Longitudinal associations of rapid eye movement (REM) sleep and slow‐wave sleep with plasma Alzheimer's disease biomarkers

Abstract

INTRODUCTION

Data on how sleep architecture relates to progression of plasma Alzheimer's disease biomarkers remains limited.

METHODS

Sixty‐four cognitively unimpaired older adults (ages 55–90) underwent plasma collection twice 2.6 years apart and polysomnography proximate to the first plasma collection. Plasma amyloid beta (Aβ) 42 and 40 were measured via mass spectrometry.

RESULTS

Frontal slow‐wave activity (SWA) was inversely associated with plasma Aβ42/40 at cross‐section in bivariate analysis and in a covariate‐adjusted model. In comparison, low rapid eye movement (REM) sleep percentage was associated with a large longitudinal decline in annualized blood‐based Aβ42/40. In exploratory analyses, the association between low REM sleep percentage and faster decline in Aβ42/40 ratio was stronger in the amyloid‐positive group.

DISCUSSION

In cognitively unimpaired older adults, non‐REM SWA correlates with cross‐sectional amyloid load, whereas REM sleep may predict accelerated amyloid accumulation in preclinical stage, suggesting both may serve as early biomarkers and modifiable interventional targets.

Highlights

• Higher frontal slow‐wave activity (SWA) was associated with a lower plasma amyloid beta (Aβ)42/40 ratio at baseline, indicating a cross‐sectional relationship between sleep physiology and Alzheimer's disease (AD) pathology in cognitively unimpaired older adults.

• Lower rapid eye movement (REM) sleep percentage predicted a faster annualized decline in plasma Aβ42/40 ratio, independent of age, sex, apolipoprotein E ε4 status, obstructive sleep apnea, and antidepressant use.

• The association between reduced REM sleep and a faster annualized decline in plasma Aβ42/40 ratio was markedly stronger in amyloid‐positive individuals, suggesting a stage‐dependent effect in preclinical AD.

• REM sleep parameters may have potential prognostic relevance for longitudinal amyloid dynamics. These findings warrant further investigation in larger longitudinal studies.

1. INTRODUCTION

Alzheimer's disease (AD) is the most common cause of dementia in adults aged ≥ 65. As the population ages, the total estimated economic burden of AD and related dementias is expected to grow from $344 billion in 2020 to > $3 trillion in 2060. ¹ AD pathology, including senile plaques and neurofibrillary tangles, is known to accumulate in the brain decades before cognitive symptoms emerge. ² To detect preclinical AD pathology, biomarkers have been developed mainly using cerebrospinal fluid (CSF) analysis or positron emission tomography (PET) imaging. Until recently, measuring AD pathology‐specific proteins in plasma has been challenging due to their low concentration, despite the accessibility and safety of plasma sampling. ³ These limitations have been overcome by advances in laboratory techniques, such as mass spectrometry, which now enable reliable detection of low‐abundance proteins in plasma. ⁴ Among newly developed blood‐based AD biomarkers, the amyloid beta (Aβ) 42/40 ratio has been validated across multiple international cohorts as a reliable marker for detecting cerebral amyloid pathology, achieving diagnostic accuracy with area under the curve (AUC) 0.84 to 0.89, which approaches established CSF and PET imaging biomarkers. ⁴ , ⁵ , ⁶ , ⁷ Systematic evaluation of analytical platforms has established that immunoprecipitation mass spectrometry methods achieve superior discriminative accuracies (AUC 82%–97%) compared to immunoassay‐based methods (AUC 62%–79%) for detecting brain amyloid pathology. ⁸

Over the last decade, increasing evidence supports the concept of a bidirectional relationship between sleep and AD. ⁹ Disruptions in sleep structure and circadian rhythms have long been recognized as common symptoms of AD. ¹⁰ , ¹¹ However, a growing body of literature has recently demonstrated that sleep disturbances begin even before the onset of cognitive symptoms ¹² , ¹³ and increase the risk of AD and dementia ¹⁴ , ¹⁵ , ¹⁶ . Efforts to disentangle this relationship in preclinical AD have focused on understanding the effect of sleep on AD‐specific biomarkers in the cognitively unimpaired population. Poor subjective and objective sleep quality, decreased slow‐wave activity (SWA), and low rapid eye movement (REM) theta power were associated with high Aβ and tau burden in cross‐sectional analyses in the preclinical stage of AD. ¹² , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ Longitudinally, low SWA < 1 Hz and reduced sleep efficiency predicted a higher speed of Aβ deposition by PET imaging in cognitively unimpaired older adults. ²⁵ These findings might indicate the clinical value of changes in macro‐ and microstructure of sleep as an early biomarker as well as a potential interventional target for AD.

Despite growing interest in sleep as a modifiable risk factor for AD, prospective evidence linking sleep architecture to longitudinal AD biomarker trajectories remains scarce. Existing longitudinal studies have relied predominantly on CSF and PET imaging, constrained by invasiveness, high costs, limited scalability, and ionizing radiation exposure, and have focused on a restricted set of sleep parameters, particularly SWA, or sleep apnea cohorts. ²⁵ , ²⁶ , ²⁷ , ²⁸ , ²⁹ Whether these associations extend to blood‐based biomarkers, or whether a broader constellation of macro‐ and microstructural sleep measures independently predicts longitudinal amyloid accumulation in cognitively unimpaired older adults has not been established.

The present study addresses these gaps by examining associations between macro‐ and microstructural sleep architecture and longitudinal changes in blood‐based AD biomarkers. First, the relationship between baseline sleep metrics, including SWA and macrostructure, and the plasma Aβ42/40 ratio was assessed cross‐sectionally. Subsequently, their longitudinal associations were explored to identify predictors of the longitudinal decline in the plasma Aβ42/40 ratio, given that lower plasma Aβ concentrations, as in CSF, reflect increased brain amyloid burden.

2. METHODS

2.1. Subjects

Participants were drawn from ongoing National Institutes of Health–supported longitudinal studies of normal aging and AD risk conducted at New York University (NYU) Grossman School of Medicine between 2009 and 2024. From this larger cohort, we identified 64 community‐dwelling, cognitively unimpaired adults who met study‐specific inclusion criteria and had completed at least one in‐laboratory polysomnogram and two longitudinal blood draws available for plasma biomarker analysis. Fourteen subjects in the current study were also evaluated in prior work on relationships between CSF biomarkers and sleep physiology. ²¹ Inclusion criteria include subjects aged between 55 and 90 years old, English speaking, with a minimum of 12 years of education, Mini‐Mental State Examination (MMSE) scores of ≥ 25, and those who were not depressed (as defined by Geriatric Depression Scale < 6). Individuals with any clinically diagnosed or documented sleep disorders, including obstructive sleep apnea (OSA), were excluded. However, participants were not excluded if OSA was identified incidentally on polysomnography (PSG). Additional exclusion criteria included history of significant medical conditions that could affect brain structure or function, such as stroke, uncontrolled diabetes, traumatic brain injury, lung diseases, drug abuse, or magnetic resonance imaging (MRI) evidence of intracranial mass or infarcts. All participants completed in‐laboratory nocturnal PSG, blood sampling, the Uniform Data Set (UDS) of the National Alzheimer's Coordinating Center (NACC), and MMSE at the baseline visit. ²⁶ Blood sampling was repeated at a follow‐up visit, a median of 2.6 years (interquartile range [IQR] 2.0–4.3 years) after the first blood draw. Written informed consent was obtained from all participants. NYU Grossman School of Medicine and Icahn School of Medicine at Mount Sinai institutional review boards approved the inclusion of human participants for this study.

2.2. In‐lab nocturnal PSG

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using traditional (e.g., PubMed) sources and meeting abstracts and presentations. While the longitudinal effect of sleep architecture on Alzheimer's disease (AD) is not yet as widely studied using the blood‐based biomarkers, there have been a few publications using other modalities. These relevant citations were reviewed and appropriately cited.

2. Interpretation: While frontal slow‐wave activity (SWA) was inversely associated with plasma amyloid beta (Aβ) 42/40 ratio at cross‐section, low percent rapid eye movement (REM) sleep appeared to be a stronger predictor of longitudinal decline in plasma Aβ42/40 ratio, with increased strength of this association among amyloid‐positive subjects. These results suggest that both non‐REM SWA and REM sleep can be an early biomarker of AD pathology and may represent a future interventional target to prevent further pathologic progression in the preclinical stage of AD.

3. Future directions: Future studies will be required to explore other plasma AD biomarkers as well as traditional imaging and cerebrospinal fluid biomarkers, which would provide a more comprehensive understanding of the relationship between sleep and AD pathology. Additionally, longitudinal evaluation of cognitive outcomes from the neuropsychology battery as a function of change in plasma amyloid levels could be explored in the future.

The sleep evaluation was completed using in‐laboratory PSG, which occurred a median of 1.1 years (IQR 0.1–2.7 years) after the initial blood draw. Attended in‐laboratory PSG was acquired at the Mount Sinai Integrative Sleep Center using Compumedics E‐series following standard American Academy of Sleep Medicine criteria. ³⁰ Lights‐off was scheduled to approximate each participant's self‐reported habitual bedtime to match natural home sleep timing. Signal acquisition included electroencephalography (EEG) with a minimum of six EEG leads (F3, F4, C3, C4, O1, O2); left and right electro‐oculography, sampled at 256 Hz and referenced to a common patient and ground electrode; bipolar chin electromyography (EMG); bilateral tibialis anterior bipolar EMG; respiration by a nasal cannula/pressure transducer and oral thermistor; effort by rib/abdomen impedance plethysmography; single‐channel electrocardiography; and oxygen saturation by pulse oximetry. PSGs were scored in 30‐second epochs according to the standard criteria for sleep and EEG arousals. Total sleep time and percent time spent in wake, REM sleep, non‐REM stage 1 (NREM 1), non‐REM stage 2 (NREM 2), and non‐REM stage 3 (NREM 3) were determined. Respiratory events were scored from the airflow signal using American Academy of Sleep Medicine criteria, ³¹ and stage‐specific (total, REM, and NREM) apnea indices were tallied. Apneas were defined as the absence of airflow for ≥ 10 seconds. Hypopnea 4% was defined as a reduction in the amplitude of breathing by 30% or more for ≥ 10 seconds with ≥ 4% decline in blood oxygen saturation, irrespective of the presence of an arousal. The apnea–hypopnea index with 4% oxygen desaturation (AHI4%) was defined as the sum of all apneas and hypopneas with ≥ 4% desaturation divided by the total sleep time in hours. All PSGs were scored by either of two registered sleep technicians, and all PSGs were reviewed by a single expert (I.A.) for respiratory event scoring that was consistent. All PSG scorers were blind to participants’ plasma amyloid data.

2.3. EEG signal processing

EEG data from the PSG‐monitored night were imported into MATLAB and epoched into 30‐second bins. Epochs containing artifacts were rejected, and the remaining epochs were filtered between 0.4 and 50 Hz. A fast Fourier transform was applied to the filtered EEG signal at 30‐second intervals with 50% overlap using a Hamming window. ³² , ³³ Absolute SWA in the EEG frontal leads (e.g., F3 and F4) was defined as the spectral power between 0.5 and 4.0 Hz during NREM sleep. To account for non‐neural anatomical factors, such as skull thickness, scalp‐to‐cortex distance, and CSF volume, that vary between individuals and globally amplify or attenuate the EEG signal, we calculated relative SWA. Relative SWA expresses delta band power as a proportion of total broadband power (0.5–50 Hz), effectively canceling out individual scaling factors and yielding a dimensionless ratio reflecting the proportion of the EEG signal dedicated to SWA band power. Relative SWA is defined as a dimensionless ratio (ranging from 0 to 100%) representing the proportion of total EEG spectral power concentrated within the delta band (0.5–4 Hz). While universally standardized “normal” clinical cutoffs do not exist, relative SWA is ideally interpreted longitudinally within subject to track potential neurodegenerative decline, or cross‐sectionally by comparing against age‐ and sex‐matched reference cohorts while accounting for clinical context.

2.4. Plasma biomarker quantification

Plasma samples were drawn at the baseline and follow‐up visits. Fasting whole blood was collected in the morning in anticoagulant EDTA‐treated tubes and centrifuged for 15 minutes at 2000 x g and 4°C to separate plasma from cellular components. The plasma from each participant's collection tubes was pooled, mixed, and aliquoted into polypropylene storage tubes. Plasma samples were stored at –80°C within 30 to 60 minutes and underwent at most one freeze–thaw cycle before analysis. Plasma Aβ 42 and 40 were measured via mass spectrometry (Araclon Biotech). A population‐specific Youden index based on subjects from NYU with both amyloid PET imaging and plasma amyloid measurement was calculated for the data‐driven clinically relevant and optimal cutoff for plasma Aβ42/40 ratio at ≤ 0.0992, indicating amyloid positivity. ³⁴ Amyloid PET positivity was determined according to a global standardized uptake value ratio (SUVR) > 1.73 in 132 participants from the NYU Alzheimer's Disease Research Center (ADRC) who had completed PET scans with Pittsburgh compound B (PiB) tracer accounting for partial volume correction. Using this amyloid PET positivity definition, a plasma Aβ42/40 ratio cutoff of 0.0992 was established. This threshold, corresponding to the maximum Youden index, yielded a sensitivity of 93.3% and a specificity of 76.1% for amyloid PET positivity after adjusting for age, sex, race, and body mass index.

2.5. Statistical analysis

For the plasma biomarker, annualized changes in plasma Aβ42/40 ratio were calculated as the absolute difference between the plasma Aβ42/40 ratio at the follow‐up visit and the plasma Aβ42/40 ratio at the initial visit, divided by the follow‐up year interval. The Shapiro–Wilk test was performed for normality evaluation. We evaluated associations between sleep stage (percentages of total sleep time) and annualized changes in the plasma biomarker divided by year interval (Δ Aβ42/40_Post‐Pre / year interval) using bivariate Spearman correlations as exploratory analyses and generalized linear models as confirmatory analyses with annualized changes in the plasma Aβ42/40 ratio divided by year interval (ΔAβ42/40_Post‐Pre / year interval) as the dependent variable, after adjusting for age, sex, apolipoprotein E (APOE) ε4 status, presence of OSA (AHI4% ≥ 5/hour), and the use of selective serotonin reuptake inhibitor and/or serotonin‐norepinephrine reuptake inhibitors (SSRI/SNRIs). We used a generalized linear model using the Gaussian family with an identity link function. Model fit was assessed using Q–Q plots and residuals versus fitted plots, which indicated normality of residuals and homoscedasticity. To maximize the sample size, APOE ε4 status was treated as a three‐level categorical variable (non‐carrier, carrier, or missing). Participants with missing genotype data (n = 10) were assigned to a “missing” indicator level within the generalized linear models. Due to technical artifacts or signal loss in the right frontal EEG channels, three participants were excluded from analyses pertaining to SWA, resulting in a final sample size of 61 individuals (including 13 amyloid‐positive subjects) for the data presented in figures demonstrating SWA. Otherwise, there was not missing data. Given the high co‐linearity between sleep variables, each statistical model includes only one sleep variable. Exploratory interaction and subgroup analyses based on the baseline amyloid positivity were conducted for the association between sleep variables and plasma Aβ42/40 ratio, to examine the moderating effect of amyloid positivity. Sensitivity analyses were performed by comparing the full dataset against models excluding an outlier for ΔAβ42/40_Post‐Pre / year interval. Pooled estimates, 95% confidence intervals (CIs), and p values are reported for the whole group and amyloid‐status subgroups. Statistical analyses were conducted using R version 4.5.2 (R Core Team, 2025).

3. RESULTS

3.1. Demographics and sleep variables

Participants were 66.1 ± 7.3 years old, 70.3% female, 60.9% White, and had 16.4 ± 2.1 years of education (Table 1). The mean MMSE score was 29.3 ± 0.8. APOE ε4 carriers comprised 32.8% of the sample, and 50.0% had incidentally identified OSA (AHI4% ≥ 5). At baseline, 21.9% of participants met the threshold for amyloid positivity, which was indicated by a plasma Aβ42/40 ratio below the Youden cutoff of 0.0992.

TABLE 1.

Demographics and sleep variables in the cognitively unimpaired older adults.

Demographics (N = 64)
Age (mean ± SD, years)	66.1 ± 7.3
Female (n, %)	45 (70.3%)
White (n, %)	39 (60.9%)
BMI (mean ± SD, kg/m2)	29.28 ± 7.97
Education (mean ± SD, years)	16.44 ± 2.08
MMSE score (mean ± SD)	29.28 ± 0.80
APOE ε4 carrier (n, %)	21 (32.8%)
AHI4% ≥5 (n, %)	32 (50.0%)
AHI4% (median [IQR])	5.04 [0.94–12.5]
Use of SSRI/SNRI (n, %)	2 (3.1%)
Amyloid positivity (n, %)	14 (21.9%)
Sleep variables
TST (median [IQR], minutes)	371.0 [317.4–429.0]
WASO (median [IQR], minutes)	62.2 [43.4–104.2]
SE (median [IQR], %)	84.9 [72.5–90.1]
SL (median [IQR], minutes)	5.9 [2.7–15.1]
REM sleep latency (median [IQR], minutes)	84.3 [60.0–139.1]
N1% (median [IQR], %)	16.6 [11.2–22.6]
N2% (median [IQR], %)	40.7 [36.0–47.7]
N3% (median [IQR], %)	20.1 [11.4–28.1]
REM% (median [IQR], %)	20.8 [16.3–24.2]
F3_SWA (median [IQR])	72.6 [69.9–75.3]
F4_SWA (median [IQR])	71.3 [67.9–74.0]
Plasma Aβ42/40 ratio
Baseline Aβ42/40 ratio (median [IQR])	0.176 [0.13–0.21]
Follow‐up Aβ42/40 ratio (median [IQR])	0.181[0.05–0.20]
Annualized ΔAβ42/40 ratio (median [IQR])	−2.07 × 10−3[−8.48 × 10−3−7.92 × 10−3]
Year interval (median [IQR])	2.57 [1.95–4.33]

Abbreviations: Aβ, amyloid beta; AHI4%, apnea–hypopnea index with 4% oxygen desaturation; APOE, apolipoprotein E; BMI, body mass index; F3_SWA, relative slow‐wave activity on electroencephalography F3 lead; F4_SWA, relative slow‐wave activity on EEG electroencephalography lead; IQR, interquartile range; MMSE, Mini‐Mental State Examination; N1%, percentage of non–rapid eye movement stage 1 sleep; N2%, percentage of non–rapid eye movement stage 2 sleep; N3%, percentage of non–rapid eye movement stage 3 sleep; REM%, percentage of rapid eye movement sleep; REM, rapid eye movement; SD, standard deviation; SE, sleep efficiency; SL, sleep latency; SNRI, serotonin‐norepinephrine reuptake inhibitor; SSRI, selective serotonin reuptake inhibitor; TST, total sleep time; WASO, wakefulness after sleep onset.

Median total sleep time was 371.0 [317.4–429.0] minutes. Median sleep efficiency was 84.9% [72.53%–90.09%]. Median sleep latency was 5.9 [2.7–15.1] minutes. Median REM sleep latency was 84.3 [60.0–139.1] minutes. The percentages of non‐REM stage 1 sleep (N1%), stage 2 sleep (N2%), stage 3 sleep (N3%), and REM sleep (REM%) were 16.6% [11.2–22.6], 40.7% [36.0–47.7], 20.1% [11.4–28.1], and 20.8% [16.3–24.2], respectively. The relative SWAs at EEG F3 (F3_SWA) and F4 (F4_SWA) were 72.6 [69.9–75.3] and 71.3 [67.9–74.0], respectively.

3.2. Relationship between sleep variables and baseline plasma Aβ42/40 ratio

Baseline plasma Aβ42/40 ratio was not normally distributed by the Shapiro–Wilk test (p = 5.66 × 10⁻⁷). Skewness and kurtosis were –0.85 and –0.48, respectively. High F4 SWA was significantly correlated with low plasma Aβ42/40 ratio (ρ = −0.29, p = 0.025; Table 2 and Figure 1A), which remained significantly associated after adjusting for age, sex, APOE ε4 status, OSA status, and SSRI/SNRI use (β = −0.0034, 95% CI [–0.0065 to –0.00026], p = 0.039). At cross‐section, REM% was marginally inversely correlated with plasma Aβ42/40 ratio in bivariate analysis (ρ = −0.24, p = 0.05; Figure 1B), but not in the model after adjusting for covariates (β = −0.0023, 95% CI [–0.0052 to 0.00066], p = 0.14).

TABLE 2.

Spearman correlation analyses and generalized linear models between baseline plasma Aβ42/40 ratio and sleep variables.

Spearman correlations—baseline plasma Aβ42/40 ratio
Sleep variables	ρ	p
TST	−0.09	0.49
WASO	0.15	0.25
SE	−0.16	0.20
SL	0.13	0.29
REM sleep latency	0.20	0.12
N1%	0.13	0.32
N2%	0.17	0.19
N3%	−0.10	0.43
REM%	−0.24	0.05
F3_SWA	−0.11	0.40
F4_SWA	−0.29	0.025 *

Generalized linear models—baseline plasma Aβ42/40 ratio
Sleep variables	T	StdE	β	95% confidence interval		p
				Lower	Upper
REM%	−1.52	0.0015	−0.0023	−0.0052	0.00066	0.14
F4_SWA	−2.13	0.0016	−0.0034	−0.0065	−0.00026	0.039 *

Note: Covariates: age, sex, ApoE4 status, AHI 4% ≥ 5, and SSRI/SNRI use.

Abbreviations: Aβ, amyloid beta; AHI4%, apnea–hypopnea index with 4% oxygen desaturation; APOE, apolipoprotein E; F3_SWA, relative slow‐wave activity on electroencephalography F3 lead; F4_SWA, relative slow‐wave activity on electroencephalography F4 lead; N1%, percentage of non–rapid eye movement stage 1 sleep; N2%, percentage of non–rapid eye movement stage 2 sleep; N3%, percentage of non–rapid eye movement stage 3 sleep; ρ, Spearman correlation coefficient rho; REM%, percentage of rapid eye movement sleep; REM, rapid eye movement; SE, sleep efficiency; StdE, standard error; SL, sleep latency; T, t statistic; TST, total sleep time; WASO, wakefulness after sleep onset; β, estimate.

^* p < 0.05.

FIGURE 1.

Scatter plots and correlations of baseline plasma Aβ42/40 ratio with frontal slow‐wave activity (A) and REM sleep percentage (B). Regression line with 95% confidence interval. F4_SWA, relative slow‐wave activity on electroencephalogram F4 lead; Aβ, amyloid beta; ρ, Spearman correlation coefficient rho; REM, rapid eye movement; REM%, percentage of rapid eye movement sleep.

3.3. Relationship between sleep variables and annualized ΔAβ42/40 ratio

The annualized ΔAβ42/40 ratio was not normally distributed by the Shapiro–Wilk test (p = 2.20 × 10⁻⁹; Figure 2). Skewness and kurtosis were –2.40 and 8.73, respectively. F4 SWA was not significantly correlated with annualized longitudinal change in plasma Aβ42/40 ratio (ρ = 0.08, p = 0.53; Table 3 and Figure 3A), whereas low REM% was significantly correlated with an annualized decline of Aβ42/40 ratio (ρ = 0.31, p = 0.013; Figure 3B). Low REM% remained significantly associated with a large, annualized decline of plasma Aβ42/40 ratio after adjusting for age, sex, APOE ε4 status, OSA status, and SSRI/SNRI use in a general linear model (β = 0.0024, 95% CI [0.00015 to 0.0047], p = 0.043; Table 3).

FIGURE 2.

Histogram of annualized ΔAβ42/40 ratio. Aβ, amyloid beta.

TABLE 3.

Spearman correlation analyses and generalized linear models between annualized ΔAβ42/40 ratio and sleep variables.

Spearman correlations—annualized ΔAβ42/40 ratio
Sleep variables	ρ	p
TST	−0.14	0.26
WASO	0.02	0.88
SE	−0.07	0.58
SL	0.01	0.95
REM sleep latency	−0.04	0.76
N1%	0.00	0.99
N2%	−0.09	0.51
N3%	−0.13	0.30
REM%	0.31	0.013 *
F3_SWA	−0.04	0.77
F4_SWA	0.08	0.53

Generalized linear models—annualized ΔAβ42/40 ratio
Sleep variables	T	SE	β	95% confidence interval		p
				Lower	Upper
REM%	2.08	0.0012	0.0024	0.00015	0.0047	0.043 *

Note: Covariates: age, sex, ApoE4 status, AHI 4% ≥ 5, and SSRI/SNRI use.

Abbreviations: Aβ, amyloid beta; AHI4%, apnea–hypopnea index with 4% oxygen desaturation; APOE, apolipoprotein E; F3_SWA, relative slow‐wave activity on electroencephalography F3 lead; F4_SWA, relative slow‐wave activity on electroencephalography F4 lead; N1%, percentage of non–rapid eye movement stage 1 sleep; N2%, percentage of non–rapid eye movement stage 2 sleep; N3%, percentage of non–rapid eye movement stage 3 sleep; ρ, Spearman correlation coefficient rho; REM%, percentage of rapid eye movement sleep; REM, rapid eye movement; SE, sleep efficiency; StdE, standard error; SL, sleep latency; T, t statistic; TST, total sleep time; WASO, wakefulness after sleep onset; β, estimate.

^* p < 0.05.

FIGURE 3.

Scatter plots and correlations of annualized ΔAβ42/40 ratio with frontal slow‐wave activity (A) and REM sleep percentage (B). Regression line with 95% confidence interval. F4_SWA, relative slow‐wave activity on electroencephalogram F4 lead; Aβ, amyloid beta; ρ, Spearman correlation coefficient rho; REM, rapid eye movement; REM%, percentage of rapid eye movement sleep.

In observing the distribution of annualized ΔAβ42/40 ratio values (Figure 2), an outlier with a Z score of ‐4.96 was detected. After excluding the outlier in a sensitivity analysis, the correlation between low REM% and large annualized decline of plasma Aβ42/40 ratio remained significant (ρ = 0.29, p = 0.02), but the strength of its association after adjusting for the covariates was marginally reduced (β = 0.0018, 95% CI [−0.000029 to 0.0037], p = 0.06).

3.4. Exploratory interaction and subgroup analyses based on amyloid positivity

Although the number of amyloid‐positive subjects was small (n = 14), as an exploratory analysis, we evaluated relationships between REM% and longitudinal annualized ΔAβ42/40 ratio values as a function of amyloid positivity. We observed a particularly strong relationship between REM% and longitudinal annualized ΔAβ42/40 ratio values among the amyloid‐positive subjects (ρ = 0.71, p = 0.0045; Figure 4 [blue]), whereas this relationship was weaker and not statistically significant among amyloid‐negative subjects (ρ = 0.18, p = 0.21; Figure 4 [gold]). Low REM% remained significantly associated with a large, annualized decline of plasma Aβ42/40 ratio after correcting for age, sex, APOE ε4 status, OSA status, and SSRI/SNRI use covariates in the amyloid‐positive group (β = 0.0062, 95% CI [0.0024 to 0.01], p = 0.033).

FIGURE 4.

Scatter plots between REM sleep percentage and annualized ΔAβ42/40 ratio stratified by the baseline amyloid positivity. Aβ, amyloid beta; REM%, percentage of rapid eye movement sleep.

4. DISCUSSION

In this study of cognitively unimpaired older adults, we first showed that at a cross‐sectional level, high SWA was associated with low plasma Aβ42/40 ratio at baseline in cognitively unimpaired older adults. This is in line with our prior observations that SWA was inversely correlated with Aβ42 levels in CSF. ¹⁹ , ²¹ Although elevated fluid Aβ42/40 ratio may be viewed as “protective” given that a low CSF and plasma Aβ42/40 ratio is thought to reflect higher intracortical amyloid burden, the interpretation at cross‐section may depend on where one is in the disease course. In mouse models of amyloidosis, CSF Aβ42 rises transiently during the earliest phase of cerebral Aβ deposition before declining as plaque burden accumulates. ³⁵ A parallel non‐linear pattern has been described in humans: our group previously reported a positive association between CSF Aβ42 and phosphorylated tau 181 in cognitively unimpaired adults aged 45 to 70 years, which inverted to a negative association in those aged 71 to 90 years. ³⁶ Furthermore, elevated baseline Aβ42/40 may itself predict subsequent decline: in the Dominantly Inherited Alzheimer Network (DIAN), mutation carriers showed a transient rise in CSF Aβ42 an estimated 10 to 25 years before expected symptom onset, followed by a sustained decrease, ³⁷ and an analogous pattern of higher baseline values preceding steeper longitudinal decline was observed in cognitively unimpaired older Asian adults. ³⁸ Together, these findings indicate that a change in plasma Aβ42/40 over follow‐up likely reflects a participant's position along this non‐linear curve rather than a simple monotonic decline.

While Aβ burden in the medial prefrontal cortex correlated inversely with impairment in SWA generation, ¹⁸ this association was only seen with power in the 0.6 to 1 Hz frequency range, whereas a positive association was observed with power in the 1 to 4 Hz range. Variables such as age, amyloid measurement tool, and precise EEG frequency band are factors that can influence the relationship between amyloid and sleep physiology.

We next show that low baseline REM% associates with a faster longitudinal annual decline in plasma Aβ42/40. To the best of our knowledge, this is the first study to show the longitudinal effect of objective sleep on plasma biomarkers in the preclinical stage of AD. Previous studies have mainly used CSF and/or imaging biomarkers for AD pathology to explore the link between sleep and AD. ²⁵ , ²⁶ , ²⁷ , ³⁹ While longitudinal changes in plasma amyloid can be challenging to interpret in amyloid‐positive individuals due to a potential floor effect after reaching a low steady‐state level, monitoring individuals within the subthreshold range remains potentially meaningful. Emerging evidence suggests that significant intraindividual declines in plasma Aβ42/40 occur well before the threshold for PET positivity is reached, making subthreshold longitudinal tracking a valuable tool for capturing early biomarker dynamics. ⁴⁰ , ⁴¹ Among amyloid‐positive individuals, longitudinal plasma Aβ42/40 trajectories did not show a unanimous unidirectional decline (Figure 5), and the increases observed in a subset of individuals are unlikely to reflect a true biological reversal of amyloid accumulation. Larger samples and longer follow‐up with serial cognitive testing will be needed to disentangle the directionality of the relationship and more accurately establish biological meaning to the observed changes in plasma Aβ42/40 among cognitively normal older adults.

FIGURE 5.

Scatter plots between annualized ΔAβ42/40 ratio and baseline Aβ42/40 ratio. Aβ, amyloid beta.

The association of low baseline REM% sleep with a faster decrease in plasma Aβ42/40 ratio from this study, suggestive of an increase in intracerebral amyloid burden, has both similarities to and differences from previous literature. Winer et al. demonstrated that slow oscillation (SO) power (< 1 Hz) in the frontal lobe longitudinally predicted the speed of Aβ accumulation with PiB PET imaging over 3.7 years in clinically normal older adults. ²⁵ In contrast, Carvalho et al. reported that SO density and slope were positively associated with longitudinal amyloid accumulation on PiB PET over 2.7 years in cognitively unimpaired OSA patients, ²⁸ suggesting that apnea‐related sleep fragmentation may shift the relationship between SO physiology and amyloid dynamics. Because OSA status was controlled for as a covariate in our models, our data suggest that a distinct cross‐sectional relationship exists in which elevated baseline SWA tracks with a lower plasma Aβ42/40 ratio even in participants where the sleep disruption from OSA may be influencing SWA in a non‐linear fashion.

Compared to studies focused on slow‐wave sleep physiology, this study showed that reduced baseline REM sleep predicts a faster decline in plasma Aβ42/40 ratio, and its association was markedly stronger in the amyloid‐positive group. This finding is consistent with data showing that low REM sleep theta power was associated with greater diffuse amyloid burden on florbetapir PET at cross‐section in cognitively unimpaired older adults. ²⁴ Jin et al. also reported that long REM latency was associated with high Aβ burden. ⁴² . Additionally, a population‐based prospective study by Pase et al. showed that low REM sleep percentage and long REM latency were associated with a higher risk of incidence of dementia. ⁴³ In the Osteoporotic Fractures in Men Study, men with reduced REM sleep showed faster annual decline in modified MMSE scores over 3.4 years. ⁴⁴ In support of our findings, these studies highlight the clinical importance of REM sleep as an indicator of the progression of neurodegenerative disease.

Notably, the association between reduced REM sleep and faster plasma Aβ42/40 decline was substantially stronger in amyloid‐positive than amyloid‐negative participants, although this was an exploratory analysis and the amyloid‐positive subgroup was small. This observation is potentially corroborated by a previous study demonstrating that low REM sleep percentage was correlated with low MMSE scores and low CSF Aβ42 in patients with AD. ⁴⁵ This relationship might become prominent once AD pathology results in the dysfunction and ultimate degeneration of cholinergic projections in the basal forebrain, which is critical to generate and maintain the REM stage. ⁴⁶ , ⁴⁷ However, the link between REM sleep physiology and AD pathology in the preclinical stage of AD needs to be elucidated further.

Although Winer et al. did not find an association between REM sleep duration and longitudinal amyloid accumulation in cognitively unimpaired older adults, ²⁵ several methodological differences may account for this discrepancy: REM was quantified as absolute duration rather than percentage of total sleep time; amyloid was assessed by PET rather than serial fluid biomarkers; the cohort had a high proportion of APOE ε4 carriers (47%) without adjustment; and antidepressant use, which is common in older adults and suppresses REM, was not documented.

The lack of a cross‐sectional association between REM% and baseline plasma Aβ42/40 ratio, contrasted with its ability to predict longitudinal accumulation, may reflect the non‐linear dynamics of plasma biomarkers in the subthreshold range. Recent evidence indicates that plasma Aβ42/40 undergoes a rapid, step‐like drop shortly before cortical amyloid plaques are detected on PET imaging, followed by a plateau. ⁴¹ Therefore, baseline REM% may capture early‐stage progressive changes that precede absolute threshold positivity, highlighting the value of longitudinal monitoring over baseline cross‐sectional assessments. Although our data support a model in which REM sleep is predicting change in amyloid load, our data are not set up to test a plausible inverse causal model—for example, accumulating amyloid triggers longitudinal changes to REM sleep.

In spite of the strengths of this study, there are several limitations. The Alzheimer's Association and National Institute on Aging support the AT(N) framework for evaluating preclinical AD pathology along axes reflecting amyloid, tau, and neurodegeneration, and only the plasma Aβ ratio was used as a primary outcome in this study. Second, we did not evaluate cognitive outcomes from the neuropsychology battery as a function of change in plasma amyloid levels. Third, PSG occurred a median of 1.1 years after the initial blood draw in this study, although sleep is being conceptualized as a stable trait marker. Future studies are necessary to measure PSG concurrently with the initial blood draw or perform sensitivity analyses constraining participants with shorter PSG‐to‐blood intervals. Additionally, relatively small sample size and short follow‐up duration (median 2.6 years) may have reduced the statistical power to detect a longitudinal association between SWA and increasing amyloid burden over time. Last, considering the well‐established, robust REM‐suppressing effects of serotonergic medications (SSRIs/SNRIs), the use of these medications was included as a covariate in our models to control for potential confounding when predicting longitudinal AD biomarker changes from baseline REM sleep. However, given that only two participants in our cohort were taking these medications, these adjusted results should be interpreted with caution.

In conclusion, this study demonstrated that frontal SWA serves as a significant correlate of plasma Aβ42/40 ratio at cross‐section, and low REM% sleep is associated with faster longitudinal decline in blood‐based plasma Aβ42/40 ratio, representative of a faster increase in cortical amyloid accumulation, in cognitively unimpaired older adults. Low REM% sleep appeared to be a stronger predictor of longitudinal decline in plasma Aβ42/40 ratio among amyloid‐positive subjects. These results suggest that both NREM SWA and REM sleep may serve as early biomarkers of AD pathology and represent potential intervention targets in the preclinical stage.

CONFLICT OF INTEREST STATEMENT

A.W.V. has served as a consultant for Axsome Therapeutics and Jazz Pharmaceuticals in the last 3 years. All other authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

Written informed consent was obtained from all participants. NYU Grossman School of Medicine and Icahn School of Medicine at Mount Sinai institutional review boards approved the inclusion of human participants for this study.

Supporting information

Supporting Information: alz71910‐sup‐0001‐ICJME.pdf