Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2025 Apr 1.
Published in final edited form as: J Neurosci Methods. 2024 Jan 30;404:110063. doi: 10.1016/j.jneumeth.2024.110063

Selective REM sleep restriction in mice using a device designed for tunable somatosensory stimulation

Dillon M Huffman a, Asma’a A Ajwad a, Anuj Agarwal b, Michael E Lhamon b, Kevin Donohue b, Bruce F O’Hara b, Sridhar Sunderam a,‡
PMCID: PMC10922658  NIHMSID: NIHMS1968129  PMID: 38301833

Abstract

Background:

Sleep perturbation is widely used to investigate the physiological mechanisms that mediate sleep-wake dynamics, and to isolate the specific roles of sleep in health and disease. However, state-of-the-art methods to accomplish sleep perturbation in preclinical models are limited in their throughput, flexibility, and specificity.

New Method:

A system was developed to deliver vibro-tactile somatosensory stimulation aimed at controlled, selective sleep perturbation. The frequency and intensity of stimulation can be tuned to target a variety of experimental applications, from sudden arousal to sub-threshold transitions between light and deep stages of NREM sleep. This device was activated in closed-loop to selectively interrupt REM sleep in mice.

Results:

Vibro-tactile stimulation effectively and selectively interrupted REM sleep – significantly reducing the average REM bout duration relative to matched, unstimulated baseline recordings. As REM sleep was repeatedly interrupted, homeostatic mechanisms prompted a progressively quicker return to REM sleep. These effects were dependent on the parameters of stimulation applied.

Comparison with Existing Methods:

Existing sleep perturbation systems often require moving parts within the cage and/or restrictive housing. The system presented is unique in that it interrupts sleep without invading the animal’s space. The ability to vary stimulation parameters is a great advantage over existing methods, as it allows for adaptation in response to habituation and/or circadian/homeostatic changes in arousal threshold.

Conclusions:

The proposed method of stimulation demonstrates feasibility in affecting mouse sleep within a standard home cage environment, thus limiting environmental stress. Furthermore, the ability to tune frequency and intensity of stimulation allows for graded control over the extent of sleep perturbation, which potentially expands the utility of this technology beyond applications related to sleep.

Keywords: real-time sleep classification, automated sleep perturbation, REM sleep restriction, somatosensory stimulation, mouse

Graphical Abstract

graphic file with name nihms-1968129-f0001.jpg

1. Introduction

Sleep is regulated by many underlying physiological processes, and largely driven by circadian and homeostatic forces (Borbély, Daan, Wirz-Justice, & Deboer, 2016). Cyclic changes in physiology and behavior during sleep reflect distinct sub-stages, mainly: rapid eye movement (REM) sleep (or paradoxical sleep) and non-REM sleep. The overall quality of sleep is dependent on the relative occurrence of these states, as well as the transitional dynamics between them. In a clinical setting, reduced sleep quality is often reported in various conditions (Alzheimer’s disease, traumatic brain injury, etc.), and is associated with a multitude of health risks (Bah, Goodman, & Iliff, 2019; Peter-Derex, Yammine, Bastuji, & Croisile, 2015). Due to the inherent difficulties associated with investigating these interactions in a clinical population, rodent preclinical models are often used as a surrogate as they allow for controlled, high-throughput experimentation and are available in a variety of genotypes and phenotypes. However, the sleep disruptions observed in clinical populations are often subtle and/or state-specific in nature, so emulating specific, clinically relevant sleep dynamics in animal models requires experimental tools to affect sleep in a controlled manner.

To this end, a number of experimental sleep perturbation methods have been devised. Gentle manual handling, while highly effective and arguably the ‘standard’ method for sleep deprivation (Colavito et al., 2013; Franken, Dijk, Tobler, & Borbely, 1991; Lemons, Saré, & Beebe Smith, 2018), relies on human intervention, making it labor-intensive and susceptible to human error or bias. Moreover, stage-specific sleep disruption is only feasible if the experimenter is able to monitor elecroencephalographic (EEG) signals, which requires visual inspection of the signals – either raw or processed – in real time (Datta et al., 2009); this is challenging and limits throughput. In contrast, the ‘platform over water’ method (the standard method for REM sleep deprivation) takes advantage of the characteristic muscle atonia in REM sleep, and is extremely effective at nearly completely extinguishing REM without human intervention or EEG implantation (Sebastien Arthaud et al., 2015; da Silva Rocha-Lopes, Machado, & Suchecki, 2018; Ma et al., 2014; Rothman, Herdener, Frankola, Mughal, & Mattson, 2013). However, it requires the animal to be in a stressful environment, and is an “all-or-none” method not suitable for partial REM deprivation (i.e. interrupting REM in certain time periods or after a set amount of REM is observed); besides it’s utility is specific to REM sleep alone.

In recent years, many labor-saving methods have been proposed that allow for automated intervention (either open- or closed-loop in design), including rotating cage floors (McCarthy et al., 2016; Puhl, Boisvert, Guan, Fang, & Grigson, 2013; Trammell, Verhulst, & Toth, 2014), motorized activity wheels (McCoy et al., 2013), or sweeping bars within the cage (Atrooz, Liu, Kochi, & Salim, 2019; Cordeira et al., 2018; Dumaine & Ashley, 2015). However, these methods require the animal to be kept in restrictive housing, and disrupt sleep by forcing the animal to move. Recent reports aimed at less stressful, non-intrusive sleep deprivation using somatosensory stimulation show a great deal of promise (Sébastien Arthaud, Libourel, Luppi, & Peyron, 2020; Feng, Vogel, Obermeyer, & Kinney, 2000; Gross et al., 2015), but still do not provide flexibility in their stimulation modality - a factor which could be critical for achieving more subtle experimental effects. Direct cortical stimulation (Kahn et al., 2022) can provide the desired flexibility, as well as temporal and spatial specificity, but it is by nature invasive. Auditory stimulation is tunable and noninvasive (Moreira et al., 2021), but may require insulation from environmental noise to be effective. To this end, we propose a novel method of sleep fragmentation that utilizes non-invasive somatosensory stimulation in the form of mechanical vibration with fine control over its timing, intensity, duration, and frequency. Here, we assess the efficacy of this approach in accomplishing selective REM sleep restriction in wild-type mice.

2. Methods

Eight C57BL/6J mice were instrumented for EEG/EMG recording, each undergoing a series of three experiments. Each experiment consisted of two 12-hour light-period recordings acquired on two consecutive days, with Day 1 serving as an undisturbed reference baseline (BSL) to compare against closed-loop REM Sleep Restriction (RSR) during Day 2 (Fig. 1). RSR was accomplished by mechanical vibration of the cage floor, which was applied upon detection of REM sleep in real time using custom EEG-based sleep classification software (D. Huffman, Ajwad, Yaghouby, O’Hara, & Sunderam, 2021). Stimulation was applied with a fixed frequency and intensity of vibration throughout the RSR period, with each experimental RSR period using one of three stimulation parameter sets of varying intensity. Following trials, recordings were segmented into 4-second epochs and manually scored as being in Wakefulness, REM sleep, or Non-REM sleep. The efficacy of stimulation was determined by comparing quantitative indices of sleep quality including percent time, mean bout duration, and number of bouts of each state between baseline and RSR periods, and across stimulation settings to gain insight as to whether efficacy was dependent on the parameters of stimulation.

Figure 1. Representation of experimental protocol.

Figure 1

Open in a new tab

D0-D9 represent days into experimental pipeline, with Bn and Sn representing the nth experiment’s BSL (B) and RSR (stimulation; S) periods respectively. Time provided for animal acclimatization and preliminary model baselines are denoted as ACC and MB respectively. A 24-hour washout period (denoted W) was allowed between successive experiments. Periods considered in this analysis are indicated in white. In practice, logistic limitations often resulted in more than one day of washout between experiments, but all respective BSL and RSR recordings took place on successive days. Times during which no data were collected are represented by black-filled boxes. The animal remained in the cage throughout the protocol.

2.1. Animal handling and surgical procedures

All procedures were performed according to protocols approved by the Institutional Animal Care and Use Committee (IACUC) at the University of Kentucky. Following methods previously described in Huffman et al. (2021), eight C57BL/6J mice (5 male, 3 female; 6–8 weeks old) were surgically instrumented with a headmount for tethered EEG/EMG recordings from bone screws attached to the skull (8201; Pinnacle Technology; Lawrence, KS). After recovering from surgery (7–10 days post-op), animals were transferred to a customized recording apparatus for experimental sleep restriction (MouseQwake cage system prototype; Signal Solutions, LLC) in a light- and temperature controlled room (lights on 7 AM – 9 PM; ambient temperature 22⁰C; relative humidity 50 ± 10%) with food and water available ad libitum. A headstage 100× preamplifier was plugged into the headmount and connected through a slipring commutator to a data acquisition and conditioning system (DACS 8206; Pinnacle Tech.; Lawrence, KS). Animals were allowed to acclimatize to the new cage environment for a minimum of 24 hours before experiments began. The MouseQwake assembly served as the homecage throughout the protocol described in Figure 1. No behavioral tests were performed that might influence sleep while the animal was in the cage.

2.2. Data acquisition and signal conditioning

The DACS 8206 conditioned the EEG and EMG signals – applying 50× amplification and bandpass filters (0.5 – 100 Hz for EEG; 10–100 Hz for EMG) – and sampled them at a rate of 400 samples per second using Sirenia Acquisition 1.7.2 software (Pinnacle Technology, Inc.). An additional signal from a piezoelectric motion sensor film laid on the cage floor (SP-78; Signal Solutions, LLC) was recorded alongside the EEG/EMG signals. This primarily helped discriminate sleep from wakefulness and confirm the timing of RSR stimulation through the presence of a visible artifact. In order to trigger stimulation in closed-loop fashion, EEG/EMG signals were also routed to an external data acquisition board (USB-6211; National Instruments) through analog output terminals on the 8206 DACS so that they could be processed for real-time sleep classification (see section 2.3). Real-time detections of REM sleep and the onset of stimulation were annotated on the data record using TTL pulses delivered through the USB-6211 board.

2.3. Real-time classification of sleep for closed-loop stimulation

Real-time identification of REM sleep was accomplished using a custom sleep classification algorithm implemented in LabVIEW™ (D. Huffman et al., 2021). After allowing the animal to acclimatize to the new cage environment, a 7-hour preliminary baseline recording was acquired. Then specific signal features were computed from 4-second non-overlapping windows and fitted to a hidden Markov Model (HMM) using custom scripts in MATLAB (Mathworks; Natick, MA). After the model parameters (i.e., conditional distributions of the feature vector for each state, and transition probabilities between states) were estimated from the baseline using a maximum likelihood estimation procedure called the Baum-Welch algorithm (Baum, Petrie, Soules, & Weiss, 1970), they were read into the LabVIEW program so that the most likely underlying sequence of states corresponding to new incoming data could be inferred by the model in real time. An assessment of the classification accuracy of this method during RSR is described in Huffman et al. (2021).

2.4. REM sleep interruption using vibrotactile stimulation

Whenever REM sleep was detected in real time by the LabVIEW program, a vibrational stimulus was applied to interrupt it. After previously demonstrating the feasibility of sleep perturbation using somatosensory stimulation (D. M. Huffman et al., 2016; Yaghouby, Schildt, Donohue, O’Hara, & Sunderam, 2014), a refined apparatus – dubbed MouseQwake – was developed in collaboration with Signal Solutions, LLC (Lexington, KY). MouseQwake consists of a cage resting on a spring-suspended platform that can be vibrated using a tactile transducer beneath the cage floor (Fig. 2). The transducer is a commercially available device (TT25–8; Dayton Audio) that uses an alternating current input to make an electromagnetically controlled mass oscillate; this allows control over the stimulation frequency, intensity, and duration (expected operating range of 20–80 Hz; < 20 VRMS). While other systems use interventions that are binary in nature (i.e. impulses that are either strictly on or off), MouseQwake can be actuated in various modes (e.g. gentle, low-frequency or intense, high frequency), thus providing great flexibility. Moreover, this system does not rely on moving parts within the cage, limiting the impact of environmental stress and allowing for a normal housing environment.

Figure 2. Cage system used for REM sleep restriction.

Figure 2

Open in a new tab

The MouseQwake system consists of a cage resting on a spring-suspended platform (Left). A tactile transducer mounted under the cage floor can be made to vibrate with specified frequency, amplitude and duration to affect sleep. Here, the device was externally triggered using a custom EEG/EMG sleep classifier. When REM sleep was detected continuously for 5 s (Lower), 1-second pulses of vibration were applied every other second until REM sleep was no longer detected. A counterexample in which stimulation failed to rouse the animal demonstrates the absence of any stimulus artifact on the EEG and EMG traces that might affect classifier accuracy (Right).

Mechanical vibration can have both tactile and auditory components. The lower limit of hearing in mice exceeds 1 kHz (Naff et al., 2007) and is still higher in many strains commonly used in laboratory research. To minimize the potential for audible output, only sinusoidal waveforms of frequency below 500 Hz were used here. The output at the onset and offset of stimulation were windowed to avoid step changes that might produce higher harmonics that exceed the audible threshold in mice. Secondary effects such as motion of the wire rack in the cage or food pellets in the hopper could add noise to the mechanical vibration that spreads the frequency into the audible spectrum. But to the human ear, which is sensitive to much lower frequencies, these effects were found to be negligible.

In this initial investigation, a set of three frequency/intensity combinations that sampled the possible output parameter space were applied (listed in order of perceived intensity): 50 Hz/200 mV (mild, mid-frequency stimulation), 100 Hz/400 mV (intense, high-frequency stimulation), and 15 Hz/500 mV (intense, low-frequency stimulation). When REM sleep was detected for a minimum of 5 seconds, a 1-second stimulation was applied with a 2-second duty cycle until REM was no longer detected. Stimulation parameters were fixed throughout the RSR period (i.e. every stimulation in a given RSR period was applied with the same frequency and amplitude). A 24-hour minimum washout period was allowed between experiments, and the sequence of stimulation parameters assigned to each RSR period was randomized for each animal to control for progressive effects. Due to health issues, one animal had to be removed from the study prior to the third RSR experiment (which was scheduled to be in the 15Hz, 500mV group), resulting in this group having a smaller sample size (n = 7 vs. n = 8).

2.5. Manual scoring to quantify sleep

For each experiment, the 12-hour segments from both baseline and RSR periods (46 records total) were manually scored in 4-second epochs as either Wake, NREM sleep, or REM sleep using Sirenia SLEEP Pro software (Pinnacle Tech.) to view signals, according to criteria described in Huffman et al. (2021). The hypnogram – i.e., the sequence of vigilance state labels – was then exported and loaded into MATLAB for further processing. For each animal, the proportion of time spent, mean bout duration, and number of bouts were computed for each state, and compared between baseline and RSR periods. Manual scores also served to assess real-time model classification accuracy and stimulation specificity.

2.6. Statistical analysis

All statistical analyses were performed in MATLAB (Mathworks, Natick, MA) using non-parametric tests. Changes in sleep metrics (state proportion, mean bout duration, and number of bouts) and transition matrices were assessed using Wilcoxon signed-rank tests. Bout duration distributions were compared using Mann-Whitney U tests. In all cases, significance was defined as a Type-I error probability p < 0.05. When reported on graphs, error bars represent the standard error of the mean. Statistics for pooled analyses (results section 3.1) are based on n=23 BSL and RSR periods, while comparisons of between stimulation settings (sections 3.2, 3.3) use either an n of seven (15Hz, 500mV) or eight (50Hz, 200mV; 100Hz, 400mV).

3. Results

3.1. Vibrotactile stimulation was effective at interrupting REM sleep

First, the overall efficacy of vibrotactile stimulation was compared against the unstimulated BSL condition without regard to differences in stimulation parameter settings: i.e., data for all RSR parameter settings were pooled and compared against BSL. Differences in sleep metrics (proportion, mean bout duration, and number of bouts) for Wake, NREM, and REM were compared between twenty-three matched BSL and RSR recordings (three per animal but for one excluded recording in one animal). Changes in sleep metrics between a given experiment’s BSL and corresponding RSR period were computed in terms of a Matched Comparison Ratio (MCR): MCR=(BSLRSR−1)*100. RSR and BSL represent mean values of a given sleep metric (e.g. REM bout duration) over an experiment’s RSR and BSL periods, respectively, and MCR represents the percent change in that particular metric relative to the baseline.

In terms of MCR, significant changes in sleep architecture were observed during closed-loop RSR. REM sleep was most affected, with mean REM bout duration being reduced by 45% compared to the previous day’s baseline (Figure 3, top). Consequently, twice as many REM bouts were also observed, which was primarily due to a significant increase in bouts less than 20 seconds in duration (accounting for 25% of baseline bouts versus 75% of bouts during RSR) (Fig. 3; bottom). While REM bouts were truncated, the reduction in the total amount of REM sleep observed over the 12-hour RSR period did not reach statistical significance. When assessed on an hour-by-hour basis however, it was found that REM sleep was reduced for approximately 4 hours after the start of RSR trials, after which more frequent transitions into REM sleep occurred (Fig. 4) – likely due to the homeostatic drive to recover lost REM sleep (this is further discussed in section 3.3).

Figure 3. Effect of 12-hour closed-loop vibro-tactile stimulation on REM sleep.

Figure 3

Open in a new tab

Compared to previous-day baselines, REM bouts were truncated during RSR trials (data for all three stimulation parameter sets pooled). As a result, animals spent less time in NREM sleep but with more frequent transitions from NREM to REM sleep (decreased NREM bout duration with increased REM bout number). While the proportion of time spent in REM was not reduced over the 12-hour period, the distribution of REM bout durations observed during all RSR periods was significantly different from that in the baseline (p<0.05, n=23 recordings; Wilcoxon signed-rank test). Note: The y-axes of the histogram plots are normalized to the highest count observed across baseline and RSR periods to better represent the differences between distributions.

Figure 4. Trends in hourly estimates of sleep metrics suggest a homeostatic drive to recover REM sleep.

Figure 4

Open in a new tab

While REM sleep is reduced over the first 4 hours of RSR (left column), sustained interruption of REM bouts during RSR (middle column) leads to progressively more frequent transitions back into REM sleep (right column) which allowed baseline levels of REM sleep to accumulate over time. The effects on REM sleep are accompanied by a slight but sustained reduction in proportion of NREM and increase in Wake. Mean NREM bout duration is dramatically reduced and the number of bouts increased in RSR due to attempted reentry into REM following stimulation, via a brief arousal (induced by stimulation) followed by a transient bout of NREM. The increased demand for REM sleep and in Wake following stimulation thus comes at the cost of some NREM sleep.

3.2. Changes in sleep were dependent on stimulation parameters

Next, it was investigated whether the extent of REM sleep perturbation was dependent on the parameters (frequency and amplitude) of the stimulation applied. When computed separately for data corresponding to each stimulation setting, MCR indicated that strong stimulation settings reduced mean REM bout duration by as much as 65%, while mild stimulation (50 Hz, 200 mV) gave a less pronounced effect (Fig. 5). Proportional changes were seen in the number of attempted REM bouts, which increased by as much as 200% (three times the baseline amount). However, no significant effects were observed on the time spent in REM sleep (discussed further in section 3.4). While mild (50Hz, 200mV) stimulation did not have a significant effect on mean REM bout duration, the distributions of REM bout duration during these RSR periods were significantly different from those during the baseline conditions (p < 0.05; Mann-Whitney U Test), primarily due to an increase in short (< 20s) REM bouts (Fig. 5). This effect was more pronounced for higher intensity stimulation settings, with as much as 80% of REM bouts being less than 10s in duration (versus 6% in the baseline).

Figure 5. The extent of REM sleep perturbation was stimulation parameter-dependent.

Figure 5

Open in a new tab

When grouped according to stimulation settings, MCR reveals changes in sleep metrics that were proportional to the stimulation intensity (top), which is also apparent in the REM bout duration distributions (bottom). Histograms are presented on a uniform vertical scale across settings; normalized by the highest count observed across experimental periods to highlight the difference between settings. Significance is defined as p<0.05; Wilcoxon signed rank test.

3.3. Vibrotactile RSR induces homeostatic recovery of REM sleep

As REM sleep was repeatedly interrupted, a dramatic increase was observed in the propensity to reenter REM sleep. Following stimulation-induced arousal, there was a significant reduction in REM sleep latency – that is, the time to the next bout of REM – as well as significantly fewer bouts of NREM between REM bouts (Fig. 6). In fact, NREM bouts in RSR periods were twice as likely to transition to REM sleep compared to the corresponding baselines; and this was accompanied by decreased NREM sleep fragmentation (lower incidence of NREM to Wake transitions) (Fig. 7). There was also a significant increase in the number of REM bouts separated by a single NREM bout, which were significantly shorter compared to such bouts in the baseline. Moreover, the extent of this was seen to be dependent on stimulation parameters, and correlated with the stimulation parameter-dependent changes in sleep metrics discussed previously: 40% of lone NREM bouts between consecutive REM bouts lasted under a minute in the baseline versus 50%, 70%, and 80% for 50 Hz, 100 Hz, and 15 Hz stimulation settings respectively. This is further illustrated by Fig. 8, which shows how stimulation at higher intensities is more effective at interrupting REM but also leads to a reduction in time taken to reenter REM due to the increased REM deficit.

Figure 6. REM sleep perturbation induces homeostatic changes in sleep.

Figure 6

Open in a new tab

Following interruption of REM sleep, animals spent less time transitioning back to REM through NREM sleep. As REM debt continues to increase, eventually the urgent need to recover lost REM sleep overpowers the efficacy of stimulation at the chosen setting.

Figure 7. Transitional dynamics of sleep are altered during RSR.

Figure 7

Open in a new tab

MCR computed from state transition probabilities highlight homeostatic drive for REM, including decreased NREM-to-WAKE probability (fewer arousals directly from NREM), increased NREM-to-REM probability, and increased WAKE-to-NREM probability for more intense stimulation settings. The increased REM-WAKE probability is a direct effect of closed-loop stimulation. Both REM-to-NREM and WAKE-to-REM transitions almost never occur, either in BSL or STM, which makes an MCR calculation meaningless; hence these were marked N/A. (Significance defined as p < 0.05; Wilcoxon signed-rank test).

Figure 8. REM sleep restriction reduces latency to REM sleep.

Figure 8

Open in a new tab

Two-dimensional probability density plots illustrate how the duration of a REM bout influences the timing of the next bout of REM sleep. In the baseline, the latency to REM is positively correlated with REM bout duration: i.e., the longer a bout of REM is, the longer it takes to see the next bout. As stimulation intensity is increased, REM sleep is interrupted more effectively and longer REM bouts become less likely. However, REM homeostasis demands that the time taken to renter REM is also drastically reduced, as illustrated by the increased height and density of the peak in the lower left corner of the graph.

3.4. RSR changes the dynamics but not the characteristics of NREM sleep

Next, we examine how NREM sleep is affected by RSR. When REM sleep is interrupted by stimulation the animal is aroused and may remain awake for some duration. With repeated RSR, REM deficit builds up and the brain works to recover lost REM. As a consequence, the duration of Wake after stimulation is reduced and the animal goes back into NREM sleep, which is the gateway to REM. Over time, the time spent in NREM sleep after stimulation gets progressively shorter. This is evident from the increase in Wake-NREM and NREM-REM transition probabilities along with a decreased likelihood of arousal from REM (see Fig. 7). The progressive effect is illustrated in Fig. 9. In the early phase of RSR (ZT 0–3 h), the probability of seeing REM after the onset of NREM increases gradually but not more than in BSL. But over the next 9 hours there is a progressive increase in the probability that NREM will give way to REM within a few seconds, especially at more intense stimulation settings. Furthermore, there is an increased probability that a NREM bout started within 20s of a REM bout that was interrupted by stimulation.

Figure 9. REM sleep probability relative to onset of NREM.

Figure 9

Open in a new tab

The plots depict the likelihood of REM sleep relative to the onset time of NREM sleep over the course of the experiment, averaged over all NREM bouts in three-hour intervals, for BSL (blue) and RSR (red) periods and for three different stimulation settings. NREM gradually gives way to REM in ZT 0–3 h with no evident difference between BSL and RSR, but there is a marked difference thereafter, with REM likelihood increasing sharply within 20s before and after NREM onset. This indicates that an increasing number of NREM bouts occur soon after REM sleep is interrupted and are increasingly under pressure to give way to REM again as the REM deficit builds up over the course of the day.

The suppression of REM by RSR therefore comes with a reduction in NREM as well, most likely due to an increase in Wake and a preferential recovery of REM as the deficit builds up (see Figs. 3–5). This is in contrast to the work of Arthaud et al. (2020), in which there was no significant change in NREM sleep with RSR using a different device. From Fig. 4, we see that the change in % NREM in the first 6 hours of RSR is not significant in the present study, but is greater from 6–12 hours. We also looked for changes in the EEG power spectrum of NREM sleep. Although the slope and amplitude of the delta peak appeared to diminish with RSR compared to BSL over the course of the day, which may reflect changes in the excitatory-inhibitory balance (Gao et al. 2017) there was no apparent shift in location (see Supplementary Figure 1) . There were no differences in the Wake spectra for RSR and BSL. Slight (but not statistically significant) differences in the REM spectra may be attributed to a greater number of transitional epochs during RSR rather than an actual change in the spectral presentation of REM sleep.

3.5. Changes in sleep were not a consequence of erroneous stimulation

Finally, in order to better understand the parameter-dependent effects of stimulation on sleep-wake dynamics, the specificity of detection was investigated. For each stimulated REM detection, the probability of each state was computed within ±1 minute of REM detection with 1-second resolution. At the time of REM detection, REM sleep was most likely – coinciding with 60–75% of REM detections (Figure 10, right; red markers). The majority of false positive detections were during NREM sleep (10–30% vs. 5–10% for wake), the number of which varied depending on the stimulation parameters (Figure 10, middle; red markers). However, the reduction in NREM bout duration during RSR trials was not proportional to the number of false positive stimulations during NREM (trials with low incidence of stimulations during NREM had the highest reduction in NREM bout duration), suggesting that the observed changes in NREM sleep were a consequence of homeostatic regulation of REM sleep (i.e. spending less time in NREM before transitioning to recover lost REM sleep). Fewer than 10% of REM bouts did not receive stimulation during RSR.

Figure 10. Stimulation was specific to REM sleep.

Figure 10

Open in a new tab

In order to gauge stimulation specificity, the probability of each state being scored was computed about model prediction of REM sleep (excluding bouts which did not meet the minimum duration criteria for stimulation). Prior to REM detection, there is an increase in the probability of scored REM (this delay is the result of detection feature smoothing). During RSR periods (red markers), the probability of REM sharply declines after stimulation (t=5s; vertical red line), with different stimulation parameters being more effective than others. Blue markers indicate average trends during each RSR trials matched baseline.

These detection-triggered averages also revealed that REM detection usually lagged behind actual REM sleep onset (average latency of 10 ± 0.2 seconds). This latency is inherent to the EEG sleep wake classifier – largely attributed to forward smoothing of classification features in a moving 20-second window to control prediction noise. This delay to detect REM sleep, combined with the programmed 5-second detection-to-stimulation delay, resulted in REM being allowed to persist for 10 to 20 seconds before stimulation was applied. Therefore, as transitions into REM sleep became increasingly frequent over time with RSR, substantial amounts of REM sleep were incrementally accumulated over the course of the RSR period.

3.6. Changes in sleep induced by RSR can affect classifier accuracy over time

We have previously shown that the real-time classifier accurately estimates baseline Wake, NREM, and REM sleep metrics throughout the course of the day (Huffman et al., 2021). However, progressive alterations in sleep-wake dynamics and/or the spectral characteristics of the EEG in the face of circadian (i.e., time of day) and homeostatic factors, could affect the accuracy of the classifier, specifically during RSR, thus reducing the specificity of closed-loop intervention. Here, we examine whether real-time classification accuracy was affected in this way.

Figure 11 shows how the real-time classification accuracy varies from the onset of each state and progressively across the experimental period during BSL and RSR. We find that Wake and NREM accuracy are slightly lower in RSR but similar to BSL. REM accuracy is more visibly affected, and the error takes longer to stabilize over the course of the day, most likely due to the increasing REM sleep debt to cause interrupted REM bouts to rebound again and again, passing through Wake and NREM on the way. Despite this tendency, the mean REM accuracy in RSR was within 10% of BSL.

Figure 11. Effect of treatment over time on real-time classifier accuracy.

Figure 11

Open in a new tab

The average classification accuracy for each vigilance state is plotted relative to the time of entry into the state and averaged across all bouts in each non-overlapping 1-hour window of the 12-hour experimental period. In general, accuracy is low in in the first few seconds of the transition but approaches a stable level as evidence in favor of the state accumulates. In the baseline, classification accuracy remain stable across the 12-hour period for all states, and shows no clear dependence on time of day. During RSR, NREM and Wake accuracy are slightly lower than the baseline, but more or less stable over 12 hours. However, REM accuracy is visibly lower during RSR and there is a progressive creep of error to the right (increasing delay in detecting REM) with hours of treatment. This may be explained by an increasing tendency for REM sleep, when interrupted, to cycle rapidly through Wake, NREM, and back into REM again as REM sleep debt accumulates, added to the inherent delay in the classifier’s response to these abrupt changes in state.

4. Discussion

Automated RSR via vibro-tactile stimulation effectively and selectively interrupted REM sleep, significantly reducing the average REM bout duration relative to corresponding unstimulated baseline recordings. While REM sleep bouts were shorter, the total REM sleep observed over the 12-hour experimental period was not reduced. However, when inspected on an hourly basis, it was found that REM sleep was reduced throughout the first 4 hours of the RSR period, following which a buildup in the homeostatic drive to recover lost REM sleep induced much more frequent transitions back into REM sleep following arousal. Interestingly, the efficacy appeared to be dependent on the stimulation parameters applied, with the most effective settings reducing 80% of REM sleep bouts to less than 10 seconds (< 3 epochs) compared to 6% in baseline conditions. Moreover, the extent of homeostasis-induced changes in sleep-wake dynamics seen during RSR periods was also dependent on the parameters of stimulation.

While closed-loop stimulation produced significant changes in REM sleep, the overall effect was likely limited by certain aspects of the study design. Since stimulation was intended to be specific to REM sleep, real-time sleep-wake classification was required. To accomplish this, we chose to develop a custom software solution based on our previous model-based sleep scoring work (D. Huffman et al., 2021; Yaghouby, Hara, & Sunderam, 2016) that was shown to track REM with acceptable accuracy. Initial investigations suggested that the incidence of false positive stimulations during NREM could be minimized by requiring detections to persist for 5 seconds before triggering stimulation. In practice, however, this 5-second delay combined with the inherent classification delay allowed relatively long durations of REM sleep to occur before stimulation was applied. During RSR, when REM bouts become increasingly more frequent, this results in a significant amount of REM sleep accumulating incrementally over the course of the experiment, which also inflates the minimum duration of REM bouts observed during this time. In retrospect, the specificity gained by avoiding stimulation during NREM is outweighed by the homeostatic accumulation of REM, as it impacts the perceived efficacy of RSR treatment. One way to counter this might be to identify an intermediate sleep state – perhaps based on a reduction in delta EEG power – in the transition from NREM to REM and interrupt it. If successful, this would prevent REM recovery in the 5-second delay to stimulation in the current protocol. However, there is at present no way to differentiate NREM-REM transitions from the brief arousals (NREM-Wake) commonly seen in mouse sleep, which is highly fragmented (see Fig. 6). Such arousal transitions might generate false positive stimulations that reduce the specificity of closed-loop RSR. Even so, it would be interesting to compare the accuracy and efficacy of such a protocol with the one implemented here.

Once REM was detected, the majority of REM bouts were interrupted shortly after stimulation onset (Fig. 10; right). However, after several successive interrupted REM sleep bouts, the likelihood of a bout persisting through stimulation increased (Fig. 6). In such cases, dynamically changing stimulation parameters may be required to combat the homeostatic drive and/or habituation. However, there may be inter- and intra-animal variation in the response to any chosen stimulation parameter set, which may further depend on homeostatic and circadian factors – posing a significant engineering challenge. While simply randomizing stimulation parameters between stimulations may be one way to address this issue, it leaves room for suboptimal settings to also be applied and potentially limit the overall effect. In future work, the possibility of promoting effective stimulation over the entire recording period using dynamic parameter adaptation will be investigated. This approach could account for the observed homeostatic and circadian modulation of sleep, as well as inter-animal variability in stimulation response to yield effective stimulation throughout the experimental period.

Furthermore, stimulation parameter-dependent changes in REM sleep observed during RSR periods indicate that the extent of REM sleep perturbation can be varied by adjusting the stimulation parameters. While other reports have demonstrated impressive amounts of REM sleep deprivation following closed-loop perturbation (Sébastien Arthaud et al., 2020; Gross et al., 2015), none have reported the ability to vary the response by altering the characteristics of stimulation. This flexibility opens the door to investigating subtle changes in sleep, which may more closely reflect certain disease conditions than depriving REM in total. Moreover, the ability to apply perturbations of varying intensity allows for broad applications, including altering sleep without arousal (e.g. transitioning from deep sleep to light sleep) (Feng et al., 2000; D. M. Huffman et al., 2016), assessing arousal thresholds for different vigilance states, or sleep enhancement through sustained and/or low-frequency vibration (Kompotis et al., 2019; Vazquez, Merchant-Nancy, Garcia, & Drucker-Colin, 1998). Moreover, high frequency (> 200 Hz) actuation results in a stimulation that is more auditory than vibrational in nature, and could be used to influence sleep through auditory stimulation (Xing et al., 2020) or to investigate startle response with pre-pulse inhibition (auditory tone followed by sudden vibration) or fear-conditioning paradigms.

The efficacy of REM sleep perturbation is often assessed by inspecting changes in sleep during the post-experimentation period (often referred to as rebound sleep). In fact, since this period is free of stimulation, it would also be a good way to decouple the accuracy of the classifier from the effectiveness of stimulation. Initially, it was our intention to investigate rebound sleep. However, due to the aforementioned limitations in the study design that caused significant REM recovery during the experimental period itself, there was no significant REM debt at the end of RSR trials, and so no clear effects are likely during rebound sleep. It is our hope that future efforts to minimize classification delay, as well as foregoing the 5-second minimum REM duration criterion, will produce a more prominent effect on REM sleep that will induce post-experimental rebound of REM sleep, and so investigating this will be a focus at that time. Yet another way to assess the efficacy of the closed-loop RSR would be to compare it with open-loop, perhaps using a yoked control design in which one animal receives the same stimulation as the other one but without regard to it’s vigilance state. Further refinements to the study design would include incorporation of dynamic stimulation parameter selection to combat homeostatic pressure and habituation, measurement of blood corticosterone levels to assess post-RSR stress levels, behavioral and cognitive assays to better characterize the effect and efficacy of stimulation, increasing the temporal responsiveness of stimulation (e.g. making a stimulation decision multiple times a second vs. once per second as was done here), and comparing efficacy in acute (i.e. 6 hours) versus chronic (> 24 hour) sleep perturbation to better understand the limitations of the method.

Overall, the method of vibrotactile stimulation presented here shows great promise in non-invasive sleep perturbation. Being completely noninvasive, it allows for sleep modification in normal housing, thus limiting environmental stress. Since the stimulation does not depend on moving fixtures (e.g. sweeping bar within the cage), stimulation can be delivered with greater temporal precision in a way that minimizes the opportunity for sleep between successive perturbation. Furthermore, the stimulus applied can be tuned in frequency and intensity, which allows graded control over the extent of sleep perturbation, potentially expanding the utility of this technology beyond applications related to sleep.

Supplementary Material

1

Highlights:

  • A closed-loop, non-invasive system was developed for rodent sleep perturbation.

  • Upon EEG detection of REM sleep, vibrational stimulation is applied to interrupt it.

  • Stimulation effectively interrupts REM sleep, reducing mean bout length by >65%.

  • Stimulation frequency and amplitude determine the extent of REM sleep disruption.

Acknowledgments

This work was supported by the National Institute of Neurological Disorders and Stroke (Grant No. NS083218).

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Disclosure of Interests:

ML, KD, and BO are co-owners of Signal Solutions, LLC, which has since commercialized the device described in this study. Since completing this research, DH has become an employee of Signal Solutions, LLC. Other authors have no conflicts of interest to disclose.

CRediT authorship contribution statement

Dillon Huffman: Conceptualization, Methodology, Software, Validation, Formal Analysis, Investigation, Writing – Original Draft, Writing – Review & Editing, Visualization.

Asma’a Ajwad: Investigation.

Anuj Agarwal: Resources.

Michael E. Lhamon: Resources.

Kevin Donohue: Methodology, Resources, Writing – Review & Editing.

Bruce O’Hara: Resources, Writing – Review & Editing.

Sridhar Sunderam: Conceptualization, Methodology, Software, Formal Analysis, Investigation, Writing – Review & Editing, Visualization, Supervision, Project administration, Funding acquisition.

References

  1. Arthaud Sébastien, Libourel P-A, Luppi P-H, & Peyron C (2020). Insights into paradoxical (REM) sleep homeostatic regulation in mice using an innovative automated sleep deprivation method. Sleep, (January), 1–16. 10.1093/sleep/zsaa003 [DOI] [PubMed] [Google Scholar]
  2. Arthaud Sebastien, Varin C, Gay N, Libourel PA, Chauveau F, Fort P, … Peyron C (2015). Paradoxical (REM) sleep deprivation in mice using the small-platforms-over-water method: Polysomnographic analyses and melanin-concentrating hormone and hypocretin/orexin neuronal activation before, during and after deprivation. Journal of Sleep Research, 24(3), 309–319. 10.1111/jsr.12269 [DOI] [PubMed] [Google Scholar]
  3. Atrooz F, Liu H, Kochi C, & Salim S (2019). Early Life Sleep Deprivation: Role of Oxido-Inflammatory Processes. Neuroscience, 406, 22–37. 10.1016/j.neuroscience.2019.02.021 [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Bah TM, Goodman J, & Iliff JJ (2019). Sleep as a Therapeutic Target in the Aging Brain. Neurotherapeutics, 16(3), 554–568. 10.1007/s13311-019-00769-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Baum LE, Petrie T, Soules G, & Weiss N (1970). A Maximization Technique Occurring in the Statistical Analysis of Probabilistic Functions of Markov Chains Author ( s ): Leonard E . Baum , Ted Petrie , George Soules , Norman Weiss Source : The Annals of Mathematical Statistics , Vol . 41 , No . 1 ( Feb. Institute of Mathematical Statistics Stable, 41(1), 164–171. [Google Scholar]
  6. Borbély AA, Daan S, Wirz-Justice A, & Deboer T (2016). The two-process model of sleep regulation: A reappraisal. Journal of Sleep Research, 25(2), 131–143. 10.1111/jsr.12371 [DOI] [PubMed] [Google Scholar]
  7. Colavito V, Fabene PF, Grassi-Zucconi G, Pifferi F, Lamberty Y, Bentivoglio M, & Bertini G (2013). Experimental sleep deprivation as a tool to test memory deficits in rodents. Frontiers in Systems Neuroscience, 7(DEC), 1–17. 10.3389/fnsys.2013.00106 [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Cordeira J, Kolluru SS, Rosenblatt H, Kry J, Strecker RE, & McCarley RW (2018). Learning and memory are impaired in the object recognition task during metestrus/diestrus and after sleep deprivation. Behavioural Brain Research, 339(August 2017), 124–129. 10.1016/j.bbr.2017.11.033 [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. da Silva Rocha-Lopes J, Machado RB, & Suchecki D (2018). Chronic REM Sleep Restriction in Juvenile Male Rats Induces Anxiety-Like Behavior and Alters Monoamine Systems in the Amygdala and Hippocampus. Molecular Neurobiology, 55(4), 2884–2896. 10.1007/s12035-017-0541-3 [DOI] [PubMed] [Google Scholar]
  10. Datta A, Bansal V, Diaz J, Patel J, Reato D, & Bikson M (2009). Gyri-precise head model of transcranial direct current stimulation: Improved spatial focality using a ring electrode versus conventional rectangular pad. Brain Stimulation, 2(4), 201–207.e1. 10.1016/j.brs.2009.03.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Dumaine JE, & Ashley NT (2015). Acute sleep fragmentation induces tissue-specific changes in cytokine gene expression and increases serum corticosterone concentration. American Journal of Physiology - Regulatory Integrative and Comparative Physiology, 308(12), R1062–R1069. 10.1152/ajpregu.00049.2015 [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Feng P, Vogel GW, Obermeyer W, & Kinney GG (2000). An instrumental method for long-term continuous REM sleep deprivation of neonatal rats. Sleep, 23(2), 175–183. 10.1093/sleep/23.2.1d [DOI] [PubMed] [Google Scholar]
  13. Franken P, Dijk DJ, Tobler I, & Borbely AA (1991). Sleep deprivation in rats: Effects on EEG power spectra, vigilance states, and cortical temperature. American Journal of Physiology - Regulatory Integrative and Comparative Physiology, 261(1 30-1). 10.1152/ajpregu.1991.261.1.r198 [DOI] [PubMed] [Google Scholar]
  14. Gao R, Peterson EJ, Voytek B. Inferring synaptic excitation/inhibition balance from field potentials. Neuroimage. 2017. Sep;158:70–78. doi: 10.1016/j.neuroimage.2017.06.078. Epub 2017 Jul 1. [DOI] [PubMed] [Google Scholar]
  15. Gross BA, Vanderheyden WM, Urpa LM, Davis DE, Fitzpatrick CJ, Prabhu K, & Poe GR (2015). Stress-free automatic sleep deprivation using air puffs. Journal of Neuroscience Methods, 251, 83–91. 10.1016/j.jneumeth.2015.05.010 [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Huffman D, Ajwad A, Yaghouby F, O’Hara BF, & Sunderam S (2021). A real-time sleep scoring framework for closed-loop sleep manipulation in mice. Journal of Sleep Research, (September 2020), 1–15. 10.1111/jsr.13262 [DOI] [PubMed] [Google Scholar]
  17. Huffman DM, Staggs KE, Yaghouby F, Agarwal A, O’Hara BF, Donohue KD, … Sunderam S (2016). Tunable somatosensory stimulation for selective sleep restriction studies in rodents. In Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS (Vol. 2016-Octob). 10.1109/EMBC.2016.7591028 [DOI] [PubMed] [Google Scholar]
  18. Kahn M, Krone LB, Blanco-Duque C, Guillaumin MCC, Mann EO, Vyazovskiy VV. Neuronal-spiking-based closed-loop stimulation during cortical ON- and OFF-states in freely moving mice. J Sleep Res. 2022. Dec;31(6):e13603. doi: 10.1111/jsr.13603. Epub 2022 Jun 3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Kompotis K, Hubbard J, Emmenegger Y, Perrault A, Mühlethaler M, Schwartz S, … Franken P (2019). Rocking Promotes Sleep in Mice through Rhythmic Stimulation of the Vestibular System. Current Biology, 29(3), 392–401.e4. 10.1016/j.cub.2018.12.007 [DOI] [PubMed] [Google Scholar]
  20. Lemons A, Saré RM, & Beebe Smith C (2018). Chronic Sleep Deprivation in Mouse Pups by Means of Gentle Handling. Journal of Visualized Experiments : JoVE, (140), 1–5. 10.3791/58150 [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Ma C, Wu G, Wang Z, Wang P, Wu L, Zhu G, & Zhao H (2014). Effects of chronic sleep deprivation on the extracellular signal-regulated kinase pathway in the temporomandibular joint of rats. PLoS ONE, 9(9), 3–10. 10.1371/journal.pone.0107544 [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. McCarthy A, Wafford K, Shanks E, Ligocki M, Edgar DM, & Dijk DJ (2016). REM sleep homeostasis in the absence of REM sleep: Effects of antidepressants. Neuropharmacology, 108, 415–425. 10.1016/j.neuropharm.2016.04.047 [DOI] [PubMed] [Google Scholar]
  23. McCoy JG, Christie MA, Kim Y, Brennan R, Poeta DL, McCarley RW, & Strecker RE (2013). Chronic sleep restriction impairs spatial memory in rats. NeuroReport, 24(2), 91–95. 10.1097/WNR.0b013e32835cd97a [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Moreira CG, Baumann CR, Scandella M, Nemirovsky SI, Leach S, Huber R, Noain D. Closed-loop auditory stimulation method to modulate sleep slow waves and motor learning performance in rats. Elife. 2021. Oct 6;10:e68043. doi: 10.7554/eLife.68043. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Naff KA, Riva CM, Craig SL, Gray KN. Noise produced by vacuuming exceeds the hearing thresholds of C57Bl/6 and CD1 mice. J Am Assoc Lab Anim Sci. 2007. Jan;46(1):52–7. [PubMed] [Google Scholar]
  26. Peter-Derex L, Yammine P, Bastuji H, & Croisile B (2015). Sleep and Alzheimer’s disease. Sleep Medicine Reviews, 19(2015), 29–38. 10.1016/j.smrv.2014.03.007 [DOI] [PubMed] [Google Scholar]
  27. Puhl MD, Boisvert M, Guan Z, Fang J, & Grigson PS (2013). A novel model of chronic sleep restriction reveals an increase in the perceived incentive reward value of cocaine in high drug-taking rats. Pharmacology Biochemistry and Behavior, 109(717), 8–15. 10.1016/j.pbb.2013.04.010 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Rothman SM, Herdener N, Frankola KA, Mughal MR, & Mattson MP (2013). Chronic mild sleep restriction accentuates contextual memory impairments, and accumulations of cortical A-Beta and pTau in a mouse model of Alzheimer’s disease. Brain Research, 1529, 200–208. 10.1016/j.brainres.2013.07.010 [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Trammell RA, Verhulst S, & Toth LA (2014). Effects of sleep fragmentation on sleep and markers of inflammation in mice. Comparative Medicine, 64(1), 13–24. [PMC free article] [PubMed] [Google Scholar]
  30. Vazquez J, Merchant-Nancy H, Garcia F, & Drucker-Colin R (1998). The effects of sensory stimulation on REM sleep duration. Sleep, 21(2), 138–142. [DOI] [PubMed] [Google Scholar]
  31. Xing F, Fang X, Gong XD, Zhao X, Du Y, Ma ZL, … Xia TJ (2020). Photoacoustic treatment mitigates cognitive dysfunction in a model of sleep-wake rhythm disturbance. Neural Regeneration Research, 15(6), 1094–1101. 10.4103/1673-5374.270415 [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Yaghouby F, Hara BFO, & Sunderam S (2016). Unsupervised Estimation of Mouse Sleep Scores and Dynamics Using a Graphical Model of Electrophysiological Measurements, 26(4), 1–15. 10.1142/S0129065716500179 [DOI] [PubMed] [Google Scholar]
  33. Yaghouby F, Schildt CJ, Donohue KD, O’Hara BF, & Sunderam S (2014). Validation of a closed-loop sensory stimulation technique for selective sleep restriction in mice. Conference Proceedings : … Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual Conference, 2014, 3771–3774. 10.1109/EMBC.2014.6944444 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

1
NIHMS1968129-supplement-1.pdf (254.5KB, pdf)

RESOURCES