NREM sleep as a novel protective cognitive reserve factor in the face of Alzheimer's disease pathology

Sixty-two cognitively normal older adults participated in the study. Thirty-one individuals were β-amyloid positive (Aβ+) and thirty-one were Aβ-, based on an established cutoff (1.065 global PiB distribution volume ratio (DVR) [29, 30], and see Table 1 and Fig. 1A). Data from twenty-six of these participants were included in a previous publication [27]. Participants were recruited from the Berkeley Aging Cohort Study (BACS) which allowed for the selection and enrollment of participants with a full range of PiB DVR values. Exclusion criteria were history of neurologic, psychiatric, or sleep disorders, current use of antidepressant or hypnotic medications, and presence of contraindications for exposure to the high magnetic field of MRI imaging. All participants scored ≥ 25 on the Mini-Mental State Examination (MMSE). Several factors motivated the inclusion of cognitively normal elderly individuals in the current study. First, in the early stages of Alzheimer’s disease (AD) pathology, the diversity and severity of different pathologies in the brain are less extensive and therefore more optimal when studying specific pathologies (i.e., Aβ) in isolation, as motivated by the current hypothesis. Second, the early stages of the AD pathological cascade, such as Aβ pathology [10], allow for cognitive reserve to the greatest extent, indicated by a higher proportion of cognitively normal elderly adults with only Aβ pathology than with Aβ pathology and neurodegeneration [34, 35]. Finally, from a treatment perspective, results in such a population allow insights for early therapeutic intervention.

All participants completed three sessions. In one session, participants underwent PET scanning following ¹¹C-PiB injection to measure Aβ burden. Within 2 years of ¹¹C-PiB PET scanning (median: 0.52 years, and see Table 1), participants completed the sleep study session. Participants spent two nights in the laboratory, where they were given 8-hour sleep opportunities monitored with polysomnography (PSG). This 8-hour interval was in accordance with age-appropriate averages [36], and the total sleep time of the participants was in the normative range [37] (Table 1). The first night served as an adaptation night to prevent first-night effects [38, 39], whereas the second served as the experimental night. Following the experimental night, structural MRI scans were obtained from all participants in the morning to measure gray matter atrophy. Subsequently, participants completed a hippocampal-dependent face-name memory task [40]. Four participants had missing data in the memory task due to not completing the task (N = 2) and insufficient number of certain response types (N = 2) resulting in an inability to compute memory measures (N = 1 missing measure for item and N = 1 missing measure for associative memory). Furthermore, BMI was not assessed for one participant and structural MRI was missing for one participant. All participants abstained from caffeine, alcohol, and daytime naps for 48 hours before and during the sleep study session. Participants kept habitual sleep-wake rhythms for at least 1 week preceding the in-laboratory sleep session and completed the in-laboratory sleep study in accordance with their individual rhythms. In a separate session (on average within a half year from the experimental sleep session, median: 0.44 years), participants also completed multiple questionnaires and cognitive tasks, including the assessment of physical activity in the last year [41] and a composite memory measure [42].

¹¹C-PiB PET imaging, quantifying Aβ burden, was conducted in 3D acquisition mode using an ECAT EXACT HR (N = 20) or BIOGRAPH Truepoint 6 scanner (Siemens Medical Systems, Erlangen, Germany). ¹¹C-PiB was synthesized at the Lawrence Berkeley National Laboratory Biomedical Isotope Facility using a published protocol and described in detail previously [43]. After intravenous injection of ∼15 mCi of ¹¹C-PiB into the antecubital vein, 90 minutes of dynamic acquisition frames were obtained (4 × 15, 8 × 30, 9 × 60, 2 × 180, 10 × 300, and 2 × 600 s). For each ¹¹C-PiB scan, a positron transmission scan or CT scan was obtained for attenuation correction. PET images were reconstructed using an ordered subset expectation maximization algorithm with weighted attenuation. Images were smoothed with a 4 mm Gaussian kernel with scatter correction.

¹¹C-PiB data were realigned, and frames from the first 20 minutes of the acquisition were averaged and coregistered to participants’ corresponding structural MRI. Structural MRI was assessed using a 1.5T Siemens Magnetom Avanto scanner at Lawrence Berkeley National Laboratory (T1-weighted MPRAGE images, TR/TE = 2110/3.58 ms, FA = 15°, 1 × 1 × 1 mm resolution). Distribution volume ratios (DVRs) for ¹¹C-PiB images were generated with Logan graphical analysis on ¹¹C-PiB frames corresponding to 35-90 minutes after injection using a cerebellar gray matter reference region [44, 45]. Global ¹¹C-PiB DVR was calculated as a weighted mean across FreeSurfer-derived native space frontal, temporal, parietal, and cingulate cortical regions as in previous studies [29, 30]. Participants were classified as Aβ+ if their global cortical ¹¹C-PiB DVR was ≥1.065 in accordance with previous studies [29, 30]. Experimenters were blinded to participants’ Aβ status during data collection.

Polysomnography on the nights of the sleep session was recorded using a Grass Technologies Comet XL system (Astro-Med, Inc., West Warwick, RI). This system included a 19-channel EEG placed according to the 10–20 system, electrooculography (EOG) recorded at the left inferior and right superior outer canthi, and electromyography (EMG) on the chin. Reference electrodes were left and right mastoid (A1, A2). Data were digitized at 400 Hz.

Data were scored using a validated automated sleep scoring software [46] and further assured with visual inspection by two trained sleep-scoring professionals (Z.Z. and O.M.) in accordance with standardized criteria [47]. In parallel to sleep scoring, 30-second epochs with major body movements were rejected based on visual inspection. Channels with marked artifact noise were also identified during visual inspection and omitted from subsequent analyses. To reduce artifacts, EEG data were re-referenced from the contralateral mastoids as needed to the average of the mastoids (N = 24) or a unilateral mastoid (N = 5). Furthermore, for any participant with slow drift artifacts due to sweating (N = 27), a semi-automatic, amplitude-based rejection was applied to their EEG data. For each affected participant, a sweat artifact-free, 30-second long epoch from deep sleep (NREM3 stage) was selected based on visual inspection. For each participant and for each EEG channel, the average bandpower between 0.2 and 0.8 Hz in this sweat artifact-free epoch was computed. Further EEG data was rejected on a 5-second basis whenever the bandpower between 0.2 and 0.8 Hz exceeded 1.5 times the individual amplitude of the clean epoch in any of 3 frontal channels (F3, F4, Fz). Frontal channels were specifically selected as the sweat artifact was expressed the strongest in those channels.

Spectral power analysis of artifact-free segments was performed using the validated open-source Python toolbox for sleep analysis (Yet Another Spindle Algorithm, YASA [48]). Data were first downsampled to 100 Hz and filtered between 0.5 and 35 Hz. Next, power spectrum was calculated for NREM and REM sleep stages separately using Welch’s method, median-averaging 4-second intervals with 50% overlap, and employing Hamming windowing. The analyzed frequencies spanned between 0.5 and 35 Hz with a 0.25 Hz bin resolution. Next, bandpower was calculated by integration (area under the curve) for the following frequency ranges: Delta (0.5-4 Hz), Theta (4-8 Hz), Alpha (8-12 Hz), Sigma (12-16 Hz), Beta (16-30 Hz), Gamma (30-35 Hz). Finally, relative power was computed by dividing the bandpowers by the total power observed in the 0.5 - 35 Hz frequency range. Analyses in the current report focused, a priori, on SWA, defined as relative delta spectral power during NREM (NREM2 and NREM3 stages) sleep. Comparing the adaptation night to the experimental night, NREM SWA showed high test-retest reliability (r = 0.76, p < 0.001).

Next-day memory was assessed using a validated, face-name hippocampal-dependent task sensitive to age and sleep effects [40, 49, 50]. This task has been proven to have good reliability in older adults [51‐53]. The task consisted of an encoding phase where 120 face-name pairs were introduced and a recognition test that followed the end of the encoding phase after a 30-minute delay [40].

During the encoding phase, each face-name pair was shown a single time. The task consisted of 4 blocks, each containing 30 trials. Participants were given the opportunity to rest between the blocks. Each block took approximately 5 minutes, with the entire encoding phase taking 20 minutes. For each trial, a fixation was first presented for 0.5 seconds (white background, black Arial font, 24 points, center screen), followed by a face-name pair for 3.5 seconds. To ensure that participants attended to each face-name pair, each pair remained on the screen for an additional 2.75 seconds during which participants had to determine if the name “fit” the face [40]. Participants were instructed to remember each face-name pair for testing. Face stimuli were drawn from a standardized database [54] and were counterbalanced based on gender (male, female) and age group (young: ages 18-49, old ages: 50-90). Names were drawn from the 1990 US census and were counterbalanced for frequency of use in the US population. The presentation order was pseudo-randomized such that no more than four faces of identical gender or age group were presented successively.

The recognition test consisted of 200 trials, with each face/face-name pair tested once. The whole test lasted approximately 45 minutes. Out of the 200 trials, 120 contained faces presented during the encoding, and 80 were novel faces (foils). Foil faces were balanced for age and gender similarly to the encoded material. For each recognition trial, a face was presented on the screen, at the same size and location as presented during encoding. In a self-paced manner, participants first had to answer whether they had seen the face during the prior encoding phase (original) or not (new) by a corresponding keyboard response [40]. Following this, with the face stimulus remaining on the screen, participants were presented with four options to choose from to test associative memory recognition: (1) the original name previously paired with that face (correct ‘hit’ response), (2) a name previously seen before at encoding, but with a different face (incorrect ‘lure’ response), (3) a new name never shown during encoding (incorrect response), or (4) an option “new” rejecting the trial as a foil trial [40]. The four response options appeared horizontally on the screen with the first three always corresponding to names in a randomized order in regards to response type. New (foil) faces were presented with all new names never shown during encoding. As with encoding, faces were presented in a randomized order.

Two distinct d-prime sensitivity memory measures were computed based on the recognition test. Item memory performance was quantified by subtracting the standardized false alarm rate for the faces (the proportion of faces falsely declared to be previously seen out of all the foil trials) from the standardized hit rate for previously encoded faces (the proportion of faces correctly recognized as previously seen out of all the encoded face trials). Associative memory was quantified by subtracting the standardized lure (miss) rate (the proportion of misses, where participants chose the lure names instead of the correct answer out of all previously encoded trials where the participant correctly identified the face as being previously studied) from the standardized hit rate of the face-name pairs (the proportion of faces and names correctly paired out of all previously encoded trials where the participant correctly identified the face as being previously studied).

Physical activity was quantified using a previously validated questionnaire [41]. A list of 15 leisure-time physical activities (e.g., cycling, dancing, swimming) was provided and participants were asked to report whether they engaged in each of the activities (or any other that was not included in the list). Further, for each activity, information was collected on the frequency (how often during the previous 2 weeks and how many months per year) and duration (time spent per session) of engagement. Frequency and duration information were multiplied using an activity-specific intensity code indicating calorie expenditure [55] and summed to represent the intensity of physical activity (total kilocalories of energy expended) during the past year.

Trait-like episodic memory function was measured by a composite score that consisted of short-delay and long-delay (after 20 min) free recall scores of both the California Verbal Learning Test [56] and the Visual Reproduction Test [57]. Individual scores were Z-transformed using mean and SD from the first cognitive session data of a larger sample of BACS participants. The composite score was calculated as the mean of the standardized individual test scores.

Selective atrophy within the medial prefrontal cortex (mPFC) has been shown to influence the relationship between sleep and cognitive performance in elderly adults [58, 59]. Therefore, to measure mPFC atrophy, structural MRI scans were obtained from all participants during the sleep session. Structural MRI was assessed using a Siemens Trio 3T scanner. High-resolution T1-weighted MPRAGE images were acquired for every participant (TR/TE = 1900/2.52 ms, FA = 9°, 1 × 1 × 1 mm resolution).

To measure gray matter volume, optimized voxel-based morphometry (VBM) [60] was performed using SPM12 [61] with the Computational Anatomy Toolbox (CAT12) [62] and the Diffeomorphic Anatomical Registration through Exponentiated Lie algebra (DARTEL) [63] toolboxes. Gray matter and white matter segmentations were used to create a study-specific DARTEL template, which was then used to normalize individual brains into MNI space to improve the registration of older brains to the normalized template [64]. Modulated gray matter maps were then smoothed using an 8 mm Gaussian kernel.

To compute gray matter volume in the mPFC, the Anatomical Automatic Labeling repository [65] within the Wake Forest University PickAtlas toolbox [66] was used to generate an anatomically-based ROI. Mean voxelwise gray matter volume within this ROI was extracted using the Marsbar toolbox [67]. Measures of total intracranial volume for each participant were estimated from the sum of gray matter, white matter, and CSF segmentation, and used to adjust mPFC gray matter volumetric measures to account for differences in head size.

Statistical analyses were conducted in R 4.1.0 [68] using the stats, sjPLot, and interaction packages [69, 70]. To compare the global PiB DVR, demographic, sleep, and cognitive measures between the Aβ+ and Aβ- groups, independent, two-sample t-tests were used. To test the hypothesis prediction that NREM SWA moderates the effect of Aβ pathology status on memory function, a multiple linear regression was employed with Aβ status, NREM relative delta power, NREM relative delta power by Aβ status interaction, and age, sex, Body Mass Index (BMI), gray matter atrophy, the time difference between the PET and sleep sessions, education, and physical activity as regressors predicting memory performance. Specifically, from this multiple regression, the interaction between NREM relative delta power (SWA) and Aβ status indicated evidence of cognitive reserve. A post-hoc statistical power analysis was performed in GPower 3.1.9.7 to estimate the achieved power of this analysis. To explore regional specificity, multiple linear regressions were computed with relative delta power for each EEG channel separately. To control for multiple testing in the topography analyses, False Discovery Rate (FDR) correction was applied. To assess frequency range and sleep stage specificity, similar multiple linear regressions were employed, by replacing NREM relative delta power with other NREM and REM frequency range spectral power measures. Similar to NREM SWA, the interaction between Aβ status and spectral power measures was evaluated. To control for multiple testing in these non-hypothesis-driven regression results, FDR correction was applied. Further, to differentiate trait and state-like memory benefit effects, a similar multiple regression was applied to evaluate the interaction between Aβ status and NREM SWA predicting the composite memory score that was assessed during a separate session from sleep. For all regressions, cognitive performance and sleep data were excluded on a case-by-case basis, when outside of the 1.5 interquartile range from the first and third quartile across participants in the given Aβ group to ensure normal distribution (see Additional file 1: Table S1 for final sample sizes for each regression by Aβ group).