Sleep spindles in the healthy brain from birth through 18 years

Abstract

Study Objective

Sleep spindles are present from birth and reflect cognitive functions across the lifespan, but normative values for this cognitive biomarker across development are lacking. This study aims to establish normative spindle features over development.

Methods

All available normal 19-channel electroencephalograms from developmentally normal children between February 2002 and June 2021 in the MGH EEG lab were analyzed. Approximately, 20 000 spindles were hand-marked to train and validate an automated spindle detector across ages. Normative values for spindle rate, duration, frequency, refractory period, and interhemispheric lag are provided for each channel and each age.

Results

Sleep EEGs from 567 developmentally normal children (range 0 days to 18 years) were included. The detector had excellent performance (F1 = 0.47). Maximal spindle activity is seen over central regions during infancy and adolescence and frontopolar regions during childhood. Spindle rate and duration increase nonlinearly, with the most rapid changes during the first 4 months of life and between ages 3 and 14 years. Peak spindle frequency follows a U-shaped curve and discrete frontal slow and central fast spindles are evident by 18 months. Spindle refractory periods decrease between ages 1 and 14 years while interhemispheric asynchrony decreases over the first 3 months of life and between ages 1 and 14 years.

Conclusions

These data provide age- and region-specific normative values for sleep spindles across development, where measures that deviate from these values can be considered pathological. As spindles provide a noninvasive biomarker for cognitive function across the lifespan, these normative measures can accelerate the discovery and diagnosis in neurodevelopmental disorders.

Graphical Abstract

Graphical Abstract.

Statement of Significance.

Work in the last two decades has identified sleep spindles, discrete “sigma band” oscillations during stage 2 sleep, as a key oscillatory mechanism required for offline memory consolidation. Although sleep spindles are known to evolve concomitantly with brain maturation and reflect cognitive function across the lifespan, the details of this developmental trajectory are unknown. We demonstrate that sleep spindle features follow distinct age-specific patterns in distribution, rate, duration, frequency, estimated refractory period, and interhemispheric spindle lag. These data expand our current knowledge of normal physiological brain development and provide a large normative database to detect deviations in sleep spindles to aid discovery, biomarker development, and diagnosis in neurodevelopmental disorders.

Introduction

Sleep spindles, discrete 0.5–2 sec bursts of 9–15 Hz “sigma band” oscillations, are a hallmark feature of electroencephalogram (EEG) recordings during stage two non-rapid eye movement (N2) sleep [1, 2]. Work in the last two decades has identified spindles as a key oscillatory mechanism required for offline memory consolidation during N2 sleep [3–5]. Human cognitive studies have linked spindles to sleep-dependent consolidation of both procedural [6, 7] and declarative [8, 9] memory tasks. Spindle rate correlates with general cognitive abilities and sleep-dependent memory consolidation in both typically developing children [10–14] and those with neurologic disorders, including autism [15], developmental delay [15], and epilepsy [13]. Confirming a mechanistic role, interventions that increase spindle activity result in improved sleep-dependent memory consolidation [16–19]. Sleep spindles gate dendritic calcium shifts required for synaptic plasticity [20] and coordinate the reactivation patterns of sharp-wave ripples [4, 21], oscillations that reflect neuronal replay [22]. Thus, sleep spindles provide a powerful noninvasive neurophysiological biomarker for thalamocortical circuitry and cognitive function across development.

Sleep spindles are generated by GABAergic neurons in the thalamic reticular nucleus and elaborated by well-delineated thalamocortical circuits [23]. In newborns, immature “pre-spindles” appear before 3 weeks of age, followed by robust and prominent asynchronous spindles between 3 and 9 weeks of age [24]. These prominent spindle oscillations evolve dynamically in rate, duration, frequency, and interhemispheric synchrony, with the most marked changes reported in the first year of life [25]. Sleep spindles become increasingly synchronous between hemispheres over the first 2 years of life, and spindle frequencies increase linearly from childhood to adolescence in the frontal regions [2]. Previous studies have suggested that different spindle populations, slow spindles (< 13 Hz) and fast spindles (>13 Hz) have different cognitive roles [26], with frontal slow spindles more consistently related to declarative memory tasks [5, 9, 27] and central fast spindles related to motor learning [5, 28]. The reported frequency border between fast and slow spindles varies across ages and studies [5, 29], and when these distinct populations emerge over development is unknown.

Although sleep spindles offer a unique, noninvasive biomarker of cognitive function across the lifespan, normative values for spindle features are lacking for pediatric age groups. To address this gap, we curated and analyzed a database of scalp EEG recordings during N2 sleep from a large cohort of developmentally normal children at the time of EEG from age 0 days to 18 years. We then trained and validated an automated spindle detector with hand-marked spindles across age groups. Using this detector, we provide normative values for several sleep spindle features—rate, duration, percentage, peak frequency, refractory period, and interhemispheric synchrony—across typical development. These data expand our current knowledge of normal physiological brain development and provide a large normative database to accelerate the discovery and diagnosis in neurodevelopmental disorders.

Methods

Participants and EEG recordings

Participants aged 0–18 years with available normal clinical EEG recordings between February 2002 and June 2021 in which N2 sleep was captured were identified from the Massachusetts General Hospital EEG laboratory. Prior to EEG appointments at our institution, participants were instructed to limit their sleep to half the typical duration the night prior to the ~1 hr EEG recording. Clinical chart review was performed and EEGs from children with term delivery (>37 weeks gestational age), no neuroactive medications during the recording period, documented normal neurodevelopment from birth through the follow-up visit following the EEG, and events leading to EEG evaluation not expected to alter EEG rhythms were included. Although longitudinal follow-up was not required for inclusion in this study, we reviewed all available clinical data and any participants who subsequently developed an autism spectrum disorder, developmental delay in any domain, or psychotic disorder were excluded. Children with chronic neurologic or psychiatric diagnoses were excluded, except those diagnosed with mild attention or depression symptoms or tics not requiring medication treatment. Children with a history of provoked seizures, including febrile seizures were excluded. In addition, EEGs from 27 children (6–18 years) recruited as population controls in research protocols were included.

Clinical EEG data were acquired following the international 10–20 system (Fp1, Fp2, F3, F4, C3, C4, P3, P4, F7, F8, T3, T4, T5, T6, O1, O2, Fz, Cz, and Pz; Xltek, a subsidiary of Natus Medical). Sampling frequencies ranged from 200 to 512 Hz. Research EEGs were acquired using a 64-channel cap (Easycap) at 2035 Hz and only the subset of 10–20 channels were selected and the data were downsampled to 407 Hz for analysis. In each case, impedances were maintained below 10 kΩ. Channels with poor recording quality and periods of significant artifact were ignored. Epochs containing N2 sleep (or trace alternant or quiet sleep patterns for <1 month-old participant) were identified [30]. The N2 EEGs were re-referenced to an average signal for subsequent analysis. This study was conducted in accordance with protocols approved by the local Institutional Review Board according to National Institutes of Health guidelines.

Manual spindle detections

Sleep spindles were manually marked by consensus in 100 s of EEG data from 19 channels in 93 healthy participants ages 0–18 years (n = 47 infants ages 0–2 years old, n = 46 ages 2–18 years). To train the automated spindle detector for subsequent use in both healthy and disease states, we included manual spindle markings from children with continuous spike and wave sleep with encephalopathy (n = 10, ages 3–18 years) and with Rolandic epilepsy (n = 12, ages 4–15 years) in our training dataset [13]. This resulted in 19 625 manually marked spindles from 115 unique pediatric participants.

Sleep spindle detector

Typical N2 sleep architecture includes both vertex waves and K-complexes, brief events with increasingly sharper slopes at younger ages [30]. Such sharp events are typically presented by the age of 5 months and produce wideband spectral features [31] that impact measures of sigma power common to many spindle detection approaches [2, 13, 32]. For accurate detection of spindles in the setting of sharp sleep architecture, we applied an automated latent state model (LS) spindle detector that we developed specifically to perform well in the setting of sharp events in the EEG [13]; code to replicate or modify this detector is available at https://github.com/Mark-Kramer/Spindle-Detector-Method.

Application of the spindle detector to this pediatric dataset consisted of three steps: (1) training, (2) validation across age groups, and (3) application. For each step and channel, we evaluate 0.5-sec intervals of data and compute three EEG features: theta band power (4–8 Hz), sigma band power (9–15 Hz), and the Fano factor of the oscillation intervals—a measure of cycle regularity. We chose 0.5-sec intervals, which are the typical minimum duration accepted for sleep spindles [2, 33–35], to maintain a 2 Hz frequency resolution, allowing reliable estimation of the theta band power. We advance each 0.5-sec interval by 0.1 sec, enabling the detection of spindles at least 0.5 sec in duration with 0.1-sec resolution. To estimate the theta and sigma power in an interval, we detrended the (unfiltered) data, applied a Hanning taper, computed the Fourier transform multiplied by its complex conjugate, and divided the power at each frequency by the summed total power from 1 to 50 Hz. To compute the variability of oscillation cycles, we considered bandpass filtered data between 3–25 Hz (FIR, stop band attenuation 40 dB at 3 Hz, stop band attenuation 20 dB at 25 Hz, passband ripple 0.1 dB) and identified the peaks and troughs (minimum peak distance 28 ms, minimum peak prominence 2 uV); here filtering reduced the impact of a high-frequency activity on peak/trough detections. Then, to characterize the variability of the times between adjacent peaks and troughs we computed the Fano factor [36]. We took the natural logarithm of each feature, shifted the interval by 0.1 sec, and repeated these computations for the entire duration of the EEG signal for each channel.

To train the LS spindle detector, we used the manual spindle detections described above. For each channel (n = 18 channels) and each participant (n = 115), we computed for each 0.5-sec interval the three features and assigned a spindle state label (“spindle” or “not spindle”). An interval was designated as “spindle” only if the entire interval lay within the bounds of a manually marked spindle. From these data, we fit empirical likelihood functions to each feature and state and estimated the transition matrix between the “spindle” and “not spindle” states. The LS spindle detector estimates the probability of a spindle in each 0.5-sec interval, providing an easily interpretable value. Briefly, the probabilities of the spindle (p_yes) and not spindle (p_no) state are first initialized at 0.5. Then, for the first 0.5-sec interval, the transition matrix is applied to [p_yes, p_no] to compute the one-step prediction [p¹_yes, p¹_no]. The three features for this interval are then computed and the likelihood of each feature is used to compute the posterior [p^*_yes, p^*_no]. The posterior is normalized so that the probability sums to one (i.e. p^*_yes + p^*_no = 1), where p^*_yes is the probability of a spindle for this interval. This process is repeated for each 0.5-sec interval in 0.1-sec steps for the duration of the signal, where the [p_yes, p_no] for each interval is set to the normalized posterior of the previous interval.

To validate the LS detector across age groups, we performed a leave-one-out cross-validation. To do so, we trained the detector with one participant omitted. We then applied this LS detector to estimate the probability of a spindle for each 0.5-sec interval of each marked electrode of the omitted participant. We repeated this process for all 115 participants. We then evaluated detector performance across all participants using standard measures (see Statistical Analysis). We found that the probability threshold value of 0.95 is optimal across ages, so spindle detections were identified when the probability of a spindle exceeded this threshold value (i.e. when the probability of a spindle exceeded 95%). Spindle detections separated by less than 1 sec were concatenated [2, 13].

Analysis of spindle features

From the detected spindles, rate (number per minute), duration (seconds), and percentage (percentage of N2 with spindles) were calculated at each electrode for each EEG recording (Supplementary Table 1). To compute peak spindle frequency, we applied a Hanning taper and 5 sec of zero padding to each spindle detection and then estimated the power spectrum using the Fast Fourier Transform. At each spindle detection, the power spectra from 8 to 16 Hz were then divided by the summed power over 1–50 Hz. The resulting power spectra were averaged in each channel and each age group, and the peak frequency was identified as the frequency in the spectrum with the highest power.

The inter-spindle interval (ISI) was calculated as the time interval (s) between the end of a spindle and the onset of the next spindle in the same channel. We computed the ISI for the central electrodes from each EEG recording. To estimate spindle refractory periods, we computed the average of the lowest 10% of ISIs for each channel in which at least 10 spindles were detected. We choose the lowest 10% of ISIs to focus on the shortest times between spindles while minimizing the impact of outliers. To compute interhemispheric spindle lag, we first identified all times when spindles were detected in at least two frontopolar or centroparietal channels on either hemisphere (FP1, F3, C3, P3 or FP2, F4, C4, and P4). The spindle lag was then calculated as the time between the first detected spindle among these channels in one hemisphere and the first detected spindle among the homologous channels in the opposite hemisphere.

Statistical analyses

To assess the performance of the spindle detector across ages, we report positive predictive value, sensitivity, and the F1 score (the harmonic mean of the positive predictive value and sensitivity) of the detector relative to manual expert classification. To provide the most conservative assessment of detector performance, for each measure, we performed a by-sample analysis, in which manual and automated detections are compared at each sample of data [13, 32].

To characterize spindle features over age, participants were categorized by year, rounding down to the most recent birthday. For visualization, topographic maps of spindle rate, duration, and percentage were interpolated using the Fieldtrip toolbox (http://www.ru.nl/neuroimaging/fieldtrip) [37].

To model spindle features across age, for each datapoint, we included the minimum number of adjacent datapoints required to pass a Shapiro Wilk test for normality [38] using the function swtest.m in MATLAB 2021b (The Math Works, Natick, MA). We then computed the mean and standard deviation for this distribution. We then removed the youngest datapoint and repeated the process through the oldest datapoint. We converted the mean and standard deviation into percentiles for each distribution. Finally, we performed a rolling average of mean and standard deviation estimates across adjacent months between 0 to 12 months and across adjacent years between 1 to 18 years.

To test for differences in the spindle features (spindle rate, duration, and percentage) across age groups, we used one-way ANOVA where p < 0.05 was significant. Post hoc two-tailed two-sample tests after Holm-Bonferonni correction (p < 0.025) were used to assess between-group differences.

To test for differences in spindle rate between males and females at each electrode and each age group, we used a two-tailed two-sampled t-test where p < 0.025 after Holm-Bonferonni correction was considered significant.

We performed a cluster-based permutation test to identify clusters of electrodes with spindle rates higher than the mean across electrodes for each age group [39]. To do so, for each participant in each age group, we (1) subtracted the mean spindle value across channels from the value at each channel. We then (2) evaluated the distribution of these deviations across participants to identify channels with higher spindle values compared to zero, using a one-tailed t-test. Channels below a critical alpha level (p < 0.05) were joined in a cluster if they were adjacent. Next, (3) a cluster-level statistic was computed as the sum of the absolute value of all t-values within a cluster. To determine which cluster-level statistics were unlikely to occur by chance, we used a cluster-based permutation test. Assuming spindle features are equally likely to occur above or below the mean, we randomly shuffled the sign of the deviations in each participant from (1). We then performed steps (2) and (3) and computed the largest cluster-level statistic from the resampled data. This process was repeated 5000 times. Clusters from unpermuted data were considered significant if their cluster-level statistic exceeded the top 5% of the permutation distribution. We did not restrict the minimum number of channels to form a cluster.

To test for differences in spindle peak frequency between frontal and central electrodes at each age subgroup, we compared peak frequency values at Fz and Cz using a one-sampled t-test where p < 0.05 after Holm-Bonferonni correction was considered significant.

Results

Database characteristics

We analyzed 567 N2 EEG recordings from 567 unique participants (291 F, 51.3%) ranging in age from 0 days to 18 years. Median longitudinal follow-up to evaluate for the development of an exclusionary subsequent neurological or psychiatric condition was 2.9 years (range: 0–20.1 years). Clinical characteristics are provided in Table 1. We note that our participants’ demographics reflect the population demographics of the state of Massachusetts [40].

Table 1.

Patient demographics

Demographics	Participants, N (%)
Total EEGs	567
Female sex	291	(51.3)
Ethnicity
Not Hispanic or Latino	293	(51.7)
Hispanic or Latino	99	(17.5)
Unknown	175	(30.9)
Race
White	365	(64.4)
Black or African American	45	(7.9)
Asian	18	(3.2)
American Indian or Alaska Native	1	(0.18)
Native Hawaiian or Other Pacific Islander	0	(0)
Other	73	(12.9)
More than one Race	14	(2.5)
Unknown	51	(9.0)

The final diagnosis for the events that led to the EEG evaluations is provided in Table 2. Among infants, The average duration of N2 sleep (TA or quiet sleep) per EEG recording was 15.67 min (IQR 9.38–21.42 min). In this dataset, 9 months old children had a longer sleep duration than 11 months old children (one-way ANOVA (F(12, 193) = 2.21, p = 0.01); p = .0004, two-tailed t-test). Among 1 to 18 year olds, the average duration of N2 sleep per EEG recording was 12.90 min (IQR 7.84–16.13 min). In this dataset, there were no differences in sleep duration between any of the 1 to 18-year age groups (one-way ANOVA (F(17, 378) = 0.95, p = 0.5)). Sleep duration is provided in Supplementary Table 1. Of the 567 participants, 34 electrodes in total were removed (0.32%) due to artifacts.

Table 2.

Indications for EEGs (N = 567)

Diagnosis	N	Diagnosis	N	Diagnosis	N
Syncope/pre-syncope	91	Respiratory event	11	Mild hypertonia during infancy	2
Migraine/headache	52	Anxiety/panic attack	7	Muscle fasciculation	2
Nonspecific movement	50	Dizziness	7	Aplastic anemia	1
Sleep phenomenon	45	Tremors	7	Bell’s palsy	1
Staring spell	37	Vomiting	6	Benign hypotonia of infancy	1
Stereotypy/tic	27	Concussion	5	Episode of unsteadiness	1
Gastrointestinal reflux	25	Altered mental status	4	Hiccups	1
Research control participants	24	Fall	4	Intussusception	1
Behavioral event	23	Fatigue	4	Lyme disease	1
Episode of unresponsiveness	23	Visual phenomenon	4	Polydipsia	1
Breath-holding spell	22	Startle reflex	3	Rigors	1
Unusual eye movement	19	Toxic exposure	3	Suspected abuse	1
Shuddering spell	17	Vertigo	3	Transient weakness	1
Brief resolved unexplained event	13	Pain	2	Trauma	1
Functional neurologic disorder	11	Hypoglycemia	2

The sleep phenomenon category includes all sleep related indications: sleep myoclonus (29), parasomnias (10), hypersomnia (1), periodic limb movement (1), somnolence (1) and other sleep phenomena (3).

Automated spindle detector validation across age groups

Balanced detector performance was achieved using a 95% probability threshold (mean F1: 0.47, 95% CI [0.10, 0.84], Figure 1). At this threshold, the detector’s sensitivity was 0.49, and positive predictive value was 0.50 against hand markings. We note that this performance is higher than previously reported automated spindle detectors, which have ranged from F1 0.2–0.43 [32].

Figure 1.

Spindle detector performance across ages. (A) Using leave-one out cross validation, the optimal F1 statistic against hand markings was achieved at the 95% probability threshold (0.47, red marker). Grey curves represent detector performance for individual participants across probability thresholds and the thick black curve indicates the mean performance across participants. (B) The sensitivity and positive predictive value of the detector at different probability thresholds (95% threshold indicated in red). (C) Example spindle detections in different age groups: 1 month, 3 months, 6 months, 1 year, 5 years, and 18 years. Each sample shows 5 sec of N2 sleep with automated spindle detections highlighted.

Spindle rates over development

Spindle distribution.

Sleep spindles have a distinct, age-specific spatial distribution. In infants, spindles are more prominent over frontopolar and central regions (p < 0.05, cluster-based statistic, Figure 2, A). In toddlers, spindles are most prominent over the central midline regions (p < 0.05, cluster-based statistic). Spindles then migrate anteriorly to become dominant in frontal regions by 5–6 years (p < 0.05, cluster-based statistic) and frontopolar regions by 9–10 years (p < 0.05, cluster-based statistic), though with increasing rates appearing diffusely. Over adolescence, spindles migrate posteriorly and dominate in central and parietal regions by late teens (17–18 years, p < 0.05, cluster-based statistic, Figure 3, A). Electrode clusters for each age bin are provided in Supplementary Table 2.

Figure 2.

Normative spindle distribution and rates over infancy. (A) Topographic maps of spindle distribution from 0 to 12 months. Top Row) Electrode clusters with spindle rates higher than average (p < 0.05). Bottom Row) The mean spindle rate (number per minute) is plotted per age group. The number of EEG recordings included in each age group is indicated below each topoplot. (B) Average spindle rate across all electrodes with age. The black thick line indicates the mean of normal distribution for each age bin. The black dashed lines from bottom to top indicate 10%, 25%, 75%, and 90% percentile of the population data. The outermost black lines indicate 5% and 95 % percentile of the population data. (C) Differences in spindle rates between age groups are indicated by a blue * for uncorrected p < 0.025 and by a red * for p < 0.025 after correction for multiple comparisons. (D) Results for each electrode, arranged according to 10–20 electrode placements. Colors indicate the brain location (Left-blue, Right-red, and Central-black).

Figure 3.

Childhood and adolescence normative spindle distribution and rates over development. (A) Topographic maps of spindle distribution over development. Top Row) Electrode clusters with spindle rates higher than average (p < 0.05). Bottom Row) The mean spindle rate (number per minute) is plotted per age group. The number of EEG recordings included in each age group is indicated below each topoplot. (B) Average spindle rate across all electrodes with age. The black thick line indicates the mean of normal distribution for each age bin. The black dashed lines from bottom to top indicate 10%, 25%, 75%, and 90% percentile of the population data. The outermost black lines indicate 5% and 95 % percentile of the population data. (C) Differences in spindle rates between age groups are indicated by a blue * for uncorrected p < 0.025 and by a red * for p < 0.025 after correction for multiple comparisons. (D) Results for each electrode, arranged according to 10–20 electrode placements. Colors indicate the brain location (Left-blue, Right-red, and Central-black).

Spindle rate.

In infancy, the mean spindle rate increases near-linearly from approximately birth to 4 months, where this increase is most prominent in the midline and parasagittal regions (Figure 2, C andD). After 4 months, there is a gradual deceleration in the mean spindle rate through 10 months of age, most prominent in the frontal channels (Figure 2, C and D). From 10 to 12 months of age, midline spindle rates increase slightly but remain stable in the rest of the channels (Figure 2, D). Spindle rates remain relatively stable throughout infancy in the temporal and occipital channels (Figure 2, D). On grouped analysis by month, spindle rates are significantly higher after 2 months compared to the newborn period (p < 0.025, two-sided t-tests, Figure 2, C, Supplementary Table 3).

In childhood, the spindle rate is approximately stable from age 1 to 3 years, a pattern observed across all channels (Figure 3, B and D). From age 3 to 14 years, there is a near-linear increase in spindle rate; this pattern is evident in all channels except CZ, where the spindle rate remains stable until approximately age 11 and then increases (Figure 3, B and D). After the age of 14 years, the spindle rate approximately plateaus across all channels (Figure 3, B and D). On grouped analysis by year, spindle rates are significantly different or trend to be different between most years of childhood and adolescence (p < 0.025, two-sided t-tests, Figure 3, C _andSupplementary Table 4).

We found no difference in spindle rate between males and females across infancy, childhood, or adolescence. However, we saw a trend of an increase in spindle rates in females compared to males from 0 to 9 months and from 10 to 14 years (Supplementary Figure 1).

Spindle duration over development

In early infancy, spindle duration follows a similar pattern as spindle rate, with increases between 1 to 4 months, then a slower decline through 12 months (Figure 4, A). This pattern is most prominent in the midline and parasagittal regions, with near-stable values observed across infancy in the temporal and occipital regions (Supplementary Figure 2, A). On grouped analysis, spindle durations are significantly different or trend to be different between most months across infancy (p < 0.025 most tests, two-sided t-tests, Figure 4, B, Supplementary Table 5).

Figure 4.

Normative spindle duration over development. (A–B) Early infancy development (0–12 months old). (C–D) Childhood and adolescence development (1–18 years old). (A,C) Average spindle duration across all electrodes with age. The black thick line indicates the mean of normal distribution for each age bin. The black dashed lines from bottom to top indicate 10%, 25%, 75%, and 90% percentile of the population data. The outermost black lines indicate 5% and 95 % percentile of the population data. (B,D) Differences in spindle durations between age groups are indicated by a blue * for uncorrected p < 0.025 and by a red * for p < 0.025 after correction for multiple comparisons.

In childhood and adolescence, spindle duration decreases slightly from 1 to 3 years followed by slow increases through age 18 years (Figure 4, C). The initial decrease between 1 to 3 years is most dramatic in centroparietal regions and the rate of change is greatest in the temporal regions from age 3 to 18 years (Supplementary Figure 2, B). On grouped analysis, spindle durations are significantly different or trend to be different between most years across childhood and adolescence (p < 0.025 on most tests, two-sided t-tests, Figure 4, D, Supplementary Table 6).

Spindle percentage over development

Spindle percentage reflects the combination of spindle rates and duration. As such, in infancy, the increases in spindle rate and duration from 0 to 4 months and subsequent slow decline through 12 months are prominently observed in spindle percentage (Figure 5, A). In the temporal and occipital regions, the mean spindle percentage is stable across infancy (Supplementary Figure 3, A). On grouped analysis, spindle percentages are significantly different or trend to be different between most months across infancy (p < 0.025 on most tests, two-sided t-tests, Figure 5, B, Supplementary Table 7).

Figure 5.

Normative spindle percentage over development. (A–B) Early infancy development (0–12 months old). (C–D) Childhood and adolescence development (1–18 years old). (A,C) Average spindle percentage across all electrodes with age. The black thick line indicates the mean of normal distribution for each age bin. The black dashed lines from bottom to top indicate 10%, 25%, 75%, and 90% percentile of the population data. The outermost black lines indicate 5% and 95 % percentile of the population data. B,D) Differences in spindle percentages between age groups are indicated by a blue * for uncorrected p < 0.025 and by a red * for p < 0.025 after correction for multiple comparisons.

In childhood and adolescence, spindle percentage decreases slightly from 1 to 3 years followed by slow increases through age 18 years (Figure 5, C). The initial decrease between 1 to 3 years is most dramatic in centroparietal regions and the rate of change is greatest in the temporal regions from age 3 to 18 years (Supplementary Figure 3, B). On grouped analysis, spindle percentages are significantly different or trend to be different between most years across childhood and adolescence (p < 0.025 on most tests, two-sided t-tests, Figure 5, D, Supplementary Table 8).

Spindle frequencies over development

Spindle frequency changes with age.

Spindle frequency follows a U-shaped developmental trajectory (Figure 6). The peak frequency of spindles is highest in infancy and mid-adolescence, and similar between these age groups (0–1 year, 13.1 Hz; 14–18 years, 13.1 Hz). Between ages 1 and 2 years, there is a sharp drop in spindle frequency (2.0 years, mean peak frequency 11.2 Hz). From ages 2–14 years, there is a consistent increase in spindle frequency with age.

Figure 6.

Normative spindle frequency over development. (A) An example recording from one channel showing spindle detections (blue) in left column. The EEG recording was broadband filtered 1–50Hz. The normalized spindle spectra and peak frequencies (red markers) were identified. Note, some spindles demonstrate two peaks which will be retained in the distributions but not peak frequency analysis. Example normalized spectra (thin gray curves) from electrode Fz from a 10 year-old participant in right column. From the average spectrum (thick black curve) the peak spindle frequency (red marker) was identified. (B) The distribution of normalized spindle frequencies across channels is plotted as a function of age. (C) Each line represents the normalized peak spindle power across channels by age. Peak frequencies are high in infancy and late adolescence.

Slow and fast spindles emerge early in development.

To identify when canonical frontal “slow” and central “fast” spindles emerge, we compared peak spindle frequency between central (Cz) and frontal (Fz) regions over age. Beginning at 6 months, the peak frequency of frontal spindles declines and by 18.0 months frontal spindles had a lower peak frequency (11.7 Hz) compared to central regions (12.7 Hz; p = 0.001, paired t-test (Holm-Bonferroni adjusted p = 0.006), Figure 7). After 2 years, peak frequencies in the frontal and central regions both increase with age through age 18 years, but remain distinct from each other, with lower frequencies always observed in frontal compared to central regions (Holm-Bonferroni adjusted p < 0.006, paired t-test, for all comparisons after 18 months, Figure 7, D, Supplementary Table 9).

Figure 7.

Emergence of distinct frontal slow and central fast spindle over 0 to 18 years. (A) Example of averaged power spectra for spindle detections in Fz and Cz in a 10 year old. (B) Peak spindle frequency in Fz and Cz over 0 to 24 months (left) and 0 to 18 years (right). (C) The distribution of normalized spindle frequencies across channels is plotted as a function of age. (D) Spindle peak frequencies are lower in Fz compared to Cz channels after 18 months (* indicate p < 0.05 after correction for multiple comparisons).

Spindle refractory periods decrease over development

To estimate spindle refractory periods over development, we computed the mean of the smallest 10% of ISIs in the central regions in each hemisphere (channels C3 and C4) (Figure 8, A). In infancy, spindle refractory periods are consistent across ages (~5s) (Figure 8, B and C; Supplementary Table 10).

Figure 8.

Minimum ISI over development. (A) The ISI is calculated as the time from the end of one spindle to the start of the next. The lowest 10% of ISIs from channels C3 and C4 were considered. An example of ISI histogram from a 10 year-old participant is shown in bottom row. (B–C) Early infancy development (0–12 months old). (D–E) Childhood and adolescence development (1–18 years old). (B,D) Average ISIs across all electrodes with age. The black thick line indicates the mean of normal distribution for each age bin. The black dashed lines from bottom to top indicate 10%, 25%, 75%, and 90% percentile of the population data. The outermost black lines indicate 5% and 95% percentile of the population data. (C,E) Differences in spindle minimum ISIs between age groups are indicated by a blue * for uncorrected p < 0.025 and by a red * for p< 0.025 after correction for multiple comparisons.

In childhood and adolescence, spindle refractory periods decrease slightly from age 1 to 18 years (Figure 8, D). On grouped analysis, spindle refractory periods are significantly different or trend to be different between most years across childhood and adolescence (p < 0.025 on most tests, two-sided t-tests, Figure 8, E, Supplementary Table 11).

Interhemispheric spindle lag over development

To characterize the coordination of spindles between hemispheres, we computed the interhemispheric lag between spindles detected in the frontal, central, and parietal regions (channels FP1, F3, C3, P3 and FP2, F4, C4, and P4) (Figure 9, A). In infancy, interhemispheric spindle lag decreases from 0 to 4 months and then remains constant by 12 months. (Figure 9, B). On grouped analysis, interhemispheric spindle lag tends to be different between the first 2 months and most other bins (Figure 9, C, Supplementary Table 12).

Figure 9.

Interhemispheric spindle lag over development. (A) Spindles detected in at least two frontocentral channels in the same hemisphere were analyzed. The spindle lag was calculated from the start of the first detected spindle in one hemisphere to the start of the first detected spindle in the contralateral hemisphere. (B–C) Early infancy development (0–12 months old). (D–E) Childhood and adolescence development (1–18 years old). (B,D) Average spindle lags across all electrodes with age. The black thick line indicates the mean of normal distribution for each age bin. The black dashed lines from bottom to top indicate 10%, 25%, 75%, and 90% percentile of the population data. The outermost black lines indicate 5% and 95% percentile of the population data. (C,E) Differences in spindle lags between age groups are indicated by a blue * for uncorrected p < 0.025 and by a red * for p < 0.025 after correction for multiple comparisons.

In childhood and adolescence, interhemispheric spindle lag decreases slightly from 1 to 18 years (Figure 9, D). On grouped analysis, interhemispheric spindle lag trend to be different between most years across childhood and adolescence (Figure 9, E and Supplementary Table 13).

Discussion

Sleep spindles are fundamental thalamocortical sleep rhythms that are mechanistically linked to essential sleep-dependent cognitive processes across the lifespan. Although this noninvasive physiological biomarker for cognitive function offers enormous potential for use in pediatric populations, the normative values for spindle spindles across development have not been determined. By training and applying a validated automated spindle detector to a large database of N2 sleep EEGs from healthy neonates, infants, and children we provide age- and regional-specific normative values for sleep spindles over development. This work reveals that, like anatomical brain maturation, sleep spindle features—including rate, duration, topography, refractory period, and interhemispheric synchrony, have predictable, age-specific, regional, nonlinear values from birth to late adolescence. These normative measures can be used to identify pathology in disease populations and accelerate the discovery and diagnosis in neurodevelopmental disorders.

The ability to utilize biomarkers to detect disease in pediatric populations is hampered by the dynamic changes in baseline measures that occur over healthy development. Using this large database, we provide robust normative parameters of spindle features for each age across development. Historical papers comparing spindle rate to age-matched cohorts have identified increased spindles in newborns with phenylketonuria [41]; decreased sleep spindles in infants with hypothyroidism [42], autism [15], and epilepsy [13, 14]; and delayed spindle development in infants with trisomy 21 [43]. Thus, we expect the developmental data provided here can be used to identify deviations from this distribution in at-risk populations with confidence to aid in the diagnosis, elucidate thalamocortical pathophysiology, and measure the efficacy of treatment responses after interventions, as in [44, 45].

We note that previous literature reports that sleep spindles are not reliably detected until about 2 months of age [30, 46–48]; however, we identified non-zero spindle values in full-term infants less than 1 month of age. This is likely because prior studies only included a small number of infants [46] or did not evaluate infants younger than 2 months [48]. The spindles we identified in neonates were generally rare (~1/min), fast (~13 Hz), short (0.5 sec), and asynchronous (> 1.5-sec delay between hemispheres), consistent with the “rudimentary” [49] or “evanescent” [46] spindles described in 3–4 week infants in small studies using visual analysis.

Using the same detection and recording techniques across ages, we found that spindles observed in neonatal periods and over the first year of life share similar spatial distributions and frequency to the mature spindles observed in late adolescents. These findings are consistent with previous clinical observations in infants and a larger study of older children, reporting that spindles in infancy and late adolescence are dominated by faster ~13 Hz frequencies over central or Rolandic brain regions [2, 30, 50, 51]. In contrast, but also consistent with earlier work, we found that between infancy and adolescence, spindles increase in peak frequency from 11 to 13 Hz and dominate over frontal regions [2, 51, 52]. As spindle frequency is a heritable trait [2], spindle measures from infancy may provide a reliable early window into the integrity of the thalamocortical circuitry required for mature spindles. Topographically, consistent with prior work (e.g. see [30]), we found that spindles are most prominent over central regions at most ages. However, we found that spindles are most prominent over the frontopolar regions from age 2 to 7 years (Figure 3), coinciding with the emergence and dominance of slow spindles (Figure 7), which several others have also found to be predominantly frontal [2, 53].

Compared to spindle frequency and topography, spindle rate, and duration follow a more complex developmental trajectory, with an initial peak in the first year followed by a rapid decline and then a steady increase from age 2 years onward. After age 2 years, we found that spindle duration increases slowly, whereas prior reports found a small decrease in spindle duration from childhood to adulthood [2, 51]. This discrepancy may be due to the different detection methods applied where we note that our detector used consistent parameters across participants, was validated across ages, and allowed spindles longer than 2 sec. Regarding sex differences, one prior large study of 4–97-year-old participants found that females had 0.16 spindles per minute more than males [2]. In contrast, we and others did not find differences in spindle rate between pediatric males and females when compared between age bins, suggesting this small effect requires larger sample sizes to detect [52]. Spindle maturation is thought to reflect the maturation of underlying thalamocortical circuitry, including neocortical dendritic growth and thalamocortical synaptogenesis [54, 55]. Notably, the early nonlinear dynamics we observe in the first year are consistent with prior reports [2, 4, 47 ]and rise and fall in a striking parallel with the topography and nonlinear trajectory of cortical growth [56–59]. Given the rich normative sleep spindle metrics identified here, future work is required to relate these oscillations to the maturation underlying anatomical networks.

How and why spindles terminate is thought to be related to after-depolarization refractory periods in the thalamocortical neurons. The burst firing of inhibitory TRN cells results in hyperpolarization of thalamocortical neurons, which conversely activates a hyperpolarization-activated cation current (I_h) and disinhibits the T-type calcium current, leading to a burst of action potentials from the glutamatergic thalamocortical cells that target the cortex and sends feedback to further engage the TRN. In ex vivo slices, spindles can persist indefinitely if the I_h current is blocked [60, 61]. Interestingly, we observed the longest spindle durations during early infancy, which may reflect an immature state of this current. Thus, normative spindle measures may provide empirical insights to guide future investigation of the developing thalamocortical networks that generate them.

Several lines of evidence suggest that sleep spindles support offline replay of learned experiences during sleep, resulting in improved consolidation of these memories [5]. Several studies have linked sleep spindle rate with cognitive functions in children [11–13, 62, 63] and adults in both health [64 ]and disease [13, 15, 65], with fast spindles more tied to procedural or motor learning [28] and slow spindles to declarative memory [9, 27]. We identified that fast spindles are present from birth and distinct slow spindles consistently emerge by 18 months. Although the presence of fast and slow spindles has been reported in children and adolescents [2, 53, 66], to the best of our knowledge, this is the first report of when slow spindles appear over neurodevelopment.

Our study draws from a database of primarily hospital-based recordings, which biases the participants included towards those that pursued medical evaluation for an unusual event. Because all children in Massachusetts have public access to medical care, the participant cohort reflects the state demographics. Our approach did allow detailed clinical chart review to confirm normal neurodevelopment at the time of the EEG and the absence of neuroactive medications, or neurological or psychiatric diagnoses expected to impact cortical function, which may improve capture of these concerns compared to self-reported reporting in population studies. We also note that the findings reported here are consistent with several smaller population-based studies. While our study included 19-channel EEG montages, which improves on the spatial sampling of EEGs used in typical polysomnograms, our EEG durations primarily included only short naps. We therefore could not evaluate the dynamics of spindles over the course of several N2 epochs, as would be expected over a full night of sleep. We note that while subtle changes in spindle rate may be present over the course of a full night [2], prior work has also found that naps can reliably estimate overnight spindle rate in adults [67] and reflect several measures of cognitive ability in children [13]. Furthermore, we evaluated spindle dynamics over age using a cross-sectional study design. Validation of the dynamics observed here within individuals will require a long-term longitudinal study. Finally, we only focused on sleep spindles here, while the relationship between spindles and other oscillations (e.g. slow oscillations) is another important area of study. How these cross-frequency relationships evolve with development is unknown, and the topic of ongoing work.

Sleep spindles are a unique neuronal rhythm that reflects and support cognitive function across the lifespan. Spindle features have a nonlinear trajectory from late childhood through late adulthood. This work fills in the critical developmental gap of spindle ontogeny and maturation from birth through adolescence. Our findings reveal that neonatal spindle frequency and topology provide early, transient views of the adolescent form of these genetically influenced traits, while early spindle rates follow a maturational trajectory that closely parallels cortical development. These data expand our current knowledge of normal physiological brain development and provide normative values to detect deviations in sleep spindles to aid discovery, biomarker development, and detection of cognitive development and neurodevelopmental disorders.

Funding

This work was supported by NINDS R01NS115868.

Disclosure Statement

Financial Disclosure: none. Nonfinancial Disclosure: none.

Preprint Repositories

An early draft of this manuscript was uploaded to the preprint BioRxiv doi: https://doi.org/10.1101/2022.03.31.486476