1 Infant low-frequency EEG cortical power, cortical tracking and 2 phase-amplitude coupling predicts language a year later

Cortical signals have been shown to track acoustic and linguistic properties of continual speech. This phenomenon has been measured across the lifespan, reflecting speech understanding as well as cognitive functions such as attention and prediction. Furthermore, atypical low-frequency cortical tracking of speech is found in children with phonological difficulties (developmental dyslexia). Accordingly, low-frequency cortical signals, especially in the delta and theta ranges, may play a critical role in language acquisition. A recent investigation Attaheri et al., 2022 (1) probed cortical tracking mechanisms in infants aged 4, 7 and 11 months as they listened to sung speech. Results from temporal response functions (TRF), phase-24 amplitude coupling (PAC) and dynamic theta-delta power (PSD) analyses indicated speech envelope tracking and stimulus related power (PSD) via the delta & theta neural signals. Furthermore, delta and theta driven PAC was found at all ages with gamma amplitudes 27 displaying a stronger PAC to low frequency phases than beta. The present study tests whether 28 those previous findings replicate in the second half of the same cohort (first half: N=61, (1); 29 second half: N=61). In addition to demonstrating good replication, we investigate whether 30 cortical tracking in the first year of life predicts later language acquisition for the full infant 31 cohort (122 infants recruited, 113, retained). Increased delta cortical tracking and theta-gamma 32 PAC were related to better language outcomes using both infant-led and parent-estimated 33 measures. By contrast, increased ~4Hz PSD power and a greater theta/delta power ratio related 34 to decreased parent-estimated language outcomes. The data are interpreted within a “Temporal 35 Sampling” framework for developmental language trajectories.

137 gestational weeks) and had no diagnosed developmental disorder. The study was reviewed by the 138 Psychology Research Ethics Committee of the University of Cambridge. Parents gave written 139 informed consent after a detailed explanation of the study and families were reminded that they 140 could withdraw from the study at any point during the repeated appointments (8 EEG recordings 141 at 2-, 4-, 5-, 6-, 7-, 8-, 9-and 11-months; 6 language follow-ups at 12-, 15-, 18-, 24-, 30-and 42-142 months).

144
A selection of 18 common English language nursery rhymes were chosen as the stimuli.
145 Audio-visual stimuli of a singing female (head only) were recorded using a Canon XA20 video 146 camera at 1080p, 50fps and with audio at 4800 Hz. A native female speaker of British English 147 used infant directed speech to melodically sing (for example "Mary Mary Quite Contrary") or 148 rhythmically chant (for nursery rhymes like "There was an old woman who lived in a shoe") the 149 nursery rhymes whilst listening to a 120 bpm metronome. Although the nursery rhymes had a 150 range of beat rates and indeed some utilized a 1 Hz rate, the metronome was used to keep the 151 singer on time. The beat was not present on the stimulus videos, but it ensured that a consistent 152 quasi-rhythmic production was maintained throughout the 18 nursery rhymes. To ensure natural 153 vocalisations the nursery rhyme videos were recorded while being sung or rhythmically chanted 154 to an alert infant.  268 This NWR task measured phonological production of both nonsense words (e.g. "punky") and 269 real words (e.g. "puppy") to name a series of toys. First, we present the power spectral density (using a periodogram; PSD), cortical tracking 283 (using multivariate temporal response functions; mTRFs), phase-amplitude coupling (normalised 284 modulation index, nMI; PAC) and theta/delta PSD ratio analyses using EEG recorded from the 285 second half of our cohort of infants. We compare these data to the analyses from the first half of 286 the cohort (1), and also merge the neural data to explore these neural factors in the combined full 287 sample of 113 infants. Due to missed recording sessions, technical issues with data files and not 288 enough trials after preprocessing, less than 113 data points are included in each analysis.

Language outcomes: statistics
289 Furthermore, outlier analysis was also conducted to remove extreme data points that would 290 compromise each of the LMEMs. The number of data points included are given separately for 291 each analysis.  (Table 1). To allow for individual variation across participants, a 319 maximum peak value per participant was taken from the participants' 60 channel averaged data, 320 within 0. 25  For the 4.05 HZ peak, the LMEM results replicated less well across the first half of the 354 sample, the second half of the sample and the full cohort. Significant effects of both recording 355 type and age were only observed when analysing the full cohort. The significant effect of age 356 was driven by the 7-month (M = 63.1 SE ± 6.0) PSD power being larger than the 11-month (M = 357 52.5 SE ± 3.2), however this result is in part due to the 0. 25 Hz window not fully encapsulating 358 the 11-month peak for all participants (Fig. 1). A significant effect of recording type was 359 observed in the second half of the sample, with a trend observed in the first half of the sample 360 (  396 alpha) and age (4-, 7-or 11-months) were investigated along with interactions between data type 397 by frequency band, data type by age and age by frequency band. A random intercept (AR1) per 398 participant was included to account for individual differences across the 3 recording sessions (4-, 399 7-and 11-months).

400
The main pattern of results replicated between the first and second halves of the sample, 401 with significant fixed effects of data type and frequency band, as well as significant interactions 402 between data type by frequency band and age by frequency band (Table 4). There were no 403 significant effects of age nor data type by age interactions when analysing the second half of the 404 sample in isolation, although these effects were present in the first half of the sample (

447
The PAC analyses (Fig. 3) Table 5 and estimates of fixed effects are provided in A significant difference between the theta/delta PSD power ratio across the three ages 491 was also found (Fig. 4), indicated by a one-way ANOVA (F(2,273) = 9.54, p = 9.389x10 -5 ). The 492 full sample submitted to the ANOVA was as follows, 4-months (N=96), 7-months (N=88) and 493 11-months (N=92). The amount of theta PSD power increased relative to the amount of delta 494 PSD power as our sample aged from 4-to 11-months. For comparison, the child modelling 495 using EEG recorded from children with and without dyslexia aged around 9 years showed that a 496 higher theta/delta ratio was associated with poorer language outcomes (27).

497
498 Figure 4, Theta / delta PSD ratio. The PSD peak per individual at 4.35 Hz and 1.95 Hz, taken 499 from the spectral decomposition of the EEG signal, was used to create a theta / delta ratio 500 calculation (4.35/1.95). Data provided at 4-months (red), 7-months (green) and 11-months (blue) 501 with standard error bars in black.  The 11-months EEG data were selected for comparison to both the parent-estimated and 523 infant-led language measures, due to the observed developmental maturation of the ~4 Hz peak.
524 The analyses for the 4-month and 7-month EEG are provided as Tables S4 to S10. As described 525 above, the 4.35 Hz peak was the more robust and consistent of the two theta peaks (see Tables 1 526 & 2) and therefore was used for predicting the language outcomes (4.05 Hz peak results are also 527 reported in Tables S6-S8). The analyses for the ~1.92 Hz peak are also reported here. 77 infants 528 provided complete data for all the parent-estimated measures and PSD power results, whilst 70 529 infants provided data for all the infant-led language measures and PSD power.

536
The multivariate analyses for ~4.35 Hz PSD power in response to nursery rhymes 537 showed that increased power led to a significant decrease (F(2,75) = 3.18, p= 0.047) in the 538 global parent-estimated language outcomes (Table 7). It is important to note that whilst the 539 posthoc univariate linear model analyses indicated that comprehension and production did not 540 individually make significant contributions to this prediction (Table 7b), the direction of the beta 541 estimates suggest that higher power was associated with fewer known words later in 542 development. Taken together, these data suggest that there is a global effect of theta power on 543 parent-estimated (UK-CDI) language outcomes, but that this is not driven by either single parent-