Sleep Modulates Neural Timescales and Spatiotemporal Integration in the Human Cortex

The study of intrinsic neural dynamics is fundamental for understanding brain function. Neural activity fluctuates across brain areas at characteristic times (neural timescales) that follow a hierarchical organization across recording modalities and species. Timescales are short in sensory areas and progressively increase along cortical hierarchies, influenced by microstructure (e.g., myelination), excitation/inhibition-related gene expression, and connectivity patterns. Interconnected areas tend to have longer timescales, supporting information integration. Although often considered intrinsic properties, single-neuron timescales vary within an area and multiple timescales can coexist, as neural oscillations at different frequencies are hierarchically organized but not unique. Timescales also change with brain states, particularly from wake to sleep. Sleep drastically alters neural dynamics, with global oscillatory signatures (theta in REM, slow waves in NREM3), yet sleep dynamics also show local, region-specific changes. Prior studies report mixed conclusions on how timescales change during sleep: increases with scalp EEG, fMRI, and spiking activity, versus decreases when measured from gamma power in iEEG. This discrepancy may reflect heterogeneity and nonuniform changes across circuits during sleep. Spatial integration across brain networks also changes with sleep; NREM is often viewed as a disconnected state, though slow waves may enable communication windows. In wake, higher timescales associate with stronger connectivity and spatial correlations, but whether this holds in sleep is unclear. Therefore, it is important to quantify both temporal and spatial correlations, and their relationship (spatiotemporal integration). Here, using human iEEG, the authors aimed to characterize neural timescales, spatial correlations, and their relationship across wake, NREM3, and REM sleep, hypothesizing that (1) timescales are regionally reconfigured in sleep; (2) timescales relate to spatial correlations; and (3) coexisting timescale hierarchies are associated with different frequency ranges.

Dataset: Open MNI Open iEEG dataset of drug-resistant focal epilepsy patients (106 patients; 48 females; ages 13–62). Stages analyzed: wake, NREM3, REM. Contacts: wake 1,772 (106 patients), NREM3 1,468 (91), REM 1,012 (65). Each stage had 1 min segments; NREM3 also had 10 min segments, used specifically for slow-wave analyses. Preprocessing (per original dataset): resampled to 200 Hz, bandpass 0.5–80 Hz, rigorous selection of non-epileptiform channels/segments. Any 2 s zero buffers between concatenated segments excluded. Original channels localized to 38 gray matter areas (MNI parcellation). Cortical parcellations: Channel-level metrics (timescales, slow-wave density) were mapped to the 180 left-hemisphere parcels of the HCP-MMP parcellation. Projection from MNI to fsaverage used NiBabel and MNE-Python. Each channel’s metric was extrapolated to neighboring voxels via a Gaussian weight w(r)=exp(-(r^2/d^2)) with d chosen so w(4 mm)=0.5. Within each patient, voxel values were weighted sums of channels; within-parcel, per-patient averages were computed; final parcel maps obtained by averaging across patients weighted by maximum parcel weight per patient. Anatomical hierarchy: Used HCP S1200 group-average myelin (T1w/T2w) map in HCP-MMP parcels (median per parcel) as an MRI-based proxy for cortical hierarchy (high in sensory, low in associative regions). Neural timescales: Computed from two signals: (a) broadband iEEG (0.5–80 Hz voltage), (b) gamma power (40–80 Hz bandpass, Hilbert transform; power as squared magnitude of analytic signal; log-transformed). Autocorrelation functions (ACFs) were computed in sliding 1 s windows with 0.5 s overlap: ACF(k) = (1/(N−k)) Σ_i (X_i − μ̂)(X_{i+k} − μ̂) / σ̂^2. Window ACFs were averaged per channel. Timescales τ were obtained by fitting an exponential f(k)=a(e^(−k/τ)+b). ACF fitting ranges: broadband [0, 500] ms; gamma [15, 300] ms (to avoid initial fast decay dominance). Validation: Compared to FOOOF knee model (periodic/aperiodic separation) to account for oscillatory peaks; high agreement across stages (Spearman p>0.67, Pcorr<0.001). FOOOF fitting was harder in sleep due to frequent knees at 0 (infinite τ). No evidence of oscillation-induced bias in fits. Slow-wave detection (NREM3): Performed on 10 min NREM3 recordings to improve detection. Used two signals per channel: low-frequency 0.5–4 Hz and gamma power 30–80 Hz (slightly different band than for timescales to improve detection). Criteria per channel: positive peak duration 0.25–1 s, negative trough duration 0.25–1 s, total duration 0.5–2 s; retained top 25% peak-to-peak amplitude events independently per channel. Polarity check: ensured the average slow-wave trough coincided with decreased gamma power; otherwise inverted signal polarity and recomputed events. Implemented with yasa. Slow-wave density: number of slow waves per minute. For time-resolved timescales around slow waves: computed ACFs in 1 s windows (0.5 s overlap) from −2 to +2 s around troughs; timescales extracted per window, then parcellated to HCP-MMP. Spatial correlations (SC): Connectivity proxy computed as maximum absolute cross-correlation across lags between pairs of channels within the same patient and hemisphere. CCF per 1 s window (0.5 s overlap): CCF_{a,b}(k) = (1/(N σ̂_a σ̂_b)) Σ_j (X_{a,j} − μ̂_a)(X_{b,j+k} − μ̂_b). Averaged across windows per pair; SC defined as max |CCF| across lags. Total pairs: wake 19,459; NREM3 14,804; REM 10,364. Distance between channels: Euclidean distance of MNI coordinates (approximation of anatomical distance). SC decay over distance fit with f(r)=a(e^(−r/d)+b). For relating SC and timescales, used 38-area MNI parcellation for larger parcels: timescales estimated via linear mixed-effects models (patients as random intercepts) and SC pooled per parcel (if either channel belonged) and modeled similarly. Correlated parcel-wise timescales with mean SC in distance bins of 20 mm, sliding by 10 mm from 10 to 120 mm. Statistics: Emphasized estimation statistics for stage differences (bootstrap, n=9,999; CIs from 2.5–97.5 percentiles). Map-to-map correlations via Spearman; corrected for spatial autocorrelation using permutation (“vasa” method; 1,000 nulls; two-tailed Pcorr) with cortical parcel permutations (excluding hippocampus and amygdala but using them in both true and permuted correlations). Multiple comparisons across distance bins controlled via FDR. Mixed-effects modeling used R nlme. Data and code: Data accessible at https://mni-open-ieegatlas.research.mcgill.ca/. Analysis code: https://github.com/cusinatr/MNI-Analysis

• Two complementary neural timescales from iEEG were characterized: broadband (0.5–80 Hz voltage) and gamma (40–80 Hz power). Both increased in sleep relative to wake, with distinct hierarchical patterns across cortex.
• Broadband timescales across stages: • Marked increase in NREM3 across all cortical areas (mean increase 105 ms; 95% CI [102, 108]). • Moderate increase in REM across 163/180 HCP-MMP areas (mean increase 16 ms; 95% CI [14, 18]). • Only the presubiculum (MTL) increased from NREM3 to REM. • Hierarchy: Negative correlation with T1w/T2w (sensorimotor-association increase) in wake (ρ = −0.51, Pcorr < 0.001) and NREM3 (ρ = −0.45, Pcorr < 0.001); absent in REM (ρ = −0.12, Pcorr = 0.60).
• Gamma timescales across stages: • Increased in NREM3 across all cortical areas (mean increase 31 ms; 95% CI [29, 34]). • Slight increase in REM across 92/180 areas (mean increase 2 ms; 95% CI [1, 3]). • V6 and V7 increased from NREM3 to REM. • Gamma timescales lower magnitude than broadband and not obviously correlated with broadband within areas (wake: ρ = −0.10, Pcorr = 0.47; NREM3: ρ = −0.09, Pcorr = 0.60; REM: ρ = 0.01, Pcorr = 0.99). • Hierarchy: Positive correlation with T1w/T2w (higher in sensory regions). Wake: ρ = 0.24, Pcorr = 0.09; NREM3: ρ = 0.58, Pcorr < 0.001; REM: ρ = 0.55, Pcorr < 0.001.
• Slow waves (NREM3) and timescales: • Time-locked to slow-wave troughs, broadband timescales increased by 171% relative to baseline; gamma by 51% (broadband effect approximately threefold larger). • Channel-level adherence: 99.7% channels showed broadband increase pattern; 65.2% for gamma. • Slow-wave density correlated strongly with broadband timescales across areas (ρ = 0.71, Pcorr < 0.001), but not with gamma (ρ = −0.03, Pcorr = 0.88). Partial correlations controlling for T1w/T2w confirmed these relationships (broadband: ρ = 0.66, p < 0.001; gamma: ρ = 0.07, p = 0.36; uncorrected p values).
• Spatial correlations (SC): • SC vs distance followed an exponential decay for both signals; NREM3 exhibited the highest long-range SC, especially for broadband. • Area-averaged SC changes (vs wake): NREM3 increased (broadband: mean diff 0.018, 95% CI [0.012, 0.022]; gamma: 0.003, 95% CI [0.002, 0.004]); REM decreased (broadband: −0.011, 95% CI [−0.016, −0.007]; gamma: −0.001, 95% CI [−0.003, −0.0005]). • When averaging SC across all distances there was no robust correlation with timescales. However, distance-resolved analyses revealed positive, distance- and stage-dependent associations:
  • Wake: Broadband SC positively correlated with broadband timescales at long distances (>50 mm; correlation coefficients > 0.40; Pcorr < 0.02 after FDR).
  • NREM3: Gamma SC positively correlated with gamma timescales at short distances (<30 mm; correlation coefficients > 0.46; Pcorr < 0.02 after FDR).
  • REM: No significant SC–timescale correlations at any distance.
• Overall, broadband timescales and SC indicate higher long-range spatiotemporal integration in wake; gamma timescales and SC suggest higher short-range integration during NREM3.

The study addressed how intrinsic neural dynamics, specifically neural timescales and spatial correlations, reorganize across wake and sleep. It revealed two mesoscopic timescale hierarchies in human cortex derived from broadband and gamma signals. Both timescales increased in sleep, particularly NREM3, yet they displayed opposite relationships to cortical hierarchy: broadband timescales lengthened along the sensorimotor-association axis (longest in associative regions), whereas gamma timescales were relatively longer in sensory regions. These opposing hierarchies suggest distinct physiological substrates and functional roles for the two signals. The pronounced influence of NREM3 slow waves on timescales—especially broadband—helps reconcile prior observations of increased sleep timescales in modalities dominated by low frequencies. Spatial correlations changed with state and co-varied with timescales in a distance- and stage-specific fashion: broadband-based spatiotemporal integration was strongest over long distances in wake, consistent with coordinated large-scale processing that supports cognition, while gamma-based integration was enhanced locally (short distances) in NREM3, aligning with a view of NREM as a relatively disconnected state with local coordination windows. Discrepancies with past reports of decreased gamma timescales in sleep may stem from species differences, methodological differences (e.g., epoch length affecting timescale ranges), or regional sampling. Overall, the findings indicate that cortical networks host coexisting timescales shaped by anatomy and state, and that their coupling with spatial correlations reflects distinct regimes of spatiotemporal integration across vigilance states.

This work demonstrates that human cortical mesoscopic activity exhibits two complementary neural timescale hierarchies—broadband and gamma—that both lengthen during sleep but follow opposite anatomical gradients. Broadband timescales increase along the sensorimotor-association axis, whereas gamma timescales are longer in sensory regions. NREM3 slow waves strongly drive timescale increases, particularly for broadband, and areas with longer timescales tend to show higher spatial correlations in a distance- and state-dependent manner (long-range broadband integration in wake; short-range gamma integration in NREM3). These results suggest distinct physiological processes underpinning broadband and gamma dynamics and distinct functional roles across states. Future work using continuous, longer iEEG recordings and multimodal approaches could delineate within-state dynamics, examine cross-frequency interactions, and directly link structural connectivity and microarchitecture to state-dependent timescales and spatiotemporal integration.

• Participant population and sampling: Data derive from epilepsy patients (though non-epileptiform tissue was carefully selected), and iEEG spatial sampling is clinically determined and uneven across cortex. Sleep was not recorded in all patients, leading to fewer contacts in NREM3 and REM than in wake.
• Anatomical distance approximation: Channel distances were estimated via Euclidean distance in MNI space; individual anatomical pathways were unavailable, and this approximation may not capture true axonal distances.
• Connectivity inference: Cross-correlation (SC) reflects signal similarity and is influenced by low frequencies; it does not establish effective or causal connectivity. Potential contributions from cross-frequency coupling (e.g., delta–gamma) were not explicitly analyzed.
• Timescale estimation constraints: Gamma timescales were estimated on 1 s windows (vs longer windows in some prior work), yielding millisecond-range timescales that may index different aspects of dynamics. FOOOF fits in sleep sometimes produced knee=0 (infinite τ), complicating validation.
• Parcellation and hemisphere coverage: Timescale parcellation and hierarchy analyses were conducted in the left hemisphere (HCP-MMP). Mapping from MNI to surface involves smoothing assumptions; coverage was highest in sensorimotor, temporal, and frontal regions.
• Stage and data constraints: Analyses emphasized 1 min segments for consistency; 10 min data were used only for slow-wave detection. Slight differences in gamma band definitions were used for timescale vs slow-wave analyses. Potential confounds from antiepileptic medications or inter-individual variability are not fully controlled beyond mixed-effects modeling.