency coupling (LFC)]) is then computed to yield a continuously varying measure of CPC. The output is thus a moving average of overlapping CPC windows. The graph of CPC at relevant frequencies (ordinate) vs. time (abscissa) provides a sleep spectrogram (Figure 1) The LFC band can be further fractionated, providing a metric, elevated-LFC (e-LFC), which is a measure of sleep fragmentation. This band can be further fractionated into broadband (e-LFC_BB, generic sleep fragmentation, sleep apnea) or narrowband (e-LFC_NB), which reflects strong respiratory chemoreflex modulatory effects on sleep state. The relatively slow human heart rate dictates the sampling windows. The analytic window is 8.5 min, and the computation moves over the signal in consecutive 2.1 min windows.

Stable non-rapid eye movement (NREM) sleep is characterized by HFC, increased absolute and relative delta power [28], a consolidated NREM sleep < 1 Hz slow oscillation, temporally stable breathing, stable arousal thresholds, normal arterial oxygen (O₂) and carbon dioxide (CO₂) concentrations, and blood pressure (BP) dipping [29]. Unstable NREM is characterized by opposite features. Ineffective (fragmented) REM sleep takes on LFC coupling signatures, while wake or effective REM sleep shows VLFC [27]. An example of the ECG-derived sleep spectrogram is below.

The concept of NREM sleep instability at the electroencephalographic (EEG) level is well established, the cyclic alternating pattern (CAP) [30]. Periods of stable sleep are designated non-CAP. The CPC output correlates poorly with conventional sleep stages, but far more closely with CAP/non-CAP, such that periods of HFC are usually NREM stage N3, but more importantly, the majority of NREM stage N2 in health [27]. LFC aligns with CAP, parts of N2 and light NREM N1 sleep. Disease states expand the proportion of LFC relative to HFC, which are mutually exclusive states.

The dates of MRI evaluation were from 10/13/2011 to 2/28/2018. The scan-sleep measurement time difference mean was −1.42 years (= scan performed about 1.4 years before the sleep study). The longest interval was 3.01 years (= brain scan performed 3 years after the sleep study).

Participants were evaluated with a 1.5-Tesla Siemens Avanto scanner for DTI. The DTI sequence parameters: repetition time (TR) = 3,600 ms, echo time (TE) = 94 ms, 25 contiguous slices total, with no gaps between sections, field-of-view (FOV) = 25 cm, acquisition matrix = 128 × 128, slice thickness = 5 mm. Diffusion weighted images were generated using 30 gradient directions, repeated four times, with total gradient diffusion sensitivity of b = 1,000 s/mm², and four unweighted images with b = 0 s/mm².

DTI images were first preprocessed using FMRIB software library (FSL) software tools [31], including correction for eddy current-induced distortions and participant head movements. Individual fractional anisotropy (FA) maps were coregistered to the FSL FA DTI template using linear and nonlinear transformations. Resulting transformation parameters were applied to the individual FA and mean diffusivity (MD) maps. FSL FA template was thresholded at 0.3 to provide a mask of WM region. For each individual, overall measures of mean FA and MD were computed by superimposing the WM mask onto the respective individual coregistered DTI-derived maps and averaging values within these WM voxels. Peak width of skeletonized MD (PSMD) is a DTI-derived measure based on skeletonization and histogram analysis and is calculated as the difference between the 95^th and 5^th percentile of MD values within the masked MD skeleton [32].

Both nights of sleep were averaged for analysis. Gender differences in sleep data were analyzed with the t-test for normally distributed data and the Mann-Whiney U test when variables were skewed. Unadjusted associations between sleep variables (sleep quality, fragmentation, and oxygenation) and DTI measures (FA, MD, PSMD) examined using regression analysis. A multivariate analyses of covariance (MANCOVA) was conducted to test the associations between sleep (HFC, LFC, e-LFC, e-LFC_BB, e-LFC_NB, actigraphic wake) and DTI (FA, MD, PSMD) variables, while adjusting for various demographic and clinical covariates (age, sex, body mass index, diabetes, hypertension, hyperlipidemia, and the time interval between polysomnography and DTI). The statistical models were, Model 1: unadjusted; Model 2: adjusted for age and gender; Model 3: adjusted for diabetes, hypertension, hyperlipidemia, body mass index.

Secondary analyses were performed using interaction terms to examine the presence of moderating effect of gender. All analyses were conducted using SAS 9.4 (SAS Institute Inc., Cary, North Carolina).

The demographics of the participants in this analysis are summarized in Table 1. Women were slightly over-represented, and the mean age was 67.92 ± 6.22 years at the time of sleep recording.

Sleep measures are summarized in Table 2, including gender differences (Table 3)

Stable sleep, as identified by the % total sleep time of HFC, was significantly associated with increased FA and reduced MD. Unstable sleep, identified by the % total sleep time of low frequency coupling, was associated only with reduced MD (Table 4). Moderating effects of gender differences were notable with only women showing the impact of protective and injurious effects.

General sleep fragmentation (e-LFC) was associated in this cross-sectional analysis with increased MD, consistent with increased deterioration of WM integrity. e-LFC_NB, a signal biomarker of high loop gain (respiratory chemoreflex activation) was significantly associated with reduced FA and increased MD (Table 5). Moderating effects of gender were noted, with only women showing significant effects.

Both measures of oxygen desaturation had significant cross-sectional associations with DTI measures (Table 6). The oxygen desaturation index was associated with reduced FA and increased MD. The time in minutes under 90% oxygen saturation was associated with reduced FA and increased MD. The peak width of skeletonized MD was positively associated with time under 90% oxygen saturation. Women showed a greater impact on the oxygen desaturation index.