Sleep and Learning in Dogs: What Polysomnography Has Established

1. Why Dogs Became a Sleep Model

1.1 The Methodological Breakthrough

The field opened when a fully non-invasive polysomnography technique was developed for dogs — surface electrodes, no restraint, no sedation, the dog asleep in a familiar setting with its owner present (Kis et al., 2014). Before that, sleep research required implanted electrodes and laboratory animals.

The significance is easy to understate. Almost everything else known about canine behavior comes from owner questionnaires or from coding video of ambiguous signals (with all the interpretive problems that carries). Here the measurement is electrical activity, and the dog's compliance is not required beyond falling asleep.

1.2 Why It Matters Beyond Dogs

Sleep-dependent memory consolidation has been studied almost entirely in humans and laboratory rodents, and there is substantial variation between species, which is the standing problem with importing mechanisms into canine science. Dogs offer something neither provides: a non-laboratory mammal sharing the human environment, learning socially transmitted tasks, testable without invasive procedures (Kis et al., 2017).

1.3 The Stages Being Scored

Recordings are visually scored in twenty-second epochs by inspecting EEG, EOG, ECG and EMG channels, blind to subject and condition. Four states are distinguished: wakefulness, drowsiness, non-REM and REM sleep (Kis, Gergely et al., 2017).

From the resulting hypnogram come the macrostructural variables that most canine sleep studies report — sleep latency, relative time in each stage, average REM period duration and REM latency. Drowsiness is worth flagging because it has no direct equivalent in standard human scoring and turns out to carry information.

1.4 What Gets Measured

Polysomnography records EEG, eye movements and muscle tone, which together allow sleep to be divided into stages (with the resulting picture of canine sleep architecture set out separately). The canine work distinguishes drowsiness, non-REM sleep and REM sleep, and examines both the macrostructure — how much of each — and the spectral content of the EEG within them.

2. The Core Experiment

2.1 The Design

Dogs were taught two commands in a language they did not know, replacing familiar cues with unfamiliar equivalents — a design that avoids testing what the dog already knew (and that treats the cue as arbitrary rather than meaningful). Performance was tested, the dog then spent three hours in polysomnography, and performance was tested again (Kis et al., 2017).

An additional group was tested with different activities during the retention interval, allowing sleep to be compared against wakefulness rather than merely described.

2.2 The Behavioral Result

Performance improved significantly after the three-hour recording compared with the pre-sleep baseline (Kis et al., 2017). Dogs were better at the commands after sleeping than immediately after learning them.

2.3 Sleep Against Wakefulness

Comparing sleep with other activities during the retention interval showed effects on both short and long timescales — the state a dog is in after learning matters, and not only immediately (Kis et al., 2017).

A later and considerably larger study addressed the duration question directly. Across 64 dogs, a performance gap present before the retention interval had disappeared after sleep, with the effect requiring a minimum of roughly 25 minutes (Kovács et al., 2025a).

2.4 The Finding That Complicates It

The performance increase was not related to sleep duration, nor to any of the macrostructural sleep variables (Kis et al., 2017).

More sleep did not mean more improvement. Whatever produced the gain was not simply the quantity of sleep, which is the version of this finding most often repeated in training contexts (a pattern of oversimplification that recurs).

3. What the EEG Showed

3.1 Learning Changed the Sleep Spectrum

The spectral analysis is where the finding becomes substantive. Learning affected the EEG spectrum during subsequent sleep: during REM sleep, relative theta activity increased after learning, with changes also appearing in the delta and alpha ranges (Kis et al., 2017).

The brain of a dog that has just learned something sleeps differently from the brain of a dog that has not — which is the kind of physiological correlate that behavioral work alone cannot deliver (and that most canine constructs still lack).

3.2 Spectrum Predicted Improvement

More usefully, specific spectral features correlated with how much the dog improved. Decreased REM delta activity and increased REM beta activity were both associated with better post-sleep performance (Kis et al., 2017).

That is a link between a physiological measure taken during sleep and a behavioral outcome measured afterwards — which is what the field needs and what questionnaire studies cannot provide (given how much canine research rests on owner report).

3.3 It Depends on the Dog

One further result constrains how far any of this generalises. In highly trained dogs, the relationship between sleep measures and performance was no longer detectable (Kovács, Szűcs & Gácsi, 2025b).

The most plausible reading is a ceiling effect: a dog with extensive training history has less room to improve on a simple command task, and the consolidation signal disappears into that ceiling (which is why baseline ability has to be accounted for). It also means findings from pet dogs should not be transferred to working dogs without checking.

3.4 The Sample

Eleven dogs contributed to the EEG analyses and fifteen to the behavioral comparison (Kis et al., 2017). These are small numbers, and section 9 returns to what follows from that.

4. Sleep Spindles

4.1 What They Are

Sleep spindles are brief bursts of thalamo-cortical activity visible in the cortex as transient oscillations in the sigma range. In humans and rodents they are associated with sleep-dependent memory consolidation and sleep stability, and their occurrence, frequency, amplitude and duration vary with age, sex and psychiatric condition (Iotchev, Kis, Bódizs, van Luijtelaar & Kubinyi, 2017).

4.2 The Canine Finding

Spindle-like activity in dogs had been described qualitatively but never quantified or linked to function. Applying a detection method previously validated in children, researchers found that the density of EEG transients in the 9–16 Hz range during non-REM sleep relates to memory — and shows a sexual dimorphism similar to that reported in humans (Iotchev et al., 2017).

Reanalysis of the learning dataset found more spindles per minute in the learning condition than in the control condition, correlating with the learning gain across sleep.

4.3 Refinements and a Correction

Subsequent work established age-related differences and sexual dimorphism in canine spindles across a larger sample (Iotchev et al., 2019), and showed that averaging spindle occurrence across recordings predicts learning performance better than any single measure — interpreted as reducing measurement error while approximating a stable individual trait (Iotchev et al., 2020).

One point of transparency: the 2017 spindle paper carries a published author correction concerning errors in the methods section (Iotchev et al., 2018). The correction does not retract the finding, and it belongs in any honest account of the evidence.

5. Testing Causality

5.1 The Problem With Everything Above

Almost all of it is correlational. Spindle density correlates with learning gain; spectral features correlate with improvement. None of that establishes that the sleep produced the learning rather than accompanying it — a limitation the researchers themselves name as the field's central challenge.

5.2 The Reactivation Approach

The most direct attempt to move beyond correlation adapted targeted memory reactivation to dogs. Sixteen dogs learned commands associated with different locations and were then re-exposed to one of the cues during polysomnography, testing whether reactivating a memory during sleep affects its consolidation.

Averaged fast spindle density at the frontal site correlated significantly with relative latency reduction, though not at the central site — a partial result rather than a decisive one.

5.3 Why This Is Hard

Establishing causality would require manipulating sleep itself, and depriving dogs of sleep to see whether learning suffers raises welfare objections that do not apply to human volunteers (and chronic sleep disruption has its own consequences). The constraint is structural, not temporary.

6. The Complication Nobody Mentions

6.1 Play Works Too

A study in Labrador Retrievers found that a single thirty-minute play session administered shortly after learning improved performance on a two-object discrimination task (Affenzeller, Palme & Zulch, 2017).

Not sleep. Play.

6.2 What That Does to the Argument

If both sleeping and playing after learning improve subsequent performance, then the specific claim — that sleep consolidates canine memory — becomes harder to isolate. Both conditions share features: a break from the task, a change of activity, and elevated arousal followed by settling (with arousal itself affecting performance). What the interval provides may be recovery rather than consolidation, and recovery is an active process in its own right (as the anxiety literature describes).

6.3 What Survives It

The EEG findings do survive, and this is why they matter. A play session cannot explain why learning changes the subsequent sleep spectrum, or why spindle density correlates with the size of the gain. The behavioral improvement alone would be ambiguous; the physiological correlates are what make the sleep interpretation more than a guess (the same logic that makes objective measures valuable throughout).

7. What Follows for Training

7.1 The Defensible Recommendation

Give a dog an opportunity to rest after a learning session rather than moving straight into the next activity. This is supported by the behavioral data, costs nothing, and carries no risk of being wrong in a damaging direction.

7.2 What Not to Claim

That a specific duration is required — with one qualification. Improvement did not correlate with sleep duration within a three-hour window (Kis et al., 2017), while the larger study found the effect required a minimum of roughly 25 minutes to appear at all (Kovács et al., 2025a).

These are not contradictory. One is a threshold below which nothing happens; the other is the absence of a dose-response relationship above it. More sleep is not better; too little is worse. Any figure beyond that — the specific rest periods quoted in training literature — is not derived from this evidence base.

That sleep will fix a training problem. Consolidation acts on what was learned; it does not repair a poorly constructed session (where the problem is usually the emotional layer rather than the memory).

7.3 What Happens Before the Nap

A separate experiment reversed the question and asked what the hours before sleep do to it. In a within-subject design, dogs were exposed either to a positive social interaction — petting and ball play — or to a negative one combining separation, a threatening approach and a still-face test, and then recorded (Kis, Gergely et al., 2017).

Sleep macrostructure differed markedly between the two conditions. After the negative interaction, sleep latency was shorter and time was redistributed across the stages, with drowsiness decreasing and REM sleep increasing. The dogs' behavior during the pre-treatment predicted the size of the difference, and individual personality modulated it further.

The authors describe this as the first direct evidence that emotional stimuli affect subsequent sleep physiology in dogs. For practice it closes a loop: a stressful session does not end when the session ends (with the physiological consequences well documented), and falling asleep faster is not necessarily a sign that a dog is fine (a reading error that recurs across welfare assessment).

7.4 The Broader Point

What this literature supports is a general principle rather than a protocol: learning continues after the session ends, and the hours afterwards are not neutral (which applies to fear learning as much as to commands). Whether the dog rests, plays or is immediately confronted with something demanding is part of the training, not separate from it (and individual differences shape how much this matters).

8. Summary at a Glance

The method is the achievement — Fully non-invasive polysomnography in unrestrained pet dogs, no sedation, owner present (Kis et al., 2014).

Performance improved after sleep — Dogs performed significantly better on newly learned commands after a three-hour sleep recording than before it (Kis et al., 2017).

But not in proportion to sleep — The improvement was not related to sleep duration or to any macrostructural sleep variable (Kis et al., 2017).

A minimum duration exists, a dose-response does not — Across 64 dogs the effect required roughly 25 minutes to appear (Kovács et al., 2025a), while within a three-hour window more sleep did not mean more improvement (Kis et al., 2017).

The effect vanishes in highly trained dogs — No detectable relationship between sleep measures and performance in that population (Kovács et al., 2025b), most plausibly a ceiling effect.

Learning changed the sleep EEG — REM theta activity increased after learning, with further changes in delta and alpha ranges (Kis et al., 2017).

Spectral features predicted the gain — Decreased REM delta and increased REM beta activity were associated with better post-sleep performance (Kis et al., 2017).

Spindle density relates to memory — Transients in the 9–16 Hz range during non-REM sleep relate to memory, with a sexual dimorphism resembling the human pattern (Iotchev et al., 2017).

Experience before sleep alters the sleep itself — After a negative social interaction, sleep latency shortened, drowsiness decreased and REM sleep increased relative to a positive interaction (Kis, Gergely et al., 2017).

Play produced comparable gains — A thirty-minute play session after learning improved subsequent performance in Labradors (Affenzeller et al., 2017), which complicates the sleep-specific claim.

Nearly all of it is correlational — The causal question remains open, and the ethical route to answering it is largely closed.

9. Research Gaps and Critical Appraisal

The samples are very small. Eleven dogs in the EEG analyses, fifteen in the behavioral comparison, sixteen in the reactivation study. The methodology is excellent and the statistical power is limited, which is the standing trade-off in this field.

The subjects are self-selected. Dogs that tolerate an electrode cap, lie still for three hours in an unfamiliar room and fall asleep on cue are not a random sample. They are likely calmer, better habituated and more owner-focused than average.

Correlation dominates. Spindle density and spectral features correlate with learning gains. The direction of causation is not established, and the researchers say so explicitly.

Sleep deprivation is not available as a tool. The manipulation that would settle the question cannot ethically be performed on companion animals, which means this limitation is permanent rather than a matter of time.

The play finding is under-explored. A single study in one breed found comparable benefit from a non-sleep activity (Affenzeller et al., 2017). It has not been systematically followed up, and until it is, the specificity of the sleep effect remains uncertain.

One paper carries a correction. The 2017 spindle study has a published author correction regarding its methods section (Iotchev et al., 2018) — worth stating rather than quietly omitting.

Single recordings may not represent typical sleep. Repeated afternoon recordings in family dogs show a first-night-effect-like adaptation process (Reicher et al., 2020), meaning a dog's first session in the laboratory differs systematically from later ones. Studies using a single recording per dog are measuring sleep under partly novel conditions.

Most of the work comes from one research network. The Hungarian group has effectively created this field, which is a considerable achievement and also means independent replication is scarce (a structural weakness that appears elsewhere in canine research too).

10. Conclusion

Canine sleep research is methodologically the strongest area in the field, and it is worth being precise about what that strength has bought. Dogs perform better on newly learned commands after sleeping, learning measurably alters the EEG spectrum of the sleep that follows, and specific spectral and spindle features predict how much a given dog improves. That is a genuine chain from behavior to physiology and back, established without a single owner questionnaire. The qualifications are equally real. The improvement did not scale with sleep duration, which undercuts any prescription of a specific rest period. A thirty-minute play session after learning produced comparable behavioral gains in a separate study, which means the sleep interpretation rests on the physiological correlates rather than on the performance data alone. The samples number in the low teens, the dogs are self-selected for calmness, and the manipulation that would establish causation cannot ethically be run. What remains is a defensible and modest recommendation: the period after a training session is part of the training, rest during it is well supported and free, and any claim more specific than that is running ahead of the evidence (much as the developmental claims do).

Key Insights (Takeaways)

• Dogs performed better after sleeping, but not in proportion to how long they slept. Performance on newly learned commands improved significantly after a three-hour polysomnography recording, while the improvement showed no relationship to sleep duration or to any macrostructural sleep variable (Kis et al., 2017). Any training recommendation specifying a number of minutes is not derived from this evidence.

• Learning changed the sleep that followed, and the changes predicted the gain. REM theta activity increased after learning, and decreased REM delta together with increased REM beta activity were associated with better post-sleep performance (Kis et al., 2017). This is the part of the finding that cannot be explained by anything other than sleep-related processing.

• Spindle density relates to memory in dogs as in humans. Transients in the 9–16 Hz range during non-REM sleep relate to memory and show a sexual dimorphism resembling the human pattern (Iotchev et al., 2017), with averaged density across recordings predicting learning performance better than single measures (Iotchev et al., 2020).

• A play session after learning worked comparably well. Thirty minutes of play following a discrimination task improved subsequent performance in Labrador Retrievers (Affenzeller et al., 2017). This rarely appears alongside the sleep findings and it should: whatever the post-learning interval provides may not be specific to sleep.

• There is a floor but no ceiling benefit. The largest canine study found the consolidation effect required roughly 25 minutes to appear across 64 dogs (Kovács et al., 2025a), while within a three-hour window the size of the improvement showed no relationship to how long the dog slept (Kis et al., 2017). Short rests are not enough and long ones are not better — and in highly trained dogs the relationship disappeared entirely (Kovács et al., 2025b).

• The causal question may remain difficult to resolve. Nearly all canine findings here are correlational, and the manipulation that would settle it — depriving dogs of sleep after learning — is not ethically available for companion animals. That constraint is structural rather than a gap awaiting the next study.

References

Affenzeller, N., Palme, R., & Zulch, H. (2017). Playful activity post-learning improves training performance in Labrador Retriever dogs (Canis lupus familiaris). Physiology & Behavior, 168, 62–73. https://doi.org/10.1016/j.physbeh.2016.10.014

Iotchev, I. B., Kis, A., Bódizs, R., van Luijtelaar, G., & Kubinyi, E. (2017). EEG transients in the sigma range during non-REM sleep predict learning in dogs. Scientific Reports, 7, 12936. https://doi.org/10.1038/s41598-017-13278-3

Iotchev, I. B., Kis, A., Bódizs, R., van Luijtelaar, G., & Kubinyi, E. (2018). Author correction: EEG transients in the sigma range during non-REM sleep predict learning in dogs. Scientific Reports, 8, 6045. https://doi.org/10.1038/s41598-018-24096-6

Iotchev, I. B., Kis, A., Turcsán, B., Tejeda Fernández de Lara, D. R., Reicher, V., & Kubinyi, E. (2019). Age-related differences and sexual dimorphism in canine sleep spindles. Scientific Reports, 9, 10092. https://doi.org/10.1038/s41598-019-46434-y

Iotchev, I. B., Reicher, V., Kovács, E., Kovács, T., Kis, A., Gácsi, M., & Kubinyi, E. (2020). Averaging sleep spindle occurrence in dogs predicts learning performance better than single measures. Scientific Reports, 10, 22461. https://doi.org/10.1038/s41598-020-80417-8

Reicher, V., Kis, A., Simor, P., Bódizs, R., Gombos, F., & Gácsi, M. (2020). Repeated afternoon sleep recordings indicate first-night-effect-like adaptation process in family dogs. Journal of Sleep Research, 29(6), e12998. https://doi.org/10.1111/jsr.12998

Kovács, E., Kosztolányi, A., Kubinyi, E., & Gácsi, M. (2025a). Sleep-dependent memory consolidation in family dogs. Animals, 15(21), 3074. https://doi.org/10.3390/ani15213074

Kovács, E., Szűcs, A., & Gácsi, M. (2025b). Sleep and learning performance in highly trained dogs. The Veterinary Journal, 314, 106483. https://doi.org/10.1016/j.tvjl.2025.106483

Kis, A., Gergely, A., Galambos, Á., Abdai, J., Gombos, F., Bódizs, R., & Topál, J. (2017). Sleep macrostructure is modulated by positive and negative social experience in adult pet dogs. Proceedings of the Royal Society B: Biological Sciences, 284(1865), 20171883. https://doi.org/10.1098/rspb.2017.1883

Kis, A., Szakadát, S., Gácsi, M., Kovács, E., Simor, P., Török, C., Gombos, F., Bódizs, R., & Topál, J. (2017). The interrelated effect of sleep and learning in dogs (Canis familiaris); an EEG and behavioural study. Scientific Reports, 7, 41873. https://doi.org/10.1038/srep41873

Kis, A., Szakadát, S., Kovács, E., Gácsi, M., Simor, P., Gombos, F., Topál, J., Miklósi, Á., & Bódizs, R. (2014). Development of a non-invasive polysomnography technique for dogs (Canis familiaris). Physiology & Behavior, 130, 149–156. https://doi.org/10.1016/j.physbeh.2014.04.004