Heterogeneity in spontaneous sleep arousals: positive and negative links with early amyloid-beta and cognition

Recent literature is pointing towards a tight relationship between sleep quality and amyloid-beta (Aβ) accumulation, a hallmark of Alzheimer’s disease (AD). Sleep arousals are considered to induce sleep disruption, and though their heterogeneity has been suggested, their correlates remain to be established. We classied arousals in sleep of 100 healthy older individuals according to their association with muscular tone increase (E+/E-) and sleep stage transition (T+/T-), and show differences in EEG oscillatory compositions across arousal types. We found that T + E- arousals, which interrupt sleep stability, were positively correlated with Aβ burden in brain regions earliest affected by AD neuropathology. By contrast, more prevalent T-E + arousals, upholding sleep continuity, were associated with lower cortical Aβ burden, and better cognition. We provide empirical evidence that spontaneous arousals are diverse and differently associated with brain integrity and cognition. Sleep arousals may offer opportunities to transiently synchronise distant brain areas, akin to sleep spindles.

Sleep arousals, de ned as transient accelerations in sleep electroencephalogram (EEG) rhythms, are usually considered as brain reactions to internal (e.g., apnoea) or external (e.g., auditory stimulus) perturbations 17 . Although key elements of sleep microstructure, they can also shape its macrostructure and lead to a shallower sleep stage 18 . They are most often considered as markers of sleep disruption, thereby a detrimental and harmful sleep feature. Several conceptual de nitions classi ed them almost exclusively in the context of sleep disorders (e.g. sleep disordered breathing -SDB, and to a smaller extent periodic limb movements syndrome -PLMS), or in experimental protocols inducing arousals through external -mainly using auditory -stimulation [19][20][21][22] . These types of studies yielded mixed results. A negative link between arousal prevalence during sleep and cognitive performance was revealed in SDB, particularly in attention, and sometimes the executive and memory domains 23 . By contrast, other investigations did not nd such a relationship, and imputed alterations in cognition in SDB to brain hypoxia (see 23 for a review). In individuals devoid of sleep pathologies, arousals evoked by auditory stimuli were reported to impact subsequent daytime alertness 20 . In addition, sleep fragmentation induced by auditory stimulation is associated with higher Aβ cerebrospinal uid (CSF) content the following day 10 .
Importantly, spontaneous arousals, i.e. not elicited by any identi able internal or external stimuli, also constitute an authentic element of undisturbed sleep in healthy individuals. Their mechanisms, cerebral correlates and functional consequences remain largely unknown 17 with some authors suggesting that there may be physiologic and pathologic arousals. Understanding their respective roles might shed light on the adaptive properties of the sleeping brain and provide insight into pathological mechanisms associated with sleep disturbances 17 .
Here, we assessed whether different types of spontaneous arousals during sleep were differentially associated with Aβ cortical deposition and cognitive performance in a cohort of healthy individuals in late midlife. We were able to tease apart different types of arousals, based on their temporal relationships with increased muscular tone and sleep stage transitions. In line with the hypothesis that arousals perturb sleep, we anticipated that arousals fragmenting sleep structure would be associated with both worse cognitive performance and Aβ deposition in brain areas that are rst affected by AD neuropathology.

EEG oscillations differ across arousal types
We recorded undisturbed sleep at habitual sleep times under EEG in 101 healthy individuals aged 50 to 70 y (59 ± 5y; 68 women), following one week of regular sleep-wake schedule. In order to evaluate the potential heterogeneity of arousals, we split them according to two criteria, which we considered as relevant in research settings as well as clinical practice. Firstly, whether arousals did trigger a sleep stage transition (T+) (when they occurred within 15s of a stage change) or not (T-); and secondly, their salience, re ected by the concomitant increase in EMG tone (E+) or its absence (E-).
Arousals heterogeneity re ects different associations with Aβ burden Aβ burden was quanti ed over the regions previously reported as the earliest cortical aggregation sites 24 in all but one participant. In a second GLMM, we tested whether associations between arousal density and early cortical Aβ burden depend on transition (T+, T-) and EMG (E+, E-) statuses, while regressing out age and sex effects. We observed a main effect of transition (F 1,196   We further computed a GLMM with early cortical Aβ burden as dependent variable to assess whether its association with T-E + and T + E-arousal were truly signi cant and different from one another in a more complex model, regressing out age and sex. Both associations were signi cant with a negative link between Aβ and T-E + arousals (F 1,95 =14.15, p = 0.0003, R² β* =0.13) and a positive association between Aβ and T + E-arousals (F 1,95 =8.16, p = 0.0053, R² β* =0.08) -together with a main effect of age (F 1,95 =13.02, p = 0.0005, R² β* =0.12), and no main effect of sex (F 1 , 95 =2.54, p = 0.11). Critically, a post-hoc contrast showed that the links between the two types of arousals and early cortical Aβ burden were signi cantly different (t 93 = 3.73, p = 0.0003). In addition, T-E + and T + E-arousals are not correlated (suppl. g. S1).
Supplementary analysis showed the same statistical picture in a GLMM including all four arousal types together, with a signi cant post-hoc contrast when considering T-E + and T + E-arousals vs. early cortical Aβ burden (suppl. table S2). Arousals linked with better Aβ status are associated with better cognitive performance We then tested whether cognition, as assessed through an extensive neuropsychological test battery, was differentially associated with the two arousal types showing opposite association with early cortical Aβ burden. Pearson's correlation revealed a signi cant positive correlation between global cognitive performance and T-E + arousal index (r = 0.22, p = 0.026), but not between T + E-arousal index and global cognition (r=-0.15, p = 0.14). In a GLMM, the association between global cognition and T-E + arousal index remained signi cant (p = 0.048, R² β* =0.04), on top of the education effect, but no relation with T + Earousals index (p = 0.25), age, or sex (Table 1; Fig. 3). Table 1 Outputs of GLMMs assessing associations between cognitive performances (global and speci c domains-dependent variables) and arousal types, while adjusting for age, sex and education (independent variables). All F tests had 1 (main effect) and 95 (error) degrees of freedom. Signi cant associations are in bold and are accompanied by their corresponding Semi-partial R² (R²β*). T: arousal associated (T+) or not (T-) with sleep stage transition; E: arousal associated (E+) or not (E-) with an increase in EMG tone. T

Discussion
Brain dynamics which buttress cerebral functions entail stationary and non-stationary interactions between neuronal populations 25 . Sleep stages, which can be seen as enduring and widespread oscillatory modes sculpting brain activity, allow recurrent brief faster oscillatory activity, which sometimes lead to stage transitions 26,27 . Here, we focused on spontaneous arousals because their functional correlates remain undetermined. They are usually considered to induce sleep disruption and its detrimental functional consequences. However, spontaneous sleep arousals might also carry positive effects on brain functions. We quanti ed the prevalence of spontaneous arousals during undisturbed sleep in healthy individuals in late midlife (N = 101), and assessed whether it was associated with early cortical Aβ deposition and cognitive performance. Based on the theoretical concept that sleep arousals are diverse 17 , we classi ed them according to their temporal association with a change in muscular tone and a sleep stage transition. Based on this straightforward phenotyping in a large data sample we provide the rst empirical evidence that different types of sleep arousals have distinct correlates in terms of cognition and brain amyloid burden. Indeed, we found that arousals associated with sleep transitions (T + E-) are associated with higher cortical Aβ deposition in brain regions affected early on by AD neuropathology, suggesting their association with sleep fragmentation and worse brain status. By contrast and unexpectedly, the more prevalent T-E + arousals, which do not result in sleep transitions, are all the more frequent as Aβ deposition is low and cognitive performance superior, particularly in the attentional domain. This arousal type is therefore associated to a more favourable brain and cognitive status. This is of particular importance since arousals have been reported to increase with age, and age represents the most important risk factor for cognitive decline and AD 4 .
Our analyses show that the main characteristic differentiating the two types of arousals is whether or not they lead to a sleep stage transition. Two hypotheses can be put forward to explain the heterogeneity in arousals. On the one hand, all arousals, triggered by a common set of brain areas, might be part of a continuum in which each arousal is characterised by the intensity in its driving neural activity, its spectral composition, its associated muscular tone and its probability of sleep stage transition. Alternatively, the two arousal types are distinct physiological events prompted by different triggering brain structures and propagation cerebral networks.
Oddly enough, the origin of spontaneous arousals remains elusive. Recent fMRI data showed that subcortical regions (including the thalamus, midbrain, basal ganglia and cerebellum) were activated during non-REM (NREM) arousals while cortical regions were deactivated 28 . A recent yet-to-be-reviewed study in rodents provides evidence that arousals leading to sleep state transition are, at least partly driven by the locus coeruleus (LC), brainstem source of norepinephrine with strong and ubiquitous in uence on distant cortical brain regions, including during sleep 29 . In addition, optogenetic stimulation of the LC causes immediate sleep-to-wake transitions, from both NREM and REM sleep and results in highfrequency EEG activity 30,31 . Hence, subcortical activity, for instance in the LC, could underlie transitionarousals while no-transition arousals could also merely be the re ection of cortico-cortical or thalamocortical interplay 17 . Identifying the brain sources of the two types of arousals would require invasive animal testing, coupling EEG to fMRI recordings in humans, or source reconstruction of high density EEG signals 27 .
The cellular and molecular underpinnings of the distinct relationship between the two types of arousals, Aβ burden, and cognition are currently unknown. We can reasonably speculate that T + E-arousals have 2 potentially deleterious impacts. Firstly, they interrupt a sleep stage and consequently all its associated cellular phenomena, like plasticity 26 . Secondly, it seems possible that they considerably increase cellular activity in diffused cerebral regions, a condition conducive to increase Aβ release. By contrast, T-E + arousals might promote Aβ clearance, hypothetically by increasing the pulsatility of cortical penetrating arteries 32 . Additionally, T-E + arousals might offer recurring opportunities to transiently synchronise distant brain areas, in frequency bands otherwise related to cognition (beta oscillations) without enduringly disrupting the underlying brain oscillations (i.e. sleep state), similarly to what sleep spindles allow over sigma band (12-16Hz) oscillations 33 . In complex dynamics wordings, T-E + arousals can be seen as distinct dynamics generated when the oscillatory trajectory is trapped in a local submanifold of an attractor 34 . These transient oscillations give rise to dynamic instability despite the fact that the global manifold does not change. Dynamic instability is a form of complexity in neuronal systems, which is critical for adaptive brain functions such as selection in self-organising systems, learning or memory 25  We emphasise that (1) our cohort only comprised healthy individuals, devoid of SDB, and (2) we focused on spontaneous arousals, which are not generated in response to an endogenous or exogenous perturbation (e.g. apnoea or noise). Therefore, our ndings probably do not apply to perturbation-induced arousals and their negative behavioural 20,23 and neurodegenerative aftermaths 10 . It is tantalising to suggest, and empirically testable, that arousals found in SDB mostly consist in transition-arousals which would contribute in part to the higher risk for AD reported in SDB 35 . We further found no signi cant link between early Aβ burden and the number of full night-time awakenings during sleep or with time spent awake after sleep onset, two markers related to the fragmentation of sleep macrostructure de ning in part sleep quality. The associations we nd with Aβ burden in healthy late midlife appear therefore to be stronger with, if not speci c to, sleep arousals, as compared to other indices of wakefulness during sleep or fragmentation of sleep. This contrast with a previous actigraphy study that reported correlations between WASO and Aβ burden in participants older than those included here (mean: 76.7 ± 3.5y). 36 Our ndings may therefore suggest that, at a younger age (~ 59y), the detrimental association between sleep quality and AD neuropathology initially concerns transition-arousals leading to sleep macrostructure fragmentation, before being subsequently detected over other markers of sleep fragmentation.
Sleep arousals may connect the sleeper's brain with the surrounding endogenous and exogenous relevant incoming information and contribute to elements of cortico-cortical information processing 17,34 . as done through sleep spindles, another fundamental feature of sleep microstructure 33 . Our ndings constitute the rst empirical evidence of the conceptual existence of different arousal types differently associated to important parameters of cognitive and brain health 17 . Sleep micro-fragmentation, as easily indexed by automatic detection of spontaneous arousals, could therefore constitute a marker of favourable brain and cognitive trajectory in clinical practice, at least in late midlife adults and/or in individuals with still early AD neuropathology.

Methods
The study was registered with EudraCT 2016-001436-35. All procedures were approved by the Hospital-Faculty Ethic Committee of ULiège.

Study design and participants
Participants signed an informed consent prior to participating.
In order to ensure the presence of at least some Aβ brain deposit 24 , we targeted healthy older individuals aged 50-70y. 208 volunteers were recruited, of which 101 participated in the actual study (Table 1), the rest being excluded due to one of the following exclusion criteria: clinical symptoms of cognitive impairment (Dementia rating scale < 130; Mini Mental State Examination < 27); Body Mass Index (BMI) ≤ 18 and ≥ 29; recent psychiatric history or severe brain trauma; medication affecting the central nervous system; smoking; excessive alcohol (> 14 units/week) or caffeine (> 5 cups/day) consumption; shift work in the past 6 months; transmeridian travel in the last 2 months. Participants were screened for sleep apnoea/hypopnoea syndrome during an in-lab night of sleep under polysomnography; volunteers with apnoea/hypopnea index ≥ 15/h were excluded. One volunteer was excluded from analyses that included amyloid-beta data due to corrupted PET-scan data caused by technical issues during acquisition. Demographic characteristics of the study sample can be found in Table 2. Sleep assessment Participants were required to follow a regular sleep-wake schedule (± 30 min) for 1 week based on their preferred bed and wake-up times. Compliance was veri ed using sleep diaries and wrist actigraphy (Actiwatch©, Cambridge Neurotechnology, UK). Participants then joined the laboratory ~ 6.5h prior to habitual sleep time and were maintained in dim-light thereafter. Undisturbed habitual sleep was recorded with N7000 ampli ers (EMBLA, Natus, Planegg, Germany) using 11 EEG derivations placed according to the 10-20 system (F3, Fz, F4; C3, Cz, C4; P3, Pz, P4; O1, O2), 2 bipolar electrooculogram (EOGs), and 2 bipolar submental electromyogram (EMG) electrodes. Recordings were sampled at 200 Hz, and rereferenced to the mean of the two mastoids.
Arousal detection Sleep stage scoring and arousal detection were carried out in separate steps by two independent algorithms. Sleep stage scoring was performed in 30s windows using a validated algorithm (ASEEGA, Physip, Paris, France) 38 . Automatic arousal detection was then computed as it is objective, reproducible and time-saving 39 . We used an individually tailored algorithm based on the American Academy of Sleep Medicine (AASM) de nition 18 .
In brief, arousal detection is performed over whole-night recordings split into 1s epoch in two successive steps computed over the power in the broad-alpha (7-13Hz), beta (16-30Hz) and lower-theta (3-7Hz) frequency bands. A xed threshold is rst applied to detect abnormal EEG activity relatively to the wholenight recording: any 1s epoch with power in any of the three frequency bands higher than the whole-night median value in each frequency band is considered as a potential arousal. The second step adapts the threshold to account for the speci c EEG background activity in a shorter time window. A speci c threshold is computed for each 30s window: all 1s epochs without concomitant EMG tone increase are selected, as well as the rst ten 1s epochs without EMG increase before and after the 30s window being evaluated; threshold of each frequency band consists in the median power over the selected 1s epochs. Events composed of at least 3 consecutive 1s epochs with changes in EEG frequencies higher than twice the local median and one median of the whole recording for that frequency band were considered as arousals. For detailed explanations on the method, see 39 .

MRI data
Quantitative multi-parametric MRI acquisition was performed on a 3-Tesla MR scanner (Siemens MAGNETOM Prisma, Siemens Healthineers, Erlangen, Germany). Quantitative maps were obtained by combining images using different parameters sensitive to distinct tissue properties. Multi-parameter mapping was based on multi-echo 3D fast low angle shot at 1 mm isotropic resolution 40 . This included three datasets with T1, proton density (PD), and magnetization transfer (MT)-weighted contrasts imposed by the choice of the ip angle (FA = 6° for PD & MT, 21° for T1) and the application of an additional off-resonance Gaussian-shaped RF pulse for the MT-weighted acquisition. MRI multiparameter maps were processed with the hMRI toolbox 41 (http://hmri.info) and SPM12 (Welcome Trust Centre for Neuroimaging, London, UK) to obtain notably a quantitative MT map as well as segmented images (grey matter, white matter, CSF), normalised to the standard MNI space using uni ed segmentation 42 . Flow-eld deformation parameters obtained from DARTEL spatial normalisation of the MT maps were applied to averaged co-registered PET images 43 . Volumes of interest were determined using the automated anatomical labelling (AAL) atlas 44   Composite scores were computed for the memory, executive function, and attentional domains, and consisted of the standardised sum of the standardised, domain-speci c scores, where higher scores indicate better performances. The memory score consisted of the FCSRT (sum of all 4 free recalls) and the recognition memory score from the MST. The executive function score included Verbal Fluency tests (letter and animals score for 2min), inverse order digit span, TMT (part B), N-back (3-back variant) and Stroop Test (interfering items errors). The attentional score comprised the DSST, TMT (part A), N-back (1back variant), D2 (Gz-F) and CRT (reaction time to dissimilar items).

Statistical analyses
Statistical analyses were performed in SAS 9.4 (SAS Institute, Cary, NC) using Generalised Linear Mixed Models (GLMMs). The distribution of dependent variables was determined by tting all parametric probability distributions to data, using the "all tdist" function in Matlab Declarations Study concept and design: E.S., P.M., C.P., C.B., F.C. and G.V. Data acquisition, analysis and interpretation: all authors. D.C. and G.V. drafted the rst version of the manuscript. All authors revised the manuscript, and had nal responsibility for the decision to submit for publication.

Data availability
The dataset, including deidenti ed participant data, can be made available upon request after approval of a proposal with a signed data access agreement. In order to access the data, the requestor may contact the corresponding author.