Developmental trajectories of EEG aperiodic and periodic power: Implications for understanding the timing of thalamocortical development during infancy

The development of neural circuits over the first years of life has long-lasting effects on brain function, yet our understanding of early circuit development in humans remains limited. Here, aperiodic and periodic EEG power features were examined from longitudinal EEGs collected from 592 healthy 2–44 month-old infants, revealing age-dependent nonlinear changes suggestive of distinct milestones in early brain maturation. Consistent with the transient developmental progression of thalamocortical circuitry, we observe the presence and then absence of periodic alpha and high beta peaks across the three-year period, as well as the emergence of a low beta peak (12–20Hz) after six months of age. We present preliminary evidence that the emergence of the low beta peak is associated with thalamocortical connectivity sufficient for anesthesia-induced alpha coherence. Together, these findings suggest that early age-dependent changes in alpha and beta periodic peaks may reflect the state of thalamocortical network development.

The infant brain undergoes dramatic structural and physiological change in the first year after birth. Rapid 4 increases in brain volume coincide with expansive synaptogenesis 1-3 , as well as interneuron migration, 5 maturation and network integration 4 . In particular, during this early period thalamocortical connections 6 are established through an intricately choreographed sequence that plays a critical role in the 7 development of sensory cortical networks 5 . However, the detailed timing of interneuron and 8 thalamocortical maturation in human development is largely unknown. In rodent models, the 9 development of thalamocortical circuitry is notable for transient inhibitory connections that drive 10 subsequent circuit formation and coincide with critical periods of plasticity present during the first 2-3 11 postnatal weeks 6 . In humans, longitudinal resting-state fMRI data suggest that while thalamus-12 sensorimotor connectivity networks are present at birth, other networks (e.g. thalamus-medial-visual,

20
The EEG power spectrum is comprised of two physiologically distinct components reflecting underlying 21 neuronal activity: aperiodic and periodic power. The aperiodic component defines the slope of the power 22 spectrum, following a 1/f power law distribution ( Fig 1A) and reflects non-oscillatory neuronal spiking 23 activity. In addition, the aperiodic slope has been linked to the excitatory-inhibitory (E/I) balance of the 24 underlying neuronal network, where a flattened, reduced slope is associated with increased excitation 25 over inhibition, and a steeply more accelerated slope with increased inhibition over excitation 10 .

26
Longitudinal studies of child-to adulthood have observed decreases in aperiodic slope with age, 27 suggestive of increases in E/I balance with age [11][12][13][14] . Changes in the aperiodic component in early infancy 28 are less well described, and we hypothesize they may be substantially different from those in childhood, 29 as the first year after birth includes rapid increases in neuronal activity, synaptogenesis, and inhibitory 30 neuron integration.

32
The periodic component of the power spectrum is defined as the portion of the absolute power spectrum 33 rising above the aperiodic slope ( Fig 1B). Periodic power reflects oscillatory activity occurring in narrow 34 frequency bands that are highly correlated with various cognitive processes and behavioral states 15,16 , 35 and provide the foundation for both local and long-range communication within the brain. The majority of neural oscillations observed in the power spectrum are the direct result of inhibitory and thalamocortical 1 network responses to sensory input. Thus, as a measure, the EEG power spectrum is well positioned to 2 shed light on the developmental timing of inhibitory and thalamocortical network maturation.  however, the precise mechanisms remain unknown.

13
Unlike theta/alpha power, little is known about the early developmental changes in periodic beta power.
14 In adults, beta oscillations are strongly associated with sensorimotor processing in addition to higher-15 order cognitive tasks such as working memory 25 . Similar to alpha oscillations, the generation of beta 16 oscillations relies on GABAergic interneuron networks and thalamocortical connectivity. In adults, low-17 dose GABA-modulating anesthetics induce a sedative state with 13-25Hz beta oscillations, whereas 18 higher doses used to maintain unconsciousness progressively slow these beta oscillations into coherent, 19 frontal specific, alpha oscillations 26-28 . However, GABA-dependent anesthesia-induced frontal alpha 20 coherence does not emerge in infants until after 10 months of age and is not consistently present until 21 15-20 months of age 29-31 . Anesthesia-induced alpha coherence is hypothesized to involve GABA-22 dependent thalamocortical loops leading to hypersynchronization between thalamic and prefrontal 23 cortices 27,32,33 . Therefore, potential covariation of developing beta oscillations and anesthesia-induced 24 counterparts may lend insight into the role and time course of developing inhibition in human 25 thalamocortical circuit development.

27
Using longitudinal EEG data collected from 592 healthy infants (yielding a total of 1335 EEGs) from 2 to 28 44 months after birth, we characterize early developmental trajectories of EEG aperiodic and periodic 29 power from 2-50Hz and to identify potential ages relevant to sequential steps in inhibitory network and 30 thalamocortical circuit development. Consistent with the transient and stepwise developmental 31 progression of thalamocortical circuitry, we observe transient periodic peaks in alpha power at 2-3 32 months and high beta power at 4-18 months. A low beta peak ) also begins to emerge in infants 33 starting as early as 6 months of age. We hypothesized that emergence of this low beta peak reflects 34 maturation of early connections between the thalamus and cortex. To test this hypothesis, we leveraged 35 a smaller dataset consisting of a cohort of infants with EEG recordings before and during clinical anesthesia. Consistent with our hypothesis, we find infants with an identifiable low beta peak have higher 1 anesthesia-induced alpha coherence than those that do not, suggesting that the emergence of this peak 2 is associated with thalamocortical loop maturation.

6
Resting-state EEG were collected longitudinally from 592 healthy infants, aged 2-44 months, across 4 7 studies occurring in the same laboratory ( Fig 1C, Table 1). Whole brain power spectra for each individual 8 were calculated by averaging across electrodes (Supplemental Figure 1. Individual spectra shown in 9 Supplemental Figure 2). Spectra were then averaged across individuals within 8 age bins (Fig 1D).

10
Notable nonlinear changes in aperiodic and periodic power spectra were observed between age bins, 11 including transient peaks in the periodic spectrum across both alpha and beta frequency ranges (Fig 1E-

12
G). To further characterize these developmental changes in the spectra, we used generalized additive 13 mixed models (GAMMs) to model non-linear trajectories of power parameters. For each model an age-14 by-sex interaction was tested for significance. If not significant, the interaction term was removed and the 15 model was refit using sex as an additive covariate. All models also included study as a covariate factor. First, we assessed age-dependent changes in the aperiodic component and observed the largest 20 developmental increases in aperiodic power between 2 and 8 months after birth ( Fig 1E). Aperiodic 21 offset, but not slope, significantly increased with age (FDR-adjusted q value <0.01), and age-by-sex 22 interactions were present for both aperiodic offset (F = 4.59, q = 0.01) and slope (F = 3.18, q = 0.02).

23
Modeled developmental trajectories of the aperiodic offset showed a sharp linear increase over the first 24 year after birth for both males and females ( Fig 1H). Modeled developmental trajectories of the aperiodic 25 slope showed a gradual increase over the first year. These findings contrast with consistent reports of 26 decreasing offset and slope across child and adulthood [11][12][13][14] , and likely reflect the known increases in   At the youngest age bin (2-4 months) two peaks with similar amplitude are observed across the 3 theta/alpha (4-12Hz) range in the majority of infants (69% ; Fig 2A,C). A lower frequency peak is 4 observed in the theta (4-6Hz) range at 5.5 ± 0.3Hz, and higher frequency peak is observed in the alpha 5 (6-12Hz) range at 9.5 ± 0.45Hz. However, by 6-months only 15% of infants have two peaks in this range, 6 and for most infants it is the higher 9.5Hz peak that is no longer observed. At 6-months fewer than 40% 7 of infants exhibit a dominant peak in the "alpha" (6.5-12Hz) range ( Fig 1D) and the average peak 8 frequency in the theta/alpha range is 6.3±1Hz. This disappearance of the higher peak after 4-months of 9 age may reflect a transient step in thalamocortical circuitry development. Previous research has 10 observed a gradual shift in peak frequency from 5 to 8Hz from infancy to early childhood, however these 11 studies started no earlier than 5 months of age 20-23 . In order to assess whether an increase in peak 12 frequency beginning at 5 months is present in our data set we modeled developmental trajectories of     Several age-dependent transient changes are observed in the low beta  and high beta (20-3 35Hz) range. First, the shape of the periodic power spectra in the low beta range is notable for a 4 prominent trough prior to 1 year of age ( Fig 3A), with only 10% of infants (24/222) exhibiting a low beta 5 peak between 6-8 months of age (Fig 3A-C). After 8 months, a low beta peak begins to emerge in some 6 of infants, with 48% (52/107) showing a peak at 18-20 months, and 70% (199/285) by 36 months (Fig   7   2C). As a low beta peak was not identified in many children across the age range, peak amplitude and 8 frequency was not modeled. In contrast, virtually all (99.5%) of the infants had an identifiable high beta 9 peak prior to 12 months of age ( Fig 3A). However, notable nonlinear shifts in frequency and amplitude of 10 the high beta peak were observed (Fig 1F, 3D and E). During the first year after birth, the high beta peak 11 amplitude rapidly increases, peaking at 229 days (7.5 months), and then substantially decreases until 12 802 days(2.2 years). High beta peak frequency trajectories are also nonlinear, with peak frequency at its 13 highest at 473 days (male 29.0Hz, female 29.4Hz), followed by a steady decline in frequency. Modeled

17
The observed nonlinear changes across the beta range are striking. While many EEG infant studies 18 group beta oscillations into a singular frequency range, the data presented here supports that low and 19 high beta have distinct developmental origins. Specifically, between 6-24 months we observe the gradual 20 emergence of a low beta peak, and simultaneously the rise and fall of a prominent high beta peak, 21 ultimately resolving into a broader beta peak by 36 months.

23
Traditionally, beta oscillations measured in children and adults are associated with sensory and motor 24 processing, where reductions in beta power are observed during the preparation or execution of motor 25 tasks 25 . However, beta activity has also been shown to be modulated during a wide range of nonmotor 26 cognitive tasks 25,34 . The developmental emergence of low beta oscillations may represent sensorimotor 27 skills (e.g., crawling, walking) gained during this period, but may also represent the developmental 28 maturation of neurobiological circuitry. For example, GABAeric interneuron networks and thalamocortical 29 connectivity are highly associated with the generation of cortical beta oscillations, as well as anesthesia-  We hypothesized that developmental changes in infant beta power measured in a resting state may 4 represent concurrent maturation of GABAergic interneuron networks and thalamocortical connectivity. To 5 explore this possibility, we assessed EEG recordings of healthy infants before and during exposure to 6 GABA-modulating sevoflurane anesthesia 35 . All infants were undergoing elective procedures (eg. 7 circumcision) and infants were excluded for prematurity, neurologic injury, epilepsy, or planned 8 intracranial surgery. Here we hypothesized that the emergence of low beta oscillations (as measured by 9 the presence of a low beta peak) before anesthesia would be associated with GABA-dependent 10 anesthesia-induced frontal alpha coherence. EEG data from 36 infants (6-15 months old), collected 11 during the awake and anesthetized state were analyzed. Developmental changes in the aperiodic-12 adjusted power spectra in this smaller dataset were qualitatively similar to those described above ( Figure   13 4A), with a low beta peak beginning to emerge after 7 months and present in roughly half the infants 14 between 7-12 months of age (11/21). As hypothesized, alpha coherence was significantly increased in 15 those with a low beta peak compared to those without (ANCOVA, with sevoflurane level as covariate; p 16 <0.05; Figure 4B). Here we present the largest-to-date longitudinal analysis of EEG data collected between 2 -44 months 5 of age. Findings provide insight into the developmental timing of inhibitory network and thalamocortical 6 circuit maturation during human infancy. Several age-dependent findings in our study contrast to 7 previous longitudinal studies of child and adulthood. First, we observe increases in both aperiodic offset 8 and slope, especially during the first year, whereas decreases in both measures are observed starting as 9 early as 4 years of age and continue to decrease with adulthood [11][12][13][14] . Second, while expected shifts in 10 the dominant peak from the theta to alpha range were observed between 5 to 44 months, in the 2-4 11 months age bin, a 9.5Hz peak was also transiently observed. Third, striking changes within the beta (12-12 30Hz) range were observed, including the emergence of a low beta peak starting after 6-months of age, 13 and age-dependent shifts in high beta peak frequency and amplitude -first increasing and then 14 decreasing with age. Below we discuss how the above age-related changes may represent sequential 15 developmental maturation in the inhibitory system and thalamocortical network connections.

17
The aperiodic offset is hypothesized to represent broad band neuronal firing 36,37 , and thus early 18 increases in aperiodic offset are consistent with established increases in neuronal number, gray matter volume, and synaptic number during the first year. Stabilization of aperiodic offset after 1-year of age is 1 also consistent with MRI findings that gray matter volume doubles during the first postnatal year and then 2 slows to 20% in its second year 38-40 . Regionally, we also observe differences between posterior and 3 frontal aperiodic offset trajectories, which either plateau after 1 year (posterior), or have a slow continued 4 increase (frontal) beyond 1 year of age (Supplemental Fig 3). Consistent with this pattern, 5 synaptogenesis differs across cortical regions, with the posterior visual cortex exhibiting a burst in 6 synapse formation between 3-4 months of age, whereas the prefrontal cortex shows peak 7 synaptogenesis around 8 months of age and continued gains during the second year of life 1 .

9
Our observed age-dependent increases in aperiodic slope in infancy also contrast with multiple studies 10 covering child to adulthood, where decreases in slope have consistently been reported. Schaworonkov 11 et. al. 41 also reported decreased slope with age in infants from 1 to 7-months-old, however the 12 parameterization of the spectra in that study was limited to 1-10Hz due to excessive muscle noise in the 13 data, and it is unclear how the shifts in 4-12Hz periodic power described below may affect modeling of 14 the underlying aperiodic component in this range. We hypothesize that observed increases in aperiodic 15 slope reflect changes in inhibitory networks that are unique to early development. Indeed, aperiodic slope 16 from EEG recorded from sleeping newborns is observed to increase with age during the first 7 weeks 17 after birth 42 . Growing evidence suggests that aperiodic slope is modulated by the balance between 18 excitation and inhibition, with increased slope associated with a reduction in E/I ratio 10,12,43 . An age-  fetal monkey and human brains suggests that the SPN in primates and humans slowly begins to 34 disappear in the 3 rd trimester but may persist until 6 months, with an overlapping period in which the 35 thalamus makes connections with both the SPN and cortical layer IV neurons 6,49,50 . We hypothesize that the 9.5Hz peak observed at 2-4 months, but not at 6 months, reflects this transient period when mature 1 excitatory subplate neurons are still receiving and relaying thalamic input to cortices, resulting in higher 2 frequency alpha oscillations. Additionally, newly established connections between the thalamus and layer 3 IV produce lower frequency theta rhythms that will later become the dominant alpha rhythm.

23
Finally, our study identified early age-dependent changes in periodic beta power that we hypothesize are 24 associated with thalamocortical loop maturation. We observe the emergence of a low beta peak in 25 infants older than 6-months of age and find that the presence of a low beta peak is associated with 26 higher anesthesia-induced frontal alpha coherence. Biophysical models demonstrate that this frontal 27 anesthesia-induced alpha coherence requires inputs from both the thalamus and cortex 27 . Together 28 these findings suggest that low beta oscillations may directly reflect thalamocortical loop maturation. Beta 29 rhythms are thought to be both generated locally in the cortex through pyramidal-interneuron loops, as 30 well as through thalamus to cortical connections that also rely on inhibitory inputs 25 . The emergence of 31 the low beta peak in awake infants may reflect the combination of newly established network connections 32 between thalamic nuclei and cortical layers, as well as the maturation of interneurons within the 33 thalamocortical pathways.

34
It is also possible that developmental changes in beta power are related to infant movement. During EEG 1 acquisition, infants are held in their parent's lap and behavioral supports are in place to keep the infants 2 calm. However, it is not possible to control the infants' movement, and movements both small (hand 3 movements) and large (head turns, leg and arm movements) ubiquitously occur across recordings -likely 4 increasing over the first year as infants become more mobile. Our preprocessing artifact removal 5 pipelines (see Methods) includes several steps for removing high frequency noise from muscle artifact.

6
However this would not remove EEG signal in response to sensorimotor processing. Infant jaw and 7 upper limb movements have been shown to increase power between 9-20Hz along frontal and occipital 8 sites, while hand and lower limb movements do not have significant effects 56 . In our dataset, increases in 9 low beta power were most prominent in central (not frontal or occipital) ROIs (Supplemental Fig 3K), 10 suggesting that age-dependent changes in beta power more likely represent underlying brain maturation 11 than sensorimotor processing or movement artifact during data acquisition. This is further supported by 12 the consistency in age-dependent shifts of both low and high beta across individuals (see Supplemental

16
Transient high beta peaks were also observed across this early period of development. Specifically, we 17 observed early increases in high beta peak amplitude, which reached a maximum at 7.5 months, 18 followed by decreases in both high beta peak amplitude and frequency, such that by 36 months the low 19 beta peak is the dominant peak across the 12-30Hz range. The neurobiological mechanism of this high 20 beta peak is unclear. As discussed above, administration of GABA agonists induces beta activity. In      family history of ASD starting as early as 3-months of age. For this analysis only infants without family history of ASD were included. Study 4 also included a group of infants with elevated social 1 communication concerns at 12 months of age, and they were also excluded from this analysis. EEGs 2 and MSEL were performed at 3, 12, 18, 24, and 36 months for both studies, as well as 6 and 9 months 3 for Study 2. Infants were specifically assessed for ASD using the Autism Diagnostic Observation 4 Schedule (ADOS) in conjunction with clinical best estimate at 24-and 36-month visits.

6
Study 3, the Emotion Project (IRB-P00002876), was a cohort/longitudinal study. Infants were enrolled at 7 either 5, 7, or 12 months, and then followed through 7 years of age. In addition to the first time point,

8
EEG data was again collected at 3 years of age. There were no developmental assessments performed 9 for this study, however parent questionnaires regarding child development, diagnoses (e.g., ASD), and 10 therapies were collected.

12
All infants had a minimal gestational age of 36 weeks, no history of prenatal or postnatal medical or 13 neurological problems, and no known genetic disorders. Infants who were later diagnosed with ASD 14 (either by assessment during the study, or by community diagnosis disclosed by parents prior to age of    to a single vertex electrode (Cz) and impedances were kept below 100kΩ in accordance with the 10 impedance capabilities of the high-impedance amplifiers inside the electrically shielded room.

11
Electrooculographic electrodes were removed to improve the child's comfort.     Figure 1E) from the absolute power spectrum ( Figure 1D). To further characterize 9 peaks and troughs within the power spectra across development, the periodic spectrum was then 10 smoothed using a savgol filter (scipy.signal.savgoal_filter, window length = 101, polyorder = 8). Individual 11 periodic power spectrum plots before and after savgol filter are shown in Supplemental Figure 2. We 12 decided to use this method instead of using the FOOOF estimated peak_fit as the high beta peak 13 appeared to have a non-gaussian shape at some ages, and thus peak_fit estimates did not accurately 14 identify the high beta peak frequency. Using the smoothed periodic spectra, maxima were identified 15 within the following frequency ranges: 4-6.5 (theta), 4-12Hz (theta/alpha), 12-20Hz (low beta) and 20-16 35Hz (high beta). A low beta trough was also identified based on the minima between 10-20Hz.

21
As Study 2 collected data with 2 net types and 2 amplifiers, data from 6-, 9-, and 12-month age 22 bins were assessed for spectra differences in total (2-50hz) aperiodic and periodic power as well as 23 aperiodic slope and intercept measure from central electrodes between either 64 and 128 channel nets,

24
or NetAmps 200 or 300 amplifiers. Of the 24 analyses performed, 3 showed significant differences. Net-25 type differences were observed for 9-and 12-month central aperiodic slope (p=0.04 for both) and an 26 amplifier difference was observed for 12-month central offset (p=0.04). None of these were significant 27 after correcting for multiple comparisons.

29
Anesthesia cohort: EEG data was also collected from infants undergoing anesthesia as part of a 30 prospective observational study approved by the institutional review board at Montefiore Medical Center,

32
circumcision, hernia repair) were recruited. Infants were excluded for prematurity, known neurologic 33 injury, epilepsy, or planned intracranial surgery. Infants less than 6 months of age, or those documented 34 to be asleep or crying during baseline (pre-anesthesia) recordings were excluded from further analysis.

35
All subjects received general anesthesia with sevoflurane.

18
A separate GAMM was fit to predict each power spectral measure, for each region of interest (e.g., whole 19 brain theta power, frontal aperiodic offset). First, to determine whether to include an age-by-sex 20 interaction, two models were fit with the following forms: where oSex represents sex stored as an ordered factor, Study represents study stored as a factor, 28 s(age_days, k = 4, fx = T) is a smoothed age term, and s(age_days, New_ID, bs = 'fs') accounts for 29 repeated observations. These models were compared by ANOVA, and model 2 (including the interaction 30 term) was chosen if the difference was significant (p<0.05). Effects were corrected for multiple 31 comparisons across measure types (e.g., theta power, beta power, aperiodic offset) within region of 32 interest using FDR correction 67 .

34
To further understand the nonlinear trajectories of change, inflection points were calculated using the 35 argrelextrema function from scipy in python with order = 100. A standardized rate of change per day was calculated to visualize developmental changes within features. The modeled value of a feature at a given 1 age (in days) was subtracted from the modeled value from the subsequent day, and this was divided by 2 the standard deviation of the modeled values of that feature across the age range.

4
To assess the differences between regions of interest, GAMM models were fit with the following form:

5
(3) Power Measure ~ s(age_days) + oSex + Study+ ROI+ s(age_days, New_ID, bs = 'fs') 6 7 where the terms have the same meanings as above, and ROI is a factor representing the four regions 8 (frontal, central, temporal, posterior). Because prior literature and preliminary visual inspection of the 9 data indicated that the posterior ROI is most unique in the time course of development, the posterior ROI 10 was set as the reference factor. Thus, the effect and significance associated with each of the other ROIs 11 is a measure of the difference between that ROI and the posterior ROI.

13
Anesthesia Statistical Analysis: ANCOVA, with sevoflurane levels as a covariate, was used to determine 14 effects of presence of low beta peak on anesthesia induced alpha coherence.