Alterations of sleep oscillations in Alzheimer’s disease: A potential role for GABAergic neurons in the cortex, hippocampus, and thalamus

Abstract

Sleep abnormalities are widely reported in patients with Alzheimer’s disease (AD) and are linked to cognitive impairments. Sleep abnormalities could be potential biomarkers to detect AD since they are often observed at the preclinical stage. Moreover, sleep could be a target for early intervention to prevent or slow AD progression. Thus, here we review changes in brain oscillations observed during sleep, their connection to AD pathophysiology and the role of specific brain circuits. Slow oscillations (0.1–1 Hz), sleep spindles (8–15 Hz) and their coupling during non-REM sleep are consistently reduced in studies of patients and in AD mouse models although the timing and magnitude of these alterations depends on the pathophysiological changes and the animal model studied. Changes in delta (1–4 Hz) activity are more variable. Animal studies suggest that hippocampal sharp-wave ripples (100–250 Hz) are also affected. Reductions in REM sleep amount and slower oscillations during REM are seen in patients but less consistently in animal models. Thus, changes in a variety of sleep oscillations could impact sleep-dependent memory consolidation or restorative functions of sleep. Recent mechanistic studies suggest that alterations in the activity of GABAergic neurons in the cortex, hippocampus and thalamic reticular nucleus mediate sleep oscillatory changes in AD and represent a potential target for intervention. Longitudinal studies of the timing of AD-related sleep abnormalities with respect to pathology and dysfunction of specific neural networks are needed to identify translationally relevant biomarkers and guide early intervention strategies to prevent or delay AD progression.

1. Sleep changes are common in Alzheimer’s Disease

Sleep abnormalities are widely reported in patients with Alzheimer’s disease (AD) and in AD mouse models. These abnormalities have been linked to the cognitive impairments characteristic of AD. Statistical modeling approaches as well as correlational studies suggest an association between AD pathology and sleep disruption, which together lead to memory impairment (Guarnieri et al., 2012; Mander et al., 2015; Wang and Holtzman, 2020; Wennberg et al., 2017). Although sleep disturbances are common in normal aging, the severity of these disturbances are exaggerated in people with AD. The sleep abnormalities in AD include changes in overall sleep pattern such as decreased total sleep time and fragmentation of non-rapid-eye-movement (NREM) sleep as well as REM sleep (Bliwise et al., 1989; Bonanni et al., 2005; Duncan et al., 2022; Liguori et al., 2014; Lim et al., 2013; Mander et al., 2016; Vitiello et al., 1990). Furthermore, as discussed in the rest of this review, alterations in specific cortical electrical oscillations during sleep are common. These changes in sleep architecture and circadian rhythms lead to excessive daytime sleepiness, sundowning, and insomnia in AD patients (Vitiello et al., 1990). Notably, the sleep changes become apparent at the preclinical stage of AD and get worse over the course of the disease (Mander et al., 2016). Sleep and AD pathology seem to have a bidirectional relationship, whereby sleep disturbances exacerbate clinical symptoms/pathology and increases in amyloid-β (Aβ)/tau burden are correlated with sleep disturbances (Ju et al., 2014; Mander et al., 2016).

A variety of brain oscillations are observed during sleep. A growing body of evidence suggests contributions of sleep oscillations to cognitive functions especially sleep-dependent memory consolidation and re-consolidation (Diekelmann and Born, 2010; Klinzing et al., 2019; Manoach and Stickgold, 2019; Rasch and Born, 2013). Of particular interest for this review are recent AD studies showing impairments in sleep oscillations including changes in slow wave activity (0.5–4 Hz) (Byron et al., 2021), sleep spindles (8–15 Hz) (Gorgoni et al., 2016; Weng et al., 2020), and sharp-wave ripples (100–250 Hz) (Sanchez-A-guilera and Quintanilla, 2021; Zhen et al., 2021). These abnormalities correlate with the severity of deposition of Aβ and cognitive impairment during later stages of the disease. It is also suggested that sleep abnormalities in AD are linked to impaired glymphatic waste clearance of amyloid and tau, which is known to occur during NREM sleep in healthy subjects (Holth et al., 2019; Jessen et al., 2015; Kang et al., 2009; Mander et al., 2015). Despite these interesting findings, the exact causal relationships between the AD-related neuronal dysfunction, sleep abnormalities, and increased Aβ and Tau burden are still to be definitively determined. Furthermore, the mechanisms underlying these changes and the identification of specific brain circuits which can be targeted therapeutically are topics of considerable importance. Studies with animal models of AD could provide further details in changes in sleep oscillations and the potential causality of these abnormalities, as well as their effect on AD pathophysiology. Thus, in this review, we focus on the changes in the NREM sleep related oscillations in both AD patients and in AD mouse models (Tables 1–5), including slow and delta oscillations, sleep spindles, sharp-wave ripples, as well as REM sleep related oscillations, which are all recognized as the key features of memory processing and potential biomarkers of AD and dementia (Fig. 1). In the final section, we consider evidence that alterations in GABAergic neurons in the cortex, hippocampus and thalamic reticular nucleus could account for sleep oscillatory changes in AD. For more general reviews on mechanisms of sleep, and sleep and circadian changes in AD, readers may refer to other comprehensive reviews (Brown et al., 2012; Byron et al., 2021; Hanke et al., 2022; Kent et al., 2021; Lee et al., 2020; Mander, 2020; Ning and Jorfi, 2019; Wang and Holtzman, 2020; Weng et al., 2020). Another article in this special issue focuses on pharmacological and brain stimulation approaches to influence sleep and arousal in brain disorders including dementia (Brown et al., 2022).

Table 1.

Studies of slow wave activities in AD mouse models.

Mouse Model Name	Amyloid/Tau Model	Frequency Band (Hz)	Age (months)	Findings	Reference
APP/PS1	Amyloid	0.1–4	8–10	↓ power at 8–10 mo	Kent et al. (2018)
APP/PS1	Amyloid	Optogenetic stimulation at 0.6	4–7	↓ Aβ, ↓ calcium overload, Restored SO power	Kastanenka et al. (2017)
APP/PS1	Amyloid	Optogenetic stimulation at 1.2	3–9	↑ Aβ, ↑ calcium overload,	Kastanenka et al. (2019)
				↓ spine density
Tg2576	Amyloid	0.5–4	2, 6, 12	↓ power at 6–12 mo	Zhang et al. (2005)
5XFAD	Amyloid	0.1–0.5, 0.5–4	3, 6	↑ 0.1–0.5 Hz power at 6 mo	Schneider et al. (2014)
				↓ 0.5–4 Hz power at 6 mo
rTG4510	Tau	0.1–4	5–10	↓ power at > 6 mo	Holton et al. (2020)
P301S (PS19)	Tau	1–4	3, 6, 9, 11	↓ power at 11 mo	Holth et al. (2017)
PLB2Tau	Tau	1.5–5	6	↓ power at 6 mo	Koss et al. (2016)
3xTg-AD	Amyloid/Tau	0.1–4	18	No change in power at 18 mo	Kent et al. (2018)
3xTg-AD	Amyloid/Tau	< 1	7, 20	Under anesthesia	Castano-Prat et al. (2019)
				Faster SO freq. and shorter UP state duration at
				7mo
				Slower SO freq. and loger UP/DOWN state duration at 20 mo
AppNL-G-F/NL-G-F	Amyloid	1–4	6, 12	↓power during NREM sleep at 6 and 12 mo	Maezono et al. (2020)

Fig. 1. A variety of neural oscillations are observed during sleep and altered in AD.

(A) Key brain regions involved in sleep oscillations discussed in this review. (B) Schematics of sleep oscillations in different frequencies. (C) An illustration of slow oscillation (SO) properties. Red dots show positive to negative zero crossings used to identify SO duration (t). Peak-to-peak amplitude is defined by adding the trough height (x) and the peak height (y). (D) Coupling of spindles to the rising phase of SO in local field potential recordings. Red vertical bar shows the spindle peak. 100 spindle-SO coupled events in one mouse were aligned at the spindle peak and averaged.

2. NREM sleep

2.1. Slow wave activity (SWA): Slow (<1 Hz) and Delta (1–4 Hz) oscillations

2.1.1. Slow wave activity

Slow wave activity (SWA) during NREM sleep commonly refers to waveforms with high amplitude and low frequency ranging from 0.5 to 4 Hz, which can be further divided into slow (<1 Hz) and delta (1–4 Hz) oscillations (Kim et al., 2019; Steriade, 2006; Steriade et al., 1993c, 1993d, 1993b, 1993a; Steriade and Timofeev, 2003; Uygun and Basheer, in this issue). SWA is one of the major components of neural oscillations characterized during sleep, especially in the deeper stages of NREM sleep (Steriade et al., 1993a, 1993c, 1993d). Furthermore, SWA is considered a marker of sleep intensity since it increases in the recovery sleep following an acute period of sleep deprivation (Achermann et al., 2001). Recent animal studies suggest that the narrow 2.5–3.5 Hz band is especially increased following sleep deprivation (Hubbard et al., 2020). Slow oscillations/waves (SO, <1 Hz) are characterized by larger amplitudes and more global occurrence across brain regions whereas delta oscillations (1–4 Hz) have smaller amplitudes and are a more local phenomenon (Bernardi et al., 2018; Genzel et al., 2014; Siclari et al., 2014). Of note, although in the sleep field SOs are typically defined as < 1 Hz and delta as 1–4 Hz, some studies use the terms slow wave activity and delta interchangeably. Here we describe findings regarding < 1 Hz EEG activity as SOs and 1–4 Hz as delta oscillations, even if the cited studies used different terminology. In this subsection we first describe the cellular and network mechanisms underlying these EEG features, as well as their potential physiological functions before reviewing changes observed in AD patients (Section 2.1.2.) and in mouse models of AD (Section 2.1.3.). We follow a similar organization for subsequent sections focused on other sleep oscillations.

Animal studies showed that cortical layer 5 pyramidal cells are the key generators of SOs (Beltramo et al., 2013; Chauvette et al., 2010; Sakata and Harris, 2009; Sanchez-Vives and McCormick, 2000). Thus, SOs are commonly observed in the cortex (Chauvette et al., 2010; Isomura et al., 2006; Luczak et al., 2007; Sakata and Harris, 2009; Steriade et al., 1993d; Timofeev et al., 2000). Slow oscillations originate locally in distinct cortical areas and then spread from the site of origin throughout the cerebral cortex as travelling waves (Riedner et al., 2011; Timofeev, 2013). Waves originate more frequently in prefrontal-orbitofrontal regions and propagate in an anteroposterior direction (Massimini et al., 2004). Triggering of travelling waves could be a potential mechanism by which local optogenetic stimulation of cortical neurons on one side of the brain caused a reduction of amyloid deposition and prevented neuronal calcium elevations on the contralateral side of the brain (Kastanenka et al., 2017). SOs can be observed in subcortical regions as well, such as the hippocampus (Isomura et al., 2006), thalamus (David et al., 2013; Hughes et al., 2002), thalamic reticular nucleus (Blethyn et al., 2006), brainstem (Mena-Segovia et al., 2008) and other regions receiving substantial cortical input. Importantly, more recent studies suggested that the thalamus actively contributes to generation of SOs as well (Crunelli et al., 2018; Gent et al., 2018). The SOs are comprised of two aperiodic alternating states known as UP and DOWN states (Steriade, 2006; Steriade et al., 1993d). The UP state represents the synchronous excitation of population of neurons whereas the DOWN state is characterized by neuronal silencing (Harris and Thiele, 2011; Luczak et al., 2007; Sakata and Harris, 2012, 2009; Steriade et al., 1993d). Although the synaptic potentials and discharge of cortical pyramidal neurons are the major contributors to SOs, this dynamic cycle of UP and DOWN states requires a balance between excitatory and inhibitory activities of neurons. Specifically, parvalbumin (PV) and somatostatin (SOM)-positive GABAergic neurons were shown to regulate UP and DOWN state transitions (Zucca et al., 2017). A series of studies showed that GABAergic neurons co-expressing neuronal nitric oxide synthase (nNOS) and SOM are sleep-active and involved in regulation of SWA (0.5–4.0 Hz) especially in the SO frequency range (0–1.5 Hz). These neurons are also involved in cortical-dependent recognition memory performance (Gerashchenko et al., 2018, 2008; Zielinski et al., 2019). Recent studies also suggested involvement of deep-layer neurogliaform cells and astrocytes in SOs and NREM sleep (Amzica and Steriade, 2000; Bojarskaite et al., 2020; Buskila et al., 2019; Hirase et al., 2004; Szabo et al., 2017; Valero et al., 2021).

Whereas SOs originate within cortical circuits, delta oscillations are generated primarily in the thalamus, although cortical mechanisms also contribute to EEG delta power (Uygun and Basheer, in this issue). At the cortical level, delta oscillations are derived from intracortical network interactions, where intrinsically bursting neurons respond to the onset of the depolarizing phase of the slow oscillation and generate cycles of short oscillations at delta frequencies (Amzica and Steriade, 1998). At the thalamic level, this rhythmic firing at delta frequency is derived from the intrinsic voltage-gated ion channels of thalamocortical neurons which promote bursting at delta frequencies when these neurons are hyperpolarized due to the withdrawal of excitatory neuromodulatory inputs and/or increased hyperpolarization from GABAergic neurons, especially from the thalamic reticular nucleus (Jahnsen and Llinas, 1984a, 1984b; Leresche et al., 1990; Lewis et al., 2015; Maquet et al., 1997; Soltesz et al., 1991; Steriade et al., 1993b; Uygun et al., 2022).

One of the well-studied functions of SWA is its role in sleep-dependent memory consolidation (Lu and Goder, 2012; Marshall et al., 2006; Steriade and Timofeev, 2003; Walker, 2009). Studies showed that cortical SOs trigger the reactivation of temporarily-stored hippocampal memories in coordination with hippocampal sharp-wave ripples and sleep spindles, which will be reviewed in more detail in the later sections (Diekelmann and Born, 2010; Klinzing et al., 2019; Rasch and Born, 2013). The synaptic homeostasis hypothesis illustrates a critical role of SWA in downscaling synaptic strength during sleep to enhance learning and memory (Tononi and Cirelli, 2006). Studies suggest that increased synaptic potentiation during wakefulness is decreased during subsequent sleep (Cirelli and Tononi, 2020). It is considered that SWA, especially the slow oscillation (< 1 Hz) with sequences of depolarization-hyperpolarization, is suitable to promote synaptic depotentiation via changes in calcium dynamics (Kemp and Bashir, 2001). Downscaling of synaptic strength during sleep conserves energy and space and is beneficial for learning and memory by increasing the signal-to-noise ratio in the relevant neural network (Tononi and Cirelli, 2006). In fact, the amount of slow wave sleep has been shown to correlate with memory performance (Backhaus et al., 2007, 2006). A study in healthy humans in which slow-oscillation-like activities were induced via transcranial stimulation demonstrated enhanced retention of hippocampus-dependent declarative memories (Marshall et al., 2006). Although the importance of slow wave sleep or slow wave activities in memory consolidation has been widely-reported, the distinct roles of SO and delta oscillations are still being revealed. Intriguingly, a recent mouse optogenetic study of motor learning suggests that SO contributes to memory consolidation processes by preserving memory reactivation whereas delta oscillations promote forgetting by weakening memory reactivations (Kim et al., 2019).

Another important event that happens during slow wave sleep is the clearance of waste in the brain. For example, a higher concentration of Aβ and tau in the cerebrospinal fluid and interstitial fluid were reported with prolonged wakefulness or sleep deprivation (Holth et al., 2019; Kang et al., 2009). Additionally, SOs are associated with the activity of the glymphatic system which removes protein waste products in the brain such as Aβ (Fultz et al., 2019; Xie et al., 2013). In fact, an optogenetic study in AD mouse model demonstrated that stimulation of cortical neurons at SO frequency reduced the rate of amyloid plaque deposition (Kastanenka et al., 2017), which will be discussed more in the following section (Section 2.3.).

2.1.2. Slow wave activity and AD in humans

As slow wave activity is a key component of NREM sleep, it is perhaps not surprising that a deficit in SWA measured in EEG is one of the most widely reported sleep abnormalities in AD (Kent et al., 2021; Liguori et al., 2020, 2014; Liu et al., 2020; Loewenstein et al., 1982; Vitiello et al., 1990). Reduced SWA has been observed in EEG recordings over midline frontal, central, and parietal regions in cognitively normal older adults with Aβ or tau pathology (Lucey et al., 2019; Mander et al., 2015, 2013; Varga et al., 2016), as well as in patients with mild cognitive impairment (MCI) (Westerberg et al., 2012), indicating that this abnormality emerges early in the AD progression. Interestingly, the decrease in number of slow wave events associated with Aβ burden in medial prefrontal cortex (mPFC) was found to be specific to the lower frequency range (< 1 Hz) of SWA (Mander et al., 2015). The power of lower frequency SWA (< 1 Hz) was also shown to have a non-linear association with longitudinal changes in cognitive performance (Lucey et al., 2021). In contrast, slow waves in the delta range (1–4 Hz) were found to be increased with Aβ burden, suggesting that slow highly synchronous events which normally occur in healthy adults might be converted to faster, lesser synchronous delta waves in AD patients. Animal studies suggest that this pattern of changes is associated with weaker memory reactivation and greater forgetting (Kim et al., 2019). However, in another study, decreased delta power (1–4 Hz) was observed with higher tau deposition (Lucey et al., 2019). Thus, delta changes are more variable and may depend on the extent of amyloid vs tau deposition. Although the mechanisms linking altered SO and prefrontal Aβ level are yet to be fully understood, studies suggest that Aβ pathology affects SO generation by disrupting the cortico-thalamic network (Kurup et al., 2010; Snyder et al., 2005; Steriade et al., 1993a, 1993c). Specifically, Aβ has been implicated in prefrontal atrophy in older adults as well as disrupted functioning of synaptic NMDA receptors which are involved in SO generation (Kurup et al., 2010; Mander et al., 2013; Snyder et al., 2005). Importantly, it has been pointed out that this relationship between Aβ burden and SWA abnormalities could be bidirectional (Lucey et al., 2019; Mander et al., 2015). Aβ levels increase during wakefulness and decrease during sleep (Kang et al., 2009), while one night of total sleep deprivation diminished this Aβ clearance (Ooms et al., 2014). Tau level is also affected by sleep but to a lesser degree than Aβ, as one night of total sleep deprivation did not affect cerebrospinal fluid (CSF) tau level (Ooms et al., 2014), whereas several nights of low-quality sleep increased the CSF tau level (Ju et al., 2017). It is possible that one factor (Aβ/tau pathology or SWA deficits) could trigger the other abnormality, which in turn initiates the vicious cycle to exacerbate disease progression.

Another hallmark of SWA is its homeostatic response, that is, it increases after prolonged wakefulness (Achermann et al., 2001; Tononi and Cirelli, 2006). Although deficits in SWA are reported in the AD patients, changes in SOs as a homeostatic response following sleep deprivation in AD patients have not yet been directly characterized as far as we are aware. Of note, changes in SWA in AD patients have commonly been measured as proportion of NREM stage 3 or spectral power of SWA combining both SO and delta range. As distinct functions of SOs and delta have been suggested in recent studies (Kim et al., 2019; Mander et al., 2015), it will be important for future studies to investigate SO and delta oscillations separately to precisely identify abnormalities in AD and MCI. Furthermore, investigating detailed properties of SOs in AD and MCI in addition to power, such as density, amplitude, and duration by detecting each SO event might further reveal distinct features of these two oscillatory phenomena and inform neural circuitry models (Fig. 1C).

2.1.3. Slow wave activity in mouse models of AD (Table 1)

Studies have shown altered SWA in AD mouse models although the changes vary among the different mouse models and stages of the disease (Table 1). Reduced NREM slow wave power (ranging from 0.1 to 5 Hz) was observed in APP/PS1 mice expressing amyloid precursor protein (hAPP) and presenilin 1 (hPSEN1), in frequency band of 0.1–4 Hz at 8–10 months (Kent et al., 2018); Tg2576 mice overexpressing APP with APPK670/671 L mutation, in frequency band of 0.5–4 Hz at 6–12 months (Zhang et al., 2005); rTG4510 mice overexpressing human P301L tau, in frequency band of 0.1–4 Hz at > 6 months (Holton et al., 2020); P301S mice (PS19) containing a human tau transgene with a P301S mutation, in delta band (1–4 Hz) at 11 months (Holth et al., 2017); forebrain specific human mutated Tau (hTauP301L + R406W) knock-in mice PLB2_Tau in frequency band of 1.5–5 Hz at 6 months (Koss et al., 2016); and App^NL-G-F homozygous mice containing a mutated human version of App singly knocked into the original App locus, in delta band (1–4 Hz) at 6 and 12 months (Maezono et al., 2020). In 5XFAD mice expressing hAPP and hPSEN1 with a total of five mutations, the power of the 0.5–4 Hz frequency band was decreased at 6 months compared to 3 months of age or when compared to age-matched non-transgenic mice (Schneider et al., 2014). On the other hand, the power of the low SO frequency range (0.1–0.5 Hz) was increased in the 6 month-old 5XFAD mice compared to 3 months or when compared to age-matched control mice (Schneider et al., 2014). In one EEG study where three different AD mouse models were compared, the 3xTg-AD mice, a model which exhibits both Aβ and tau pathology, showed no change in NREM or Wake SWA (0.1–4 Hz) power at 18 months while APP/PS1 and Tg2576 mice showed reduced NREM or Wake SWA power, respectively (Kent et al., 2018). This difference among the mouse models could be attributed to the fact that the level of Aβ and tau pathology in the 3xTg-AD mice is milder compared to the APP/PS1 and Tg2576 mice. Notably, another study which performed local field potential recordings on the anesthetized 3xTg-AD mice showed slower SO (< 1 Hz) frequency and more irregular SO cycle at 20 months of age (Castano-Prat et al., 2019). Thus, detection of changes in SWA could also depend on the recording methods, states of animals (wake/sleep/anesthetized), and SWA parameters which are examined.

Intriguingly, some studies in AD mouse models reported that the direction of SWA alterations changes over the course of the disease. For example, the same study where 3xTg-AD mice showed slower SO frequency at 20 months of age compared to the age-matched control also demonstrated faster SO frequency at 7 months compared to the control (Castano-Prat et al., 2019). This finding is accompanied by shorter UP state duration at 7 months and longer UP and DOWN state duration at 20 months of age. Another study in P301S tau model exhibited increased delta (1–4 Hz) power at an early disease stage (6–9 months), then, showed decreased delta power at a later stage (11 months), which corresponds to decreased cortical volume (Holth et al., 2017). Increased delta power at an early stage could be due to a compensatory response to tau aggregation before cortical atrophy occurs. Longitudinal studies in AD mouse models investigating change in SWA along with pathological change will be invaluable for revealing mechanisms of stage-dependent alteration of SWA.

Analysis of SWA in many of the studies listed above included both SO and delta frequency range, however some recent studies demonstrated that SWA changes in the AD mouse models are frequency-specific. A study by Kastanenka and colleagues showed that the power of SO (< 1 Hz) was altered in APP/PS1 mice, and optogenetic excitation of cortical pyramidal neurons at SO frequency (0.6 Hz) restored the SO power and halted amyloid deposition (Kastanenka et al., 2017). On the contrary, when the authors doubled the optogenetic stimulation frequency to 1.2 Hz, it resulted in increased Aβ production, neuronal calcium overload, and decreased spine density (Kastanenka et al., 2019). Collectively, these findings suggest that SWA affected in AD mice is disease stage-dependent and frequency-specific, which indicates importance of identifying a proper timing and stimulation parameters including frequency as well as targeted brain regions to develop effective treatment. Furthermore, it will also be beneficial to assess impacts of this type of SO enhancement on other AD-related impairments, including glymphatic waste clearance, which has been shown to typically take place during slow wave sleep (Iliff et al., 2012; Xie et al., 2013).

2.1.4. Summary of slow wave changes in human and animal studies of AD

SWA abnormalities in NREM sleep are common in AD. When studies separate out SO and delta oscillations it is found that reductions in SOs (0.1–1 Hz) are consistently reported in studies of patients and in AD mouse models whereas changes in delta (1–4 Hz) activity are more variable. Although slow wave frequency band selected to compute power varies among studies, both amyloid and tau-based mouse models exhibit alteration in power in the SWA frequency band covering both SO and delta range. As mentioned above, the direction of these alterations could be dependent on the pathological stage of the disease. Thus, it is critical to identify the time-course of alterations. Both human and animal studies suggest that decreases in SO low frequency range power are more relevant to memory consolidation and Aβ burden and therefore, a potentially more important target of treatment (Kastanenka et al., 2019, 2017; Mander et al., 2015). However, to date, the number of studies that specifically investigated or targeted the SO range (< 1 Hz) in AD mouse models or in AD patients is relatively few. Future studies investigating SO in both amyloid and tau mouse models will be invaluable to examine SOs as reliable biomarkers as well as a target for intervention. Interestingly, in healthy older humans, artificially boosting SO activity by applying transcranial current stimulation at a slow frequency (0.75 Hz) during sleep improved declarative memory (Westerberg et al., 2015). Thus, it would be of interest whether similar non-invasive stimulation studies targeting SO might be beneficial in AD patients, as seen in animal models (Kastanenka et al., 2017). It is important to note that there is a tight connection between sleep and epileptiform activity which is frequently observed in patients with AD and AD mouse models at earlier stages of disease (Brown et al., 2018; Hanke et al., 2022; Miranda and Brucki, 2014). Epileptiform activity is more prevalent during sleep, especially in slow wave sleep, than wakefulness (Bazil, 2000; Horvath et al., 2018; Vossel et al., 2016). Thus, increasing the amount of slow wave sleep could potentially come with a risk to induce more epileptiform activity and negatively affect cognition. Development of a balanced therapeutic regimen to improve SO/slow wave sleep without increasing epileptiform activity would be crucial.

2.2. Sleep spindles

2.2.1. Sleep spindles

Sleep spindles are brief (~1 s) waxing and waning oscillations (typically, 9–16 Hz in humans and 10–15 Hz in rodents; Fig. 1), most abundantly observed in cortical EEG during light NREM sleep (Manoach and Stickgold, 2019). The thalamic reticular nucleus (TRN), more specifically, GABAergic neurons in TRN which contain the calcium binding protein PV, are the generators of sleep spindles (Buzsaki et al., 1988; Clemente-Perez et al., 2017; Halassa et al., 2011; Kim et al., 2012; Steriade et al., 1985; Thankachan et al., 2019). Sleep spindles are implicated in protecting sleep by filtering out sensory input (Dang-Vu et al., 2010; Yamadori, 1971), as well as in promoting memory formation (Luthi, 2014). Over the last decade, an increasing number of studies provided evidence on the involvement of sleep spindles in memory processes (Diekelmann and Born, 2010; Manoach and Stickgold, 2019; Ngo et al., 2013; Staresina et al., 2015). Many neuropsychiatric and neurodegenerative conditions including schizophrenia (Astori and Luthi, 2013; Davies et al., 2017; Keshavan et al., 2011; Manoach et al., 2016), autism (Limoges et al., 2005), dementia (Latreille et al., 2015; Montplaisir et al., 1995) and stroke (Hermann et al., 2008) demonstrate both sleep spindle abnormalities and cognitive deficits. The role of sleep spindles in sleep-dependent memory consolidation has been investigated in humans and animal models. Human studies showed correlations between the spindle activity and memory performances in both declarative memory (Baran et al., 2018; Cairney et al., 2018; Goder et al., 2015; Marshall et al., 2011; Schabus et al., 2004) and non-declarative/procedural memory tasks (Demanuele et al., 2017; Fogel and Smith, 2011; Walker et al., 2002; Wamsley et al., 2012). In animal studies, it has been reported that there was an increase in sleep spindle density or coordination between spindles, SOs, and hippocampal ripples following learning (Fogel et al., 2010; Latchoumane et al., 2017; Maingret et al., 2016; Molle et al., 2009; Siapas and Wilson, 1998; Sirota et al., 2003), although not all studies support this association. Moreover, normal aging negatively affects sleep spindle properties including density, duration, amplitude, and temporal coordination of SOs and spindles (Crowley et al., 2002; Helfrich et al., 2018). Of note, growing evidence points to the importance of precise coordination between spindles and other sleep oscillations such as slow wave activity and sharp-wave ripples in the memory processing and their disturbances in AD, which will be reviewed separately in a later section (Section 5).

2.2.2. Sleep spindles in AD patients

Changes in sleep spindles occur with normal aging (Nicolas et al., 2001). This includes decreases in spindle counts, density, duration, amplitude, and disrupted temporal coordination of SOs and spindles (Crowley et al., 2002; Helfrich et al., 2018). However, the degree of sleep spindle alteration in AD patients is greater than seen in healthy aging controls (Nicolas et al., 2001).

Spindle density is one of the key features evaluated in the sleep spindle analysis. It is typically calculated as the number of spindles divided by NREM sleep minutes (Uygun et al., 2019). Decreased spindle density was observed in AD patients compared to the healthy controls (Kam et al., 2019). Additionally, a longitudinal study in Parkinson’s disease patients showed that the patients with lower spindle density were more likely to develop dementia later compared to the healthy controls or the patients with less altered spindle density (Latreille et al., 2015). Negative correlations between the spindle density and the AD biomarkers, including the levels of Aβ42, P-tau, and T-tau in CSF, were also reported (Kam et al., 2019). Studies also revealed that spindle density correlated with cognitive performance, shown as lower spindle densities with decreased Mini-Mental Status Examination (MMSE) and Montreal Cognitive Assessment (MoCA) scores (Gorgoni et al., 2016; Liu et al., 2020) and impaired cognitive task performance (Kam et al., 2019).

Spindle duration, which is shown to positively correlate with learning in healthy subjects in both procedural (Fogel et al., 2007; Morin et al., 2008) and declarative memory (Schabus et al., 2004), is shorter in AD as well as in MCI patients (Liu et al., 2020). The shorter spindle duration also correlates with decreased MMSE and MoCA scores (Liu et al., 2020).

A larger spindle amplitude, typically observed at the UP phase of SO, was positively associated with declarative memory consolidation (Ladenbauer et al., 2017). The greater spindle amplitude was found to coincide with higher activity in the hippocampus (Bergmann et al., 2012), which supports the notion that coordination between cortical spindles and hippocampal ripples is a key factor for the reactivation of memory traces during sleep. In AD patients, spindle amplitude was reduced (Liu et al., 2020; Taillard et al., 2019). Moreover, it was reported that patients with early stages of dementia who exhibited smaller spindle amplitude tend to develop greater cognitive decline later (Taillard et al., 2019). Thus, spindle amplitude could be a potential early biomarker to predict risk of cognitive decline in older adults. However, it was also suggested that spindle density and duration could be used as more reliable biomarkers of dementia compared to the amplitude (Latreille et al., 2015; Liu et al., 2020). Of note, correlations with T-tau levels in CSF which were found with spindle density and duration were not observed with spindle amplitude (Kam et al., 2019).

Sleep spindles are often classified into two groups based on their frequency ranges: fast (12–16 Hz) and slow (9–12 Hz) spindles (de Gennaro and Ferrara, 2003; Molle et al., 2011; Schabus et al., 2007). EEG recordings showed that slow spindles are predominantly found in the frontal areas while fast spindles are observed more in centro-parietal areas (Werth et al., 1997; Zeitlhofer et al., 1997). Although the underlying mechanisms and functional distinctions of the two groups of spindles are yet to be understood, studies in AD and MCI found decrease in centroparietal fast spindles but not in frontal slow spindles (Gorgoni et al., 2016; Rauchs et al., 2008; Westerberg et al., 2012). Moreover, it was reported that fast spindles produced larger responses in the hippocampus and medial prefrontal cortex than slow spindles in healthy humans (Anderer et al., 2001; Schabus et al., 2007), which are known to be critical for memory processing and the highly affected brain regions in AD (Park and Reuter-Lorenz, 2009; Sole-Padulles et al., 2009). Collectively, these findings suggest that fast spindles rather than slow spindles could be more sensitive to AD pathology.

2.2.3. Sleep spindles in mouse models of AD (Table 2)

Table 2.

Studies of sleep spindles in AD mouse models.

Mouse Model Name	Amyloid/Tau Model	Frequency Band (Hz)	Age (months)	Findings	Reference
APP/PS1	Amyloid	10–18	5–6	↓ hippocampal ripple-centered mPFC spindle power	Zhurakovskaya et al. (2019)
3xTg-AD	Amyloid/Tau	8–18	6–7	↓ spindle density	Benthem et al. (2020)

Compared to the human AD studies, there are relatively few studies of sleep spindles in AD mouse models (Table 2), potentially because detection of sleep spindles is more challenging in mouse EEG or LFP data (Uygun et al., 2019). Some studies investigated the sleep spindles in AD mouse models by analyzing the spindle spectral (sigma) power. For example, a study in APP/PS1 mice showed that the hippocampal ripple-centered spindle power was decreased in the transgenic mice compared to the wild-type mice (Zhurakovskaya et al., 2019). In contrast, there was no significant difference in the spindle power off-centered from the hippocampal ripple between the transgenic and wild-type mice. In a recent study in 3xTg-AD mice, which express mutant APP, presenilin and Tau and exhibit similar plaque and tangle pathology to the human AD patients, detection of spindle events was performed as in human studies and decreased spindle density was demonstrated in 3xTg-AD mice at 6–7 months of age compared to the wild-type mice (Benthem et al., 2020). Interestingly, the spindle count did not differ across genotype due to the increased NREM time found in 3xTg-AD mice. The same study also investigated hippocampal cortical coupling in this mouse model which will be reviewed later (Section 5).

2.2.4. Summary of sleep spindle changes in patients and AD mouse models

Current findings in AD mouse models indicate the sleep spindle alterations are consistent with those found in the human AD patients i.e. spindle densities and other spindle features tend to be reduced. However, analysis of spindles in the mouse models tends to be less detailed compared to the human studies. Furthermore, current rodent studies have not identified distinct slow and fast spindle events as seen in human studies. In addition to looking at the spectral power of spindle ranges, detecting spindle events and analyzing spindle properties as in human studies would be helpful to further identify the mechanisms and relationship between sleep spindles and AD progression in the different mouse models.

2.3. Hippocampal sharp-wave ripples (100–250 Hz)

2.3.1. Sharp-wave ripples

Hippocampal sharp-wave ripples (SWRs) are large field potentials (sharp waves) generated by synchronous activation of CA3 pyramidal inputs to CA1 and accompanied by high frequency (typically 100–250 Hz) oscillatory activities (ripples) generated by local interneurons (Skelin et al., 2019; Tang and Jadhav, 2018). SWRs can occur in both waking and sleeping states (Carr et al., 2011; O’Neill et al., 2010). SWRs especially observed during slow wave sleep are implicated in memory consolidation, commonly described with the two-stage model of spatial memory encoding (Buzsaki, 1989; Frankland and Bontempi, 2005; Squire and Alvarez, 1995) where newly acquired information is transferred from the initial storage location, hippocampus, to the more permanent storage space, the neocortex, via separate processes occurring in NREM and REM sleep. Studies demonstrated that memory reactivation/replay occurs in the hippocampus and various cortical areas during SWRs (Buzsaki, 2015; Tang and Jadhav, 2018). Precise coordination among hippocampal SWRs, cortical spindles, and slow wave activity has been shown to promote communication between the hippocampus and cortex (Siapas and Wilson, 1998; Staresina et al., 2015) which is crucial for proper memory formation, and disruption of this coordination could impair the memory (Ego-Stengel and Wilson, 2010; Girardeau et al., 2009). Although the exact neural mechanisms and networks involved in SWRs are still to be clarified, numerous studies support involvement of SWRs in memory formation. For instance, a recent study in awake behaving rats demonstrated that optogenetically prolonged ripples enhanced memory during a maze learning (Fernandez-Ruiz et al., 1979). Thus, there is no surprise that changes in SWRs are considered as one of the key phenomena and potential functional biomarkers in the disorders with memory impairment including AD. In the following sections, we will review how SWRs are altered in AD.

2.3.2. Sharp wave-ripples in humans

Investigation of SWRs in humans is challenging because it requires a means to detect signals in deep brain structures. Thus, most of the SWR data in humans have been obtained from hippocampal electrodes implanted for identifying epileptogenic tissue in epileptic patients (Bragin et al., 1999; Helfrich et al., 2019). It is crucial to be able to detect SWRs reliably and non-invasively in order to assess SWRs as a functional biomarker of AD and other dementias. Fortunately, recent technological advancement started opening the door to promising non-invasive methods. It has been demonstrated that magnetoencephalography (MEG) can detect oscillations at ripple frequency range in the hippocampus (Liu et al., 2019) and in deep brain structures (Yin et al., 2019). Other studies also introduced the possibility of detecting SWRs indirectly using the scalp EEG. For example, high-density scalp EEG was able to detect signals that correlated with the SWRs recorded by electrodes implanted in deep brain structures (Seeber et al., 2019). Another study demonstrated that SWRs detected from the temporal association cortices was highly associated with SWR signals detected in the medial temporal lobe, thus cortically detected SWRs signal could act as a proxy for hippocampal SWRs (Vaz et al., 2019). To the best of our knowledge, there is no published data investigating SWRs in AD patients. Hopefully, current and future advancement of technologies will make it possible to closely investigate associations between SWRs and AD progression non-invasively in humans, and to develop another potential preclinical assessment of AD risk.

2.3.3. Sharp wave-ripples in mouse models of AD (Table 3)

Table 3.

Studies of sharp-wave ripples (SWRs) in AD mouse models.

Mouse Model Name	Amyloid/Tau Model	Frequency Band (Hz)	Age (months)	Findings	Reference
APP/PS1	Amyloid	100–250	5–6	↓ number, oscillation freq., power of SWRs	Jura et al. (2019)
APP/PS1 (in vitro)	Amyloid	120–400	3	↓ number, amplitude of SWRs ↓ slow gamma power during SWRs	Hollnagel et al. (2019)
5XFAD	Amyloid	150–250	3	↓ slow gamma power during SWRs	Iaccarino et al. (2016)
5XFAD (in vitro)	Amyloid	120–220	1, 3	↑ number, amplitude of SWRs at 3 mo ↑ slow gamma power during SWRs	Caccavano et al. (2020)
C57BL/6 J mice injected with soluble a Aβ oligomers solution	Amyloid	100–250	3–4	↓ learning induced SWRs	Nicole et al. (2016)
hAAP/Nptx2−/− (in vitro)	Amyloid	150–250	3–4	↑ number of SWRs No change in slow gamma power during SWRs	Xiao et al. (2017)
rTg4510	Tau	100–250	2–4, 7–9	↓ learning induced SWRs at 2–4 & 7–9 mo, ↓ number, amplitude of SWRs at 7–8 mo	Ciupek et al. (2015); Witton et al. (2016)
apoE4-KI		150–250	12–18	↓ number of SWRs ↓ slow gamma power during SWRs	Gillespie et al. (2016);Jones et al. (2019)

While it is challenging to investigate SWRs in humans, SWRs have been well studied in animal models which allow the investigators to record neural activities directly from the hippocampus and deep brain structures using implanted electrodes. Deficits in SWRs were observed in a variety of AD mouse models (Table 3). With in vivo studies, fewer SWRs were found in APP/PS1 mice that overexpress human genes for amyloid precursor protein (APP) and presenilin1 (PS1) (Jura et al., 2019), and apoE4-KI mice that have the human AD genetic risk factor, Apolipoprotein (apo) E4, knocked in (Gillespie et al., 2016; Jones et al., 2019). Reduction in learning induced SWRs was also found in a mouse model injected with soluble amyloid beta oligomers solution (Nicole et al., 2016), as well as in both young and old rTg4510 mice overexpressing tau in the hippocampus (Booth et al., 2016; Ciupek et al., 2015; Witton et al., 2016). Along with learning-induced SWRs, slow gamma (20–50 Hz) power is elevated during SWRs in wild type animals (Carr et al., 2012). In AD mouse models including 5XFAD and apoE4-KI mice, the slow gamma power during SWRs was reduced in AD mice compared to the wild type mice (Gillespie et al., 2016; Iaccarino et al., 2016).

Contrary to these in vivo findings of SWR deficits, in vitro recordings showed less consistent SWR results in different AD mouse models (Sanchez-Aguilera and Quintanilla, 2021). For instance, the number of SWRs were increased in 5XFAD mice (Caccavano et al., 2020) and in hAAP/Nptx2−/− mice, an amyloid pathology mouse model (Xiao et al., 2017), while the abundance of SWRs was decreased in APP/PS1 mice (Hollnagel et al., 2019). In vitro findings regarding the power of slow gamma during SWRs were not always consistent with those of in vivo recordings. In vitro recordings in 5XFAD mice showed increased slow gamma power (Caccavano et al., 2020) whereas slow gamma power was reduced in APP/PS1 mice (Hollnagel et al., 2019) and no change compared to controls was found in hAAP/Nptx2−/− mice (Xiao et al., 2017). These inconsistent and contradictory findings among in vivo and in vitro experiments could stem from alterations in inputs to the hippocampus which are severed in in vitro recordings. Other differences in in vitro recording conditions, such as the angle which the slices are cut, differences in the oxygenation levels or the composition of the extracellular medium could affect SWRs recorded in slices (Aivar et al., 2014; Hajos and Mody, 2009; Sanchez-Aguilera and Quintanilla, 2021). Although the in vitro setup does not always match with in vivo environment, in vitro studies provide insights on how specific cell types are behaving in the microcircuit. The involvement of specific cell types such as PV-positive interneurons in SWRs and other sleep oscillations will be discussed more in a later section (Section 4).

2.4. NREM oscillatory coupling in AD

2.4.1. NREM oscillatory coupling

In addition to each sleep-related oscillation reviewed above, it has been widely reported that temporal associations of these oscillatory events, in particular among SOs, sleep spindles, and hippocampal SWRs, play a major role in memory consolidation (Fig. 1D) (Bergmann and Born, 2018; Cairney et al., 2018; Diekelmann and Born, 2010; Genzel et al., 2014; Helfrich et al., 2018; Latchoumane et al., 2017; Maingret et al., 2016; Miyamoto et al., 2017; Navarro-Lobato and Genzel, 2019; Ngo et al., 2013; Oyanedel et al., 2020; Peyrache et al., 2009; Staresina et al., 2015). This “triple-coupling” of oscillations can be observed during slow wave activities in which the hippocampal SWRs are nested in the troughs of the spindles and the spindles themselves are nested in the UP states of SOs (Klinzing et al., 2019; Oyanedel et al., 2020; Rasch and Born, 2013; Skelin et al., 2019; Watson and Buzsaki, 2015). A recent study in rats investigating the directionality of this interaction demonstrated that the cortical SOs drive thalamic spindles which subsequently regulate SWR occurrence (Oyanedel et al., 2020). The study also suggested that the hippocampal SWRs contribute to the emergence of cortical SOs, indicating the loop-like interaction among thalamus, cortex and hippocampus on the NREM oscillatory regulation. This precise temporal coordination among brain regions during sleep has been suggested to promote the two-stage model of memory encoding where memory information is transmitted from the hippocampus to the cortex for more long-term storage (Buzsaki, 1989; Clemens et al., 2007; Latchoumane et al., 2017; Oyanedel et al., 2020; Siapas and Wilson, 1998; Sirota et al., 2003; Staresina et al., 2015). Indeed, augmenting SO-spindle coupling with closed-loop transcranial alternating current stimulation (tACS) during NREM in young adults enhanced sleep-dependent memory consolidation (Ketz et al., 2018). There are some recent studies in humans and animal models investigating how temporal associations of sleep oscillatory events are affected in AD or MCI in humans and animal models, which we review next.

2.4.2. NREM oscillatory coupling in humans with AD

In humans, SO-spindle coupling was impaired in older adults and linked to overnight forgetting (Helfrich et al., 2018). This disrupted SO-spindle coupling in older adults also correlated with atrophy in the medial frontal cortex. A recent positron emission tomography (PET) study in older adults without cognitive deficits showed that weaker SO-spindle coupling during NREM was associated with greater accumulation of tau in the medial temporal lobe (Winer et al., 2019). Interestingly, this correlation was unique to tau and not found with Aβ burden, while the decreased SO activity predicts Aβ burden but not tau burden. Thus, these results suggest a dissociable relationship of tau and Aβ, two major neuropathological features of AD, to the sleep oscillatory change. Another recent study in MCI patients showed that enhancement of the temporal coupling between SO and spindles could be a potential intervention to improve memory (Ladenbauer et al., 2017). The study demonstrated slow oscillatory transcranial direct current stimulation (so-tDCS) enhanced temporal coupling of SO and spindles and improved visual declarative memory consolidation in MCI patients. Currently, there is no study investigating the triple-coupling of SOs, spindles, and SWR in AD patients. Further characterization of oscillatory coupling during NREM in AD and MCI patients would help establish these changes as predictors for AD and potential targets for preventative treatment.

2.4.3. NREM oscillatory coupling and AD in mouse models (Table 4)

Table 4.

Studies of NREM oscillatory coupling in AD mouse models.

Mouse Model Name	Amyloid/Tau Model	Frequency Band (Hz)	Age (months)	Findings	Reference
APP/PS1	Amyloid	10–18 (spindles), 150–250 (SWRs)	5–6	↓ hippocampal ripple-centered mPFC spindle power	Zhurakovskaya et al. (2019)
3xTg-AD	Amyloid/Tau	0.5–4 (SWA), 8–18 (spindles), 75–300 (SWRs)	6	↓ SWR-SWA coupling Impaired phase coupling of spindle-SWA	Benthem et al. (2020)

Impaired NREM oscillatory coupling was also reported in recent studies in animal models of AD (Table 4). A study in APP/PS1 mice showed a significant reduction in coupling of peak spindle band power in the medial prefrontal cortex to the hippocampal ripples, whereas local connectivity within the hippocampus remained intact (Zhurakovskaya et al., 2019). Of note, the study was conducted in the APP/PS1 mice at an age where there is Aβ pathology in the cortex and hippocampus but no apparent memory impairment. Another recent study in 3xTg-AD mice expressing mutant APP, presenilin and Tau, demonstrated reduction in SWR-slow wave (0.5–4 Hz) coupling and impaired phase coupling of peak spindle power (8–18 Hz) to slow wave phase along with spatial learning deficit (Benthem et al., 2020). These sleep oscillatory deficits were observed at the early stage when extracellular Aβ and tau tangles were absent in the 3xTg-AD mice. Interestingly, the 3xTg-AD mice showed increased sleep time possibly to compensate for reduction in SWR density, however, this increase in sleep time was evidently not enough to augment deficits in cortico-hippocampal oscillatory coupling. To further understand the associations between changes in oscillatory coupling and AD progression, future longitudinal studies tracking these specific sleep changes would be helpful.

3. REM sleep

3.1. REM oscillations

REM sleep is characterized by EEG spectral features resembling the fast activities seen during the waking state, accompanied by tonic muscle atonia recorded in EMG signals which is interrupted by phasic bursts/twitches (Brown et al., 2012). Theta and alpha oscillatory activities are prominent features of REM sleep in humans (Landolt et al., 1996), with theta observed over frontocentral areas, and alpha over occipitoparietal areas (Scarpelli et al., 2019). In rodents, regular theta (4–8 Hz) activity dominates in EEG recordings during REM sleep, due to the location of the hippocampus close to the neocortex. In addition to theta and alpha, higher frequency activities, i.e. beta and gamma oscillations, are also detected during REM, often coupled with theta oscillations (Liscombe et al., 2002).

Although pharmacological and REM sleep deprivation studies have suggested a correlational link between REM sleep and spatial and emotional memory consolidation, direct contribution of REM related EEG activities in learning and memory remain controversial (Buzsaki, 2002; Datta et al., 2008; Diekelmann and Born, 2010; Klinzing et al., 2019). A more direct causal relationship between REM oscillations and memory consolidation was investigated in an optogenetic study in which inhibition of theta rhythm during REM sleep after learning by silencing of medial septum GABA neurons disrupted memory consolidation in mice (Boyce et al., 2016). Although further investigation is needed to identify the physiological function of REM sleep, changes in REM sleep have been characterized in patients with dementia (Mander, 2020). We will review currently known REM related changes in AD below.

3.2. REM oscillations in humans with AD

3.2.1. Overall REM sleep change in humans with AD

Changes in REM sleep are observed at early stages of cognitive impairment (Brayet et al., 2016; Liguori et al., 2016; Rodrigues Brazete et al., 2013) and reliably distinguish patients with AD versus those without AD (Hassainia et al., 1997). These REM changes include reduced time spent in REM (Liguori et al., 2016; Montplaisir et al., 1995; Prinz et al., 1982; Westerberg et al., 2012) and shift in EEG power (Brayet et al., 2016; D’Atri et al., 2021), which will be reviewed in more detail in following sections. The most robust REM EEG changes associated with MCI and AD were found on central and posterior regions (Brayet et al., 2016; Hassainia et al., 1997; Rodrigues Brazete et al., 2016, 2013). Notably, the change in frequency power during REM can predict not only AD but also onset of non-AD dementia and cognitive decline (Rodrigues Brazete et al., 2016). Thus, REM deficits could be a common characteristic of dementia rather than a specific feature of AD.

3.2.2. REM theta change in humans with AD

Loss of central-parietal theta activity during REM was reported in patients with amnestic MCI (Westerberg et al., 2012). The study also showed positive correlation between REM theta power and declarative memory performance in both MCI patients and age-matched controls (Westerberg et al., 2012). On the other hand, increased theta power was observed in AD or MCI patients during REM (Brayet et al., 2016; Petit et al., 1993) along with increased delta power and decreased alpha and beta power. This shift in EEG power is often computed as the ratio of low frequency (delta + theta) to high frequency (alpha + beta) power and described as the “EEG slowing” index which is shown to distinguish between amnestic from non-amnestic MCI patients (Brayet et al., 2016; D’Atri et al., 2021; Petit et al., 1993).

3.2.3. REM delta, alpha, and beta change in humans with AD

Consistent with studies described above, a recent study reported a change in EEG signals during REM sleep in AD and MCI patients compared to controls and linked the changes to brain structural alterations (D’Atri et al., 2021). This EEG change was characterized by the increase in delta power over frontotemporal areas and the decrease in beta power over the temporal regions. A subsequent MRI study in AD patients by the same authors revealed significant correlations between cortical thickness and this EEG alteration in REM sleep (D’Atri et al., 2021). In the study, increased delta power in the frontotemporal regions during REM sleep was strongly correlated with atrophy in temporal, parietal, and frontal cortices, the major brain regions known to be damaged in AD (Whitwell, 2010; Yang et al., 2019). The EEG study also revealed some distinctions between AD and MCI groups where only AD patients exhibited the decrease in alpha, sigma, and beta activities in the occipital areas (D’Atri et al., 2021). This could be attributed to the fact that the neurodegeneration occurs later in the occipital areas compared to the temporal areas (Braak and Braak, 1997). Thus, REM changes in temporal regions may identify cognitive decline as early as MCI phase, whereas changes in occipital areas could be more specific features for the advanced dementia including AD. Overall, the studies indicate that EEG changes during REM sleep strongly and consistently correlate with cognitive decline and dementia, which suggests that this EEG change could be used as a practical marker to identify or even predict cognitive decline before the AD symptoms become apparent.

3.3. REM oscillations in mouse models of AD (Table 5)

Table 5.

Studies of REM sleep in AD mouse models.

Mouse Model Name	Amyloid/Tau Model	Frequency Band (Hz)	Age (months)	Findings	Reference
rTG4510	Tau	N/A	5–10	↓ number of REM bouts at > 6 mo	Holton et al. (2020)
P301S (PS19)	Tau	4–8	3, 6, 9, 11	↓ amount of REM sleep at 9–11 mo ↑ REM theta power at 9 mo ↓ REM theta power at 11 mo	Holth et al. (2017)
J20	Amyloid	5–7	11–12	↓ amount of REM sleep	Filon et al. (2020)
Tg2576	Amyloid	13–20 20–30	12	↑ REM beta power ↑ REM gamma power	Kent et al. (2018)

3.3.1. Overall REM sleep change in mouse models of AD

REM sleep has been studied in different AD related mouse models (Table 5). A recent study in a mouse model of tauopathy overexpressing human P301L tau (rTg4510 mice) performed longitudinal EEG recordings from the age of ~20 weeks to ~44 weeks (Holton et al., 2020). The study found that the number of REM bouts and the bout duration were both decreased in the transgenic mice over time. Another study in the P301S human tau transgenic mice showed a decreased amount of REM sleep (Holth et al., 2017). The decrease in REM was correlated with the increase in tau pathology in brainstem. A study in the J20 mouse model that expresses human amyloid protein precursor (hAPP) mutations (Palop et al., 2007) exhibited decreased REM sleep time (Filon et al., 2020). Interestingly, treatment with a mGluR5 inhibitor, which has been shown to reduce Aβ levels and prevent cognitive impairment in mouse models of AD (Hamilton et al., 2016; Westmark et al., 2018), reversed the REM sleep deficit although Aβ levels did not significantly decrease with mGluR5 treatment in this study.

3.3.2. REM EEG power changes in mouse models of AD

A study in the P301S human tau transgenic mice showed reduced theta power during REM (Holth et al., 2017). Moreover, it was shown that decreased REM EEG power was associated with reduced cortical volume in these mice. Another study in Tg2576 mice reported increased frontal EEG power in the beta (13–20 Hz) and low gamma (20–30 Hz) frequency bands during REM compared to the wildtype mice, although there was no significant difference in time spent in REM between groups (Kent et al., 2018). Overall, both tau and amyloid overexpressing mouse models exhibited some REM sleep deficits (Table 5). However, the REM sleep properties that showed deficits varied, and their correlations with pathology or behavioral impairment were not as consistent as human studies.

3.4. Ponto-geniculo-occipital (PGO) waves during REM

Another hallmark of REM related activity is known as pontogeniculo-occipital (PGO) waves, a field potential propagated from the pons to the lateral geniculate nucleus, and to the occipital cortex (Brown et al., 2012). Single, high-amplitude PGO waves can be observed right before the onset of REM sleep, which then reappear in bursts of lower amplitude waves during REM (Bizzi and Brooks, 1963; Brooks and Bizzi, 1963; Datta, 1997; Jouvet et al., 1965). The pontine component of PGO waves (P-waves) has been widely studied in cats and rats (Datta, 1997), but is not well-characterized in humans due to the location deep in the brainstem (Diekelmann and Born, 2010). A study investigated the presence of PGO waves in human primary visual cortex suggested that sharply contoured theta waves could be a potential human correlate of PGO waves (Frauscher et al., 2018).

To the best of our knowledge, there is no study investigating PGO waves during REM in AD patients or AD animal models. Future research on PGO waves in animal models and/or human correlates of PGO waves could provide a fuller picture of REM oscillatory changes in AD and dementia.

4. Alterations in GABAergic neurons and sleep changes in AD

As reviewed in the earlier sections, detecting abnormalities in different types of sleep oscillations could be utilized as early-stage biomarkers to predict AD or MCI, as well as to develop intervention strategies to prevent or delay AD progression. Specifically, if sleep oscillatory changes can be tied to changes in specific groups of neurons or brain circuits then therapies may be developed which target those neurons/circuits. Neuronal network dysfunction is one of the early characteristics of AD observed in people at risk of developing AD (Palop and Mucke, 2010a, 2010b; Styr and Slutsky, 2018). Basic science work described in the previous sections has closely linked GABAergic neurons in the cortex, hippocampus and thalamus to the generation or regulation of specific sleep oscillations and emerging evidence described in this section has revealed alterations in the activity of these GABAergic neurons in AD patients or in mouse models. GABAergic neurons in the basal forebrain also play an important role in controlling REM theta and gamma oscillations (Brown et al., 2012; Kim et al., 2015) and NREM sleep spindles (Thankachan et al., 2019). These GABAergic neurons closely interact with neighboring cholinergic neurons (Yang et al., 2017, 2014) which degenerate in AD (Mesulam, 2012), but they have not yet been closely studied in AD mouse models so are not considered further here.

A balanced interaction between excitatory pyramidal neurons and inhibitory interneurons is crucial for normal cognitive functions. Specifically, GABA released from the inhibitory interneurons has been shown to play a major role in mediating neural excitation/inhibition balance (Booker and Vida, 2018; Isaacson and Scanziani, 2011). Approximately 10–20% of cortical neurons are GABAergic interneurons (Hu et al., 2014; Meyer et al., 2011; Moore and Wehr, 2013; Xu et al., 2010), which can be further divided into several subtypes, including PV interneurons and SOM interneurons (Defelipe et al., 2013). Of those, PV interneurons account for 40–50% of cortical GABAergic interneurons and SOM interneurons 20–30% (Wonders and Anderson, 2006). Although much attention was directed to the effect of Aβ on excitatory neurons in early studies, an increasing number of reports in recent years suggest involvement of GABAergic interneurons in AD progression (Hazra et al., 2016; Jagirdar et al., 2021; Palop and Mucke, 2016, 2010a, 2010b; Styr and Slutsky, 2018). Thus, in the following sections we will review recent evidence that links these GABAergic neurons to pathological changes in AD and alterations of EEG oscillations during sleep (Fig. 2).

Fig. 2.

A summary of potential causal interactions among changes in GABAergic activities, sleep oscillations, and AD progression. Alterations in GABAergic interneurons in the cortex, hippocampus, and thalamus could lead to disrupted excitatory/inhibitory balance in the neuronal network, which in turn causes abnormalities in variety of sleep oscillations critical for cognitive performance. The impaired sleep oscillations could also impede waste clearance from the brain such as Aβ and tau which further disrupts GABAergic activity or causes cell death. Decreases in normal network activities as well as reduced number of cells will exacerbate cognitive impairment in AD. It is important to note that many of the components in this chart could interact bidirectionally, such as bidirectional causal relationships between Aβ/tau level and altered GABAergic activities, or Aβ/tau level and sleep abnormalities. PV: parvalbumin interneurons, SOM: somatostatin interneurons, SWA: slow wave activities, SWR: sharp-wave ripples.

4.1. GABAergic neurons in cortex and hippocampus in AD

Alterations in GABA levels and GABAergic interneurons have been reported in both AD patients and animal models of AD (Xu et al., 2020). These changes have been typically observed as a decrease in the number of GABAergic neurons as well as changes in their activity (Xu et al., 2020). Abnormal activity of fast-spiking GABAergic interneurons which contain the calcium binding protein PV, has been suggested to directly contribute to AD-related cortical network dysfunction leading to progressive AD pathology (Palop and Mucke, 2016, 2010a, 2010b; Styr and Slutsky, 2018; Xu et al., 2020). Reduced PV neurons in cortical regions and hippocampus were reported in AD patients (Brady and Mufson, 1997; Mikkonen et al., 1999; Sanchez-Mejias et al., 2020; Solodkin et al., 1996). In animal models, decreases in the number of cortical and hippocampal PV neurons were reported in 5XFAD mice (Flanigan et al., 2014; Giesers and Wirths, 2020), APP/PS1 mice (Cheng et al., 2020; Saiz-Sanchez et al., 2012; Takahashi et al., 2010), 3xTg mice (Zallo et al., 2018), and Tg2576 mice (Huh et al., 2016). In addition to PV neurons, reduction in the number of SOM interneurons in cortex and hippocampus has also been reported in AD patients (Beal et al., 1986, 1985; Candy et al., 1985; Davies et al., 1980; Davies and Terry, 1981; Sanchez-Mejias et al., 2020) and AD mouse models, including APP/PS1 mice (Ramos et al., 2006; Saiz-Sanchez et al., 2012; Sanchez-Mejias et al., 2020) and TgCRND8 mice (Ma and McLaurin, 2014).

Changes in the activity of cortical and hippocampal PV neurons varies among the models and according to the ages of the animals. In vitro work in APP/PS1 mice has shown that hippocampal PV neurons are hyperexcitable at an early stage (3–4 month) in disease progression, prior to observable changes in pyramidal neuronal excitability, and that this is linked to abnormal neuronal network activity and memory impairment (Hijazi et al., 2019). Contrary to this finding, another in vitro study in 5XFAD mice reported reduced activity of hippocampal PV basket cells especially during SWRs in early (~3 month) amyloid pathology (Caccavano et al., 2020). Importantly, these findings have not been confirmed in vivo and the reason for different results in different models is unknown. Hijazi et al. reported that reactivation of PV neurons in 6 month-old APP/PS1 mice had similar benefit on cognition as inhibition of PV neurons in 4 month-old mice. Thus, one possibility is that abnormal activity of hippocampal PV neurons occurs in a biphasic manner, that is, hyperexcitability precedes hypoexcitability (Hijazi et al., 2019). In another study, it was reported that disrupted activity of PV neuron caused by reduced Nav_1.1 level results in abnormalities in hippocampal SWRs and impaired memory in AD animals (Verret et al., 2012).

It remains unclear what causes early-stage PV neuron hyperexcitability. One potential cause is the deposition of extracellular Aβ. Correlation between loss of PV interneurons and increased level of colocalized Aβ was observed in AD mouse models (Saiz-Sanchez et al., 2012; Takahashi et al., 2010). Whether Aβ plaques are a direct cause of changes in PV excitability or viability is still unknown (Carter and Lippa, 2001). Regardless, there is an interaction (possibly bidirectional) between Aβ pathology and PV neuronal dysfunction, which in turn disrupts proper regulation of pyramidal neurons in AD. Recent studies show that optogenetic stimulation of hippocampal PV neurons at 40 Hz or presenting 40 Hz auditory/visual stimulation reduces Aβ levels in AD mice likely by recruiting glial responses (Iaccarino et al., 2016; Martorell et al., 2019), which seems to contradict with the finding that suppressing hyperexcitability of PV interneurons at an early stage reduces Aβ levels (Hijazi et al., 2019). However, these apparently contradictory findings may result from different effects on network function of phasic synchronized (40 Hz) activation versus tonic activity.

Cortical neurons may adapt dynamically during disease progression. Thus, neuronal circuits may compensate for an initial insult by adjusting the neuronal properties (e.g., hyperexcitability) early in the disease, which could temporarily restore the network balance, but also results in greater pathological changes, which ends up facilitating progression of disease. Thus, either maintaining excitatory/inhibitory balance before the drastic pathophysiological change occurs, or acutely restoring impaired balance under the already altered network may work as potential treatments. While an involvement of PV interneurons appears clear, the exact changes which occur in these neurons, may depend on the cortical/hippocampal region, the stage of the disease and the mouse model. Thus, longitudinal in vivo experiments to monitor their activity are essential for understanding their role in disease progression and the timing of therapeutic interventions.

4.2. GABAergic neurons in the thalamic reticular nucleus in AD

The thalamus is another key player for regulation of sleep related oscillations (Brown et al., 2012). An MRI study in AD patients showed decreased volume in thalamus which significantly correlates with a poorer cognitive performance (de Jong et al., 2008). TRN, a narrow net of GABAergic neurons enveloping the anterior and lateral parts of the thalamus, is a key thalamic structure in generation of spindles (Manoach and Stickgold, 2019; Thankachan et al., 2019), regulating delta oscillations (Uygun et al., 2022) and coordinating sleep oscillations (Crunelli et al., 2018) as well for maintaining consolidated NREM sleep (Kim et al., 2012). A recent study showed reduced activity of TRN neurons in heterozygous transgenic mice expressing human APP carrying Swedish (K670N, M671L) and Indiana (V717F) familial AD mutations (Line J20). This reduced activity could be reversed by chemogenetic activation, which in turn reduced sleep fragmentation, increased time spent in slow wave sleep, and reduced Aβ accumulation (Jagirdar et al., 2021). Furthermore, immunohistochemical evidence for reduced TRN activity was observed in postmortem brains of AD patients compared to controls, with intermediate levels in MCI patients (Jagirdar et al., 2021). Thus, both cortical and hippocampal PV interneurons and TRN PV neurons are potentially involved in pathological changes and sleep oscillatory abnormalities.

4.3. Potential link between altered GABAergic neurons and sleep oscillations in AD (Fig. 2)

Although alterations of PV neuronal activity and sleep abnormalities are both described at an early stage of AD and suggested to be potential causal factors for the subsequent AD pathogenesis, a potential relationship between these two changes in AD is not yet established. A possible explanation is as follows: Precisely timed cortical, hippocampal, and TRN PV activity is critical for the correct generation and nesting of sleep oscillations; Abnormally increased activity of PV interneurons will lead to disruption of the timing of cortical PV discharge with respect to sleep oscillations and impair calcium dynamics in pyramidal neurons during sleep (Averkin et al., 2016; Niethard et al., 2018; Seibt et al., 2017). In fact, a study demonstrated that chemogenetic excitation of PV interneurons in frontal cortex of mice greatly decreases SWA (0.5–4.5 Hz) during NREM sleep (Funk et al., 2017), indicating that hyperexcitability of PV neurons in AD may reduce SWA vital for Aβ clearance (Mander et al., 2015).

Along with SO disruption, cortical GABA levels as well as the expression of GABA_A and GABA_B receptors were decreased in APP mice at 4 months of age prior to plaque deposition, compared to the wildtype controls (Kastanenka et al., 2017). Additionally, this study demonstrated that the power of SOs was decreased by topical application of GABA_A inhibitor onto cortices in healthy wild-type animals (Kastanenka et al., 2017). A study in APP23/PS45 transgenic mice showed impaired SWA and the long-range coherence of slow waves in the neocortex, thalamus and hippocampus, which was restored by topical applications of a GABA_A receptor agonist (Busche et al., 2015). Thus, impaired SWA in AD could be attributed to deficits in cortical inhibitory circuits involving alterations of GABA release as well as GABA_A and GABA_B receptors.

Cortical PV interneurons are also involved in REM sleep oscillations and plasticity. An in vivo calcium imaging study in mice showed that global cortical activity was reduced during REM sleep which was accompanied by reduced activity of pyramidal neurons and increased activity of cortical PV neurons (Niethard et al., 2016). Together with the other results described in this section, these results suggest that disrupted excitatory/inhibitory interactions between pyramidal cells and interneurons could be key factors that alter sleep oscillations in AD (Fig. 2).

Cortical interneurons other than PV interneurons may also be involved in AD disease processes or a potential target for therapeutic interventions. Funk and colleagues showed that chemogenetic activation of SOM positive interneurons increased SWA, whereas chemogenetic inhibition of SOM neurons decreased SWA (Funk et al., 2017). Furthermore, optogenetic excitation of cortical SOM neurons expressing nNOS evoked a response that resembles a physiological slow wave (Gerashchenko et al., 2018). Another study demonstrated that mice lacking nNOS expression in cortical SOM neurons showed significantly lower spectral power in the lower slow wave frequency range (0.5–3.0 Hz) during NREM sleep compared to the control mice (Zielinski et al., 2019). Moreover, the same nNOS knock-out mice exhibited significantly impaired cortical-dependent recognition memory performance. Thus, both PV and SOM/nNOS interneurons are highly promising therapeutic targets to enhance sleep and prevent AD progression. Further studies in AD mouse models will be necessary to determine whether and when modulation of PV and/or SOM/nNOS neurons could effectively enhance sleep oscillations and reduce AD pathology.

5. Conclusions

One in ten adults over the age of 65 are diagnosed with AD, and it is recognized as one of the largest public health challenges worldwide. Thus, there is an urgent need to identify reliable biomarkers to detect AD at preclinical stages or early stages of the disease, as well as to develop early intervention strategies to prevent or delay AD progression. As we reviewed here, a growing body of evidence supports links between sleep abnormalities, GABAergic abnormalities and AD pathophysiology in both AD patients and AD mouse models (Fig. 2). Each sleep oscillation reviewed in this article showed some level of abnormalities in AD patients or AD mouse models. Box 1 gives an overview of outstanding research questions in this field. Variabilities in timing and level of abnormalities among types of oscillations in AD can be attributed to the difference in brain regions that are involved in these oscillations which have different level of vulnerability and timing to be affected by the disease. Thus, longitudinal and comprehensive investigation of sleep oscillations at different stages of AD would be helpful to identify reliable early biomarkers to predict AD progression. Based on currently available findings, a variety of sleep oscillations reviewed here are characterized especially in APP/PS1 mouse model and 3xTg-AD mouse model (Tables 1–5). Although there is no perfect AD model, these mice models could be especially valuable to further examine effects of manipulation of sleep oscillations to test potential treatments via neural modulation. It is also important to note that most of the studies in AD animal models have been done in the aggressive early-onset AD models which better reflect familial AD than sporadic AD. Early-onset AD accounts only for 5–6% of AD population. Recent years, there has been a great amount of effort put in developing late-onset AD animal models which will be able to provide a better representation of majority of AD population in humans. It would be interesting to see how sleep changes in these late-onset animal models and how early intervention to prevent or improve sleep abnormalities affects AD progression. There is still a great deal to investigate before we understand the causal relationships of sleep abnormalities, GABAergic neuronal changes and cognitive impairments in AD, as well as the underlying mechanisms which can be targeted to ameliorate these abnormalities. A variety of behavioral, pharmacological, non-invasive brain stimulation and invasive approaches are available to modulate sleep and sleep oscillations (Brown et al., 2022). Thus, understanding these relationships may prove to be a fruitful avenue of translational research.

Box 1. Outstanding research questions:

• What are the distinct roles of SO (< 1 Hz) and delta oscillations (1–4 Hz) in AD progression in patients and in mouse models?

• What properties of sleep spindles (density, amplitude, duration, frequency etc.) are altered in AD mouse models?

• Are there any changes in SWRs in AD/MCI patients and how are they associated with AD progression?

• Are there any changes in triple-coupling of SOs, spindles, and SWR in AD/MCI patients and AD mouse models, and how are they associated with AD progression?

• Are there any changes in PGO waves during REM in AD animal models and human correlates of PGO waves in AD/MCI patients, and how are they associated with AD progression?

• What is the time course and relationship between alteration of GABAergic (PV, SOM/nNOS) neuronal activity and sleep disruption in subsequent AD pathogenesis?

• How do manipulations (enhancement/disruption) of each sleep oscillation or relevant neural circuit affect AD progression?

• Are there any changes in sleep oscillation and effects of sleep manipulation in late-onset AD animal models?

Abbreviations:

Aβ

amyloid-β

AD

Alzheimer’s disease

CSF

cerebrospinal fluid

EEG

electroencephalogram/electroencephalography

GABA

Gamma-aminobutyric acid

hAPP

human amyloid protein precursor

MCI

mild cognitive impairment

mGluR5

metabotropic glutamate receptor subtype 5

MMSE

Mini-Mental Status Examination

MoCA

Montreal Cognitive Assessment

mPFC

medial prefrontal cortex

NMDA

N-methyl-D-aspartate

nNOS

neuronal nitric oxide synthase

NREM

non-rapid-eye-movement

PGO

ponto-geniculo-occipital

PV

parvalbumin

REM

rapid-eye-movement

SO

slow oscillation

SOM

somatostatin

SWA

slow wave activity

SWR

sharp-wave ripple

TRN

thalamic reticular nucleus