Averaging sleep spindle occurrence in dogs predicts learning performance better than single measures

Sleep spindles are thalamocortical oscillations observed in mammalian non-REM sleep (typically 0.5–5 s bursts in the 9–16 Hz range). In humans, many studies have linked spindle occurrence to post-sleep learning success, but most positive results come from small samples and different automated detection methods yield divergent estimates. The largest human study failed to find the association, leaving the reliability of the effect in question. Animal work suggests plausible mechanisms and some causal evidence for spindle contributions to memory consolidation. This study asks whether the association between sleep spindle occurrence and learning success is reliably present in dogs, and whether averaging across multiple measurements (to approximate trait-like spindle density and reduce measurement error) better predicts learning than single-session measures. The dog model also allows standardization by applying a single validated detection algorithm across datasets, minimizing detector-related variability and potential publication bias.

Prior literature in humans often reports positive correlations between spindle density and subsequent recall across declarative and procedural tasks, though findings are inconsistent and sometimes null in larger samples. Automated spindle detectors vary in their outputs, complicating comparisons. Invasive animal studies indicate mechanistic roles for spindles in memory consolidation (e.g., LTP induction; triple phase-locking with hippocampal ripples and cortical slow waves), supporting functional relevance beyond correlation. Spindle activity can increase after learning in humans, rats, and dogs. Spindle traits show night-to-night stability and heritability, and reductions in spindle activity are linked to cognitive impairments and aging. Publication bias and measurement error are concerns; averaging across sessions may better capture trait-like properties and improve reliability.

Design: Replication analysis across three canine datasets using a common learning paradigm (novel words paradigm) and a single automatic spindle detection algorithm previously validated for dogs. Each dataset included an adaptation sleep followed by two counterbalanced learning conditions manipulating social and reinforcement context.

Datasets and conditions: Data set 0 (N=15; previously published) included adaptation, control, and learning conditions; Data set 1 (N=19) and Data set 2 (N=13) included adaptation plus two learning conditions each: 1a supportive training (food + social reward, no scolding) and 1b controlling training (food reward only, scolding on incorrect); 2a owner-led socially relevant training (food + social reward) and 2b unfamiliar experimenter, socially irrelevant (food only, no social reward). One dog appeared in datasets 0 and 2; otherwise samples were independent.

Participants: 46 pet dogs (23 females), age 1–9 years; 28 purebreds from 16 breeds. Missing spindle density values were assigned where dogs did not reach non-REM sleep or data were corrupted; missing cases were excluded from analyses and from averages.

Learning tasks and outcomes: Before sleep, dogs learned associations between novel English words and previously trained actions (Hungarian verbal commands). After sleep, performance was tested (18 trials). Outcomes: final performance (percent correct at re-test) and learning gain (re-test percent minus test percent). For datasets 1 and 2, learning outcomes were averaged across the two learning conditions; dataset 0 had only one learning condition, so no averaging of learning outcomes was possible.

EEG acquisition and preprocessing: Non-invasive polysomnography with electrode placement per Kis et al. Recordings manually scored into sleep stages per validated canine criteria. Spindle search restricted to non-REM epochs. Spindles detected at frontal (Fz) and central (Cz) midline electrodes; primary analyses focused on Fz because only Fz was available in dataset 0.

Automatic spindle detection: Implemented as in Iotchev et al. Signals filtered 5–16 Hz; two-step individualized detector. Step 1: initial detections within 9–16 Hz and amplitude >1 SD above mean. Step 2: maximum likelihood estimates for individual amplitude and frequency distributions; final detections constrained within ±2 SD of estimated means. Spindle density computed as spindles per minute of non-REM sleep. Analyses also considered slow vs fast spindle subtypes (see Supplementary), but main text emphasizes 9–16 Hz detections.

Statistical analysis: Pearson correlations tested associations between spindle density and learning measures using SPSS v25. Analyses examined single-session measures and averages across sessions/conditions to estimate trait-like spindle density. Primary focus on Fz and full 9–16 Hz band, with subtype/electrode details in Supplementary Tables.

Single-session analyses (Fz, 9–16 Hz):

• Dataset 0: Positive association between spindle density and learning gain (e.g., r ≈ 0.647, p = 0.009; N=15, figure reference). Effect specific to slow spindles; trend on Cz for condition 1a in other datasets.
• Dataset 1, condition 1a (supportive): Positive association with learning gain (r ≈ 0.560, p = 0.019; N=17 after excluding non-sleepers).
• Dataset 1, condition 1b (controlling): No association (r = -0.130, p = 0.595; N=19).
• Dataset 2, condition 2a: No significant association (r = 0.188, p = 0.539; N=13).
• Dataset 2, condition 2b: No significant association (r = -0.437, p = 0.156; N=12 after excluding a non-sleeper). Across these, significant effects were specific to the slow spindle subtype and the frontal (Fz) electrode.

Averaged (trait-like) measures:

• Averaging spindle density across recordings increased detectability of associations.
• Dataset 0: Averaged density positively associated with learning gain (r = 0.575, p = 0.025; N=15).
• Dataset 2: Averaged density positively associated with average final performance (r = 0.595, p = 0.032; N=13).
• Dataset 1: No significant associations with averaged learning variables; condition 1b showed a floor effect with 57.9% (11/19) of dogs performing worse after sleep, versus 36.8% (7/19) in 1a. Significant averaged associations were also specific to slow spindles and Fz.

The study replicated positive associations between sleep spindle occurrence and post-sleep learning success in dogs across three datasets when using a single, standardized detection algorithm and a common learning paradigm. Importantly, averaging spindle density across sessions to estimate a trait-like measure improved the visibility of the spindle–learning relationship compared to single measurements. The specificity of effects to slow spindles at frontal sites parallels human findings for verbal learning tasks. Null results in some conditions likely reflect low power and contextual constraints (e.g., a controlling training condition producing floor effects). Other modulators such as coupling with hippocampal ripples and slow waves, emotional arousal, and specific non-REM substages may influence whether effects emerge, but are difficult to precisely quantify in non-human studies. The results support the view that stable, trait-like spindle occurrence is a better predictor of cognitive performance than single-session measures and align with evidence for the stability and heritability of spindle metrics and their reduction in conditions associated with cognitive deficits.

Using a standardized automatic detector and a consistent learning paradigm, this work adds support to a positive association between sleep spindle occurrence and learning success in dogs. Averaging across sessions to estimate trait-like spindle density predicts learning performance better than single measurements, with effects driven by slow frontal spindles. Future research should include larger samples to address power, employ designs with control (no-learning) sleep conditions to disentangle memory consolidation from general learning potential, and further examine state-dependent modulators (e.g., sleep stage specificity, oscillatory coupling) that may condition the spindle–learning relationship.

• Small sample sizes in each dataset and condition (often N < 20), increasing the risk of Type II errors.
• Missing data due to some dogs not achieving non-REM sleep; exclusions reduce power and may bias averages.
• Potential floor effects in certain training conditions (e.g., controlling condition 1b), limiting detection of associations.
• Difficulty distinguishing non-REM substages in dogs constrains analysis of stage-specific effects.
• Single research group and one detection algorithm improve consistency but limit generalizability across methods.
• Publication bias noted in human literature; while multiple unpublished datasets were included here, broader preregistered replications would strengthen inference.