Teaching methods shape neural tuning to visual words in beginning readers

The impact of global vs. phonics teaching methods for reading on the emergence of left hemisphere neural specialization for word recognition is unknown in children. We tested 42 first graders behaviorally and with electroencephalography with Fast Periodic Visual Stimulation to measure selective neural responses to letter strings. Letter strings were inserted periodically (1/5) in pseudofonts in 40sec sequences displayed at 6Hz and were either words globally taught at school, eliciting visual whole-word form recognition (global method), or control words/pseudowords eliciting grapheme-phoneme mappings (phonic method). Selective responses (F/5, 1.2Hz) were left lateralized for control stimuli but bilateral for globally taught words, especially in poor readers. These results show that global method instruction induces activation in the right hemisphere, involved in holistic processing and visual object recognition, rather than in the specialized left hemisphere for reading. Poor readers, given their difficulties in automatizing grapheme-phoneme mappings, mostly rely on this alternative inadequate strategy.


Introduction
that strictly alphabetical approaches are more effective than mixed or non-alphabetical 129 approaches in word reading, spelling, and in text comprehension (Deauvieau & Terrail, 130 2018). Alphabetical approaches also induce the highest improvements in children at risk of a 131 reading disorder or with low socio-economic background, by increasing their self-teaching 132 ability (Ehri et al., 2001;Goigoux, 2016;Rayner et al., 2001). On the contrary, mixed or non-133 alphabetical approaches give rise to the highest proportion of poor readers and generate a 134 higher heterogeneity of performance within a class (Braibant & Gerard, 1996;Deauvieau et 135 al., 2013;Goigoux, 2000). In agreement with the self-teaching hypothesis (Share, 1995) Share, 1999). Also, kindergarten children trained to memorize 139 an artificial script or one-syllable words with a global strategy, have a lower ability to read 140 novel stimuli than children trained with a GP mapping strategy (Jeffrey & Samuels, 1967) and 141 have difficulties to infer GP mappings that have not been trained explicitly (Byrne, 1991(Byrne, , 142 1996. 143 Adults' behavioral studies confirmed the lower efficiency of the global method 144 concerning the transfer to novel stimuli and the implicit acquisition of GP mappings (Bishop,145 1964; Bitan & Karni, 2003;Byrne, 1984;McCandliss, Schneider, & Smith, 1997;Yoncheva, 146 Blau, . In neuroimaging, fMRI training studies used unfamiliar 147 stimuli (e.g. artificial script, pseudowords) to investigate neural correlates of orthographic, 148 phonological and semantic processes in learning to read novel words. In comparison to 149 global training (e.g. visual shape recognition, direct mapping with meaning), training based 150 on phonological recoding led to a response modulation in the L-VOTC (Xue, Chen, Jin, & 151 Dong, 2006; Sandak et al., 2004). Furthermore, the specific involvement of the VWFA in 152 mapping letter shapes to phonology has been demonstrated in a study which contrasted the 153 association of an artificial script with speech sounds and non-speech sounds (Hashimoto & 154 Sakai, 2004). EEG studies showed that training GP mappings led to a left-lateralized N170 155 response sensitive to trained and untrained artificial script (between subjects design: 156 Yoncheva et al., 2010) ; within-subjects design : Yoncheva et al., 2015), while global training 157 led to a right-lateralized N170 response. These results highlight that the unit's size on which 158 the learner focuses, and therefore the processes engaged in word recognition after learning, 159 directly influence brain mechanisms. It also suggests that only a phonic method engages the 160 typical left lateralized brain circuitry of reading. 161 However, we currently do not know how (different) teaching methods impact brain 162 activity in children. The current study aimed at filling this gap, by studying neural changes 163 induced by the global and phonic teaching methods in a natural school context. We tested 164 forty-four children (of whom 2 were excluded, see Material and Methods) during the first 165 trimester of Grade 1. At that time of the school year, some words have been taught with a 166 global method, in parallel to the teaching of GP mappings with a phonic method. This gave 167 us the opportunity to compare neural responses to letter strings that should trigger only GP 168 mappings (control words/pseudowords), to letter strings that should trigger visual whole form 169 recognition processes (global words). The comparison of words globally learnt to control 170 letter strings includes an inevitable and intrinsically confounded variable, which is familiarity 171 (only global words are familiar to children). However, as explained in our hypotheses below, 172 a differential impact of reading level on the use of a strict visual recognition strategy is 173 expected, helping us to disentangle these confounded factors. 174 More precisely, in the two schools where we conducted the study, a list of global 175 words was provided by the teachers, from which we extracted a common sub-part of twenty 176 4-and 5-letters words. These words have been globally taught in the classroom (mapping of 177 whole-visual word form to spoken form, without knowledge of the inner letter-speech sounds 178 associations), then printed on small cards given to the children in the so-called "words' box", 179 and home-trained every day. In parallel, teachers taught GP mappings at the rate of one 180 letter per week, with a variety of exercises: write the letter, recognize it within words, learn its 181 case variants, etc. At the time of our testing, 9 letters had been taught in classrooms (a, é, e, 182 è, I, o u, r, l). The other letters have been encountered visually by children and were variably 183 recognized (as attested by performance in behavioral tests, see Table 1). automatic visual discrimination of letter strings) and rapid approach, FPVS-EEG is ideal to 207 use in young children. 208 In the current study, we tested the hypothesis that, at an early stage of reading 209 acquisition, words learnt globally as unique visual forms would be represented differently 210 bottom rows), and target stimuli were either control words (W, top row), pseudowords (PW, middle row) or globally taught words (GW, bottom row) appearing every fifth item. Each sequence lasted 40 seconds, during which stimuli were presented by sinusoidal contrast modulation at 6 Hz, each stimulus reaching full contrast after 83 ms (i.e., one cycle duration = 166.66 ms). Stimulation alternated between base (B) and target (T) stimuli such as BBBBTBBBBTBBB. Target stimuli therefore appeared at 6Hz/5, so at 1.2Hz. Stimuli were randomly presented with no immediate repetition and appeared continuously on the screen. In total, 240 stimuli were presented per sequence (48 target stimuli and 192 base stimuli), and each condition was repeated 3 times. (B) Timeline of a sequence: each sequence started with a fixation square (for 2-5 s) after which the stimulation faded in (for 2 s) then reached full contrast (for 40 s) and then faded out (for 2 s); see Methods.
than control letter strings (words/pseudowords) which would trigger GP mappings. More 211 specifically and based on our previous work in kindergarten children (Lochy et al., 2016), we 212 expected that responses at 1.2Hz and harmonics, reflecting discrimination of letters among 213 pseudofonts, would vary according to the condition. We hypothesized that GP mappings (in 214 W/PW control letter strings) would engage the typical left brain circuitry of reading, while GW 215 would engage a right-lateralized circuitry (Yoncheva et al., , 2015. Most importantly, we 216 expected a differential effect of the global teaching method as a function of children's reading 217 level, which if confirmed also precludes interpretation of our results as simple familiarity 218 effect. Indeed, poor readers having difficulties to automatize GP mappings (e.g. unstable 219 knowledge of letters' sounds (Perfetti, 1992), deficit in phonological awareness (Ziegler et al., 220 2010) and/or visuo-attentional processing (Ans et al., 1998)), compensate their difficulties by 221 using alternative strategies for reading, like visual memory among others (e.g. salient visual 222 features; Campbell & Butterworth, 1985;Vellutino, 1987). Therefore, we hypothesized that 223 they might be more, or exclusively, influenced by the global method inciting them to visually 224 encode the whole-word forms. In other terms, they might show a different response pattern 225 to GW than good readers (e.g. more right-lateralized). Because children of both groups have 226 been familiarized explicitly and to the same degree with the GW, group differences regarding 227 responses to GW would strongly suggest that familiarity per se is not the key factor inducing 228 a right-lateralized neural response. Indeed, if familiarity induces engagement of the right 229 hemisphere, then this should be the case whatever the children's reading level. 230 Results 231

Base rate responses 255
The base stimulation frequency reflects the general synchronization of the visual 256 system to the periodic stimulation (or level of attention to the stimuli). Z scores computation 257 revealed significant responses in all conditions at exactly 6Hz and several harmonics (F to 258 5F) at six medial occipito-parietal electrodes (O1, O2, Oz, PO3, PO4, Pz). In order to 259 determine the electrodes of interest, all electrodes were ranked according to their largest 260 amplitude values for the sum of baseline corrected amplitudes computed on five significant 261 harmonics (6Hz, 12Hz, 18Hz, 24Hz, 30Hz, see Methods) as determined by grand averaged 262 data. In all conditions, the largest response was recorded at three occipito-medial (OM)

Analysis by reading level 269
First, we computed a composite score of reading for each child by averaging 270 accuracy scores for graphemes, syllables and words reading. Since our objective was to 271 compare subgroups, we assigned children in subgroups on the basis of the group's mean 272 composite score (34.57% of accuracy). Children who performed above the group's mean 273 composite score (> 35%) were considered as « good readers » (N=18), and those below (< 274 . Histograms represent the same data, with standard errors of the mean. For letter strings-selective responses, stars indicate significant difference between the left (O1) and the right (O2) occipital electrodes (**: p<0.01, ns : not significant). (C) SNR EEG spectra on O1 (dark grey), O2 (in light grey) and Oz (in black) for each condition, stars indicate significant responses at letter strings-selective frequency and harmonics (1.2Hz, 2.4Hz, 3.6Hz, 4.8Hz).  Figure 2 are provided as a compressed folder which is available through the following source data. Source data 1. EEG data in frequency domain and behavioral data. https://doi.org/10.5061/dryad.0k6t3s0 34%) as « poor readers » (N=19). We excluded from analysis 5 children whose score was at 275 the group's mean composite score (34-35%). 276

Word-discrimination responses 277
An ANOVA was performed on word discrimination responses (sum of baseline 278

Brain-behavior correlations 304
We assessed if reading scores correlated with lateralization scores (LS; calculated as 305 LH -RH) for the responses to letter strings on the 37 children retained in our subgroups 306 analysis (Fig. 4), and this, separately for the response to global words, control words, control 307 pseudowords, as well as the average of W/PW (as in Lochy et al., 2016). All these 308 correlations were significant (respectively: Spearman Rho=0.44, P = 0.003; Rho=0.43, P <

Discussion 321
Our study reveals that the methods used to teach children how to read impact the 322 development of neural responses to letter strings. We took advantage of a natural school 323 context using a mixed approach in the beginning of the first grade to compare, in 42 children, 324 neural responses to letter strings that trigger GP mappings (phonic method), with letter 325 strings that induce whole-word visual form recognition processes (global method) in a within-326 subjects design. In 2 minutes of recording only, we found different and specific neural 327 responses to these two categories of letter strings, which were moreover modulated by the 328 reading level of children. These results lead to two critical conclusions that will be discussed 329 in turn. Remarkably, the knowledge of all letters is not necessary to observe this LH lateralization 360 pattern in response to letter strings. In the current study, control words/pseudowords were 361 constituted of 18 different letters, of which only 9 had been formally taught with the phonic 362 method at school, and on average 14 were known (see Table 1). Similarly, in our previous 363 kindergarten study (Lochy et al., 2016), children who knew more than 9 letters already 364 showed this typical LH pattern of responses to letter strings. 365 The RH, on the contrary, is known to support object and face recognition ( 2000; Sergent, 1982), as evidenced by neuropsychological deficits of patients with unilateral 373 brain damage (review in Ivry & Robertson, 1998). Our results reveal that globally learnt words enhance RH responses, which we interpret as a greater involvement of purely visual 375 processes in their recognition. This atypical neural pattern during the processing of these 376 letter strings suggests that whole-word instruction exposes children to a deviant (i.e., not 377 optimal) reading strategy. here with other behavioral tests than in our previous study (single letters reading in 393 kindergarten vs. letters, syllables, words and pseudowords here), we can conclude that the 394 relationship between response amplitude to letter strings measured in FPVS and 395 performance at the beginning of the learning process, can be generalized to other age-396 groups and other behavioral tests. 397 Our second crucial finding is that poor readers were more influenced by the global 398 teaching method than good readers. In fact, only the poor readers showed the atypical 399 bilateral response to global words. This finding agrees with behavioral training studies 400 showing that sufficient GP mappings knowledge is necessary for being able to infer GP mappings from globally trained words (Byrne, 1984(Byrne, , 1991(Byrne, , 1996 or even to switch from 402 global to local attentional focus (Yoncheva et al., , 2015. This means that good readers, 403 having sufficiently automatized GP mappings, could process even the global words with an 404 orthography-to-phonology type of process while poor readers relied on whole-word 405 recognition. This finding supports the view that non-alphabetical approaches are significantly 406 less efficient in children with learning difficulties (Braibant & Gerard, 1996 confusion between visually similar words) and reading acquisition in general, for instance by 413 impeding self-teaching of novel words (Share, 1995), as well as development of phonological 414 awareness and letter knowledge. The fact that good readers processed all stimuli types 415 predominantly with their LH nevertheless shows the dominance of the most efficient process, 416 once acquired. 417 The phonic and global methods that we compared differ by definition on an important 418 aspect which is item familiarity. Indeed, the phonic method aims at providing the ability to 419 transfer GP mappings on new words, while the global method aims at creating 420 representations for the learnt items. Therefore, an intrinsic confounded factor inherent to the 421 two methods is that global words were highly familiar to children, while control words and 422 pseudowords have not been explicitly taught at school. However, the RH engagement for 423 global words is unlikely to be due to familiarity as such. Indeed, all children were as familiar 424 to the set of global words, but good readers presented a left lateralized response in all 425 conditions, while poor readers presented RH engagement for global words only. These by-426 group and by-condition differences confirm that global words are processed differently than 427 control letter strings in poor readers only, and are not simply more familiar. The inter-conditions correlations by group provide another argument supporting the view that in good 429 readers, the reading processes triggered by both control letter strings and global words were 430 similar, while for poor readers, they were different. For good readers, the lateralization scores 431 correlated highly between all three conditions (above Rho=0.75), while for poor readers, 432 correlations between the two control letter strings (words/pseudowords) were significant 433 (Rho=0.64), while they were not between global words and control letter strings. The RH 434 engagement for global words is also unlikely to be due to a fluctuation of attention given that, 435 first, base rate responses (i.e., 6Hz and harmonics) were similar across conditions and 436 groups; and second, detection of the color change on the fixation square did not vary 437 according to conditions and groups. All these findings thus confirm the specificity of the 438 processes triggered by global words, which are different than the GP mappings triggered by 439 control letter strings, and which are modulated by the mastery of GP mappings. 440 Since the persistence of the effects has not been assessed, this study does not allow us 441 to conclude that the global method has a negative impact on the long-term development of 442 neural circuits for reading. It may well be that before acquiring the mastery and 443 automatization of GP mappings, good readers presented the same atypical neural pattern 444 than poor readers for global words. Thus, it is also possible that poor readers, after 445 improvement of GP mappings ability with formal instruction, could progressively present a left 446 lateralized response despite the whole-words instruction. On the other hand, some children 447 having a severe or specific disorder and unable to reach a sufficient level of GP mappings 448 automatization, could maintain an atypical neural pattern for these global words. 449 Furthermore, as suggested by behavioral studies (Campbell & Butterworth, 1985;Vellutino, 450 1987), they could even mostly attempt to transfer this strategy to process other written 451 words, in order to compensate their persistent difficulties in acquiring GP mappings. 452 Therefore, the atypical neural circuitry that we find here for global words and only in poor 453 readers, seem to us sufficient to strongly encourage education professionals not to use the 454 alphabetical approaches. Visual memorization of global word shape is not involved in (adult) 456 expert recognition, which results from the automatization of analytical processes performed 457 on written words. We are aware that the aim of teachers is to vary the approaches in learning 458 to read in order to motivate children, because the process of learning all the GP mappings 459 and to automatize them is long and laborious. However, only strict alphabetical approaches 460 can provide the indispensable foundations for the development of expert reading skills, while 461 non-strict alphabetical approaches penalize pupils' learning and this effect is probably even 462 enhanced in children with learning difficulties. 463

Conclusion 464
To our knowledge, our study is the first to assess the impact of two widely used teaching 465 methods, global and phonic, on the neural circuits for reading in children. We showed, in a 466 natural school context and at an early stage of reading acquisition, that the unit's size 467 involved in mapping written information to phonological forms (graphemes to phonemes vs 468 visual whole-words to its spoken form) directly influences neural mechanisms recruited 469 during written word processing, as previously found in adults with an artificial script 470 (Yoncheva et al., , 2015. Words taught globally as whole visual shapes induced 471 bilateral responses, while control letter strings (words/pseudowords), triggering GP 472 mappings, led to typical LH responses. Importantly, this pattern was modulated by reading 473 level and was present only in poor readers, which strongly suggests that the difficulty in 474 automatizing GP mappings induced an increased reliance on an alternative visual strategy. 475 Given that our data were collected in the context of a mixed approach and after only 2 476 months of school instruction, we conclude that the use of a global method very rapidly 477 exposes children to a deviant reading strategy, which might impede the acquisition of reading 478 especially in children having difficulties to acquire GP mappings. Therefore, in agreement 479 with the view predominantly defended, our data is a strong argument in favor of the use of 480 strict alphabetical approaches. 481

Material and Methods 482
Participants 483 First-grade children (N=44) from two different Belgian schools (20 boys, mean age = 484 6.08 years; range = 5.11-7.11, 41 right-handed) were tested at the very beginning of formal 485 reading instruction (in the first trimester of grade 1, i.e., in November-December). Two of 486 these children were excluded because of abnormal performances in behavioral tests (see 487 below). All children had normal or corrected-to-normal vision. They were tested after the 488 parents gave a written informed consent for a study approved by the Biomedical Ethical 489 Committee of the University of Louvain. They were unaware of the goal of the study and that 490 a change of stimulus type occurred at a periodic rate during stimulation. The testing took 491 place in a quiet room of the school in two or more sessions (EEG, behavioral).

EEG testing 505
Stimuli 506 Four categories of 20 stimuli were used for this experiment (Fig. 1.): pseudofonts 507 (PF), words learned by a global method at school (GW) and control words (W) or pseudowords (PW). The natural school context of learning provided us with the "global 509 words". We chose 4-and 5-letters words (N=10 of each) from the « words' box » that 510 teachers had provided children at the beginning of the school year and which contained the 511 words learnt item by item by mapping the whole-word form to its phonological counterpart. 512 Control words were words for which children did not receive any explicit instruction and they 513 were selected from the Manulex database (Lété, Sprenger-Charolles, & Colé, 2004)  letters words were rearranged (e.g., the words "joli" ("cute") and "fête" ("party") could give rise 525 to the pseudowords ''jote'' and ''fêli'') and the letters position of the 5-letters words was 526 changed (e.g., the words "route" ("street"), "table" ("table") and "forêt" ("forest") give rise to 527 the pseudowords "reuto", "ablet" and "frêto"). Pseudowords were matched with all words (W 528 and GW) in bigram frequency, identity of letters and in number of letters (four or five). ascending/descending features and close-up shapes. Therefore, each word (W or GW) had 541 a corresponding pseudoword and pseudofont, containing the exact same amount of black-542 on-white contrast, so that all conditions were comparable in terms of low-level visual 543 properties. 544 These different stimuli allowed to create three conditions (Fig.1.). In each condition, 545 base stimuli were pseudofonts (PF), and target stimuli were either global words (PF-GW 546 condition), control words (PF-W condition) or pseudowords (PF-PW condition). Stimuli were 547 presented centrally in Verdana font with a height between 47 and 77 pixels and a width 548 between 103 and 271 pixels, depending on the shape of the individual letters. At a viewing 549 distance of 1 m with a screen resolution of 800 × 600 pixels and a refresh rate of 60 Hz, 550 stimuli ranged from 2.69 to 7,07 (width) and 1.32 to 2.18 (height) degrees of visual angle. 551

Procedure 552
The stimulation procedure was very similar as in our previous FPVS-EEG studies on 553 word recognition (Lochy et al., 2015(Lochy et al., , 2016. Each stimulation sequence started with a fixation 554 square displayed for 2-5 s (randomly jittered between sequences), after which the stimulation 555 gradually faded-in during 2 s (progressive increasing modulation depth from 0% maximum 556 contrast level to 100%). Then, the sequence of stimulation was presented for 40 s, after 557 which the stimulation faded-out during 2 s. This procedure was used to avoid abrupt eye-558 movements or blinks at the beginning or near to the end of a sequence. Stimuli were 559 presented by means of sinusoidal contrast modulation at a base frequency rate of 6 Hz (i.e., 560 one item every 166.66 ms, from a grey background to full contrast and back in 166.66 ms 561 thus, each item reached full contrast at 83 ms). Given that the stimulus can be recognized at 562 low contrast (i.e., 20% or less), the actual duration of stimulus visibility was close to 140 ms. 563 Every fifth stimulus (1/5) of the sequence (frequency of 1.2 Hz thus, every 833 ms), a global word (PF-GW sequence) or a control letter strings (PF-W or PF-PW sequence) was 565 presented. Stimuli were presented with a software (Sinstim) running over a JavaScript (Java 566 SE Version 8). Each condition was repeated three times. Considering a total of 40 s 567 (sequence duration) × 3 (repetitions) × 3 (conditions), 6 min of stimulation were presented in 568 total. There was a pause of approximately 30 s between each sequence, which was initiated 569 manually to ensure low-artifact EEG signals. 570 During the stimulation, children continuously fixated a central square and were 571 instructed to press the space bar when they detected any brief (200 ms) color change of the 572 fixation square (blue to yellow ; six changes randomly timed per sequence) (see video 1). 573 words (PF-W condition) appearing every five items (i.e., 1.2Hz).

Acquisition 593
During EEG recording, children were seated comfortably in a quiet room in the school 594 at a distance of 1 m from the computer screen. EEG signal was acquired at 1,024Hz by 595 using a 32-channel Biosemi Active II system (Biosemi, Amsterdam, Netherlands), with 596 electrodes including standard 10-20 system locations. The magnitude of the offset of all 597 electrodes, referenced to the common mode sense, was held below 50 mV. 598

Preprocessing 599
All EEG analyses were carried out by using Letswave 5.c 600 (http://nocions.webnode.com/letswave) and Matlab 2014 (The Mathworks) and followed 601 procedures validated in several studies using letter strings or faces and objects stimuli (see 602 e.g., Retter & Rossion, 2016). After band-pass filtering between 0.1 and 30 Hz, EEG data 603 were segmented to include 2 s before and after each sequence, resulting in 44 s segments. 604 Data files were then downsampled to 256 Hz to reduce file size and data processing time. 605 Artifact or noisy channels were replaced by using linear interpolation. All channels were re-606 referenced to the common average. EEG recordings were then segmented again from 607 stimulation onset until 39.996 s, corresponding exactly to 48 complete 1.2Hz cycles within 608 stimulation. This is the largest amount of complete cycles of 833 ms at the target frequency 609 (1.2 Hz) within the 40 s of stimulation period. 610

Frequency domain analysis 611
To reduce EEG activity that is not phase-locked to the stimulus, the three repetitions 612 of each condition were averaged in the time domain for each individual participant. Then, to 613 convert data from the time domain into the frequency domain, a Fast Fourier Transform 614 (FFT) was applied to these averaged time windows and normalized amplitude spectra were 615 extracted for all channels. Thanks to the long time-window (39.996 s), this procedure yields 616 EEG spectra with a high frequency resolution (1/39.996 s = 0.025Hz), increasing SNR 617 (Regan, 1989;Rossion, 2014) and allowing unambiguous identification of the response at 618 the exact frequencies of interest (i.e., 6Hz for the base stimulation rate and 1.2Hz and its 619 harmonics for the target stimulation rate). All of the responses of interest, and thus all the potential differences between conditions, can be concentrated in a discrete frequency band 621 around the stimulation frequency. This frequency band occupies a very small fraction of the 622 total EEG bandwidth. In contrast, biological noise is distributed throughout the EEG 623 spectrum, resulting in a SNR in the bandwidth of interest that can be very high (Regan, 1989;624 Rossion, 2014). To estimate SNR across the EEG spectrum, amplitude at each frequency of 625 interest (bin) was divided by the average amplitude of 20 surrounding bins (10 on each side) 626