Abstract
Chronic pain patients frequently suffer from sleep disturbances. Improvement of sleep quality alleviates pain, but neurophysiological mechanisms underlying sleep disturbances require clarification to advance therapeutic strategies. Chronic pain causes high-frequency electrical activity in pain-processing cortical areas that could disrupt the normal process of low-frequency sleep rhythm generation. We found that the spared-nerve-injury (SNI) mouse model, mimicking human neuropathic pain, had preserved sleep-wake behavior. However, when we probed spontaneous arousability based on infraslow continuity-fragility dynamics of non-rapid-eye-movement sleep (NREMS), we found more numerous local cortical arousals accompanied by heart rate increases in hindlimb primary somatosensory, but not in prelimbic, cortices of SNI mice. Closed-loop mechanovibrational stimulation revealed higher sensory arousability in SNI. Sleep in chronic pain thus looked preserved in conventional measures but showed elevated spontaneous and evoked arousability. Our findings develop a novel moment-to-moment probing of NREMS fragility and propose that chronic pain-induced sleep complaints arise from perturbed arousability.
Introduction
Pain causes functional impairment, displeasure and stress and can impede even the simplest daily life routines, including sleep. If not treated, pain has the ability to outlast its original cause, producing chronic pain that is generally difficult to treat (Finnerup et al., 2015; Treede et al., 2019). Current estimates are that more than two out of three individuals suffering from chronic pain also show diverse symptoms characteristic for insomnia disorders, such as lower sleep efficiency, more time awake after sleep onset and frequent brief awakenings during the night (Bjurstrom & Irwin, 2016; Mathias et al., 2018). The relation between chronic pain and sleep disruptions is complex and bidirectional, but accurate assessment of sleep problems is considered critical to antagonize the perpetuation of pain (Bjurstrom & Irwin, 2016). Therefore, key mechanisms associating pain with sleep disturbance need to be clarified.
Animal models of chronic pain that mimic clinical symptoms of human patients have been critical to understand the pathophysiological mechanisms producing chronic pain states (Burma et al., 2017). Neuropathic pain is caused by damage to the somatosensory nervous system (Finnerup et al., 2021) and induced in rodents by surgically lesioning peripheral nerves, such as the sciatic nerve (Decosterd & Woolf, 2000; Bourquin et al., 2006). Neuropathic pain causes maladaptive structural and functional remodeling of the central and peripheral nervous systems, shifting brain circuits towards pain hypersensitivity and aversive behavioral states (Kuner & Kuner, 2020). Hyperexcitability and an abnormal activity in a broad range of gamma frequencies (30—100 Hz) in pain-processing cortical areas were found to be primary culprits for the elevated sensitivity to painful stimuli and for aversive behaviors (Tan et al., 2019), a finding that is in line with observations in human (Ploner et al., 2017). In contrast to these advances, sleep studies on chronic pain models are scarce, used relatively simple sleep measures, and produced variable results (Andersen & Tufik, 2003; Kontinen et al., 2003; Tokunaga et al., 2007; Cardoso-Cruz et al., 2011; Leys et al., 2013). Therefore, it is currently open whether these animal models are also suited to address the sleep complaints of human patients.
One possibility is that current approaches have so far failed to uncover the full profile of the sleep disruptions caused by chronic pain. Studies in insomnia disorders indeed suggest that changes in traditional sleep parameters often seem not in line with the severity of the sleep complaints (Feige et al., 2013; van Someren, 2020). Standard polysomnography describes sleep as a sequence of discrete states and distinguishes between non-rapid-eye-movement sleep (NREMS) and rapid-eye-movement sleep (REMS), with the former further subdivided into transitional (N1), light (N2) and deep (N3) stages (Iber et al., 2007). Many reports on human patients find little change in the absolute or relative times spent in these stages and/or their principal spectral characteristics (Salin-Pascual et al., 1992; Perlis et al., 2001b; Buysse et al., 2008; Wei et al., 2017; Feige et al., 2018; Christensen et al., 2019; Lecci et al., 2020). Instead, cortical activity patterns are abnormally enriched in the alpha (8—12 Hz) (Krystal et al., 2002; Riedner et al., 2016), beta (18—30 Hz) (Krystal et al., 2002; Spiegelhalder et al., 2012; Maes et al., 2014; Riedner et al., 2016; Lecci et al., 2020) and/or low-gamma bands (30—45 Hz) (Perlis et al., 2001b; Lecci et al., 2020), in one or more NREMS stages and/or in REMS (Spiegelhalder et al., 2012; Christensen et al., 2019; Lecci et al., 2020) and/or in restricted brain areas (St-Jean et al., 2012; Riedner et al., 2016; Lecci et al., 2020). Such high-frequency electrical rhythms during sleep are part of a physiological state referred to as “hyperarousal” (Feige et al., 2013; van Someren, 2020; Vargas et al., 2020) that has been related to less restorative sleep (Moldofsky et al., 1975; Krystal & Edinger, 2008), to misperceiving sleep as wakefulness (Perlis et al., 2001b; Lecci et al., 2020), and to higher heart rates (Maes et al., 2014), all of which are key features of insomnia disorders in humans. Other studies applied various metrics and proposed more spontaneous arousals and/or easier wake-ups in response to sensory stimulation (Parrino et al., 2009; Forget et al., 2011; Wei et al., 2017; Feige et al., 2018). Taken together, the presence of high-frequency electrical activity, combined with diverse measures of arousability, has been useful in clarifying the pathophysiological mechanisms underlying insomnia disorders. To date, however, such measures have not been applied to study sleep in chronic pain in humans and mice, and there is still a paucity of comparative studies between sleep in chronic pain patients and in primary insomniacs (Bjurstrom & Irwin, 2016). Therefore, it is currently unclear whether chronic pain is accompanied by high-frequency cortical activity during NREMS and whether this affects spontaneous and evoked arousability (Mathias et al., 2018; Kuner & Kuner, 2020).
This study pursues this question through implementing a real-time tracking method for spontaneous and evoked arousability in the mouse spared-nerve-injury (SNI) model of neuropathic pain (Bourquin et al., 2006). We start from previously described fragility-continuity dynamics of NREMS in mice and humans that indicate variable arousability on the ∼50-sec time scale (Lecci et al., 2017). The fragility-continuity dynamics are present while NREMS remains polysomnographically continuous and manifest in fluctuating activity of several brain and peripheral parameters, notably in the power of sleep spindles (10—15 Hz) in the global EEG and the local field potential (LFP) signals. On this close-to-minute timescale, we show here that we are capable of tracking spontaneous and evoked arousability across NREMS in the resting (light) phase. We find that the sleep disruptions in SNI animals concern both, altered spontaneous and evoked arousability. In particular, we identify a novel, previously undescribed cortico-autonomic arousal that pairs EEG desynchronization with increased heart rate and that occurs more frequently in SNI animals.
Results
SNI mice show normal sleep-wake behavior
SNI mice and Sham controls (n = 18 each) were first analyzed for their sleep-wake behavior using standard polysomnography. EEG/EMG measurements were carried out prior to SNI and Sham surgery and at post-surgical days 22-23 (D20+), a time point at which chronic pain is established (Decosterd & Woolf, 2000; Bourquin et al., 2006). At both time points, recordings lasted for 48 h under undisturbed conditions. Based on these data, SNI and Sham controls spent similar amounts of time asleep in the 12 h-light and dark phases, during both baseline and at D20+ (Figure 1A). Both treatment groups showed minor increases (2.4-3 %) in NREMS time at the expense of wakefulness at D20+ compared to baseline in both dark and light phases (mixed-model ANOVAs with factors ‘treatment’ and ‘day’, p = 0.0013 in light phase, p = 8.5×10−6 in dark phase for ‘day’, p > 0.8 for ‘treatment’ for either light or dark phase, no interaction). Moreover, cumulative distributions of NREMS and REMS bout lengths at D20+ were similar for Sham and SNI, with only a minor shift toward smaller values in SNI for both NREMS (Figure 1B, −2.3 s; Kolmogorov-Smirnov (KS) test, p = 0.015) and REMS (−3 s; KS test, p = 0.026). When subdivided into short, intermediate and long bouts for the light phase, there were no significant differences between Sham and SNI for both NREMS and REMS (Figure 1B, mixed-model ANOVAs for ‘treatment’ x ‘bout length’, p = 0.79 for NREMS, p = 0.23 for REMS).
Furthermore, sleep onset latency (Figure 1C) and NREMS fragmentation by brief movement-associated microarousals (MAs, defined in mouse as <= 16 s awakenings accompanied by movement activity seen in the EMG, measured over 48 h) (Figure 1D) (Franken et al., 1999), were not altered by treatment or time post-surgery (mixed-model ANOVA with factors ‘treatment’ and ‘day’, for sleep onset latency, p = 0.42 and p = 0.94, no interactions; for number of MAs, p = 0.79 and p = 0.43, no interactions).
We next investigated the mean spectral properties of each vigilance state through constructing normalized power spectral densities (Vassalli & Franken, 2017) for the full 48 h-long recordings. Both NREMS and REMS showed the respective characteristic spectral peaks at delta (1—4 Hz) and at theta frequencies (5—10 Hz), respectively. These were indistinguishable between the two groups of animals and from baseline to D20+ (Figure 1E).
We specifically evaluated power in the high-gamma frequency (60—80 Hz) range, a frequency band linked to pain sensations when optogenetically induced in mouse (Tan et al., 2019). We found that relative gamma power was increased in SNI at D20+ compared to baseline both in wake and in NREMS (Figure 1F, 1-sample t-test for wake and NREMS in SNI, p = 0.011 and p = 0.0092, respectively). The heart rate was also higher in NREMS of SNI animals at D20+ compared to baseline (Figure 1G, mixed-model ANOVA with factors ‘treatment’ x ‘state’ x ‘day’ with interaction, p = 0.02, post-hoc paired t-test for SNI in NREMS, p = 0.002, with Bonferroni-corrected α = 0.0125). A tendency was also evident in REMS, during which heart rate was already elevated (Figure 1G, effect of ‘state’ in the ANOVA, p = 0.003, paired t-test in SNI in REMS, p = 0.027). There were no correlations between relative changes in gamma power and alterations in sleep architecture in individual mice (change in the number of MAs per h of NREMS x change in gamma power; pairwise linear correlation R2 = 0.09, p = 0.08; change in total NREMS time x change in gamma power; R2 = 0.02, p = 0.36).
These data indicate that SNI animals do not suffer from major alterations in sleep-wake behaviors. Still, pain-related pathological changes in brain and periphery continued to be present in sleep. This is consistent with a state of “hyperarousal” whereby high-frequency power components are disproportionately elevated during sleep that is normally dominated by low-frequency rhythms (van Someren, 2020; Vargas et al., 2020), and where heart rate also remains elevated. As such alterations could affect arousability, we asked when and where in the brain this abnormal activity appeared. Furthermore, we developed an approach to systematically quantify alterations in both, spontaneous and evoked, types of arousability from NREMS during the resting phase.
The 0.02 Hz-fluctuation allows to probe variations in spontaneous arousability during NREMS
Arousability in sleeping rodent, measured via external stimuli or through spontaneous arousals, changes across the night (Neckelmann & Ursin, 1993; Wimmer et al., 2012), and with variations in sleep pressure (Franken et al., 1999). For NREMS in early phases of the resting phase, we described a 0.02 Hz-fluctuation during NREMS that provides a minute-by-minute time raster to measure arousability driven by sensory stimuli. This fluctuation subdivides NREMS bouts into ∼25 s-long periods of continuity and fragility that show low and high sensory-evoked arousability, respectively (Lecci et al., 2017; Yüzgeç et al., 2018). To evaluate the utility of this fluctuation for measures of spontaneous arousability across the entire light phase, we first tested whether MAs associated with muscular activity (Figure 2A), well-established correlates for spontaneous arousability, were phase-locked to the 0.02 Hz-fluctuation in healthy mice (n = 30 mice with 9,476 MAs). The onset of MAs coincided with declining or low sigma power levels that followed a pronounced sigma power peak (Figure 2B, C), which is characteristic for a fragility period (Lecci et al., 2017; Fernandez & Lüthi, 2020). A spectral band typical for NREMS, such as delta (1—4 Hz) power, showed a rapid decline preceding the MAs, indicating the momentary interruption of NREMS. The phase values of the 0.02 Hz-fluctuation, calculated via a Hilbert transform (Figure 2 – figure supplement 1), showed that MA onset times clustered around a mean preferred phase of 151.6° ± 1.1°, with 180° representing the sigma power trough (Rayleigh test, p < 1×10−16). The majority of MAs (89 %) was clustered between 90—270°, which narrows the fragility period to the low values of sigma power around the trough (Lecci et al., 2017). The phase-locking was also observed when time points at 4, 8 and 12 s before the onset of a MA were quantified (Figure 2D). This shows that the onset of the fragility period preceded the MA. Fragility periods thus constitute moments during which MAs preferentially occur.
These phase relations persisted for all 1-h intervals across time-of-day (Figure 2E), although the density of MAs showed a characteristic increase towards the end of the light phase and was higher during the dark phase (Figure 2F). The peak frequency of the 0.02 Hz-fluctuation also remained relatively constant, with a minor decrease in power during the dark phase (Figure 2G). Across the 24-h cycle, a median of 33.6 % of all fragility periods were accompanied by a MA (Figure 2H). In sum, fragility periods are permissive windows for MAs. This means that MAs appeared predominantly during fragility periods, while a majority of fragility periods occurred with NREMS remaining consolidated.
NREMS in SNI conditions shows normal phase-coupling of MAs to the 0.02 Hz-fluctuation
We next evaluated SNI and Sham animals regarding the MAs and their coupling to the 0.02 Hz-fluctuation. The 0.02 Hz-fluctuation was not different between Sham and SNI (n = 18 for both groups) across the light phase. Thus, neither its amplitude nor frequency (Figure 3A-C), or, equivalently, the number of its cycles per h of NREMS, were different between the groups (Figure 3D). The phase-coupling to MAs was also unaltered (Figure 3E, mean angle ± 95% CI: 152.3 ± 1.4 for Sham and 150.4 ± 1.3 for SNI) and the distribution of fragility periods containing transitions to MAs, to REMS, or with continuation into NREMS was indistinguishable (Figure 3F).
It has been shown that sleep loss exacerbates pain (Alexandre et al., 2017). Sleep could thus be relatively more disrupted in SNI animals after a period of sleep loss. We therefore carried out a 6 h-sleep deprivation (SD) at the beginning of the light phase as done previously in the lab (n = 12 for Sham and SNI each) (Kopp et al., 2006). We confirmed a characteristic rebound of delta power (Figure 3G,H) and a decrease in the frequency of MAs (Figure 3I,J, mixed-model ANOVA with factors ‘treatment’ and ‘SD’, p = 0.35 and p = 1.23×10−7 with no interaction). The phase-coupling of MAs to the 0.02 Hz-fluctuation remained unaltered in both groups even with high sleep pressure (Figure 3K,L). Conditions of SNI thus left spontaneous MAs, their coupling to the 0.02 Hz-fluctuation, as well as homeostatic regulation of spontaneous arousability unaltered.
NREMS in SNI conditions shows a novel type of cortical local arousal
In human NREMS, spontaneous arousals are an important measure for the severity of sleep disorders and are primarily described by EEG desynchronization (Bonnet et al., 1992; Azarbarzin et al., 2014). In contrast, the scoring of a MA in mice requires concomitant muscular activity by convention (Franken et al., 1999). We hence tested whether the 0.02 Hz-fluctuation could serve to identify previously undescribed arousal types in mice with characteristics distinct from conventional MAs. For this, we generated spectral profiles of all cycles of the 0.02 Hz-fluctuation that were devoid of MAs. To take into account the possibility that there were local events delimited to certain cortical regions (Nobili et al., 2011; St-Jean et al., 2012; Riedner et al., 2016; Lecci et al., 2017), we combined polysomnography with stereotaxically guided local field potential (LFP) recordings, as done previously in the lab (Figure 4A,B) (Fernandez et al., 2018). We chose the S1 hindlimb (S1HL, 5 Sham and 9 SNI) cortex (Figure 4C,D) that is the site of sensory discrimination of pain and the prelimbic (PrL, 6 Sham and 8 SNI) cortex (Figure 4I,J) that is concerned with aversive pain feelings in rodents (Kuner & Kuner, 2020) and in its homologue in humans (Moisset & Bouhassira, 2007).
Local field potential recordings reliably reported on the 0.02 Hz-fluctuation in these two areas. Consistent with its predominant expression in sensory cortices (Lecci et al., 2017), the 0.02 Hz-fluctuation showed a higher peak in S1HL than in PrL (Figure 4A,B). The cycles of successive continuity and fragility periods were extracted (Figure 2, figure supplement 1) and their spectral dynamics plotted separately for the relative contribution of power in the low-frequency delta (1—4 Hz) and the beta (16—25 Hz), low-(26—40 Hz) and high-(60—80 Hz) gamma bands (Figure 4E-H for S1HL, Figure 4K-N for PrL). Average values for the infraslow phase angles between 90—270°, corresponding to the fragility period enriched in MAs (see Figure 2), and for the continuity period (from 270—90°), were calculated. Such analysis revealed SNI- and region-specific alterations in the contributions of these bands to total power that were clearly present in S1HL, but not detectable in PrL. In S1HL, delta power levels were decreased compared to Sham (Figure 4E) whereas high-frequency components in the beta and the low-gamma range were elevated (Figure 4F-G). Remarkably, delta power differences between Sham and SNI varied between fragility and continuity periods (Figure 4E, mixed-model ANOVA with factors ‘treatment’ and ‘period’, p = 0.001 for the interaction). In Sham, there was a distinct rapid upstroke of power in this frequency band that reached a peak during the fragility periods (Figure 4E, post-hoc t-test for delta power in fragility vs continuity period in Sham, p = 0.001). Fragility periods continuing into NREMS were thus clearly distinct from the ones associated with MAs during which there is muscular activity and a decrease in EEG delta power (see Figure 2C). In SNI animals, in contrast to Sham, there was no detectable elevation in delta power during fragility periods continuing into NREMS (Figure 4E, post-hoc t-test in SNI, p = 0.44). The high-frequency bands in the beta and low-gamma range instead showed a tonic increase in SNI that was present throughout continuity and fragility periods (Figure 4F-G, mixed-model ANOVA with factors ‘treatment’ and ‘period’, for beta, p = 0.01, p = 1.3×10−9 and for low gamma, p = 0.0053, p = 1.4×10−10) and that was also present, although to a milder extent, in the high-gamma range (Figure 4H).
We calculated an “activation index” (AI), defined by the ratio between the summed spectral power in the beta and low-gamma bands and the delta band power, to quantify alterations in spectral balance between high- and low-frequency power components, similarly to what has been done previously in studies on insomnia disorders (Lecci et al., 2020). The AI is a measure for the degree of EEG desynchronization and increases when NREMS moves closer to wakefulness. In the fragility periods continuing into NREMS and devoid of EMG activity, the AI decreased, consistent with NREMS remaining consolidated (Figure 5A-C). In SNI animals, however, the AI was higher compared to Sham specifically in the fragility periods (Figure 5B, mixed-model ANOVA with factors ‘treatment’ and ‘period’, p = 0.039 for interaction, post-hoc t-tests Sham vs SNI in fragility period, p = 0.005, in continuity period, p = 0.027, not significant with α = 0.0125). Fragility periods during uninterrupted NREMS are thus specific moments during which the AI in SNI conditions was significantly higher compared to continuity periods.
Can such mean differences in cortical activation profiles during NREMS qualify as differences in cortical arousals? To address this, we compared the AI in fragility periods with a MA (associated with EMG increase). As expected, the AI showed an intermittent phasic peak (Figure 5D-F) in most cases (75.2 ± 4.1 %), which is explained by the strong decline in delta power (see Figure 2C) and the appearance of higher frequencies associated with MAs. Therefore, we inspected individual fragility periods continuing into NREMS (without a MA) for the presence of similar phasic increases in AI. Indeed, we noticed that a subset of these did indeed contain an intermittent peak resembling the one found during MAs (Figure 5G) and not evident in the mean AI in Figure 5B. The amplitudes of these peaks were higher in SNI, in accordance with the tonically higher AI in these animals, but in size comparable to the ones of MAs (Figure 5H, mixed-model ANOVA with factors ‘treatment’ and ‘MA’, p = 0.002 for ‘treatment’, p = 2.1×10−7 for ‘MA’, no interaction). Moreover, the half-widths of these peaks were only moderately smaller than the ones of MAs (Figure 5I, mixed-model ANOVA with factors ‘treatment’ and ‘MA’, p = 0.62 for ‘treatment’, p = 3.28×10−7 for ‘MA’, no interaction). These events could thus qualify as a local cortical arousal based on phasic spectral properties reminiscent of a MA. To further support our assumption that these AI peaks constituted arousals, we looked at heart rate increases known to accompany cortical arousals in human (Sforza et al., 2000; Azarbarzin et al., 2014). The heart rate was distinctly higher during the fragility period for cycles containing an AI peak as opposed to the ones without such peak (Figure 5J, mixed-model ANOVA with factors ‘treatment’, ‘period’ and ‘peak’, p = 0.007 for the ‘peak’ x ‘period’ interaction). These events were more frequent in SNI animals and followed a similar time-of-day dependence as the classical MAs (Figure 5K,L, t-test Sham vs SNI, p = 0.02). Moreover, their increased occurrence was specific for S1HL while absent in PrL and in the contralateral EEG (Figure 5 – figure supplement 2). The presence of a subgroup of fragility periods continuing into NREMS, yet showing a cortical arousal, is noteworthy for several reasons. First, it demonstrates that rodent NREMS shows local cortical intrusion of wake-related activity in the absence of muscular activity. Second, these local cortical arousals in SNI showed intermittent peaks in AI that were close the ones of MAs, indicating comparable cortical desynchronization at the local level. Third, they were accompanied by heart rate increases that are sensitive hallmarks of arousal in human (Azarbarzin et al., 2014). Fourth, neuropathic pain goes along with a specific increase in the relative occurrence of fragility periods with such AI peaks specifically in the S1HL area. The systematic classification of fragility periods helped unravel these novel cortico-autonomic arousals and their similarity to MAs. Still, other arousal-like events outside fragility periods could exist.
The 0.02 Hz-fluctuation allows to anticipate elevated spontaneous arousability during NREMS
We finally examined sensory arousability in SNI conditions, focusing on the somatosensory modality. To anticipate fragility and continuity periods in real-time in the sleeping animal, we trained a machine learning software to predict online periods of continuity and fragility based on EEG/EMG recordings (Figure 6A-E). For the training, we used online-calculated 0.02 Hz-fluctuation estimates onto which fragility and continuity periods were labelled using peak-and-trough detection of sigma power dynamics (Figure 6 – figure supplement 3). To control for the accuracy of the online prediction, we visually scored MAs in 12 C57Bl/6J animals implanted only for polysomnography and verified their position in either online detected peak-to-trough (‘online fragility’) or trough-to-peak phases (‘online continuity’) (Figure 6F). We compared the online prediction to that generated by chance through randomly shuffling both online fragility and continuity point positions in the recordings. This showed that the MA proportions obtained with the real detection exceeded those obtained by chance prediction (Figure 6G, for online fragility periods, p = 0.0004, for online continuity periods, p = 0.0028). Online detection of peak-to-trough and trough-to-peak periods of the 0.02 Hz-fluctuation is thus a versatile method to probe variations of evoked arousability from NREMS.
SNI conditions produce elevated somatosensory-evoked arousability from NREMS
Evoked arousability was probed through applying sensory stimuli either during online detected fragility or continuity periods. To deliver somatosensory stimuli remotely while the animals were asleep, we attached vibrational motors to their head implant that could be triggered to briefly vibrate (for 3 s) to test the chance for wake-up (Figure 7A). These motors were calibrated to vibrate with the same low intensity (∼ 30% of full power) across animals (Figure 7 – figure supplement 4). Vibrations were applied randomly with 25% probability during either online detected fragility or continuity periods, for at least two complete light phases per condition (Figure 7B). Intensity was chosen such that Sham animals showed approximately equal chances for wake-up or sleep-through in online continuity periods (Figure 7C). Moreover, these vibrations produced wake-ups that were short, indicating that the sleeping animal felt only mildly perturbed. Consistent with prior findings, similar stimuli applied during online fragility periods showed consistently higher chances for wake-up (Lecci et al., 2017). In SNI, sensory arousability was elevated for both continuity and fragility periods, leading to highest values during the online fragility periods (Figure 7D, mixed-model ANOVA with factors ‘treatment’ and ‘online period’, p = 0.0049 for ‘treatment’, p = 1.31×10−8 for ‘online period’, no interaction). Interestingly, consistent with the tonic increase in AI, this increase in sensory arousability in SNI was present across the whole light phase with a conserved time of day dependence (Figure 7E).
Discussion
Chronic pain is a widespread and complex condition compromising sleep. As poor sleep further aggravates pain, therapeutic approaches to improve sleep quality have potential to attenuate disease progression. Here, in efforts to tease apart pain-sleep associations, we decided to focus on mechanisms of sleep disruptions at early stages of chronic pain. This study progresses on the sleep-pain association in four essential ways. First, we show that brain and autonomic signatures of pain states during the day intrude in a persistent manner into sleep in early phases of the disease. Second, sleep appeared nevertheless preserved in architecture, in dominant spectral band power, and in homeostatic regulation. Third, a previously undescribed spontaneous arousal during NREMS, showing local cortical activation with concomitant heart rate increases, appeared more frequently in SNI conditions. Fourth, we also demonstrate that fine mechanovibrational stimuli triggered brief wake-ups from NREMS more easily in SNI animals. In summary, chronic pain impacts on NREMS in terms of arousability, more specifically in the probability that NREMS transits towards diverse levels of wakefulness, either spontaneously, or with external stimuli. Chronic pain additionally instates a tonic regional elevation of high-frequency electrical activity. Both these consequences are not detectable in conventional measures of sleep. However, they bear resemblance to pathophysiological markers of insomnia disorders and, as we show here, produce elevated responsiveness to fine vibrational stimuli. The study further suggests that, amongst the many peripheral and central circuit alterations caused by chronic pain, the ones affecting the primary nociceptive sensory areas could serve as a site of entry to treat pain-related sleep disturbances directly at early states of the disease. For example, targeted interference by transcranial stimulation techniques has been proposed to modulate pain-related oscillations and could thus be probed to modify abnormal arousals (Shirvalkar et al., 2018; Hohn et al., 2019).
The quantification and classification of arousals from NREMS, both in terms of their physiological correlates and in their intensity, is central to estimate the severity of a sleep disorder. The fragmentation of sleep by arousals is the primary cause for daytime fatigue and for cognitive deficits and, in more severe cases, may constitute long-term risks for cardiovascular health (Bonnet et al., 1992; Silvani, 2019). Criteria for arousal scoring in humans are comparatively well established and there are strong indications that arousal intensity is graded, with EEG desynchronization and concomitant heart rate acceleration occurring independently of muscular activity (e.g. leg movements) (Sforza et al., 2000; Azarbarzin et al., 2014). In chronic pain patients, few systematic analyses on spontaneous arousals are currently available and a need for more polysomnographic assessments in well-controlled patient populations has been highlighted (Bjurstrom & Irwin, 2016). In rodents, only few studies have described cortical desynchronization events without EMG activity, and these were not characterized with respect to autonomic correlates (Bergmann et al., 1987; Franken, 2002; Léna et al., 2004; Fulda, 2011). This relative lack of arousal characterization in mouse NREMS could have hampered the identification of models to replicate sleep disorders, as we found here to be the case for chronic pain models. To the best of our knowledge, our analysis of the SNI mouse is the first that qualifies as a rodent model replicating physiological correlates of insomnia disorders that are hidden behind a comparatively normal sleep and that could raise awareness for refined analysis of the diverse forms of sleep disruptions in chronic pain patients (Bjurstrom & Irwin, 2016). Our work also adds a novel variant to proposed insomnia models that provoked severe macroscopic sleep disruptions either through acute stress (Cano et al., 2008; Li et al., 2020) or by optogenetically enforced full awakenings to fragment NREMS (Rolls et al., 2011).
This study proposes a novel type of arousal from NREMS in mouse that pairs cortical desynchronization with heart rate increases. Inclusion of this type of arousal is crucial to identify the exact sleep disruptions for the case of neuropathic pain. We departed here from previous observations on infraslow variations in sensory arousability during NREMS that take place over the minute time-scale (Lecci et al., 2017). We first demonstrated that spontaneous MAs, the only spontaneous arousal in rodent for which scoring criteria are widely established, occur remarkably clustered at phases for which sensory evoked arousals were most likely (Lecci et al., 2017). The high number of MAs and their remarkable clustering at phases > 90° and < 270°, allowed us to allocate the fragility periods to the phases with low sigma power. We consider this basic finding on MA timing in rodent NREMS significant for several reasons. First, it suggests that the infraslow 0.02 Hz-fluctuation is part of an overarching process that periodically sets a fragility of NREMS towards wake-promoting inputs, whether they arise from internal processes or external stimuli. Mechanistically, this points to an involvement of widely projecting neuromodulatory brain areas such as the locus coeruleus that remains active during NREMS (Aston-Jones & Bloom, 1981; Kjaerby et al., 2020) and regulates sensory arousability (Hayat et al., 2020). Second, the finding contributes to a long-standing uncertainty about the origin and the stochastic time-scales on which MAs are thought to occur (Lo et al., 2004; Dvir et al., 2018). We now calculate that MAs in consolidated NREMS occur with a mean ∼35% probability on ∼50-sec intervals, thus providing a temporal raster on which these can be anticipated with a measurable degree of certainty. MAs depend on activity in wake-promoting brain areas (Dvir et al., 2018), specifically on the histaminergic hypothalamus in mice (Huang et al., 2006) and on cholinergic nicotinic receptors (Léna et al., 2004). This mechanistic origin of MAs is consistent with our observation that they occur at moments of NREMS during which sensory wake-ups occur preferentially. We therefore consider the identification of fragility periods as time raster for MAs as key to reinforce mechanistic investigations into the origins of spontaneous arousals. This not only concerns MAs, but opens the opportunity to search for other arousal-like events that could be relevant to model pathological conditions of human patients.
The majority of this study was dedicated to providing a proof-of-concept for the usefulness of the infraslow fragility periods to scrutinize arousability. The chosen SNI model appeared particularly appropriate for this purpose because it produced a sensory deficit that could be exploited to specifically test somatosensory arousability. The separation of NREMS into fragility and continuity periods throughout the resting phase allowed us to sample and scrutinize the many fragility periods continuing apparently uninterrupted into NREMS. The fragility periods were also critical to determine when AI became most disparate between SNI and Sham and to identify previously undescribed cortico-autonomic arousals. During these, the activation index increased because of a phasic decrease in low-frequency delta power and an increase in high-frequency power. These events were more pronounced and more frequent in SNI because both these phasic power alterations were disrupted, with the deficits in delta power most pronounced. Without the raster provided by the fragility periods, the phasic differences amidst the tonically elevated high-frequency power would easily have gone undetected. To further ascertain that these detected events nevertheless constituted true arousals rather than accidental spectral fluctuations, we sought for independent physiological correlates. Inspired by the human literature (Sforza et al., 2000; Azarbarzin et al., 2014), we found that heart rate increases were consistently higher when calculated for cortical arousals with AI peak than for the ones without AI peak. Moreover, their distribution across the resting phase was similar to the one found for MAs and they were present in both S1HL and PrL. This result supports our interpretation that we have identified here a novel cortical-autonomic arousal subject to similar time-of-day-dependent regulatory mechanisms. Still, we cannot exclude that arousal subtypes outside the fragility periods went undetected that would require further characterization. We also remark here clearly that, aside from more frequent cortico-autonomic arousals, SNI animals suffered from a tonically elevated high-frequency power in S1HL that likely underlay the more elevated sensory arousability throughout continuity and fragility periods.
The lack of major sleep disruptions in the SNI mouse model was initially unexpected but seemed in line with other studies. We analyzed these animals at a time point when pain from the wound and associated inflammations are largely over (Guida et al., 2020), both of which can strongly disrupt sleep (Landis et al., 1989; Andersen & Tufik, 2003; Silva et al., 2008). Moreover, the animals showed a preserved time spent in REMS, suggesting that they did not suffer from chronic mild stress-inflicted sleep disruptions (Nollet et al., 2019). Other studies on chronic pain also report diverse moderate effects on sleep (Kontinen et al., 2003; Leys et al., 2013). One study on rats at 2 and 10 days after SNI surgery suggested that brain states intermediate between NREMS and wakefulness during the resting phase exist (Cardoso-Cruz et al., 2011), which could in part reflect our observations. We also found no alterations in theta power or for shifts in theta peaks in wakefulness, as reported for other animal models of chronic pain (LeBlanc et al., 2014) or for humans with severe neurogenic pain or arthritis (Sarnthein et al., 2006). Analysis of sleep disruptions at later stages in the disease will help decide whether distinct phases of sleep disruptions mark distinct phases of pain chronicity when anxiety- and depression-related behaviors appear more strongly (Guida et al., 2020).
The spectral dynamics in two cortical regions we present here delineate possible areas of pathological neuronal activity that underlie the cortical arousals. The continued presence of high-frequency activity in S1HL is reminiscent of the cortical oscillatory activity evoked with acute painful stimuli, suggesting that nociceptive input continues to arrive in cortex during NREMS to generate excessive excitation. Indeed, it has been suggested that the SNI model does show spontaneous ectopic electrical activity in peripheral sensory neurons as a result of nerve injury (Wall & Devor, 1983; Devor, 2009). NREMS is thought to protect relatively weakly from nociceptive inputs (Claude et al., 2015), therefore possibly allowing continued processing of spontaneous nociceptive activity that could explain the cortical spectral changes we detected. It has also been shown that optogenetic stimulation of the thalamic reticular nucleus, known to be implied in the balanced occurrence of delta and spindle waves during NREMS (Fernandez et al., 2018), can alleviate pain in SNI (LeBlanc et al., 2014). Suppressed TRN activity during NREMS could be implied in the attenuated delta dynamics observed in NREMS of SNI mice. In contrast, we found unperturbed local spectral dynamics in PrL during NREMS, although this area is concerned with signaling emotional discomfort in several forms of chronic pain in humans (Schulz et al., 2015; Nickel et al., 2017; May et al., 2019) and is known to undergo strengthened synaptic inhibition in SNI (Zhang et al., 2015; Radzicki et al., 2017). Further cellular studies will be necessary to understand why these alterations seem not to perturb oscillatory activity in this area during NREMS.
Do SNI animals suffer from insomnia? Our objective measures of NREMS’s spectral composition point to regionally restricted but tonic imbalances in the contribution of low-vs higher frequencies. Patients with insomnia show such imbalances over widespread brain regions that include sensorimotor areas (Lecci et al., 2020). Furthermore, higher power in the beta frequencies has been related to the patients remaining hypervigilant or excessively ruminating at sleep onset (Perlis et al., 2001a), preventing the deactivation of cortical processes required for the loss of consciousness. Although insomnia also needs subjective assessments that are not possible in animals, this phenomenological comparison suggests that SNI might suffer from similar experiences due to the tonically enhanced high-frequency oscillations. This interpretation is supported by the elevated wake-up rates in response to mild vibrational stimuli throughout the infraslow cycles, suggesting hyperalertness to environmental disturbance. On top of these tonic changes, there were more frequent cortico-autonomic arousals. Although these do not seem to elevate daytime sleepiness based on the mostly unchanged delta power dynamics across time-of-day, frequent increases in heart rate during the night could augment cardiovascular risk in the long-term (Silvani, 2019). To further analyze the animal’s conditions during daytime, tests on their cognitive abilities in memory-dependent tasks while locally manipulating sleep in the affected hindlimb area could be considered. Deficits in working and declarative memories in rodents with SNI have been documented from early periods of chronic pain (Guida et al., 2020). Chemogenetic manipulation of neuronal populations proposed to be responsible for the gamma activity in chronic pain, restricted to sleep periods (Tan et al., 2019), seems a feasible approach to specifically suppress abnormal pain-related activity during sleep while testing performance in such tasks during wakefulness.
We provide here novel approaches to classify arousals in mouse NREMS that will help in the examination and validation of future candidates for rodent models of sleep disorders. We noted a remarkable stability of the 0.02 Hz-fluctuation across the resting phase that provided us with a temporal raster to screen the characteristics of fragility periods. These led us to identify a previously undescribed cortico-autonomic arousal in mouse NREMS that we also found more frequently in a chronic pain model. Together, this study presents NREMS as a state that is interwoven with arousals showing diverse combinations of physiological parameters with different graduations in intensity that can be the target of pathophysiological changes. Recognizing NREMS as a fluctuating state between fragility and continuity will thus further heighten awareness to arousability as a core component of sleep quality. In this study, we unraveled a so far undescribed sleep disruption in chronic pain that we hope will facilitate further research into the treatment of this devastating condition.
Materials and methods
Animal housing and experimental groups
Mice from the C57BL/6J line were singly housed in a temperature- and humidity-controlled environment with a 12-h/12 h light-dark cycle (lights on at 9:00 am, corresponding to ZT0), with access to food and water ad libitum. We first used 36 mice, 10-14 weeks-old and bred in our colonies in a conventional-clean animal house, for polysomnography (combined EEG (ECoG)/EMG electrodes), followed by SNI or Sham surgery (18 animals per Sham or SNI group). Mice were transferred from the animal house into the recording room 2-3 d before surgery for polysomnography recording. We recorded a 48 h-long baseline before SNI or Sham surgeries, followed by recording at 22-23 d after surgery (D20+). These data were used for Figures 1 and 3. Total sleep deprivation in Figure 3 was done on 24 of these 36 animals (12 SNI, 12 Sham) within one day following the recording at D20+. The baseline data for Figure 2 were obtained from the baseline recordings of 23 randomly selected animals from the previous 36, completed with 7 more animals from previous baseline recording in the lab. For EEG (ECoG)/EMG/LFP recordings, 33 C57BL/6J male mice of the same age were first operated for SNI or Sham (17 and 14, respectively) and 5 d later, implanted for recordings from S1HL (4 Sham, 6 SNI) or PrL (3 Sham, 4 SNI) or both (3 Sham, 4 SNI). The misplaced or non-functional electrodes were excluded. Recordings were carried on from day 20 to 35 after SNI or Sham surgery. These data were used for Figures 4 and 5. The data of 13 animals previously recorded in the lab and otherwise not included in any dataset in this study were used to train the neural network (EEG/EMG implantation, in Figure 6). The experiments on sensory evoked arousals (Figure 7) were done on 16 animals (8 Sham, 8 SNI) out of which some (4 sham, 6 SNI) were used for Figures 4-5. All experimental procedures complied with the Swiss National Institutional Guidelines on Animal Experimentation and were approved by the Swiss Cantonal Veterinary Office Committee for Animal Experimentation.
Surgery for the SNI model of neuropathic pain
The Sham and SNI surgeries were performed as previously described (Decosterd & Woolf, 2000). Briefly, mice were kept under gas anesthesia (1—2 % isofluorane, mixed with O2). The left hindleg was shaved and the skin incised. The muscles were minimally cut until the sciatic nerve was exposed. Just below the trifurcation between common peroneal, tibial and sural branches of the nerve, the common peroneal and tibial branches were ligated and transected. The Sham animals, as controls, went through the same surgery without the transection. The muscle and the skin were then stitched closed and the animals were monitored via a score sheet established with the Veterinarian Authorities.
Surgery for polysomnographic and LFP recordings in mice
Surgeries were performed as recently described (Lecci et al., 2017; Fernandez et al., 2018). Animals were maintained under gas anesthesia (1—2 % isofluorane, mixed with O2). Small craniotomies were performed in frontal and parietal areas over the right hemisphere and 2 gold-plated screws (1.1 mm diameter at their base) (Mang & Franken, 2012) were gently inserted to serve as EEG electrodes. Careful scratching of the skull surface with a blade strengthened the attachment of the implant by the glue, so that additional stabilization screws were no longer necessary. Two gold wires were inserted into the neck muscle to serve as EMG electrodes. In the case of LFP recordings, small craniotomies (0.2-0.3 mm) were performed to implant high-impedance tungsten LFP microelectrodes (10–12 MΩ, 75-μm shaft diameter, FHC, Bowdoin, ME) at the following stereotaxic coordinates relative to Bregma in mm, for S1HL: anteroposterior −0.7, lateral −1.8, depth from surface −0.45; for PrL: +1.8, −0.3, −1.45). For the neutral reference for the LFP recordings, a silver wire (Harvard Apparatus, Holliston, MA) was placed in contact with the bone within a small grove drilled above the cerebellum. The electrodes were then soldered to a female connector and the whole implant was covered with glue and dental cement. The animals were allowed 5 d of recovery, while being monitored via a score sheet established with the Veterinarian Authorities, with access to paracetamol (2mg/mL, drinking water). The paracetamol was removed when the animals were tethered to the recording cable for another 5 d of habituation prior to the recording.
Polysomnographic recording
For sleep recordings, recording cables were connected to amplifier boards that were in turn connected to a RHD USB interface board (C3100) using SPI cables (RHD recording system, Intan Technologies, Los Angeles, CA). For EEG/EMG and/or LFP electrodes, signals were recorded through homemade adapters connected to RHD2216 16-channel amplifier chips with bipolar input or RHD2132 32-channel amplifier chips with unipolar inputs and common reference, respectively. Data were acquired at 1000 Hz via a homemade Matlab recording software using the Intan Matlab toolbox. Each recording was then visually scored in 4-s epochs into wake, NREMS, REMS, as described (Lecci et al., 2017) using a homemade Matlab scoring software.
Total sleep deprivation protocol
Total SD was carried out from ZT0—ZT6 using the gentle handling method used previously in the lab (Kopp et al., 2006), while animals remained tethered in their home cage. At ZT3, the cages were changed and, from ZT5 to ZT6, new bedding material was provided. At ZT6, the animals were left undisturbed. The recordings carried out during SD were visually scored to assure the absence of NREMS from ZT0 to ZT6. There were no detected NREMS epochs during SD in the mice included in the analysis.
Probing sensory arousability with vibration motors and automatic wake-up classification
An online detection of continuity and fragility period (described below) was used in a closed-loop manner to time vibration stimuli during NREMS such that sensory arousability could be probed. Small vibrating motors (DC 3—4.2 V Button Type Vibration Motor, diameter 11 mm, thickness 3 mm) were fixed using double-sided tape, at the end of the recording cables, close to the animals’ heads. The motors were driven using a Raspberry Pi 3B+ through a 3.3 V pulse-width modulation (PWM) signal. Each motor was calibrated to find the necessary PWM duty-cycle to output the same amount of mild vibration using a homemade vibration measurer equipped with a piezo sensor. A Python script was running on the Raspberry Pi to detect the voltage change sent by the digital-out channels on the Intan RHD USB interface board. Upon detecting a change from low to high, the Python script waited for an additional 4 s, and assessed the voltage again. In case the voltage was still high, it launched a 3 s vibration with 25% probability. To close the loop, the PWM signals from the Raspberry Pi driving the motors were as well fed into the analog-in channels of the Intan RHD USB interface board to detect the stimuli time-locked to the EEG/EMG signals. In the experiments, the voltage values were set to high during either continuity or fragility, using online detection as described in Data analysis. Four animals could be tested in parallel for their sensory arousability.
Histological verification of recording sites
After the in vivo LFP recordings, the animals were deeply anesthetized with pentobarbital (80 mg/kg) and electrode positions were marked through electro-coagulation (50 µA, 8—10 s). The animals were then transcardially perfused with 4 % paraformaldehyde (in 0.1 M phosphate buffer). After brain extraction and post-fixation for 24 h, 100 µm-thick coronal brain sections were cut and imaged in brightfield microscopy to verify correct electrode positioning.
Data analysis
Scoring and basic sleep measures
Scoring was done blind to the animal treatment according to standard scoring procedures (Fernandez et al., 2018). A MA was scored whenever the EEG presented a desynchronization time-matched with a burst of EMG activity lasting maximally 4 consecutive epochs (16 s). Latency to sleep onset was defined from ZT0 to the first appearance of 6 consecutive NREMS epochs (24 s). The bout size binning in short, intermediate and long bouts for NREMS and REMS was obtained from the pulled distribution of the bout sizes from all the animals. The edges of the intermediate bin were defined as: mean -½ standard deviation to mean + 1 standard deviation.
Spectral power was computed on the raw EEG signal using a FFT on scored 4-s windows after offset correction through subtraction of the mean value of each epoch. The median power spectrum for each state was obtained for epochs non-adjacent to state transitions. The normalization was done through dividing by the average of mean power levels (from 0.75–47 Hz) for each vigilance state, ensuring that each state had the same weight in the averaging (Vassalli & Franken, 2017). This normalization was done separately for Baseline and D20+ recordings. Gamma power at D20+ was extracted through calculating mean power levels between 60—80 Hz. Data were normalized to corresponding baseline values.
The heart rate was extracted from the EMG signal as described previously (Lecci et al., 2017). Briefly, the EMG signal was highpass-filtered (>25 Hz) and squared. The R peaks of the heartbeats were detected using the Matlab ‘Findpeaks’ function. Only animals with clearly visible R peaks present in the EMG in NREMS were included in this analysis (Fernandez et al., 2017).
For delta power time course, raw delta power (mean power between 1—4 Hz from FFT on mean-centered epochs) was extracted for each NREMS epoch non-adjacent to a state transition. Total NREMS time was divided into periods of equal amounts of NREMS (12 in light phases, 6 in dark phases) from which mean values for delta power were computed. The position in time of these periods was not different between groups. Normalization was done via mean values between ZT9—12, when sleep pressure is the lowest.
Wake-up and sleep-through events after vibration were scored automatically as follows. For each trial, the EEG and EMG signals were analyzed within time intervals from 5 s prior to 5 s after the vibrations. To distinguish wake-up and sleep-through events, three values were calculated: 1) The ratio theta (5-10 Hz)/ delta (1-4 Hz) for the 5 s before stimulation, 2) the difference in the low-/ high-frequency ratios (1-4 Hz/ 100—500 Hz), before and after the stimulation, 3) the squared EMG amplitude ratio after/ before stimulation.
A trial was rejected when the ratio theta/delta was > 1 before stimulation or the EMG amplitude was larger before than after stimulation. In this way, trials starting in REMS or wakefulness were excluded. Wake-up events were scored when the difference in low/high ratios mentioned above decreased markedly after stimulation together with EMG activity. Occasionally, some wake-ups were also scored when EEG or EMG activity was very high while the other channel showed moderate changes. Appropriate thresholds were set upon visual inspection blinded to the animal’s condition.
Analyses related to the 0.02 Hz-fluctuation
Extraction
The 0.02 Hz-fluctuation in sigma power (10–15 Hz) was extracted from EEG or LFP signals using a wavelet transform (Morlet wavelet, 4 cycles), calculated over 12 h recordings in 0.5-Hz bins. The resulting signal was down-sampled to 10 Hz and smoothed using an attenuating FIR filter (cutoff frequency 0.0125 Hz, order of 100, the low order allowing for frequencies above the cutoff). The mean of the datapoints within NREMS and MA epochs was used for normalization (Figure 2-figure supplement 1D). The peak and frequency of the 0.02 Hz-fluctuation were calculated through a FFT on continuous NREMS bouts as described (Lecci et al., 2017). FFTs from individual bouts at frequency bins from 0 to 0.5 Hz were interpolated to 201 points before averaging across bouts to obtain a single measure per mouse. The angles of the phase of the 0.02 Hz-fluctuation were obtained through the Hilbert transform (Matlab signal processing toolbox). We set the troughs of the 0.02 Hz-fluctuation at 180°, the peaks at 0° (Figure 2-figure supplement 1G).
In several instances (Figures 3D, 4, 5), instead of calculating FFTs in the infraslow frequency range, we needed to detect individual cycles of the 0.02 Hz-fluctuation. To do this, we applied the Matlab “Findpeaks” function, with the conditions that the peak values were > mean and the trough values < mean, each separated by > 20 s. With such parameters, the sequence trough-peak-trough appears only in NREMS and allows to count individual cycles.
Band-limited power dynamics during the 0.02 Hz-fluctuation
To calculate the power dynamics in different frequency bands, Morlet wavelet transforms were down-sampled to 10 Hz to match the sampling of the 0.02 Hz-fluctuation and normalized by the sum of their means in NREMS. The mean power of each band was then binned in 18 bins of 20° and a mean across cycles (with or without MAs) of power activity per phase bin was obtained per animal.
Analysis of activation index
AI was computed by the natural logarithm (ln) of the ratio between beta (16– 25 Hz) + low gamma (26–40 Hz) over delta power (1–4 Hz), extracted as described above. Individual cycles from peak-to-peak were classified whether a MA was present in the fragility period or whether it continued into NREMS. To assess the presence of peaks in activation indices, the “findpeaks” function was used at phase values of 90–270°, with mean values used as a threshold.
Online detection of continuity and fragility periods
For the online detection of fragility and continuity periods during closed-loop sensory stimulation, a homemade software was generated with two layers of decision. The first one determined the likely current state of vigilance (wake, NREMS or REMS), whereas the second one made a machine learning-based decision between a continuity or a fragility period.
1 – Determination of vigilance state. This assessment was based on power band ratios characteristic for wake, NREMS and REMS using appropriate thresholds (Figure 6B,C). Every s, a FFT was calculated on the mean-centered last 4 s of EEG values and the power ratio between the delta (1–4 Hz) and the theta (5– 10 Hz) was calculated.
Transitions out of wake
1) Switch to NREMS if the last 3 s of EMG were below a high threshold and at least 2 out of the 3 last s of ratio were above a high threshold. 2) Switch to REMS if the full 5 s of EMG were below a low threshold and the full 5 s of ratio were above a high threshold.
Transitions out of NREMS
1) Switch to wake if the last s of EMG was above a high threshold. 2) Switch to REMS if the last five s of EMG were below a low threshold and if among the last 5 s of ratio, at least 4 were below a low threshold and all 5 were below a high threshold.
Transitions out of REMS
1) Switch to wake if the last s of EMG were above a high threshold. 2) Switch to NREMS if the ratio was above a high threshold for at least 4 out of 5 s.
2 – Continuity and fragility detection: From the previous step, the value of sigma (10—15 Hz) was kept every second. The mean sigma value in NREMS was dynamically updated if the likely state was determined as NREMS and used to normalize the incoming sigma power values. The last 200 s of sigma power regardless of the likely state were kept in memory. We heuristically found that a 9th-order polynomial fit (Matlab ‘polyfit’ and ‘polyval’ functions) best approximated the 0.02 Hz-fluctuation. To train the network, we first generated online-estimated 0.02 Hz-fluctuation at 1 Hz for the 12 h of the light phase. We next applied offline cycle detection in NREMS periods. For simplicity, and in agreement with previous measures of sensory arousability (Lecci et al., 2017), we set the continuity periods from trough to peak and fragility from peak to trough. Then, we subdivided these recordings in chunks of 200 s (moving window of 1 s, as they would appear online) and keeping the label continuity, fragility or none for each of them. We could thus obtain 43,000 labelled chunks per 12 h of recording. We used 642,000 of these chunks from 13 animals to train a neural network (pattern recognition ‘nprtool’ from Matlab Statistics and Machine Learning Toolbox) 70 % of the for training, 15 % for validation and 15 % for testing. The network was composed of one hidden layer with 10 neurons and one output layer with the 3 different output. We then used the generated neural network online to take the decision between continuity, fragility or none.
Statistics
The statistics were done using Matlab R2018a and the R statistical language version 3.6.1. The normality and homogeneity of the variances (homoscedasticity) were assessed using the Shapiro-Wilk and the Bartlett tests, respectively to decide for parametric statistics. In the cases where normality or homoscedasticity were violated, a log transformation was assessed at first and finally, non-parametric post-hoc tests were used (Wilcoxon rank sum test for unpaired and signed-rank test for paired data). The degrees of freedom and residuals for the F values are reported according to the R output. Post-hoc analyses were done only when the interaction between factors were significant (p < 0.05). Bonferroni’s correction for multiple comparisons was applied routinely, and the corrected α values are given in the legends. The factors used in the ANOVAs are depicted with pictograms once the corresponding effects were significant. The factors used in the analysis were: ‘treatment’ with two levels: Sham and SNI; ‘day’ with two repeated levels: baseline and D20+; ‘size’ with three repeated levels: small, intermediate or long bouts; ‘period’ with two repeated levels: continuity or fragility; ‘SD’ with two repeated levels: control or recovery after sleep deprivation; ‘state’ with three repeated levels: wake, NREMS or REMS; ‘MAs’ with two levels: with or without MA in the fragility period; ‘peak’ with two repeated levels: cycles with or without a peak in AI during fragility periods. The circular statistics were done using the CircStat for Matlab toolbox (Berens, 2009).
Competing interests
The authors declare no competing interests.
Author contributions
Romain Cardis, Conceptualization, Data collection, Data curation, Formal analysis, Software, Validation, Investigation, Visualization, Methodology, Figures, Writing—review and editing; Sandro Lecci, Conceptualization, Data collection, Data curation, Formal analysis, Methodology; Laura Fernandez, Methodology, Data curation; Alejandro Osorio-Forero, Methodology; Paul Chu Sin Chung, Editing; Stephany Fulda, Conceptualization, Data Validation, Writing—review and editing; Isabelle Decosterd and Anita Lüthi, Conceptualization, Data curation, Supervision, Funding acquisition, Validation, Visualization, Writing—original draft, Project administration, Writing—review and editing.
Supplementary figures
Acknowledgements
All lab members provided critical input at all stages of this manuscript. The excellent animal caretaking headed by Michelle Blom and the Team of Animaliers, in particular Titouan Tromme, is highly appreciated. Expert veterinary support and advice was provided by Drs. Gisèle Ferrand and Laure Sériot. We thank Christiane Devenoges for support in histological analysis and Marie Pertin and Guylène Kirschmann for excellent technical support with SNI surgeries. Dr. Simone Astori and Dr. Marc Suter provided insightful comments on pre-final versions of the manuscript and Laura Solanelles Farré helped with careful proofreading. The useful discussions with Raquel Sandoval Adaia, Paul Franken, Thomas Nevian, Francesca Siclari and Raphaelle Winsky-Sommerer are gratefully acknowledged. This study was funded by The Swiss National Science Foundation (n° 310030_184759 to AL, n° 310030_179169 to ID, n° 320030-179194 to SF), and Etat de Vaud.