Infant mismatch responses to speech-sound changes predict language development in preschoolers at risk for dyslexia

(cid:1) A larger positive mismatch response (P-MMR) at birth is associated with better development of prereading skills. (cid:1) A more mature pattern of neural auditory speech discrimination is associated with a better development of pre-reading skills. (cid:1) Familial dyslexia risk affects the development of pre-reading skills and their associations with MMRs.


Introduction
Adequate language skills are vital for communication, cognitive development, and academic achievement.Language development starts when the infant is exposed to spoken native-language speech and learns to extract meaningful fragments from it (Kuhl, 2004), and this forms a basis for oral and written language (Gervain, 2015).Disturbances in phonetic learning may lead to difficulties in language acquisition in developmental dyslexia, a reading-skill disability (Eden et al., 2015;Peterson and Pennington, 2015).In this context, the change-detection-related responses of the auditory event-related potentials (ERPs) derived from the electric signal of the brain (electroencephalography [EEG]; Kuhl, 2010) provide a useful method for studying auditory processing in infancy and early childhood (Kushnerenko et al., 2013), since they have been associated with future language abilities (Kujala and Leminen, 2017).Auditory ERPs, being sensitive to phonological deficits in developmental dyslexia (Ozernov-Palchik and Gaab, 2016;Peterson and Pennington, 2015) and related conditions (Kujala and Leminen, 2017), could potentially provide a measure for learning impairments (e.g., Choudhury and Benasich, 2011;Volkmer and Schulte-Körne, 2018).Most studies have focused on language skills at school age, but many elemental principles of human communication are already learned before the skills of reading comprehension.In this regard, there is a lack of longitudinal studies assessing whether infant ERPs are predictive of language development during the pre-reading stage.In this study, we use well-established methods to determine the associations of change-detection-related responses (here referred to as mismatch responses, MMRs) during the first years of life with the development of pre-reading skills.We also aim to assess the impact that familial dyslexia risk has on MMRs.This may provide early markers for neural deficits which could lead to future language and reading problems.

Neural auditory change detection in infancy and early childhood
MMRs are an infant counterpart of adult mismatch negativity (MMN), a task-related pre-attentive ERP that reflects the discrimination accuracy of an unpredicted (deviant) stimulus along any presented auditory stream of repeating stimuli (standard stimulus; Näätänen et al., 2007;Winkler, 2007).MMN is calculated as the difference between the averages of responses to standard and deviant stimuli, and in adults it consists on a fronto-central negativity on the scalp at around 150-250 ms after deviance onset (Näätänen et al., 2007).In infants, instead of or in addition to MMN, a positively displaced response, often called positive-MMR (P-MMR), is elicited (e.g., He et al., 2007He et al., , 2009;;Trainor, 2012).
Studies assessing the maturation of change-detection responses during the first years of life (Choudhury and Benasich, 2011;Fellman et al., 2004;Virtala et al., 2022) have reported that P-MMR is the most prevalent MMR in infancy, but it starts to diminish in both amplitude and latency during the first years of life.Furthermore, P-MMR was reported to arise between emerging early (immature MMN) and late (late discriminative negativity, LDN) negativities, both growing in amplitude with maturation.LDN appears to have a different neural origin than MMN (C ˇeponiene et al., 2005) and is thought to reflect additional higher-order processing of subtle changes in stimuli (Bishop et al., 2011;Kuuluvainen et al., 2016).Thus, LDN and MMN are thought to be distinct components from birth (Kushnerenko et al., 2001;Martynova et al., 2003).

MMRs and pre-reading skills
The development of reading skills is often described as a continuum where reading is preceded by pre-reading skills (i.e., the different abilities, knowledge, attitudes, and environments that support the development of actual reading later; Whitehurst and Lonigan, 1998).According to previous studies, the pre-reading skills that most consistently predict subsequent reading development are letter recognition (Gallagher et al., 2000;Lyytinen et al., 2006;Torppa et al., 2010) and phonological processing (Gallagher et al., 2000;Snowling et al., 2003;Vellutino et al., 2004).Letter recognition is defined as the knowledge of letter-to-speechsound correspondence (Ozernov-Palchik and Gaab, 2016) and becomes stronger with age, making the ability to read more automatic, accurate, and fast (Ehri, 1987).On the other hand, phonological processing is usually divided into phonological awareness, serial naming, and auditory working memory (Wimmer and Schurz, 2010).The associations of these subskills with future reading abilities have been demonstrated in several studies (Christopher et al., 2015;Eberhard-Moscicka et al., 2021;Helland and Morken, 2016;Kirby et al., 2003;Torppa et al., 2010).
MMR amplitudes show associations with later oral and written language skills (for review, see Kujala and Leminen, 2017;Oh et al., 2019;Volkmer and Schulte-Körne, 2018), but only few studies have investigated these associations during the pre-reading stage.Larger and left-lateralized ERPs in newborns reflecting change detection of tones varying in pitch were linked to better phonological skills and letter recognition at ages 3-5 (Guttorm et al., 2005;Leppänen et al., 2010).In Chinese-and Finnish-speaking children aged 5-6, larger MMN and LDN responses to changes in mothertongue syllables have been associated with better phonological awareness and auditory working memory (Hong et al., 2018;Linnavalli et al., 2017).In addition, a left-side dominance of MMN and LDN responses to speech changes in 6-year-olds was found to be associated with better scores in different language tests (Kuuluvainen, et al., 2016).Furthermore, a larger LDN to consonant change at 2 months of age was found to be associated with better communication skills at 12 months and strong language at 24 months (Maitre et al., 2013).Also, the P-MMR to a change in stress pattern at 5 months was associated with word production at 12-24 months (Weber et al., 2005).However, despite these findings, there is a need for longitudinal designs with optimal paradigms and cognitive test batteries that could create a more comprehensive picture of the potential of ERPs as predictive markers of language development in the pre-reading stage.

MMRs, reading-skill learning and dyslexia
Children do not always develop language abilities as expected, and challenges in speech and language development can have a serious impact on a child's psychosocial and educational outcome (Schoon et al., 2010).Developmental dyslexia is a prevalent neurobiological disorder in which the acquisition of abilities related to reading is compromised (Hulme and Snowling, 2016).It typically manifests as reading inaccuracy, lack of fluency and recognition of words from a text, and difficulty or slowness of reading (Hulme and Snowling, 2016).However, problems in future reading are often preceded by challenges in pre-reading skills and their development in early childhood (Hulme and Snowling, 2016;Ozernov-Palchik and Gaab, 2016;Snowling and Melby-Lervåg, 2016).Concretely, children who later develop dyslexia exhibit difficulties in phonological processing, letter knowledge, grammatical competence, articulation, and understanding of words (Snowling and Melby-Lervåg, 2016;Torppa et al., 2010).
Taking into account the negative impact that dyslexia has on an individual's life, it is important to identify reliable markers of the deficient neural processes underlying this disorder.In this regard, ERPs revealed that dyslexia is understood to stem from a largely auditory-system-based deficit in acquiring and adequately processing native-language phonemes (Eden et al., 2015;Peterson and Pennington, 2015;Vellutino et al., 2004).In line with this theory, MMRs have been found to be diminished (Choudhury and Benasich, 2011;Leppänen et al., 2002;Thiede et al., 2019;Van Leeuwen et al., 2006;Virtala et al., 2022Virtala et al., , 2023) ) or atypically lateralized (Leppänen et al., 2002;Thiede et al., 2019;Van Leeuwen et al., 2006) in children at risk of familial dyslexia.Thus, investigating the connections between infant neural speech processing and the development of pre-reading skills may offer opportunities for early detection of reading deficits.

Research questions and hypothesis
Studies are scarce on the associations between infant MMRs and speech changes and pre-reading skills across ages within the same population.To our knowledge, only two studies have followed the same group of children prospectively from birth to school age (Choudhury and Benasich, 2011;Leppänen et al., 2010), but the studies presented sinusoidal tones as stimuli instead of speech sounds.Our study aims to fill this gap in research by using a longitudinal setting to study how neural speech discrimination during the very first years of life is associated with the subsequent development of pre-reading skills.We measured the amplitudes and hemispheric lateralization of MMRs (MMN, P-MMR, and LDN) to vowel-duration, syllable-frequency,and vowel-identity changes in the same group of children at birth (age between 1 and 17 days) and at 28 months of age, and investigated their associations with the children's pre-reading skills at 28 months and 4-5 years of age.As shown above, despite being a crucial period for language development, this age range has previously been remarkably understudied.28 months is the earliest age to study pre-reading skills with standardized tests, and 4-5year-old children already possess important pre-reading skills that are known to predict future literacy (Byrne, 1998;Gallagher et al., 2000;Lyytinen et al., 2006).Thus, a follow-up in this age window can produce novel information regarding the early development of pre-reading skills during early childhood, before children can read but can already speak.In addition, we also aimed to analyze the differences in the development of pre-reading skills and their connections with MMRs between infants with or without familial dyslexia risk.
We hypothesized that larger and left-lateralized MMR amplitudes during the first years of life would predict better prereading skills at 28 months and 4-5 years of age, in line with the aforementioned studies (Kuuluvainen et al., 2016;Leppänen et al., 2010;Oh et al., 2019;Volkmer and Schulte-Körne, 2018).Based on previous findings of P-MMR being the most robust MMR at birth but decreasing in amplitude with maturation (Choudhury and Benasich, 2011;Slugocki and Trainor, 2014;Shafer et al., 2012;Virtala et al., 2022), we expected that this response would be the strongest predictor of future pre-reading skills at birth, whilst at 28 months, these associations would become more complex.In addition, based on studies reporting an increase in MMN and LDN amplitudes with age (Cheng et al., 2013(Cheng et al., , 2015;;Choudhury and Benasich, 2011;Virtala et al., 2022), we hypothesized that larger amplitudes of these responses at 28 months would predict better development in subsequent prereading skills.Furthermore, based on previous literature (Hulme and Snowling, 2016;Snowling and Melby-Lervåg, 2016), we hypothesized that children at risk of dyslexia would exhibit deficient development in pre-reading skills in comparison with control children.Finally, in keeping with leading theories (Eden et al., 2015;Hulme and Snowling, 2016;Peterson and Pennington, 2015;Vellutino et al., 2004) and our previous results (Thiede et al., 2019;Virtala et al., 2022) suggesting auditory-processing deficits in children with familial dyslexia risk, we expected that the observed associations between infant MMRs and pre-reading skills may differ between children with and without dyslexia risk.

Data
This study includes a subsample of the infants participating in the longitudinal DyslexiaBaby (2014-present) project (described in Thiede et al., 2019;Virtala et al., 2022;Virtala and Partanen, 2018).The families were recruited via traditional media appearances, social-media advertisements, advertisements at Finnish maternity clinics, and via the DyslexiaBaby study website.The DyslexiaBaby study protocol was approved by the Ethics Committee for Gynaecology and Obstetrics, Pediatrics, and Psychiatry (Hospital District of Helsinki and Uusimaa).The study is conducted in compliance with the Declaration of Helsinki.One or both parents of the newborn participants gave written informed consent to participate when the infant was enrolled in the study at birth.

Participants
The original sample consisted of 210 healthy full-term gestational age (at least 37 weeks, and birth weight at least 2,500 g) infants with normal hearing.Infants were considered healthy when they had a 5-minute Apgar score of 7-10.In two infants, the Apgar score was only 6, but they were both in good health at newborn EEG measurement; and in two infants, the Apgar score was missing, but there was no indication of health issues at time of newborn measurement.To test hearing, evoked oto-acoustic emissions (EOAE) were recorded from newborns routinely at the hospital.In two infants in the present sample, this information was missing, but their hearing was later screened at a maternity clinic and was normal.In addition, Finnish had to be (one of) their native language(s).The DyslexiaBaby sample was selected to be over-representative of infants with a heightened risk of developing language difficulties due to parental dyslexia, but also infants of non-dyslexic parents were recruited using the same strategy.Infants included in the risk group of the study had one or two biological parents with a diagnosis of dyslexia, confirmed by a recent (<5 years) diagnostic statement from a professional, or if missing, by researchers of the DyslexiaBaby study.A detailed description of the testing protocol, as well as of the inclusion criteria, can be found in Virtala et al (2022).
For the current study, we collected EEG data at birth (0.5-17 days after being born), EEG and neurocognitive data at 28 months of age, and neurocognitive data at 4-5 years of age (Table 1).To be included in the analysis, children had to have usable EEG data (without excessive artifacts, no more than five bad channels based on visual inspection at birth and at 28 months, and a sufficient number of trials; see section 2.6.1.)for at least one age point.In addition, all children with usable EEG data had to have usable language-test data for at least one age point.Infants were excluded from the analysis due to the following reasons: failure to meet inclusion criteria (18 infants), families withdrawing from the study (4), contact lost (1), children with no EEG data at 28month follow-up (27), children with no usable language-test data at any age point (15), and EEG-data-quality issues (15; see section 2.6.1.).The final samples consisted of 175 infants with complete data at birth, 145 infants with complete data at 28 months, and 162 children with complete data through 4-5 years.Of these children, around 75% had a heightened risk of dyslexia, and 25% had no familial risk of language difficulties (Table 1).Furthermore, the atrisk children participated in a music-listening intervention between birth and 6 months of age, 2/3 of them being exposed to music material and 1/3 being unexposed (for a description of the intervention, see Virtala and Partanen, 2018).In the present study, we utilized the complete samples while controlling for parental dyslexia and intervention group in our analyses.

Experimental stimuli and paradigm
The experimental stimuli and paradigm have been previously described in Thiede et al. (2019).The stimuli (first used by Pakarinen et al., 2014) were a bi-syllabic pseudoword /tata/ (original stimulus) and its three variants, uttered by a female native Finnish speaker (Fig. 1).The original /tata/ stimulus was spoken naturally with stress on the first syllable.The total stimulus duration was 300 ms, of which $250 ms were audible.The second syllable started at $168 ms, and the second /a/ at $181 ms.Deviant stimuli were constructed by editing the second syllable of the standard /tata/ with Adobe Audition CS6 (version 5.0; Adobe Systems Inc.) and Praat (version 5.4.01;Boersma and Weenik, 2013) software.The duration deviant was constructed by lengthening the duration of the second syllable from 71 ms to 158 ms (/tataː/, total length 400 ms, of which $327 ms were audible).The frequency deviant was composed by lifting the fundamental frequency F 0 of the second syllable from 175 hertz (Hz) to 225 Hz, and the vowel deviant was constructed by replacing the second syllable with a naturally uttered syllable /tato/, with the same start, duration, and F 0 as the original /tata/ sound file.The paradigm also contained very rarely presented novel human (e.g., sigh, cry, laugh) and non-human (e.g., telephone ring, electric drill) sounds, the data for which are not included in the analyses of the current study.
The stimuli were presented in a multi-feature paradigm with four blocks of seven minutes each containing 472 stimuli, of which on average 70.1% were standard stimuli and 25.3% deviants (the remaining 4.6% were novel sounds).Each deviant type was presented with a probability of approximately 8.5%.The stimuli were presented otherwise in random order, but so that four standards started each block and a standard always followed a deviant or novel stimulus.An inter-stimulus interval of 850-950 ms alternating in 10-ms steps randomly was used, thus minimizing expectancy effects related to the predictability of stimulus onset.

EEG recordings
EEG data were recorded using an electrode cap (ActiCap, Brain Products GmbH, Gilching, Germany) with 18 (newborn) and 32 (at 28 months) electrodes placed according to the international 10/20 system.In all EEG recordings, we used a BrainProducts Quick-Amp amplifier (v.10.08.14; software: BrainVision Recorder 1.20.0801,Brain Products GmbH, Gilching, Germany) at a sampling rate of 500 Hz, a low-pass filter of 100 Hz, and a high-pass filter of 0.5 Hz.During the recordings, the data were referenced to the average of all electrodes.For stimulus presentation, we used Presentation 17.2 software (Neurobehavioural Systems Ltd., Berkeley, CA, USA) and two Genelec speakers with a stimulus intensity of $65 dB (sound pressure level, SPL) at the child's head.The measurement, with preparations included, took approximately 1-2 hours.
At birth, recordings were carried out in a quiet room at Jorvi Hospital, Helsinki University Hospital, Espoo, Finland (n = 157/175), or in a sound-proof laboratory at the University of Jyväskylä, Finland (n = 18/175).Infants were lying in a crib on their back, with the speaker placed $40 cm from their heads, and their alertness state (''active sleep", ''quiet sleep", ''intermediate sleep stage" or ''awake"), based on Grigg-Damberger et al. (2007), was monitored by a trained professional who conducted the recording.
In the present study, data from all alertness states were combined, and dyslexia-risk and control groups did not differ in the proportions of different alertness states during the recordings.
At 28 months, EEG recordings were conducted after a neuropsychological assessment in a sound-proof, electrically shielded booth in a laboratory at the University of Helsinki (n = 130/145) and in a similar laboratory at the University of Jyväskylä (n = 15/145).The same models of equipment and recording protocol were used at both recording sites.During EEG recording, we tried to direct the children's attention away from the presented stimuli by presenting them muted cartoons during the measurement at 28 months.At birth (0.5-17 days of age), the infants were mostly asleep, and therefore no attention manipulation was possible or necessary.Each child ´s informed consent was ensured, and children were awake and sitting on a parent's lap 160 cm from the speakers.The child was asked not to talk or move, and time stamps with any talking or moving during the recording were saved and considered during manual rejection (see Section 2.6.1.).

Neurocognitive assessment
Children ´s pre-reading skills at 28 months and 4-5 years were tested with the Phoneme processing subtest (NEPSY II, Korkman et al., 2008), Rapid automatic naming task (LUKIVA, Puolakanaho et al., 2011), and an auditory working-memory task developed by the research group.The Letter knowledge subtest (LUKIVA, Puolakanaho et al., 2011) was also used in the 4-5-year age group's assessment.The neuropsychological testing was performed in a testing room at the University of Helsinki or in a similar room at the University of Jyväskylä, and it took around 2 hours.The neuropsychological testing was blinded, so that no researcher testing a child knew which group the child belonged to.Sometimes tasks were not fully completed.The most common reason for this was the child's refusal to continue with the task, and in some cases a researcher's mistake (i.e., concluding the task before fully meeting the age-specific finishing criteria).The task score's eligibility for analysis despite partial completion (e.g., the interruption was most likely due to reaching the limit of a child's abilities) was evaluated case-by-case.
The Phoneme processing (PP) subtest evaluates a child ´s phonological awareness and auditory memory.It is composed of a first section that requires identification of words from word segments, and a second section that assesses phonological processing at the level of syllables and phonemes by asking the child to first repeat a word and later omit or change one or more phonemes to form a new word.The task requires the ability to recognize the word when presented in syllables or when only a part of the word is presented.Responses are given by pointing, and no verbal response is required.
The Rapid automatic naming (RAN) task measures naming speed.It is used regularly by speech therapists in Finland and has been found to differentiate children with developmental dyslexia from typically developing children (Puolakanaho et al., 2011).Because of its relevance, we were interested to see if RAN has a correspondence with neural measures.In this test, the child is asked to retrieve the names of familiar items as fast and accurately as possible.The child ´s time as well as correct answers during naming are recorded.
The Verbal auditory working memory (VAWM) task has been developed by our research group for the DyslexiaBaby study, and it was used to evaluate auditory working-memory capacity and span.It was designed to resemble the Digit span test Note: In the 28-month sample, children must have usable EEG data at 28 months and language test data at 28 months or 4-5 years.In the 4-5 years sample, usable language test data at 4-5 years and EEG data at 0 or 28 months were required.(Weschler, 2006) but without the requirement for verbal responses or knowing digits.In this task, the child was asked to put certain toys in a box in any order under the examiner's oral instruction.
In the first part, only one toy was picked up and put into a box, then two, three, and so on, until the child reached a maximum of five (at the 28-month assessment) or seven (at the 4-5-year assessment) toys or failed to put the right toys into the box in two consecutive attempts.The task requires adequate verbal comprehension skills, auditory working-memory capacity, attention, and concentration.The Letter knowledge (LK) subtest measures a child's knowledge of upper-case alphabetic letters.It is an oral-response task, requiring the child to say aloud the sound and name of 23 letters of the alphabet one by one.The child is given 1 point for each correct answer (max.23 points).

EEG preprocessing
Preprocessing was conducted using Matlab 2017a-2020a (The MathWorks, Inc., USA), with Toolboxes EEGLAB 14.0.0b and 2019_0 (Delorme and Makeig, 2004) and ERPLAB 7.0.0(Lopez-Calderon and Luck, 2014).Firstly, the EEG was filtered with a 0.025-40-Hz band pass in order to get rid of large low-and highfrequency artifacts, and to allow for visual inspection.The visual inspection consisted of searching for electrodes with a flat or continuously noisy signal, and it resulted in no more than five (28%) or six (19%) electrodes marked as ''bad" in the newborn and 28month EEG, respectively.Peripheral electrodes (newborn: Fp1, Fp2, F7, F8, Oz; 28-month: T7, T8, Po9, Po10, O1, O2) with bad data were excluded from the analysis, whereas central electrodes (newborn: F3, Fz, F4, C3, Cz, C4, P3, Pz, P4; 28-month: FC5, FC1, FC2, FC6, CP5, CP2, CP3, CP6) with bad data were further interpolated according to the signal of the valid electrodes.In addition, confirmed muscle-related artifacts were manually excluded from the data, and eye-movement and heart-beat artifacts were marked for a later artifact-removal stage.In the newborn data, eyemovement, muscle-related, and heart-beat artifacts were not clearly identifiable and therefore not searched for.
EEG data were filtered with a 0.5-25-Hz band pass and rereferenced to the average of two mastoid electrodes (LM, RM) and two posterior scalp electrodes (P7, P8).Following this, nonperipheral bad channels were subjected to spherical interpolation using ERPLAB's interpolation algorithm which takes into account electrode positions and their distances from the interpolated electrode on a 3D-sphere modeling of a human head.Interpolation of central channels marked as ''bad" enabled the same electrodes to be used for quantifying the MMR amplitudes across all partici-pants.In order to minimize problems caused by interpolation, it was only performed when absolutely necessary (i.e., when signal quality was very poor), and strict instructions and limits (a maximum of 2 electrodes interpolated in the newborn data and a maximum of 3 in the 28-month data; channels adjacent to each other were not interpolated) were followed to ensure the enhancement of data quality and consistency.Eye-movement and heart-beat artifacts were corrected for with independent component analysis (ICA) with fastica (Hyvarinen, 1999) or runica algorithms in EEGLAB.The found independent components were compared to artifacts in the raw data and their expected scalp distributions to decide whether they should be omitted from the data.After this, continuous EEG data were segmented into epochs starting at À100 ms and ending 840 ms after stimulus onset.To omit eyemovement-related artifacts, epochs with an absolute amplitude exceeding ± 120 lV in electrodes close to the eyes (Fp1, Fp2) were rejected.After that, epochs with data points ± 3 standard deviations (SD) from the mean amplitudes of all epochs and epochs with a drift of >100 lV were excluded.The remaining epochs were separated by stimulus type for each stimulus block and electrode.Thereafter, epochs of all stimulus blocks of the same stimulus type, separate for each participant, were combined, resulting in one dataset per participant and stimulus type.Data of infants/children with less than 30 accepted epochs for more than one deviant were rejected (newborn n = 0; 28-month n = 7).The mean numbers of accepted trials for each deviant were 113 at birth, and 78 at 28 months.Subtraction waveforms were calculated separately for each deviant by subtracting the standard response from the deviant one.The baseline correction was shifted to À100-0 ms from deviation onset, resulting in a baseline correction window of 125-225 ms for the duration deviant and 80-180 ms for frequency and vowel-identity deviants.

Extracting ERPs
Cluster-based mass-permutation tests were implemented using the FieldTrip toolbox (Maris and Oostenveld, 2007;Oostenveld et al., 2011) in order to define spatiotemporal windows of significant deviant-standard differences between deviance onset and epoch end (840 ms) for each deviant type and measurement time point.Thereafter, individual mean ERP amplitudes of the responses that were significantly elicited in the permutation tests were extracted separately for each deviant and measurement point using the toolboxes EEGlab (version 14.0.0;Delorme and Makeig, 2004) and CBRUPlugin (version 2.0b) in MATLAB (Release 2018b; The MathWorks, Inc., Natick, Massachusetts, USA).For individuals lacking a peak value in the search window, mean amplitudes were calculated from a window centered at the group-average peak latency.In the 0-month quantification, individual peak latencies Fig. 1.Experimental paradigm and stimuli.The stimulus types are separated by colors, with separate fundamental frequencies (F 0 , in Hz) for the frequency deviant and the other stimulus types, and with separate durations (in ms) for the duration deviant, the novel stimuli, and the other stimulus types, and for the inter-stimulus interval.The color label and probabilities (in %) for each stimulus type are illustrated in the bottom-left corner.Adapted from Thiede et al., 2019.were searched from broad time windows in a large region-ofinterest (ROI) of 6 electrodes (F3, Fz, F4, C3, Cz, and C4) chosen in accordance with the fronto-central predominant scalp distribution of the MMRs (Näätänen et al., 2007).For the mean-amplitude quantification in the 28-month data, four additional electrodes (FC1, FC2, FC5, and FC6) were added to improve the signalto-noise ratio.Furthermore, additional left and right ROI mean amplitudes were quantified from the left-and right-hemisphere electrodes (all large ROI electrodes except for midline Fz, Cz) to study the hemispheric distribution of the responses.

Statistical analysis
To test whether the MMRs were statistically significant at an alpha level of 0.05, the mean MMR amplitudes in the chosen time windows were compared to zero using one-sample t-tests (Bonferroni-corrected for multiple comparisons).Mean amplitudes of the significant MMRs at the large, left, and right ROIs are listed in Supplementary Table S1.
Associations between variables of interest were preliminarily explored using univariate Pearson correlations in SPSS (reported in Supplementary Table S2, p < 0.05).Considering the sample size and the exploratory nature of the analysis, no corrections for multiple comparisons were performed at this stage.The variables of interest were the MMR amplitudes from the large ROI, the amplitude differences between the left and right ROIs (as a measure of lateralization, more positive values meaning larger responses on the left hemisphere), and the scores in the different pre-reading tests.The amplitudes of the negative MMRs (MMN and LDN) were reversed to keep the correlation directions constant across all three MMRs.We used raw scores from the PP, RAN, and LK subtests and standardized the scores for the VAWM task.For the RAN task, we also used naming speed (in s).In addition, to assess the progression over time in pre-reading skills, we used the difference between the scores at 4-5 years and 28 months for tests implemented at both time points.
In order to test if the infant ERPs significantly predicted prereading skills across the sample, the significant correlations between neural and behavioral measures were further examined in the main multiple linear regression (MLR) models using the jmv module in Jamovi software v2.3.18(Sydney, NSW, AU).In comparison with correlation analyses, MLR analyses quantify the relationship between more than two variables and determine which predictor variables can predict the outcome variable and to what degree.Therefore, they can bring us closer to establishing causal relationships.In addition, MLR analyses allow for the inclusion of control variables.Separate MLR models were created with each test measure as the outcome variable and with amplitudes from the large ROI and amplitude differences between left and right ROIs as predictors.MLR models were created across the whole sample for the MMR amplitudes at birth and at 28 months separately.Furthermore, in order to determine the impact of dyslexia risk on the development of pre-reading skills, we introduced it as a between-subjects factor.In all models, we controlled for the intervention group and accounted for possible multi-collinearity by calculating a variance inflation factor (VIF) for each variable, as well as for possible non-normal data distribution.The statistical significance was set at an alpha level of 0.05.

Associations between ERPs at birth and pre-reading skills across the sample
Larger duration-MMN amplitudes in the left rather than right hemisphere were associated with a greater increase in PP test scores with age (Table 2).Larger vowel-P-MMR amplitudes in the left hemisphere were associated with greater RAN task scores at 28 months (Fig. 2A), and larger frequency-P-MMR amplitudes were associated with greater RAN task scores at 4-5 years (Fig. 3B).In addition, larger and left-lateralized vowel-P-MMR amplitudes predicted a larger improvement in RAN scores with age.Finally, larger vowel-P-MMR amplitudes in the left rather than right hemisphere were associated with better scores in the LK subtest at 4-5 years (Fig. 2B).No associations were found in the MLR models for the rest of the test measures (Supplementary Table S3).

Associations between ERPs at 28 months and pre-reading skills across the sample
Larger vowel-LDN amplitudes predicted a larger improvement with age in PP-subtest scores (Table 3).Left-lateralized duration-MMN and larger duration-LDN amplitudes were associated with better scores in the VAWM task at 28 months and at 4-5 years, respectively (Fig. 3).In addition, larger duration-P-MMR, duration-MMN, and frequency-MMN amplitudes were associated with more improvement in VAWM task performance.Smaller duration-P-MMR and larger and left-lateralized frequency-LDN amplitudes were predictive of a smaller RAN time at 28 months (Figs.2A, 3A), and larger vowel-LDN amplitudes predicted better RAN scores at the same age (Fig. 3A).For RAN performance at 4-5 years, larger vowel-LDN and right-lateralized duration-P-MMR amplitudes were associated with a faster RAN time (Figs.3B, 2B), and larger frequency-LDN amplitudes predicted better RAN scores.Finally, larger duration-P-MMR and frequency-MMN amplitudes were associated with more improvement with age in RAN time.No significant associations were found in the MLR models for the LK subtest (Supplementary Table S4).

Influence of familial dyslexia risk
The dyslexia-risk effect became significant in some of the MLR models (Tables 2 and 3), whereas no significant effects were found for the intervention group.At 28 months, children in the risk group exhibited worse scores in the RAN and VAWM tasks in comparison with control children (Fig. 4A, Table 4).At 4-5 years, the risk group performed significantly slower in the RAN task and obtained worse scores in the LK subtest than the control group (Fig. 4B, Table 4).No significant effects of dyslexia risk were found in rest of the MLR models (Supplementary Tables S3 and S4).
In addition, the MLR analyses yielded significant interactions between dyslexia risk and MMRs at 28 months (Table 3).Concretely, the associations between larger frequency-LDN amplitudes and faster RAN task times at 28 months, and between larger duration-P-MMR and duration-MMN amplitudes and a greater increase in VAWM scores, were only observed in the control group, but not in the risk group (Fig. 5).

Discussion
According to predominant views, efficient neural auditory processing of speech is crucial for language acquisition and development (Gervain, 2015), and the development of reading skills is preceded by pre-reading skills (Whitehurst and Lonigan, 1998).In addition, phonological deficits seem to underlie languagelearning disabilities in dyslexia (Vellutino et al., 2004).Based on these theories, the current study used a large longitudinal sample and well-established methods to determine the predictive value of infant neural speech-discrimination responses on the development of pre-reading skills.Furthermore, we aimed to investigate the influence of familial dyslexia risk on the development of pre-reading skills and on their connection with MMRs.As hypothesized, we found that a larger and left-lateralized P-MMR at birth was associated with better development of pre-reading skills.In addition, larger and more-left-lateralized amplitudes of the LDN and MMN responses to all three deviants at 28 months predicted better pre-reading skills at the investigated time window, but in the case of P-MMR, these associations were more intricate.When examining pre-reading skills in children with or without dyslexia risk, we found that at-risk children exhibited poorer auditory working memory, serial naming, and letter recognition than control children.Moreover, for serial-naming and auditory-workingmemory scores, we found that some of their associations with MMRs at 28 months were only present in the control group, but not in the at-risk one.Together, the results suggest that neural auditory-change detection during infancy predicts subsequent pre-reading skills, thus highlighting the importance of early auditory abilities for language acquisition and development.The differences between the at-risk and control groups suggest that deficits in phoneme discrimination may play a role in developmental dyslexia and related conditions.

The relationship of P-MMR at birth and the development of prereading skills
At birth, the three deviants elicited significant P-MMRs but did not elicit rest of the MMRs (except for duration-MMN), which is consistent with previous research reporting that P-MMR is the most prevalent change-detection response during the first months  of life (e.g., Cheng et al., 2015;Choudhury and Benasich, 2011;Slugocki and Trainor, 2014).Consequently, we expected that larger and left-lateralized amplitudes of this response would be associated with better pre-reading skills in the future.Consistent with this hypothesis, larger and left-lateralized amplitudes of P-MMRs, elicited mainly by the vowel-deviant, predicted better serial naming at 28 months and 4-5 years, as well as greater improvement in this skill between both ages.These results are similar to those obtained by Leppänen et al. (2010) who used similar behavioral outcomes and reported that a larger positive response to a deviant sinusoidal-tone stimulus in newborn infants was related to faster naming and better phonological processing at the age of 5.5 years.Furthermore, we found that a left-side dominance of this response was associated with better development in phonological awareness.In addition, we found that a leftward dominance of duration-MMN was related to better development of phonological awareness.This is consistent with the results of Guttorm et al. (2005) who found that a larger negative response in the right rather than left hemisphere at birth was related to poorer receptive language skills at 2.5 years, measured with language tests different Fig. 3. Averaged deviant-standard subtraction waveforms at 28 months (A) and 4-5 years (B).For illustrative purposes, the sample has been divided into higher-and lower-performing groups for each test based on median scores.Shown are only those waveforms that enclose an ERP that significantly predicted pre-reading skills.Stimuluschange onset is marked with a dashed vertical line and the windows for ERP extraction with color boxes.For illustrative purposes, children were divided into two groups based on median scores for each test.Associations of the responses with performance in pre-reading skills suggest that the speech processing skills reflected by the responses are relevant to the developing language and reading skills.Those mismatch responses (MMRs) that showed significant associations with the test score are marked with asterisks (*).28mo = 28 months old; RAN = rapid automatic naming; VAWM = verbal auditory working memory; Duration = duration deviant; Frequency = frequency deviant; Vowel = vowel deviant; P-MMR = positive mismatch response; MMN = mismatch negativity; LDN = late discriminative negativity.from those used in the present study.These results suggest that a larger and left-lateralized P-MMR may be a predictive marker of future language development already at birth.Observing these associations already at birth, when environmental experiences are modest, supports the idea that early biological/genetic factors may contribute to the development of auditory abilities and consequently of subsequent pre-reading skills.This hypothesis is consistent with various studies suggesting that variations in early reading development skills are moderately heritable (Christopher et al., 2013;Logan et al., 2013) and that genetic factors may affect brain development before birth through different pathways (such as neuronal volume, neuronal migration, and axonal connectivity) that could contribute to enhanced or diminished neural processing of speech stimuli (Peterson and Pennington, 2015).

The relationship between MMRs at 28 months and the development of pre-reading skills
As we hypothesized, larger and left-lateralized amplitudes of the MMRs at 28 months predicted better development of prereading skills, particularly of those abilities that form phonological processing (i.e., phonological awareness, serial naming, and auditory working memory, Wimmer and Schurz, 2010).This makes sense from the point of view that generating efficient speech-   sound representations is crucial for phonological awareness, and speech information has to be retained in auditory working memory in order to develop adequate language skills (Vellutino et al., 2004).Contrary to our hypothesis and previous results (Guttorm et al., 2010;Leppänen et al., 2010), we did not find any significant association between MMRs and letter recognition.However, this connection has not always been observed (Maurer et al., 2009), and although this ability is also partly dependent on speech perception (Vellutino et al., 2004), it seems to be a heterogeneous skill particularly affected by the home environment (Torppa et al., 2006).We found that larger MMNs to duration and frequency deviants predicted better development of auditory working memory.These results are in line with previous research by Grossheinrich et al. (2010) who found that a larger MMN at 4-5 years was associated with better short-term memory in a sample of German-speaking children.In addition, larger amplitudes of frequency-MMN predicted better development over time in automatic serial naming, which is consistent with Eberhard-Moscicka et al. (2021), who also reported this association, but in a sample of school-age children.However, contrary to previous research (Hong et al., 2018;Linnavalli et al., 2017), the correlations between MMN amplitudes and phonological awareness did not reach statistical significance.For LDN, we found similar results as for MMN.Larger frequencyand vowel-LDN predicted better serial naming at 28 months and 4-5 years, and a larger LDN elicited by the duration deviant predicted better auditory working memory at 4-5 years.These findings have been previously reported by Hämäläinen et al. (2015), but in a smaller and older sample and using sinusoidal tones as stimuli instead of speech sounds.Furthermore, as Hong et al. (2018) stated using an older-aged sample of Chinese-speaking preschoolers, larger vowel-LDN predicted better development of phonological awareness.Associations between MMN and LDN amplitudes during the first years of life and language tests different from the ones used in the present study have also been described in previous literature (Grossheinrich et al., 2010;Maurer et al., 2005).
Larger MMN and LDN amplitudes might reflect more-mature processes.Previous studies investigating the maturation of these responses (Cheng et al., 2013(Cheng et al., , 2015;;Choudhury and Benasich, 2011) support the idea that these negative responses increase in amplitude during childhood.Furthermore, Virtala et al. (2022) assessed the maturation of MMRs in the same dataset as included in this study, and described an increase in MMN and LDN amplitudes from 6 to 28 months across the whole sample.Following this line of thought, our results suggest that larger amplitudes of these MMRs during the first years of life may reflect a more-mature auditory cortex, and consequently may be predictive markers of better development of pre-reading skills.Conversely, children with problems in language development may exhibit less-mature auditory processing, and consequently will exhibit diminished MMN and LDN during the first years of life.This reasoning might be limited to MMN and LDN, since P-MMR usually first diminishes in amplitude after 6 months of age and then turns to a negative response or a response complex with different polarities (He et al., 2009;Trainor, 2012).
This study also had the aim of determining the effect of MMR lateralization on the development of pre-reading skills.As hypothesized, we found that at 28 months, a larger left-hemispheric dominance of duration-MMN and frequency-LDN was associated with better auditory working memory and serial naming, respectively.Some of these associations have been described previously in an older sample by Kuuluvainen et al. (2016) and are in line with studies reporting that preschoolers exhibit larger MMN responses to sound changes on left scalp electrodes (Partanen et al., 2013).Considering this, our results suggest that an atypical distribution of MMRs may be associated with challenges in the maturation of pre-reading skills at this age.In the brains of children with better pre-reading skills, there may be an increased contribution of the language-dominant left hemisphere to neural auditory processing, while in children with poorer performance, the responses may be more evenly distributed.However, literature assessing the associations between language development and lateralization of brain responses is scarce and inconsistent, so future studies should investigate this question further.
Regarding P-MMR, our findings were more intricate.Larger and more right-lateralized amplitudes of this response to the duration deviant predicted slower serial naming at 28 months and 4-5 years, respectively.These results may be explained by the inversion of polarity that infant P-MMR seems to experience with maturation.P-MMR is the most robust index of auditory change detection in infancy, but it decreases in amplitude after 6 months of age (Choudhury and Benasich, 2011;Shafer et al., 2011;Slugocki and Trainor, 2014;Virtala et al., 2022, using an identical dataset).Thus, the infant P-MMR observed in this study may represent an ''immature MMN" that would shift polarity with age (as suggested by He et al., 2007He et al., , 2009;;Trainor et al., 2003;Trainor, 2012).Following this line of thought, our results suggest that an enlarged duration-P-MMR at 28 months of age may be interpreted as reflecting an immature maturational state of the auditory cortex, and consequently may be associated with delayed development of pre-reading skills.However, contrary to this idea, larger amplitudes of the duration-P-MMR response were associated with better improvement over time in serial naming and auditory working memory.A possible explanation for this result might be a ceiling effect in which children with better pre-reading skills, and consequently with diminished duration-P-MMR, reached high scores on the tests already in the 28-month measurement, thus limiting the potential increase experienced in the 4-5-year evaluation.

Influence of dyslexia risk on the associations between infant MMRs and pre-reading skills
The diminished scores in serial naming, auditory working memory, and letter recognition observed in the dyslexia-risk group were an expected result, based on previous literature (for a review, see Snowling and Melby-Lervåg, 2016).For example, Torppa et al. (2010) found that children with reading difficulties were outperformed by typical readers on letter naming from 2 years of age onward.Also, the slower development of serial-naming skills in children at risk of dyslexia have been previously reported by Cardoso-Martins and Pennington (2004), and various studies have described the impact of dyslexia risk on phonological-memory skills (for a review, see Melby-Lervåg et al., 2012).However, these results have been obtained in samples of school-aged children, whereas our study focused on children aged between 28 months and 4.5 years, an understudied age range.Therefore, our study provides novel findings that support the hypothesis that in children with a familial dyslexia risk, genetic vulnerability manifests early as deficient pre-reading skills, which in turn may lead to subsequent problems at school age.
Interestingly, we found that larger duration-P-MMR and MMN amplitudes at 28 months predicted an increase in auditory working memory over time in the control group, whereas this association was not present in the at-risk one.Similarly, larger frequency-LDN at 28 months was associated with faster serial naming at the same age point only in the control group, but not in the at-risk one.These results are consistent with the findings obtained by Virtala et al. (2022) who used a dataset partly overlapping the one in the present study and found that children at dyslexia risk exhibited enlarged duration-P-MMR and diminished duration-MMN at 28 months.In addition, several studies reported enlarged P-MMR (Leppänen et al., 2002;Pihko et al., 1999) and diminished MMN and LDN (Leppänen et al., 2002;Thiede et al., 2019, partly overlapping newborn data;Van Zuijen et al., 2012) in dyslexia-risk infants and small children.As MMN emerges and P-MMR diminishes with increasing age (Choudhury and Benasich, 2011;Slugocki and Trainor, 2014), the observed enlarged duration-P-MMR coupled with a diminished MMN may be interpreted as immature neural auditory discrimination in the dyslexia-risk group.Thus, we suggest that delayed maturation of auditory processing may have a connection with the deficient development of pre-reading skills in children at dyslexia risk, which is in keeping with the dominant premise of a core phonological dysfunction associated with learning disabilities (Ozernov-Palchik and Gaab, 2016;Peterson and Pennington, 2015).Atypical auditory-processing abilities may be associated with cognitive processes important for reading development and may help to predict language disabilities.However, more research is needed, and it is unlikely that atypical speech processing alone is enough to lead to dyslexia, but is rather a risk factor leading to cumulative effects in processes critical to reading acquisition.

Limitations
The current study holds a few limitations.First, interpolation is an issue that can compromise the data, particularly when the overall signal quality is not optimal.However, in the present study, interpolation was done rarely and never in adjacent channels.Second, our methodological choices may lead to a risk of false positives.When analyzing the data, we included a large number of variables, including several responses elicited by each deviant and various tests for each age point; consequently there is a high risk of observing multi-collinearities that could compromise the reliability of the observed associations.However, our large sample size may help to evade this problem, and conducting a preliminary analysis using Pearson correlations allowed us to reduce the number of variables included in the final analyses.In addition, when conducting the final analyses using MLR models, we accounted for possible multi-collinearity by excluding those variables with a high VIF.In order to find the most robust and reliable associations, future studies may consider the use of methods designed specifically for analyzing longitudinal data (for instance, structural equation models) rather than the most common correlation analysis.Another methodological limitation that must be taken into account is the fact that we calculated the difference between the amplitude of left and right ROIs to assess the lateralization of the responses.Values obtained by subtraction may depend on the magnitude of the amplitude at both hemispheres, and therefore observing a significant effect of lateralization may be more difficult for smaller responses than for larger ones.Future studies should take this into account and may consider the use of a more accurate laterality index that could avoid this bias.
Also, it is important to point out that our sample has a large proportion (around 75%) of infants at dyslexia risk, which may have affected the results of our study.According to previous literature and results based on a partly overlapping dataset with the present study (Virtala et al., 2022), MMRs may be diminished or absent in the dyslexia-risk group, and therefore it may be possible that some of their observed associations with pre-reading skills would be less pronounced or even veiled in comparison with what we could observe in a sample with only control infants.However, we chose to analyze the whole sample in order to maintain a large sample size.

Summary and conclusions
The present study, with its large longitudinal sample, provides evidence on how infant auditory-change-elicited ERPs can predict the development of pre-reading skills, as well as on the effect that familial dyslexia risk has on them.Based on the observed associations, we were able to demonstrate that a larger P-MMR is associated with better development of pre-reading skills already at birth, when family environment has a modest effect.Also, a more mature pattern of neural auditory speech discrimination (consisting of a diminishing P-MMR and increasing MMN and LDN) during the first years of life was associated with better development of pre-reading skills during the studied age range.We also found that children at risk of dyslexia exhibited worse pre-reading skills than control children, and that dyslexia risk had an impact on some of the observed associations between MMR amplitudes and test scores.Overall, the present study suggests that neonatal speech-sound discrimination abilities and the maturity of the auditory system are relevant to the development of pre-reading skills.Also, it supports the idea that difficulties in auditory processing may contribute to the deficits in language acquisition observed in dyslexia, which may serve as a basis for planning effective preventive and rehabilitation interventions already at early stages, when the child is not yet reading.However, further studies are needed to determine whether the auditory speech-discrimination ability may be used to predict language-related skills at an individual level.

Declaration of competing interest
None of the authors have potential conflicts of interest to be disclosed.

Fig. 2 .
Fig.2.Significant relationships between the lateralization of responses and test performance at 28 months (A) and 4-5 years (B).For illustrative purposes, children were divided into quartile groups (G1-G4) according to test performance, so that children in G1 were the lowest 25% and children in G4 the highest 25% of scorers for the skill in question (note that groups are reversed for naming speed, as good performers are faster).The images show the amplitude difference between left and right scalp locations, positive values meaning larger responses on the left than on the right.Error bars represent ± 1 standard error of the mean.28mo = 28 months; L = left; R = right; RAN = rapid automatic naming; VAWM = verbal auditory working memory; P-MMR = positive mismatch response; MMN = mismatch negativity; LDN = late discriminative negativity.
Fig. 5. Scatter plots illustrating the significant associations between mismatch-response (MMR) amplitudes at 28 months and outcome measures in the 2 groups: atrisk children with developmental dyslexia (DD-risk) and control children with typical reading skills (Control).A) Outcome measures at 28 months.B) Difference in outcome measures between 4-5 years and 28 months; P-MMR = positive mismatch response; MMN = mismatch negativity; LDN = late discriminative negativity.

Table 2
Summary of the statistically significant (p < 0.05) main results of the multiple linear regression (MLR) model analyses with test measure as outcome variable, with MMRs amplitudes and lateralization at birth as predictors and with the dyslexia risk (DD-Risk, when appropriate: Risk vs.Control groups) and the intervention group as fixed factors.

Table 3
Summary of the statistically significant (p < 0.05) main results of the multiple linear regression (MLR) model analyses with test measure at 28 months as outcome variable, with MMRs amplitudes and lateralization at 28 months as predictors and with the dyslexia risk (DD-risk, when appropriate: Risk vs.Control groups) and the intervention group as fixed factors.

Table 4
Mean scores and standard deviations (SD) in each test for the whole sample and for the control and at-risk (DD-risk) groups.The test scores that significantly differed between groups are bolded.