Timely sleep coupling: spindle-slow wave synchrony is linked to early amyloid-β burden and predicts memory decline

Sleep alteration is a hallmark of ageing and emerges as a risk factor for Alzheimer’s disease (AD). While the fine-tuned coalescence of sleep microstructure elements may influence age-related cognitive trajectories, its association with AD processes is not fully established. Here, we investigated whether the coupling of spindles and slow waves is associated with early amyloid-beta (Aβ) brain burden, a hallmark of AD neuropathology, and cognitive change over 2 years in 100 healthy individuals in late-midlife (50-70y; 68 women). We found that, in contrast to other sleep metrics, earlier occurrence of spindles on slow-depolarisation slow waves is associated with higher medial prefrontal cortex Aβ burden (p=0.014, r2β*=0.06), and is predictive of greater longitudinal memory decline (p=0.032, r2β*=0.07). These findings unravel early links between sleep, AD-related processes and cognition and suggest that altered coupling of sleep microstructure elements, key to its mnesic function, contributes to poorer brain and cognitive trajectories in ageing.

hallmark of AD neuropathology, and cognitive change over 2 years in 100 healthy individuals in late-23 midlife (50-70y; 68 women). We found that, in contrast to other sleep metrics, earlier occurrence of 24 spindles on slow-depolarisation slow waves is associated with higher medial prefrontal cortex Aβ 25 burden (p=0.014, r²β*=0.06), and is predictive of greater longitudinal memory decline (p=0.032, 26 r²β*=0.07). These findings unravel early links between sleep, AD-related processes and cognition and 27 suggest that altered coupling of sleep microstructure elements, key to its mnesic function, contributes to 28 poorer brain and cognitive trajectories in ageing. 29

INTRODUCTION 30
Alterations in sleep quality are typical of the ageing process with a more fragmented and less 31 intense (or shallower) sleep detected as early as the fifth decade of life 1 . Beyond healthy ageing, 32 alterations in sleep are predictive of the risk of developing Alzheimer's disease (AD) over the next 5 to 33 10 years 2,3 . Similarly, sleep disorders such as insomnia and obstructive sleep apnoea syndrome are 34 associated with increased odds for AD diagnosis 4,5 . Brain burdens of aggregated amyloid-β (Aβ) and 35 tau proteins, hallmarks of AD pathophysiology, have been linked with a reduced sleep intensity, as 36 indexed by the overall production of slow waves during sleep, but also to worse objective sleep 37 efficiency and subjective sleep quality, in healthy and cognitively normal older individuals aged > 70y 6-38 11 . Sleep alteration may in turn contribute to the aggregation of AD proteins: experimental sleep 39 deprivation and sleep fragmentation (disturbance in the production of slow wave during sleep) lead to 40 increased concentration of Aβ in the cerebrospinal fluid (CSF), both in animal models and in healthy 41 human populations 7,9,12,13 . Overall, a bidirectional detrimental relationship between sleep quality and the 42 neuropathology of AD is emerging in the literature. Sleep may therefore constitute a modifiable risk 43 factor which one could act upon to prevent or delay the neuropathological processes associated to AD 44 and favour successful cognitive trajectories over the lifespan [14][15][16] . Hitherto, however, sleep is not yet 45 widely recognised as an independent risk factor for AD and the mechanistic associations between sleep 46 and early AD neuropathology are not yet fully established. 47 Sleep microstructure elements, such as sleep spindles and slow waves (SW), are essential 48 correlates of cognitive function of sleep, as higher densities of both elements during post-learning sleep 49 have been linked to a better overnight memory consolidation 17-21 . Furthermore, SW activity (i.e. a power 50 measure combining the density and amplitude of sleep SW over sleep cycles in the 0.75 to 4 Hz band) 51 was reported to modulate the regression between prefrontal cortex Aβ burden and a lower overnight 52 memory consolidation in cognitively normal older individuals 6 . The fine-tuned coupling of spindles and 53 SW has further been reported to be altered in ageing, with an earlier occurrence of the spindle relative 54 to the SW depolarisation phase in the older compared to younger individuals, and to predict overnight 55 memory retention 22 . Whether this spindle-SW coupling in ageing is associated to AD-pathological 56 processes and cognitive trajectories is not fully established, however. A study in 31 individuals, aged 57 around 75y, found a link between the brain deposit of tau protein and the coupling of spindles and SW 23 . 58 By contrast, another research failed to find a link between this coupling and Aβ brain burden 24 . Here, 59 we argue that, on top of potential statistical power issues, the difficulty to detect this link may be due to 60 the fact that the assessments were carried out late over the lifespan (i.e. > 70y), when subtle associations 61 may be masked by concurrent brain alterations, and by the heterogeneity of sleep SW. 62 The low frequency oscillations of the electroencephalography (EEG) have been divided in slow 63 oscillations (≤ 1 Hz) and delta waves (1-4 Hz) for decades in humans, notably based on the theoretical 64 framework of the generation of SW 25-27 . As one of the main features of the SW resides in their transition 65 from down-to up-state, reflecting synchronised depolarisation, a recent work proposed the transition 66 frequency of the down-to-up state as a way of distinguishing between slow and fast switcher SWs 28 . 67 Compared to young adults, on top of exhibiting a typical overall lower density of SW, older individuals 68 reportedly exhibit higher probabilities of producing slow as compared to fast switcher SWs, providing 69 important insights into age-related changes in sleep microstructure. 70 Investigating the coupling of spindles and SW, appropriately split between the slow and fast 71 switchers, in late middle-aged healthy adults may be the best approach to gain insight into the biology 72 underlying the early relationship between sleep and AD-related processes. In a longitudinal study, we 73 therefore tested whether the coupling of spindles with the slow and the fast switcher SWs is associated 74 with the early brain burden of Aβ and memory performance in a large sample (N=100) of healthy and 75 cognitively normal individuals in late midlife (50-70y concomitantly to EEG and PET measurements (N=100) as well as at follow-up, 2 years later, in a 83 substantial subsample (N=66). We hypothesised that our large sample of individuals, positioned 84 relatively early in the ageing process, would allow to detect subtle associations between the impaired 85 fine-tuned coupling of spindles and slow and fast switcher SWs, and both the early Aβ burden and 86 memory performance decline over 2 years. 87

RESULTS 88
Slow but not fast switcher SW show preferential coupling with spindles 89 We first assessed whether spindles were specifically associated with a particular phase on the 90 down-to-up state transition of the SWs, considering the two types of SWs (slow and fast switchers switcher SWs ( Figure 1B) (U² = 71.143, p <0.001). Indeed, spindles were preferentially anchored only 97 onto slow switchers, as spindles were most often initiated on the ascending phase of the depolarised 98 state of the slow switcher SWs, while no such preferred coupling was detected for fast switcher SWs 99 ( Figure 1C). As the slow switcher SWs are the only SW to exhibit a preferential coupling with spindles, 100 they remained our main focus of interest for the remaining analyses.

Spindle onset on slow switcher SW is linked to prefrontal Aβ burden 112
Statistical analysis revealed that the anchoring of the spindle onset onto slow switcher SWs was 113 significantly linked to the burden of Aβ over the medial prefrontal cortex (MPFC) (Figure 2A) (main 114 effect of Aβ PET uptake: F1,96=6.2, p=0.014, r²β*=0.06), where earlier onset of the spindle relative to 115 the SW phase was associated with higher Aβ PET uptake ( Figure 2B). This effect was detected while 116 controlling for the differences between sexes (main effect of sex: F1,96=5.01, p=0.028, r²β*=0.05), as a 117 later spindle onset was found in men relative to women, and controlling for age (main effect of age, 118

144
To test the apparent difference in the association between the coupling of spindles onto slow 145 and fast switcher SWs with the accumulation of Aβ protein, we further computed a statistical model 146 with spindle-SW coupling as the dependent variable, while including the SW type together with the Aβ 147 MPFC burden as independent variables. Statistical analysis yielded a significant interaction between the 148 burden of Aβ in the MPFC and the type of SWs (Aβ burden by SW type interaction: F1,96=7.05, p=0.009; 149 r²β*=0.07), while post-hoc tests indicated that the link between the coupling of the spindle onto the SW 150 and the MPFC Aβ burden was significant for the slow switcher type (t149.8=-2.00, p=0.047) and not for 151 the fast switcher type (t149.8=0.56, p=0.57). This finding reinforces the idea that slow and fast switcher 152 SWs constitute distinct realisations of NREM oscillations that are differently associated with brain 153 aggregation of Aβ during the ageing process. This has likely contributed to previous failures to detect 154 links between the coupling of spindles and SW and the deposit of Aβ. In fact, when testing the 155 association between the coupling of spindles and SWs, irrespective of the type of SWs, and PET Aβ 156 burden over the MPFC, the statistical analysis only yields a weak negative association between the phase 157 of the coupling of the spindle onto the SW and Aβ burden (main effect of Aβ uptake: F1,96=3.96, 158 p=0.049, r²β*=0.04; main effect of sex: F1,96=4.33, p=0.04, r²β*=0.04; main effect of age: F1,96=0.05, 159 p=0.83), which could arguably go undetected in a smaller or different sample. 160 Slow switchers spindle phase coupling is associated to memory change over two years 161 We tested whether the coupling of spindles with slow switcher SWs was associated with 162 memory performance and its decline over 2 years using the mnemonic similarity task (MST) (Figure  163 3A). The MST consists in a pattern separation task targetting the ability to distinguish between highly 164 resembling memory events, a hippocampus dependent task which is very sensitive to early cognitive 165 decline 32,33 . Across the sample, we observed an overall decline in performance between the baseline and 166 follow-up performance at the MST (t65 = 2.19, p=0.032). We found no significant link between the 167 coupling of the spindles onto both SW types and the performance on the task at baseline (i.e. assessed

DISCUSSION 196
In order to unravel early associations between the microstructure of sleep and the burden of Aβ 197 in the brain, and their cognitive implications, we collected polysomnography, PET and behavioral data 198 in a relatively large sample of individuals without cognitive impairments or sleep disorders. To this end, 199 we recruited individuals in late middle age (50-70y), that could in most cases only present limited age- reports in a smaller sample of individuals older (> 70y) than our sample. Altogether, these discrepancies 275 reinforce the idea that the alteration in the microstructure of sleep, consisting in the coupling of the 276 spindles onto a specific subpopulation of SWs, as reported here, but also in the occurrence of micro-277 arousals during sleep we previously reported based on the same sample 41 , shows a prior association with 278 AD-related processes compared with the amount of slow brain oscillations generated during overnight 279 sleep. The latter may only be significantly associated at a later age, when the pathophysiological changes 280 are already more substantial. In addition, the coupling of spindles onto slow switcher SWs is predictive 281 of the future change in memory performance. Sleep microstructure could therefore constitute a 282 promising early marker of future cognitive and brain ageing trajectory 41 . We did not evaluate whether 283 distinct links between the SW types and slow and fast spindles are observed. As some reports describe 284 that fast spindles are rather coupled to the up-state of the SWs and slow spindles tend to occur on the 285 waning depolarisation phase of the SWs 42 , we could hypothesise that the associations we observe are 286 probably rather driven by fast spindles. Future investigations are, however, needed to confirm this 287 hypothesis 288 One should bear in mind the potential limitations of our study. First, although we collected data 289 in a relatively large sample, we may have insufficient power to detect some associations with other sleep 290 measures. One can nevertheless frame our findings in relative terms such that association between 291 spindle-SWs coupling and the early accumulation of Aβ protein is at least stronger than the association 292 with the coupling of spindles onto fast switchers SW, the density of SWs and spindles, the SWE, etc. 293 Also, the longitudinal aspect of our study is relatively short-termed and only concerned the performance 294 to a sensitive mnesic task while it did not include sleep EEG and the PET assessments. Further studies 295 should evaluate the predictive value of such parameters on longer longitudinal protocols, and the 296 evolution of the sleep EEG and the PET parameters as well as their generalisability over other 297 precociously impacted cognitive abilities. Finally, given that our protocol does not include manipulation 298 of the coupling of the spindles onto the SWs, it precludes any inference on the causality of one aspect 299 onto the other. 300 Together, our findings reveal that the timely occurrence of spindles onto a specific type of SWs 301 showing a relative preservation in ageing seems to play a determining role in ageing trajectory, both at 302 the cognitive level and with regards to structural brain integrity. These findings may help to unravel 303 early links between sleep, AD-related pathophysiology and cognitive trajectories in ageing and warrants 304 future clinical trials attempting at manipulating sleep microstructure or Aβ protein accumulation. Committee of ULiège. All participants signed an informed consent prior to participating in the study. 321

Slow waves and spindle detection 338
Only the frontal electrodes were considered because the frontal cortex is an early site showing 339 Aβ deposit and is the primary generator of the SWs during sleep 6,29-31 as well as to facilitate 340 interpretations of future large-scale studies using headband EEG restricted to frontal electrodes 8 . SWs 341 were automatically detected during N2 and N3 epochs of NREM sleep devoid of artefacts/arousals >5s 342 long, using a previously developed algorithm 48 . Data were first band-filtered between 0.3 and 4.0Hz 343 with a linear phase Finite Impulse Response (FIR) filter. Following recent work, SW detection criteria 344 were adapted for age and sex 48 : peak to peak amplitude ≥70µV (resp. ≥60.5µV) and negative amplitude 345 ≤ -37µV (resp. ≤ -32µV) was used for women (resp. for men), instead of the standard ≥75µV and ≤ -346 Centre for Neuroimaging, London, UK) to obtain notably a quantitative MT, which was segmented into 371 grey matter, white matter, and CSF using unified segmentation 54 . Flow-field deformation parameters 372 obtained from DARTEL spatial normalisation of the individual MT maps were applied to the averaged 373 co-registered PET images 55 . The volumes of interest were determined using the automated anatomical 374 labelling (AAL) atlas 56 . 375

PET-scan 376
Aβ PET imaging was performed using [ 18 F]Flutemetamol, except for 3 volunteers for which 377 [ 18 F]Florbetapir was used. PET-scans were performed on an ECAT EXACT+ HR scanner (Siemens, 378 Erlangen, Germany). Participants received a single dose of the radioligand in the antecubital vein (target 379 dose 185±10% MBq); image acquisition started 85min after the injection and consisted of 4 frames of 380 5 minutes, followed by a 10 minutes transmission scan using 68 Ge line sources. Images were 381 reconstructed using a filtered back-projection algorithm including corrections for the measured 382 attenuation, dead time, random events, and scatter using standard software (Siemens ECAT -HR + 383 V7.1, Siemens/CTI Knoxville, TN, USA). Individual PET average images were produced using all 384 frames and were then manually reoriented according to MT-weighted structural MRI volumes and 385 coregistered to the individual space structural MT map. Standardised uptake value ratio (SUVR) was 386 computed using the whole cerebellum as reference region 57 . As images were acquired using 2 different 387 radioligands, their SUVR values were converted into Centiloid Units 57 (the validation of the procedure 388 in our sample was previously published 58 ). The Aβ burden was averaged over a mask covering the 389 medial prefrontal cortex previously reported to undergo the earliest aggregation sites for Aβ pathology 34 . 390

Cognitive assessments 391
As part of an extensive neuropsychological assessment, participants were administered the 392 Mnemonic Similarity Task (MST) 59 , a visual recognition memory task. After an incidental encoding 393 phase during which participants were randomly presented 128 common objects for a period of 2s, and 394 were instructed to determine whether the object presented on the screen was rather an 'indoor' or 395 'outdoor' item, the recognition memory phase consisted in the presentation of 192 objects (64 old, 396 presented previouslytarget items; 64 similar but not identical to the previously presented stimuli -397 lure; and 64 new objectsfoil items). In this phase, participants were instructed to determine whether 398 the presented object was new (foil), previously presented (old), or similar but not perfectly identical 399 (lure). For statistical analyses, the recognition memory score was used (RM), computed as the difference 400 between the rate of calling a target item "old" minus the rate of calling a foil item "old" [P("old"|target)-401 P("old"|foil)] 32,43 . 402 The MST was administered at two timepoints: the first time, the day preceding the baseline 403 night, during a cognitive evaluation performed ~ 6.5h before habitual bedtime. The second 404 neuropsychological evaluation was carried out ~24 months after the first one (mean 767±54 days). The 405 memory decline score was computed as the baseline performance minus the follow-up performance, 406 divided by the baseline performance, so that a higher score indicates a higher decline over the 2 years. 407 = RM baseline − RM follow − up RM baseline 408

Statistics 409
Statistical analyses were performed using Generalised Linear Mixed Models (GLMMs) in SAS 410 9.4 (SAS Institute, Cary, NC). The distribution of dependent variables was verified in MATLAB 2013a 411 and the GLMMs were adapted accordingly. Subject was treated as a random factor and each model was 412 corrected for age and sex effects. Kenward-Roger's correction was used to determine the degrees of 413 freedom. Cook's distance was used to assess the potential presence of outliers driving the associations, 414 and as values ranged below 0.45 no datapoint was excluded from the analyses (a Cook's distant > 1 is 415 typically considered to reflect outlier value). Our main analysis concerned the coupling between SW 416 types and spindles, and as two analyses were performed (one with slow switcher and one with fast 417 switcher SW), the significance threshold is set at p<0.025 for these analysis, to account for multiple 418 comparisons. The reader should note that we performed a separate statistical test with Aβ burden as 419 dependent variable with both SW-spindle coupling and SW type as independent variables, which 420 confirmed the associations reported in the separate models. The remaining analysis were exploratory as 421 they arise from the main analyses and do not require correction for multiple comparisons. Semi-partial 422 R² (R²β*) values were computed to estimate the effect sizes of significant fixed effects and statistical 423 trends in all GLMMs 60 . P-values in post-hoc contrasts (difference of least square means) were adjusted 424 for multiple testing using Tukey's procedure. Watson's non-parametric two-sample U² test for circular-425 normal data was performed in MATLAB 2019 to assess the difference between the distribution of 426 spindle onset on the phase of slow waves for slow and fast switcher SW. For analyses using the phase 427 of spindle onset on the slow waves, as the phase of all subjects were in the quarter between the zero 428 crossing (-π/2) and depolarisation, the cosine of the phase was used instead of the phase, in order to 429 perform linear statistics. Statistics with the phase yielded the same results.