Infant EEG theta modulation predicts childhood intelligence

Intellectual functioning is a critical determinant of economic and personal productivity. Identifying early neural predictors of cognitive function in infancy will allow us to map the neurodevelopmental pathways that underpin individual differences in intellect. Here, in three different cohorts we investigate the association between a putative neurophysiological indicator of information encoding (change in frontal theta during a novel video) in infancy and later general cognitive outcome. In a discovery cohort of 12-month-old typically developing infants, we recorded EEG during presentation of dynamic movies of people and objects. Frontal theta power (3–6 Hz) significantly increased during the course of viewing each video. Critically, increase in frontal theta during viewing of a video was associated with a differential response to repetition of that specific video, confirming relation to learning. Further, individual differences in the magnitude of change in frontal theta power were related to concurrent nonverbal cognitive level. We then sought to extend this association in two independent samples enriched for variation in cognitive outcome due to the inclusion of infants at familial risk for autism. We observed similar patterns of theta EEG change at 12 months, and found a predictive relation to verbal and nonverbal cognitive skills measured at 2, 3 and 7 years of age. For the subset of high-risk infants later diagnosed with autism, infant theta EEG explained over 80% of the variance in nonverbal skills at age 3 years. We suggest that EEG theta change in infancy is an excellent candidate predictive biomarker that could yield substantial insight into the mechanisms that underlie individual differences in childhood intelligence, particularly in high risk populations.


Further participant information:
Infants were reported to be developing typically by their parents. Additional exclusion criteria assessed through a parental screening form included: physical signs (e.g., dysmorphic features) of known genetic syndromes, serious medical or neurological conditions (e.g., encephalitis, concussion, seizure disorder, diabetes, congenital heart disease), neurocutaneous markings, or sensory impairments such as vision or hearing loss; serious motor impairment; birth weight < 2000 grams and/or gestational age < 37 weeks, history of intraventricular hemorrhage, exposure to neurotoxins (including alcohol, drugs), and maternal gestational diabetes. In addition, variables that may impact family functioning (e.g., serious parental substance abuse, bipolar disorder, or psychosis) were exclusion criteria.

Methods for Artifact Detection:
Continuous EEG data was segmented into consecutive one-second segments with no overlap. Artifact detection of EEG data was accomplished with both automatic artifact-detection software (NetStation 4.3) and through hand-editing (EJ).
Segments were rejected if the signal amplitude exceeded 250 µV, or if electro-ocular, movement or muscular artifact occurred. Channels with noisy data were interpolated by an algorithm incorporated within NetStation 4.3 (segments were excluded from analysis if more than 20% of channels were subject to interpolation, or if there were more than 5 interpolated channels within a scalp region). Data was then re-referenced to the average reference, and the resulting segmented data was imported into Matlab.
Within Matlab (using in-house algorithms), segments were detrended and subjected to an FFT, producing power spectra for electrodes grouped within a priori regions approximately equating to F3/F4 (frontal left: 24,28,29,25,21,20; frontal right: 3,4,124,123,119,118). For each segment, data from electrodes with a power value of more than 3 standard deviations from the mean of the remaining electrodes in a topographical group in the frequency bands of interest were dropped. Power values were then averaged across artifact-free segments and electrodes within topographical groups and within each half of each video repetition; natural logs were calculated to reduce skew. Finally, logged power values were averaged across the theta (3 to 6Hz) frequency range. Participants were only included in analyses if they provided at least 10 artifact-free trials per video half/condition (e.g. first half of the social video set); this represents a minimum cut-off of approximately a third of the duration of each video (20 seconds out of the total video duration of 60 seconds). We have previously published grouplevel data on averaged theta power (but not change in theta power or relation to IQ) from this cohort 1 .

Methods -Final sample:
For the first set of analyses (theta change by condition, region and half) we included all infants who had sufficient EEG data for each half of both the social and non-social video. Of the original sample of 106 typically developing infants, 36 had sufficient data for this analysis. For examination of relation to cognitive skills, we collapsed across condition and region, yielding greater availability of EEG data because infants with data from one condition could be included. 67 infants had sufficient EEG data for this analysis (an inclusion rate of approximately 63%), of whom 33 infants did not have data available on the Mullen Scales of Early Learning. This is because we originally only administered the Mullen to approximately 50% of this large sample, due to time constraints.   but not significantly to verbal cognitive level (r(31) = 0.31, p = 0.083).

Results-Sex:
We also examined whether any of the main analyses differed between male and female infants. For the analysis of theta power by condition, region and half, the main effect of half remained (F(1,34) = 6.58, p = 0.015, η 2 = 0.16). The only significant effect of gender was an interaction between condition, half, region and gender (F(1,34) = 5.33, p = 0.027, η 2 = 0.14). Follow-up analyses suggested that increase in theta power between the first and second half of the video was bigger for boys over the right vs left hemisphere for non-social videos (F(1,14) = 6.14, p = 0.027, η 2 = 0.31). However, since this was not predicted we did not interpret this pattern further. The relation between nonverbal scores and frontal theta change during the video remained significant (F(2,33) = 5.29, p = 0.028, η 2 = 0.15) and did not vary by gender (interaction F(2,33) = 0.51, p = 0.48, η 2 = 0.016).

Cohort 2 2.1 Further participant information:
Infants were from a larger sample of 43 high-risk infants followed longitudinally as part of the Early Connections Study at the University of Washington 2 . Low-risk controls in this study did not receive cognitive assessments at 24 months. Inclusion criteria for high-risk infant siblings included age (< 6 months), presence of autism in a full biological older sibling, and anticipated residence in the region (within 1.5 hours driving distance from the University) for the next 2 years. To confirm the diagnosis of ASD in an older sibling, the Autism Diagnostic Interview-Revised (ADI-R) was administered by phone and medical records were collected to confirm the diagnosis was based on DSM-IV criteria from a psychologist or physician. Additional exclusion criteria included: physical signs (e.g., dysmorphic features) of known genetic syndromes, serious medical or neurological conditions (e.g., encephalitis, concussion, seizure disorder, diabetes, congenital heart disease), neurocutaneous markings, or sensory impairments such as vision or hearing loss; serious motor impairment; birth weight < 2000 grams and/or gestational age < 37 weeks, history of intraventricular hemorrhage, exposure to neurotoxins (including alcohol, drugs), and maternal gestational diabetes. In addition, variables that may impact family functioning (e.g., serious parental substance abuse, bipolar disorder, or psychosis) were exclusion criteria. Based on this information, infants were classified as having "Autistic Disorder", "Pervasive Developmental Disorder-Not Otherwise Specified" (collapsed into ASD) or "no diagnosis".

Methods -Autism
Clinicians judged their confidence in the classification as "Very confident", "Somewhat confident", or "Not confident".

Methods -Artifact detection:
Collection and processing of infant EEG data were identical to those described in Experiment 1. Trained clinicians administered the Mullen Scales of Early Learning at 24 months. We have previously published group-level data on averaged theta power (but not change in theta power or relation to IQ) from this cohort 3 .

Methods -Final sample:
Of the original sample of 43 high-risk infants, 28 did not provide sufficient EEG data at 12 months and one did not receive a 24-month cognitive assessment. Thus, the final included sample was 14 infants with an older sibling with a clinical diagnosis of ASD (confirmed with the ADI-R) who provided both sufficient data in an EEG assessment at 12 months and a cognitive assessment at 24 months. For visualization (but not analysis, given the small sample size), infants within the HR group were further divided based on their diagnostic outcome at 24 months. Of the included group of 14 HR infants, infants in the HR-ASD (n=5) group all met DSM IV criteria for ASD at 24 months.
Where clinicians had judged that they were "Not confident" in this judgment, infants were additionally required to meet cut-off on the ADOS for ASD (n=1). Infants in the HR-ASD-No ASD group were judged to have "no diagnosis" on DSM-IV criteria (n=9).

Methods -Analysis strategy:
We focused on percent change in frontal theta between the first and second half of the first presented video set (collapsed across condition), and its relation to nonverbal cognitive skills at 24 months (the oldest age to which children were followed in this sample). We required infants to have at least 12 segments in the first and second half of the social/non-social video. We computed simple correlations between change in theta power and later verbal/nonverbal skills.

High Risk (n=14)
N trials 1 st half 27.4 (9.2) N trials 2 nd half 29.2 (9.9) 12 month Nonverbal t-score 58.9 (6.3) 24 month verbal t-score 49.9 (12.9) 24 month nonverbal t-score 52.4 (7.7)  confirm absence of ASD in these older siblings, with no child scoring above instrument cutoff (> 15; n = 1 missing data). developmental history (no ADI-R was administered) no formal clinical diagnoses were considered but none had a community clinical ASD diagnosis. Infants' behaviour (looking, gross body/head/arm movements, crying) and distracting events (e.g. parent's speech, sucking of pacifier, etc.) were coded off-line. The infant was rated as watching the video when she/he looked at the screen, did not move and was not distressed.

Methods -EEG
All participants included in the analysis looked at the screen for at least 85% of time and did so without motion or negative affect for more than 65% of the time.
EEG was recorded using a 128-electrode Hydrocel Geodesic Sensor Net (EGI, Eugene, OR) with respect to the vertex and sampled at 500 Hz. Twelve ridge electrodes most often contaminated by artifacts were excluded from analysis resulting in 116-electrode layout. Data preprocessing and analysis was performed using FieldTrip (http://fieldtrip.fcdonders.nl/) as well as in-house software. The behavioural coding results were synchronised with EEG and the periods when the baby was not looking at the screen, moved, or cried, as well as the periods of interference were excluded from analysis. EEG was visually inspected for artifacts. The bad channels were interpolated and data were segmented to 1-s segments with 50% overlap and rereferenced to the grand average of 116 channels. Full details on data pre-processing are also reported elsewhere 4,5 . Fast Fourier transform (FFT) was used to calculate theta (3)(4)(5) spectral power in the first and second halves of the videos. Frontal theta power values were obtained by averaging the power in a selection of frontal channels (3,4,5,9,10,11,12,14,15,16,18,19,21,22,23). Note that electrode numbering for this cap is slightly different to the Geodesic Sensor Nets (2.1) used in Cohorts 1 and 2, but electrode regions were substantially overlapping. The power values were then log-transformed to reduce skew. EEG was divided into the first and second halves of the first presentation of a particular stimulus condition (20 second segments); percent change was computed as above. Participants were included if they provided at least 10 segments of artifact-free EEG for each of the time intervals. We have previously published group-level data on averaged theta power (but not change in theta power or relation to IQ) from this cohort 4 .

Methods -Final sample:
Of the original sample of 54 high-risk infants and 50 low-risk infants, 35 high risk and 34 low risk infants did not provide sufficient EEG data at 12 months.
The higher attrition rate in this sample is likely because the videos were shorter than those used for Cohorts 1 and 2. One infant (with ASD outcome) was considered an outlier; see below for analyses including this child. Thus, the final included sample was 16 low risk infants and 18 high risk infants, of whom 7 were considered to have ASD at 36 months.

Methods -Analysis strategy:
Following the results of Cohort 1 and 2 and to avoid multiple comparisons, we focused on percent change in frontal theta between the first and second half of the video (collapsed across condition), and its relation to nonverbal cognitive skills at 36 months (the oldest age to which children were followed in this sample). This was analysed in an ANCOVA in which nonverbal cognitive skills were entered as the dependent variable; with outcome group (HR-ASD, HR-no ASD, LR), theta change, and the interaction between theta change and group as predictors. We then examined whether relations remained if nonverbal cognitive skill at 12 months were covaried, and whether relations were also observed for verbal skills, following the findings of Experiment 2.    Figure S1 illustrates the bivariate outlier for theta change and nonverbal skills (circled). This child was considered a "borderline" case in the 3-year diagnostic assessment and did not receive a diagnosis of autism in a clinical setting; the child did not return at age 7 years and hence stability of his profile cannot be ascertained. , and full-scale IQ (FSIQ) were used in analyses. One HR child was unable to complete the assessment due to intellectual disability.

Methods -Final sample:
Of the original sample of 16 low risk infants and 18 high risk infants with EEG data at 12 months, 11 low risk and 14 high risk infants had data available on the WASI at age 7 years. Of note, we used risk group rather than outcome group in this analysis because of the small sample size and because some children changed diagnostic category between age 3 and age 7 6 . Specifically, of the 14 children in the high-risk group four were considered to have ASD at both 3 and 7, one at 3 not 7, and one at 7 not 3. In Figure 3E children with an ASD diagnosis at 7 (concurrent with the IQ data) are shown in red; children who changed diagnostic status are indicated with arrows.