Neuromodulation of sleep rhythms in schizophrenia: Towards the rational design of non-invasive brain stimulation

Abstract

Brain function critically depends on oscillatory synchronization of neuronal populations both during wake and sleep. Originally, neural oscillations have been discounted as an epiphenomenon. More recently, specific deficits in the structure of brain oscillations have been linked to psychiatric diseases. For example, schizophrenia is hallmarked by abnormalities in different brain oscillations. Key sleep rhythms during NEM sleep such as sleep spindles, which are implicated in memory consolidation and are related to cognitive functions, are strongly diminished in these patients compared to healthy controls. To date, it remains unclear whether these reductions in sleep oscillations are causal for the functional impairments observed in schizophrenia. The application of non-invasive brain stimulation permits the causal examination of brain network dynamics and will help to establish the causal association of sleep oscillations and symptoms of schizophrenia. To accomplish this, stimulation paradigms that selectively engage specific network targets such as sleep spindles or slow oscillations are needed. We propose that the successful development and application of these non-invasive brain stimulation approaches will require rational design that takes network dynamics and neuroanatomical information into account. The purpose of this article is to prepare the grounds for the next steps towards such rational design of non-invasive stimulation, with a special focus on electrical and auditory stimulation. First, we briefly summarize the deficits in network dynamics during sleep in schizophrenia. Then, we discuss today’s and tomorrow’s non-invasive brain stimulation modalities to engage these network targets.

1. Sleep network dynamics deficits in schizophrenia

Patients with schizophrenia are commonly experiencing sleep disturbances, which is reflected in longer time needed to fall asleep, problems staying asleep and overall reduced sleep efficiency (Chan et al., 2017; Kaskie and Ferrarelli, 2020; Robertson et al., 2019). Furthermore, reductions in the amount of deep slow wave sleep (SWS) have been reported (Chan et al., 2017). These changes in macroscopic structure of sleep architecture or quantity of sleep have been known for a long time. Yet, a more recent understanding of the brain oscillations during sleep that can be observed in the EEG has further helped to unravel changes in the quality of sleep in patients with schizophrenia. These sleep oscillations, specifically during non-rapid eye movement (NREM) sleep, might be a key target for sleep enhancement in schizophrenia.

NREM sleep in humans is characterized by two dominant sleep oscillations, sleep spindles and slow waves. Sleep spindles are transient bursts of waxing and waning amplitude with a frequency between 11-16 Hz that are observable in the scalp EEG. They originate from an interplay between thalamus and cortex (reviewed in Lüthi, 2014). The slow waves specifically characterize deep sleep and are observed as high-amplitude, low-frequency (e.g., <4.5 Hz) waves that mainly originate in the neocortex but can also have thalamic origin (Neske, 2016). Slow waves measured on the scalp surface are related to the slow oscillation, which rests on the low-frequency oscillation (1 Hz) in the membrane potential of cortical neurons (e.g. Amzica and Steriade, 1998; Contreras and Steriade, 1995). Slow waves are characterized by synchronized firing of a very large population of neurons (up-states, ON windows), followed by synchronized silence across many neurons (down-states, OFF windows) leading to the typical low frequency, high amplitude waveforms that can be captured in the EEG (Steriade et al., 2001).

Slow wave impairments have been reported in patients with schizophrenia, of which most consistently a reduced slow wave density/number during NREM sleep was observed. However, several other sleep studies failed to establish slow wave impairments in those patients (for an extensive review see Castelnovo et al., 2018). More consistently, the literature points to a sleep spindle deficit in patients with schizophrenia (Castelnovo et al., 2018, 2016; Ferrarelli, 2015; Manoach et al., 2016; Wilson and Argyropoulos, 2012). Specifically, a reduction of different sleep spindle characteristics such as the density (number/min), duration, amplitude, activity (e.g. sigma power, integrated spindle activity), and coherence has been repeatedly reported. High-density EEG systems (e.g. with 256 electrodes) have further revealed that the effects of spindles might be localized to specific cortical regions (based on sensor-level analysis, e.g. prefrontal and centro-parietal regions (Ferrarelli et al., 2010, 2007)). While sleep spindle deficits have been mainly observed in chronically medicated patients, there is evidence that these spindle alterations are not explained by the use of antipsychotic medication (Ferrarelli et al., 2010, 2007; Wamsley et al., 2012). Along these lines, sleep spindle deficits have been reported in unaffected first-degree relatives of patients with schizophrenia (Agostino et al., 2018; Manoach et al., 2014; Schilling et al., 2017). Therefore, sleep spindles have been proposed as a candidate endophenotype for schizophrenia (Manoach et al., 2016). Furthermore, sleep spindles (specifically fast sleep spindle density [num/min]) have been negatively correlated with schizotypal personality traits in healthy participants, indicating a neurobiological overlap between nonclinical schizotypy and schizophrenia (Lustenberger et al., 2014).

Large-scale neural oscillations as measured by EEG arise from a coordinated interplay of neuronal activity within or across different brain regions (Buzsáki et al., 2012). In schizophrenia this interplay is impaired leading to deficient network dynamics and therefore schizophrenia has often been described as a “disconnection syndrome”(Jones, 2010). In terms of the aberrant sleep spindles found in schizophrenia, deficits or a disconnection in the thalamo-cortical system might be at play (Ferrarelli and Tononi, 2010; Jones et al., 2014). Specifically, it has been hypothesized that the circuit between the thalamic reticular nucleus (TRN; pacemaker of spindles, (Fuentealba and Steriade, 2005)), the mediodorsal (MD) thalamus, and the prefrontal cortex (PFC) is dysfunctional (reviewed in Ferrarelli and Tononi, 2017). While the TRN is the main generator of spindles, the thalamo-cortical and cortico-thalamic loops are essential for spindle synchronization and amplification (Fuentealba and Steriade, 2005). Thus, reduced sleep spindle densities might be explained by abnormal functioning of the TRN while reductions in amplitude/duration and coherence might point to deficient cortico-thalamic loops. Current research specifically points to a hypo-connectivity between PFC and the MD thalamus (Welsh et al., 2008), and reduced MD thalamus volumes that relate to impaired sleep spindles in schizophrenia (Buchmann et al., 2014). Reported reductions in slow wave parameters both in patients with schizophrenia but also first-degree relatives further point towards deficits in (functional) cortical connectivity (Agostino et al., 2018).

Sleep spindle deficits correlated with symptom severity in schizophrenia (Ferrarelli et al., 2010; Tesler et al., 2015; Wamsley et al., 2012) but whether they are causal drivers for the behavioural symptoms still needs to be established. Overall, sleep spindles likely mirror the integrity of the thalamo-cortical system and in turn, an efficient thalamo-cortical system is essential for cognition, learning efficiency, and sensory processing (Crick, 1984; Lustenberger et al., 2012; Moustafa et al., 2017; Pinault, 2004; Zikopoulos and Barbas, 2012). These abilities are impaired in patients with schizophrenia along with sleep spindles, therefore reflecting the same diminished underlying thalamo-cortical network (Ferrarelli and Tononi, 2017; Manoach and Stickgold, 2019). However, there is evidence from a study in healthy control participants that these sleep spindles are not just mirroring thalamo-cortical system integrity but also play a causal role in memory consolidation, most convincingly in motor memory consolidation (Lustenberger et al., 2016). Along the same line, deficits in motor memory consolidation have been reported in patients with schizophrenia (Genzel et al., 2015; Manoach et al., 2010; Manoach and Stickgold, 2019; Wamsley et al., 2012). Also slow waves are related to restorative brain functions, such as synaptic rescaling and memory consolidation (Klinzing et al., 2019; Tononi and Cirelli, 2019). Therefore, targeted interventions that promote sleep spindles and slow waves carry the potential to reduce functional impairments in these patients.

2. Rational design of non-invasive brain stimulation during sleep

A portfolio of non-invasive, non-pharmacological stimulation methods exist to modulate brain oscillations. Some of them have also been applied during human sleep to modulate sleep oscillations including different transcranial stimulation modalities (electrical stimulation, magnetic stimulation, radiofrequency electromagnetic fields, e.g. (Barham et al., 2016; Huber et al., 2002; C. Lustenberger et al., 2015b; Lustenberger et al., 2016; Marshall et al., 2006; Massimini et al., 2007)) and sensory stimulation modalities (auditory, somatosensory, olfactory, visual, and vestibular, e.g. (Bellesi et al., 2014; Omlin et al., 2018; Pereira et al., 2017; Perrault et al., 2019; Tononi et al., 2010)). In this review, we will focus on two of them – transcranial electrical stimulation and auditory stimulation (Figure 1). They are promising methods to modulate sleep oscillations and carry therapeutic potential because (1) they have already been tested in a multitude of studies which confirm their potential in augmenting sleep spindles and/or slow waves, and (2) these methods can be safely applied in in-home settings with portable solutions, are therefore cost effective and can possibly be applied over long time periods.

Figure 1:

Overview of selection parameters for a rational design of auditory and electrical stimulation.

2.1. Transcranial electrical stimulation

Brain oscillations are measured non-invasively by EEG capturing electrical fields that are generated by the synchronized firing of many neurons (Buzsáki et al., 2012). For a long time these recorded brain oscillations were considered an epiphenomenon of underlying neuronal activity on the ground that the underlying endogenous electric fields are small in magnitude (around 1V/m) and likely not powerful enough to modulate neuronal firing (e.g. the small change elicited in membrane voltage is not strong enough to induce an action potential in a neuron with a membrane voltage close to the resting potential). However, recent seminal papers have convincingly shown that this is a major misconception because neuronal network are susceptible to weak electric fields and can modulate neural activity as long as neurons are close to the threshold (Anastassiou et al., 2011; Bikson et al., 2004; Deans et al., 2007; Fröhlich and McCormick, 2010; Radman et al., 2007; Reato et al., 2010). With that idea in mind, transcranial electrical stimulation has become a prominent non-invasive brain stimulation approach. Transcranial electric stimulation comes in different flavours including transcranial direct stimulation (tDCS, constant anodal or cathodal current), transcranial random noise stimulation (tRNS, current with random noise waveform), transcranial alternating current stimulation (tACS, application of an oscillatory current), or a combination of these waveforms such as oscillatory tDCS (Paulus, 2011). In this review we will focus on oscillatory applications of electrical current, since they carry the potential to induce frequency-specific effects on brain dynamics, such as modulating sleep spindles and slow waves. TACS refers to the application of weak current that oscillates at a specific frequency of interest. Previous studies have applied tACS mainly during wake, both in health and disorders such as schizophrenia and depression providing mixed evidence for effects on behaviour, symptom severity, and neurophysiological effects (e.g. (Ahn et al., 2019; Alexander et al., 2019; Liu et al., 2018; Mellin et al., 2018)). A few studies also applied tACS and slow-oscillatory tDCS (so-tDCS) to modulate sleep oscillations during NREM sleep (reviewed in (Barham et al., 2016; Cellini and Mednick, 2019; Koo et al., 2018)). A majority of these studies successfully modulated sleep spindles and slow waves as assessed in stimulation-free windows of EEG immediately after stimulation along with enhanced motor or declarative memory consolidation, respectively. Only one study so far applied so-tDCS in patients with schizophrenia, which found enhanced declarative memory similar to previous studies in healthy participants (Göder et al., 2013). However, effects on the network dynamics of sleep were not reported in this publication.

The proposed mechanism of action for tACS was hypothesized to be neural entrainment of the cortex that is reflected in an aligning of spike timing of the neurons to the externally applied stimulation waveform. This has been shown in-vitro and in-vivo models (Deans et al., 2007; Fröhlich and McCormick, 2010; Ozen et al., 2010; Reato et al., 2010; Schmidt et al., 2014). The translation of these findings to human models is difficult for several reasons, including the missing (long-range) connectivity in in-vitro models and the differences in resulting electric field in lissencephalic versus gyrencephalic brains. Furthermore, since in humans the electrical fields are causing artefacts in the EEG, the assessment of direct effects of the stimulation on EEG are to date not feasible and the efficacy of tACS with the currently mostly applied intensities (e.g. around 0.5 - 2mA) has been questioned (Filmer et al., 2019). A new study by Johnson et al. (Johnson et al., 2019) now provides first in-vivo, mechanistic evidence in awake non-human primates that with currently applied field intensities in humans, neuronal entrainment indeed occurs in cortical neurons in a dose-dependent fashion. Specifically, they illustrate that neuronal entrainment can occur at field strengths < 0.5mV/mm that is attainable in humans for tACS intensities around 1-2 mA. Furthermore, a recent study (Negahbani et al., 2019) that combined multisite electrophysiological recordings in ferrets and biophysical models of thalamo-cortical network dynamics provides strong evidence for the proposed mechanism of action of tACS (Ali et al., 2013), the so-called Arnold tongue, which describes how frequency-tuning of the stimulation waveform is critical for causing entrainment by tACS. The critical discussion about efficacy in humans is essential to push for more mechanistic insights but also to emphasize the importance of mechanistically oriented, rational designs of tACS (Frohlich, 2018; Kurmann et al., 2018). In order for tACS to be effective in modulating sleep oscillation in both healthy individuals as well as in patients with psychiatric disorders, such as schizophrenia, that exhibit oscillopathies we provide here some considerations for a rational design of tACS..

Field modelling and electrode montage

A necessary ingredient for effective neuronal entrainment is the achievement of a strong enough field intensity in the region(s) of interest. Therefore, it is necessary to assess the induced field strength across the cortex with the planned stimulation electrode positions (“montage”). To date, we can approximate these electric field intensities in humans with electric field modelling (Huang et al., 2017; Saturnino et al., 2019; Windhoff et al., 2013). Thus, for designing a study one should optimize the electrode position, form and size for optimal field strength in the target location(s) by using simulations of different combinations. Specifically for electrode size, to achieve localized effects ring electrodes could be used, however one should also consider that tACS relies on network amplification, which likely requires a large enough network to achieve an effect (Fröhlich, 2014). In addition, very small electrodes will have a higher current density and could lead to a more painful experience and tingling, which is suboptimal for several reasons. Especially during sleep, such applications could lead to arousals and un-blinding. Overall, much effort should be given to optimizing the dosage of the applied field (magnitude but also duration) to maximize effects, both in terms of modifying the oscillations of interest and behavioural outcomes. Yet, it is important to note that knowing the electric field delivered to different brain regions is not sufficient to predict the effect of stimulation since the resulting perturbation of the membrane voltage can only modulate network dynamics in synergistic interaction with ongoing network patterns.

Target engagement and nested oscillations

In order to promote sleep oscillation a rational design should take their spatial localization and dominant frequency into account. In adults, sleep spindles have a very specific frequency and amplitude distribution across the cortex that is observable both in high-density EEG recordings and intracranial recordings in epilepsy patients (Andrillon et al., 2011; Lustenberger et al., 2015; Nir et al., 2011; Piantoni et al., 2017). Specifically, slower frequency spindles (e.g. around 9-12 Hz) are predominantly recorded in temporal and frontal regions whereas fast spindles (e.g. around 13-16 Hz) are more abundant in centro-parietal regions. There is also evidence that these different frequencies might serve different functions and possibly point to natural resonance frequencies that rely on the anatomical structure of the cortico-thalamic modules (Rosanova et al., 2009). In schizophrenia, a global reduction in both slow and fast frequency spindles have been observed (Ferrarelli et al., 2010). In addition, it was multiple times reported in these patients that they have reduced motor memory consolidation (Genzel et al., 2015; Manoach et al., 2010; Manoach and Stickgold, 2019; Wamsley et al., 2012). Interestingly, in one of our recent studies we observed that specifically fast sleep spindles (e.g. around 15 Hz) predicted motor memory consolidation and that feedback-controlled spindle tACS with a field strongest over motor regions promotes motor memory consolidation in healthy young adults (Lustenberger et al., 2016). The improvement was predicted by the amount of enhanced fast spindle activity in these subjects. Thus, a noteworthy target to improve motor memory impairments in these patients could be the application of fast spindle frequency tACS with a montage that maximizes the field over centro-parietal regions. Alternatively, prefrontal stimulation might also be of interest in schizophrenia because brain imaging studies specifically point to PFC-MD synchronization deficits (reviewed in (Ferrarelli and Tononi, 2017)).

Another important aspect to consider in terms of spindle modulation in schizophrenia is their coherence across regions. Coherence across different electrode locations (e.g. central, frontal and occipital electrode sites) in the spindle frequency range (fast sigma between 12-15 Hz) was found to be pronouncedly reduced (Wamsley et al., 2012). Thus tACS could be designed in order to enhance coherence across regions, for instance by applying tACS in-phase over the regions of interest. Such an in-phase stimulation could for instance be achieved by having a high-density tACS setup (e.g. see (Helfrich et al., 2014)) that allows for a localized but similar polarity over the regions that should be synchronized in order to enhance coherence among them. Alternatively, a setup with three electrodes (Alexander et al., 2019; Lustenberger et al., 2015a; Lustenberger et al., 2016; Mellin et al., 2018; Polanía et al., 2012), where two electrodes provide in-phase stimulation may suffice to enhance functional connectivity. However, the coherence was not predictive for overnight memory consolidation nor schizophrenia symptoms (Wamsley et al., 2012) and therefore it needs to be elucidated whether improved spindle coherence translates into a functional benefit.

Even though less consistent, slow wave deficits have also been reported in patients with schizophrenia (Castelnovo et al., 2018). Application of so-tDCS (tDCS oscillating at a frequency of ~0.75 Hz, compared to tACS has a DC offset), repetitive so-tDCS or tACS around the same frequency showed mixed results on post-stimulation slow wave activity and declarative memory (Antonenko et al., 2013; Cellini et al., 2019; Del Felice et al., 2015; Eggert et al., 2013; Göder et al., 2013; Ladenbauer et al., 2016, 2017; Marshall et al., 2006; Paßmann et al., 2016; Prehn-Kristensen et al., 2014; Sahlem et al., 2015; Westerberg et al., 2015). Possibly, different montages, frequency mismatch to endogenous frequency, and stimulation waveform types might have been responsible for these mixed effects (Koo et al., 2018). Thus, a rational design taking mechanistic insights into account could also help for more effective slow wave modulation. Slow waves are predominant over frontal regions in adults and are generated by a synchronized firing pattern of a large amount of neurons, which leads to the characteristic ON (synchronized firing) and OFF states (synchronized silence). Thus, montages should try to optimize frontal synchronization of neuronal firing that is matching the endogenous oscillation frequency in order to align with the ON and OFF (see more discussion in the feedback controlled approaches). Stimulations should also take into account where the hubs of slow wave occurrences are (e.g. according to intracranial recordings in the medial prefrontal regions (Murphy et al., 2009; Nir et al., 2011)) and specifically target these areas. If there is interest to modulate slow waves more locally ring electrodes could be used to achieve stronger fields over the regions of interest. However, synchronization between frontal hemispheres might also be decisive to augment slow waves and therefore it should be carefully considered where anodal and cathodal electrodes are placed. For instance, Garside et al. (2015) placed the anodal and cathodal electrodes on the left vs. right frontal cortex (F3, F4) and as they hypothesized this montage disrupted post-stimulation slow wave activity in a nap and reduced memory consolidation.

Finally, slow waves and sleep spindles are not isolated from each other but rather tightly interact. Indeed, many of the studies discussed above modulated both oscillations simultaneously either in the same or opposite direction rather than selectively (Aeschbach and Borbély, 1993; Andrillon et al., 2011; De Gennaro and Ferrara, 2003; Dijk et al., 1993; Himanen et al., 2002; Steriade and Amzica, 1998; Uchida et al., 1991). Along the same lines, spindles are nested in the up-states of slow waves, and this nesting seems to be important for declarative memory consolidation (Rasch and Born, 2013; Staresina et al., 2015). Furthermore, first hints point to a possible decoupling of these brain oscillations in schizophrenia that might be related to memory deficits (Bartsch et al., 2019). Thus, cross-frequency tACS (e.g. nested spindle frequency during the peak of the slow oscillatory phase) could be applied with the goal to promote this coupling of slow waves and spindles. Another approach to investigate is the application of spindle like tACS time-locked to the up-states of slow waves.

Feedback controlled approaches

An important aspect of transcranial electrical current stimulation to keep in mind is that the resulting changes in membrane voltage are too weak to trigger action potentials in neurons that are not close to threshold by virtue of their endogenous dynamics. In terms of tACS, the success of modulating brain oscillations will rely on the vigilance states, which are defined by different endogenous oscillations (Fröhlich, 2014; Thut et al., 2017). TACS will specifically and maybe only (at the current intensity level) be successful if the perturbed system has an ongoing oscillation at the frequency of interest and is close to threshold so that small changes in membrane voltage make a difference (Ali et al., 2013). This idea is reflected in the principle of the Arnold tongue. Shortly summarized, the principle postulates that weak forces (e.g. subthreshold, such as tACS) can only amplify the intrinsic frequencies whereas stronger forces (e.g. suprathreshold, transcranial magnetic stimulation) can possibly override the frequency of the intrinsic oscillator (see (Fröhlich, 2014) for a more detailed and intuitive description).

Taking this principle into account when modulating sleep spindles, one has to consider that they are transient events (e.g. bursts in sigma range reoccurring after a certain amount of time). Therefore, feedback-controlled approaches might be important in order to detect times when spindle activity is present and stimulate at these specific times to further enhance the ongoing activity (Lustenberger et al., 2016). Since spindle frequencies are variable across nights/cycles and across subject, advanced closed loop approaches could even dynamically detect the current spindle frequency in each individual and adapt the stimulation frequency accordingly in the course of the sleep period.

Slow waves do not show this transient behaviour and are more continuously present during deep sleep. However, slow waves correspond to synchronized states of neural firing (synchronized depolarization of cortical neurons, ON states or up-states) and silence (synchronized hyperpolarization of cortical neurons, OFF states or down-states). Therefore, a rational design of slow wave modulation with tACS or so-tDCS might optimize enhancement of slow waves by taking this slow wave phase into account (Ali et al., 2013; Jones et al., 2018; Ketz et al., 2018; Wilde et al., 2015). Furthermore, the closed-loop approach may also include an individualized and dynamic frequency matching.

The key hurdle in the development of a feedback-controlled or closed-loop system of electrical stimulation in combination with the EEG is the strong stimulation artefact that makes the EEG unusable during stimulation. Currently, only pre-stimulation and post-stimulation EEG can be considered and the reported effects are restricted to stimulation-free windows (e.g. Lustenberger et al., 2016; Marshall et al., 2006). What exactly happens with sleep spindles and slow waves while stimulation occurs is still unclear. The removal of the artefact is particularly challenging because the artefact is non-linearly modulated by factors such as respiration and heartbeat (Noury et al., 2016). Very recently, the method of signal-space projection was evaluated on a phantom head and in human recordings in removing the tACS artefact and provides potentially promising results in recovering the EEG for subsequent analysis (Vosskuhl et al., 2019). However, future studies will be needed to confirm and evaluate its potential. When designing a study with the potential to remove tACS artifacts in the future technical details such as that the amplifier is not saturating when stimulation occurs (e.g. that the complete and undistorted artefact is present in the recorded data) and that the stimulation signal form the stimulator is perfectly synchronized with the EEG amplifier need to be considered.

2.2. Auditory stimulation

Acoustic stimuli (e.g. rhythmic amplitude modulated sounds) have historically been used to identify the underlying network state because different conditions such as vigilance state and different disorders affect the response to these stimuli (e.g. (Griskova et al., 2007; Lustenberger et al., 2018; Picton et al., 2003; Plourde et al., 1998; Thuné et al., 2016; Van Deursen et al., 2011)). Today, the importance of acoustic stimuli has moved beyond probing brain states towards modulating brain activity.

Studies of auditory stimulation have mostly focussed on modulating slow waves and only a few targeted sleep spindles (e.g. (Antony and Paller, 2017; Bellesi et al., 2014; Cellini and Mednick, 2019; Fattinger et al., 2017; Lustenberger et al., 2018; H.-V. V Ngo et al., 2013; Ong et al., 2018; Papalambros et al., 2017)). Importantly, these studies provide promising evidence that auditory stimulation can significantly modulate slow waves, sleep spindles, and their interaction along with memory consolidation as well as peripheral body functions such as the release of specific hormones, and markers of the autonomic control of the heart. To date, auditory stimulation during sleep has not been applied in patients with schizophrenia.

The mechanism underlying the effects of acoustic stimuli on sleep oscillation in not clearly known and presumably depends on the type of stimulation applied (e.g. single short stimuli at specific times vs. amplitude modulated sounds, cueing words). It has been hypothesized that non-lemniscal pathways, covering brainstem nuclei, thalamic nuclei and thalamo-cortical circuits are involved (reviewed in Bellesi et al., 2014), which could upon applied acoustic stimuli synchronize large populations of neurons. In contrast to electrical stimulation, auditory stimulation is a bottom-up and supra-threshold (directly lead to action potentials) approach.

Characteristics of the applied stimulus

For the rational design of auditory stimulations during sleep the choice of the applied acoustic stimulus is likely of high importance. This includes applied volume, shape/type and duration of the stimuli (see also Bellesi et al., 2014). The choice of stimuli will be different for whether slow waves or spindles should be enhanced.

For targeting sleep spindles, only very few studies exist. Two of them have used an amplitude modulated white noise at the frequency of sleep spindles (Antony and Paller, 2017; Lustenberger et al., 2018), and one has applied pink noise clicks with a temporal spacing that resulted in a fast spindle frequency (Ngo et al., 2019). All methods were able to enhance sleep spindle activity in a window around the applied stimuli. Interestingly, control frequencies at 50 Hz or 40 Hz did not enhance sleep spindles or had a more widespread responses on different frequency bands, respectively (Antony and Paller, 2017; Lustenberger et al., 2018). Thus, for spindle enhancement, it seems to be important to have rhythmic stimuli in the range of the sleep spindle range, possibly to entrain the thalamo-cortical system in this frequency range. Yet, it is unclear whether pink/white noise amplitude modulated stimuli or clicks are the more effective, or whether other kind of oscillatory signals, such as pure tone clicks in the spindle frequency range or other oscillatory patterns such as binaural beats (e.g. with mismatch frequency in the spindle frequency range) are as or even more effective. The previous studies used different durations of stimuli from 0.5 s - 2 s. Interestingly, in all these studies spindle enhancement did not occur at the start of the stimulus but only a few hundred milliseconds later (e.g. in Ngo et al., 2019 the stimulus had already ceased after the effect was observable). The reason for this delay is unclear but its occurrence represents an important feature that should be considered when designing the experiment. One important aspect to consider is that the functional significance of these spindles enhancements is unclear because two studies had designs within the sleep period to compare spindle activity without assessing behavioural outcomes (Antony and Paller, 2017; Lustenberger et al., 2018) and one study that had the spindle stimulus phase locked to up-states failed to see a benefit for declarative memory (Ngo et al., 2019).

The modulation of slow waves has gained much more attention than the modulation of sleep spindles. Here only one study has applied amplitude modulated pink noise in the slow wave frequency range (Simor et al., 2018) but more studies have applied single, short bursts of pink noise (around 50 ms) either specifically locked to the phase of slow waves (e.g. Fattinger et al., 2017; Grimaldi et al., 2019; Ong et al., 2016; Santostasi et al., 2016), in a rhythmic manner to mimic slow wave rhythms (e.g. every ~1s, e.g. Ngo et al., 2013; Tononi et al., 2010) or a combination of both (Leminen et al., 2017; Ngo et al., 2013; Ngo et al., 2015). The rhythmic stimuli (amplitude modulated or repetitive bursts) may rather promote entrainment to the external frequency applied whereas phase-targeted applications possibly modulate synchronization of neuronal activity, either promoting (e.g. up-phase targeted stimulation, Ngo et al., 2013; Ong et al., 2016; Santostasi et al., 2016)) or disrupting the synchrony (e.g. down-phase targeted stimulation (Fattinger et al., 2017)). Whether rhythmic stimuli or phase-locked stimuli are superior is unclear and would need specific experiments that compare different stimulation modalities within the same sleep period. It is also unclear whether amplitude modulated signals or burst of pink noise are superior. Interestingly, most studies applied pink noise. A recent abstract has shown that pink noise seems to be superior to other stimuli such as “A” vowels and pure sounds (E Debellemaniere et al., 2018) and Acosta et al. (2019) further hypothesized in a theoretical opinion piece that biological systems are possibly more perceptive to pink noise stimuli according to general Weber’s law. Thus, a rational design may consider using pink noise stimuli for slow wave modulation (and possibly spindle modulation) until more effective acoustic stimuli have been identified.

Another important factor for a rational design is the volume or intensity of the acoustic stimulus. So far, studies applied sounds during sleep of around 38-65 dB SPL. The acoustic stimulus has to be loud enough to elicit a response but at the same time should not arouse the patient. This is specifically important taking into account that the non-lemniscal pathways have an anatomical overlap with arousal-promoting neuromodulatory systems (Hu, 2003; Jones, 2003) and therefore the volume represents a fine balance that should be carefully taken into account. First, the sound should be adapted to the hearing threshold of the individual and a hearing test is necessary. In this regard also frequency content should be considered. For instance, with ageing higher frequencies (e.g. above 4000 Hz) are harder to perceive (Liu and Yan, 2007), therefore pink noise that has more power in lower frequencies might be an ideal stimuli to overcome this problem. Second, sound intensity can be increased with sleep depth because awakenings are less likely. Therefore, a dynamic volume that scales with sleep depth can be implemented. Third, arousal should be monitored when performing stimulation and the volume dynamically lowered when arousals are detected as a consequence of the stimuli. Fourth, specifically for amplitude modulated designs, stimuli can be embedded in continuous white/pink noise that is delivered throughout the stimulation period of interest (Antony and Paller, 2017; Lustenberger et al., 2018). This could be a promising approach to reduce the possibility of arousal since there is not a sudden increase in sound volume from complete silence. Finally, one should carefully consider how the sound is delivered to the subject. Using loudspeakers is not ideal since there will be a clear variability how much of the sound is perceived by the subject and depending on the distance and head position relative to the loudspeaker the sound might be too low in intensity. Therefore, comfortable in-ear headphone or flat headphones that could be taped to the ear are recommended. Along the same lines, output of headphones should be carefully tested and it should be ensured that the used headphones are not inducing artefacts in the EEG recordings.

Of note, in this review we have focussed on meaningless sounds, compared to meaningful sounds, which relate to targeted memory reactivation paradigms (TMR, extensively reviewed in (Cellini and Capuozzo, 2018; Cellini and Mednick, 2019; Schouten et al., 2017)). Thus, specific words or type of sounds (cues) are first match with a target during wakefulness (e.g. learning phase) and these cues are replayed during sleep. Such stimulation has also shown to modulate sleep oscillations along with memory consolidation. How these approaches differ in terms of mechanism and effect on sleep oscillations is unclear and needs further investigation. Some drawbacks of TMR are that they require a matching phase during wakefulness, the timing and specificity is more difficult to achieve, and out of lab approaches might be less feasible.

Target engagement and nested oscillations

Auditory stimulation is a bottom-up approach that likely engages thalamic nuclei that project to many different cortical regions (e.g. matrix cells). Therefore, it is questionable how locally targeted such an approach can be. Fattinger et al. (2017) showed that slow waves can be focally suppressed in left sensory-motor regions when the electrode that informed the algorithm about underlying phase (here downphase was targeted) was in the same area. Such local modulation is likely not achievable when amplitude modulated or rhythmic sounds are applied to enhance slow waves or sleep spindles. But is local specificity even desired in enhancing sleep oscillations in schizophrenia with auditory stimulation? Maybe more important than local specificity is engaging thalamo-cortical loops, as these circuits are likely disrupted and dysfunctional. Auditory stimulation possibly targets these circuits and therefore might enhance target engagement. Nevertheless, these circuits might also be less susceptible to auditory stimulation than in healthy participants and the potential to engage them in schizophrenia has to be further explored.

Of note, enhancement of sleep spindles or slow waves with auditory stimulation are often not selective to either type of oscillations (e.g., Krugliakova et al., 2019; Ngo et al., 2019; Ngo et al., 2013). As discussed before in this review, sleep spindles and slow waves tightly interact and therefore their interaction has to be acknowledged when performing sleep modulation. For instance, single pink noise bursts during up-phases clearly enhance slow waves but at the same time also engage fast sleep spindles. Along the same lines Krugliakova et al. (2019) recently reported that single bursts of pink noise enhance cross-frequency coupling of slow waves and spindles. Such enhanced interaction could be desirable, as it provides optimal conditions for synaptic plasticity in neocortex (Niethard et al., 2018; Seibt et al., 2017) along with the potential to promote memory consolidation. In a recent study, Ngo et al. (2019) applied bursts of pink noise at the spindle frequency range that were phase locked to the up-state of slow waves to further boost this interaction. Yet, they failed to see direct increased phase-locking of the two oscillations compared to sham and control stimulation. One caveat of rhythmic spindle stimulation is that the spindle enhancement does not start with the stimuli itself but a few hundred milliseconds later. The spindle specific and control stimulus both induced a slow oscillation that had spindle activity phase-locked to the up-state of the induced slow wave. However, this enhancement did not promote declarative memory consolidation.

Feedback controlled approaches

A rational design of auditory stimulation will further benefit from feedback-controlled and closed-loop approaches to (1) optimize timing of stimulation by taking underlying network activity into account and (2) allow for more individualized approaches.

The timing seems one of the most essential aspects to consider when performing a rational design of the auditory stimulation during sleep. Thus, the stimulation should be restricted to NREM sleep to avoid unwanted modifications of REM sleep or stimulations during wake. To overcome the need of visual online detection of stages by an experimenter, feedback controlled stimulation systems can help to automatically detect NREM sleep (Debellemaniere et al., 2018; Ferster et al., 2019; Leminen et al., 2017). This is also an important prerequisite to move from sleep lab studies to in-home applications (e.g., Ferster et al., 2019). For slow wave enhancement, phase-targeting should be considered and augmentation of slow waves would be best achieved by targeting the up-phase of the slow waves while avoiding the down-phase. Such phase precision necessitates a closed-loop algorithm that assesses phase in real-time, for instance by implementing a phase-locked loop (e.g., Ferster et al., 2019; Santostasi et al., 2016). For spindles, the timing of the rhythmic stimuli might also be of particular relevance. Ngo et al. (2019) hypothesized that application during slow wave up-states might be less susceptible to rhythmic stimuli because at this time endogenous spindles could be generated in the thalamo-cortical system and may override the immediate response to the stimuli. Whether spindles are blocking the sensory gating of acoustic stimuli has been critically discussed (Dang-Vu et al., 2010; Schabus et al., 2012; Sela et al., 2016) and future research is needed, including mechanistic animal models and computational models, to identify the reason why delayed responses occur.

Closed-loop systems are also particularly helpful to avoid arousal, as this could be a main issue with auditory stimulation during sleep. Thus, closed-loop approaches can identify arousal and as a consequence reduce the intensity of the stimulus. Along the same lines, the stimulus intensity can be enhanced as a function of sleep depth to promote more effective stimulation.

3. Translating sleep modulation into functional outcomes

Cognitive impairments in schizophrenia that have been directly linked to sleep oscillations are deficits in sleep-dependent memory consolidation. On one hand, a decline in procedural memory consolidation (e.g. sequence tapping, mirror tracing) seems to be most consistently reported in these patients regarding sleep-dependent memory deficits (Genzel et al., 2015; Manoach et al., 2010; Manoach and Stickgold, 2019; Wamsley et al., 2012). On the other hand, most of the tES and auditory stimulation studies in healthy participants during sleep have reported effects on declarative memory. Nevertheless, deficits in sleep-dependent declarative memory consolidation have also been reported in patients with schizophrenia. Furthermore, we demonstrated that a targeting of sleep spindles in healthy using tACS might enhance motor memory consolidation (rather than declarative memory, (Lustenberger et al., 2016)). Yet, most of the reported studies regarding non-invasive, non-pharmacological sleep modulation have been performed in healthy controls. The only non-invasive brain stimulation study during sleep in patients with schizophrenia was performed by Göder et al. (2015), They reported that SO-tDCS significantly enhanced declarative memory in patients with schizophrenia, but how sleep modulation was linked to this cognitive improvement was not reported in this study. To date, pharmacological sleep modulation studies might provide more insights regarding the link of sleep modulation and functional outcome in schizophrenia. A few studies tried to modulate sleep (or specifically sleep spindles) alongside cognition. Specifically, Wamsley et al. administered eszopiclone (non-benzodiazepine hypnotic that acts on GABAergic neurons in the TRN) to a small group of patients with schizophrenia (Wamsley et al., 2013). Even though this drug enhanced spindle density and number in these patients compared to a placebo-control group, they did not find an effect on motor sequence memory consolidation. The authors highlighted that this study might have been underpowered to detect an effect on memory (around 10 participants per group), yet a recent abstract reported similar results in a larger trial and a more powerful within-subject design (Baran et al., 2017). This study did not find an effect on motor memory consolidation due to eszopiclone in both patients with schizophrenia and controls. As the authors bring up in their discussions, this missing effect on memory consolidation might be related to a missing temporal link to slow oscillations that could also be necessary for sufficient memory consolidation. Alternatively, as we illustrated in our recent study using closed-loop tACS in healthy controls, specifically the modulation of the very fast spindle frequencies seem to be related to the enhancement of motor memory consolidation (Lustenberger et al., 2016). These findings further highlight that a rational design of brain stimulation should clearly identify, which aspects of sleep spindles (and slow waves) need to be targeted in these patients to optimize functional outcome.

Overall, there is a clear gap for the translation of the observed efficacy of sleep oscillation modulation on functional outcomes in patients with schizophrenia. Besides its translation on sleep-dependent memory, it is unclear how sleep optimization in these patients further benefits cognitive impairments that are observed in their daily life. For instance, patients with schizophrenia suffer from reduction in processing speed, attention or vigilance, working memory, and verbal learning (Green, 2006). Sleep loss and insufficient sleep (e.g. insufficient amount of slow waves) also negatively impact these functions. Furthermore, insufficient sleep in patients with schizophrenia has been linked the severity of negative symptoms (Chemerinski et al., 2002; Kato et al., 1999), lower psychosocial function (Goldman et al., 1996), and lower quality of life (Hofstetter et al., 2005; Ritsner et al., 2004). Along these lines, a recent study in patients with schizophrenia and insomnia administered eszopiclone and reported enhanced working memory and subjective sleep quality (reduced insomnia severity index) compared to a placebo-control group (Tek et al., 2014). However, they did not find an effect on symptoms and did not report PSG data (no information on sleep spindles and other objective sleep markers).

In summary, there is very limited knowledge and research directly linking sleep modulation with functional outcome in patients with schizophrenia. Therefore, in order to fully establish the potential of non-invasive sleep modulation, studies investigating the impact on real-life function, including overall cognition, quality of life and symptom severity, will be needed, both in the short- and the long-term.

4. Conclusion and Outlook

We have presented a brief discussion of the deficits in network oscillations during sleep in people with schizophrenia and delineated the main aspects of rational design of the next generation of brain stimulation protocols to modulate these network targets. We conclude by highlighting conceptual considerations that we believe will require more scientific investigation to advance non-invasive brain stimulation paradigms for restoring sleep network oscillations in schizophrenia.

First, we point out the lack of translation research that can connect the numerous in-vitro studies of electrical stimulation to the flurry of recent human studies. The field needs more mechanistic research in the intact animal, ideally in intermediate model species that share important neuroanatomical and function features with humans, to address the mechanism of action. Establishing the Arnold tongue as the main mechanism has helped to emphasize the importance of frequency-tuning of the stimulation waveform and has provided an explanation of why such weak perturbations can entrain neuronal oscillations. Yet, there are numerous questions that will require more detailed investigations. For example, it remains unknown why repeated application of tACS strengthens network oscillations in a way that these changes persist up to potentially several weeks after conclusion of stimulation.

Second, researchers have recently distinguished two types of slow waves, Type I and Type II slow waves (Bernardi et al., 2018; Siclari et al., 2014) (or sometimes referred to as slow oscillation and delta waves, (Kim et al., 2019)), which are distinct in their spatial distribution and amplitude (but not frequency). Interestingly, Type I slow waves might arise from a bottom-up, subcortico-cortical synchronization and represent high amplitude and global slow waves. These types of slow waves might also be evoked by auditory stimulation (since it is a bottom-up approach). In contrast, Type II slow waves are lower in amplitude and might rather result from cortico-cortical spreading and origin within local cortical regions. These types of slow waves might be modulated by tACS because it directly targets cortical regions. Both types of slow waves might serve different functions. Kim et al. (2019) recently showed in a rat model that they are differentially involved in memory consolidation. Therefore, electrical and auditory stimulation might target different causal mechanisms. Yet, the underlying sources and functional roles of these two types of waves are still largely unknown and further studies are needed to better understand this.

Third, the critical question if the diseased brain responds the same to stimulation as the healthy brain has remained unaddressed. Thus, larger studies that compare the effect of stimulation between healthy controls and patients are urgently needed. An important factor in any such study is of course medication use. The resulting complexity is challenging to unpack. First, presence of pharmacological agents may acutely alter the effect of stimulation. Of note, these effects could be specific to individual medications, which makes their investigation even more difficult. Second, the acute effects of stimulation may not depend on medication. However, the extent to which these acute changes such as entrainment lead to long-term changes by neural plasticity may be altered by medication. Third, chronic use of medications may alter brain circuits in a different way from the actual underlying pathology of the disease itself. Thus, ideally the stimulation would need to address both deficits at the same time.

Fourth, we have here discussed both electrical and auditory stimulation as two separate approaches with distinguishing features. However, it may be the case that the synergistic combination of these two approaches is the most successful one. Given the numerous parameters that need to be considered for the investigation of both parameters in isolation, investigating the interaction will need to be constrained by a mechanistic understanding of the two individual modalities.

Fifth, efficacy of non-invasive brain stimulation during sleep was only tested in a single use manner with different durations, e.g. for a few minutes, during specific sleep cycles or the whole sleep period. Yet, it remains unclear what dose is needed/optimal for sleep oscillation modulation. Along the same lines, we do not know whether these effects can be obtained over longer periods of time (e.g. weeks or month) and might be cumulative. Furthermore, these studies have been performed in well-controlled lab settings, but whether they are effective enough to significantly modulate sleep oscillations in the field (e.g. when variability due to daily life confounders are more pronounced) is not known. It will be of fundamental relevance to answer these questions in order to define the therapeutic potential of non-invasive sleep oscillation modulation in schizophrenia. Thanks to recent advances in portable and mobile solutions, such as for closed-loop auditory slow wave stimulation (e.g. (Ferster et al., 2019)) such studies can be performed in the near future.

Fifth, and finally, sleep represents a unique target for stimulation since we have a clear understanding of the overall patterning of oscillations during the different sleep stages. Activity of the awake brain of course lacks this predictability, which may make successful stimulation in the awake state more difficult. However, we would be amiss not to mention that considering modulation of both sleep and awake network dynamics may provide unique, additional benefits.

In conclusion, sleep network oscillations represent a promising target for non-invasive modulation by targeted brain stimulation approaches. Given the deficits in activity patterns in patients with schizophrenia, this approach may ultimately become of therapeutic value. A long road of research lies ahead of us before the full potential of this modality will be established. We call on the field to collaborate, build bridges between animal and human research, and use best practices of open science to ensure that the gains in understanding will be permanent and not ensnared in the reproducibility crisis in biomedical research.