{"paper_id":"06a4cd72-b7b8-451e-b47f-81e03a623d32","body_text":"1 \n \n 1 \n 2 \n 3 \n 4 \nTowards a Brain-Computer Interface (BCI) for Improving  5 \nPhonological Processing in Developmental Dyslexia: An Exploratory Study 6 \n 7 \n 8 \nXuanci Zheng*, João Araújo, Quentin Busson and Usha Goswami 9 \nCentre for Neuroscience in Education, University of Cambridge 10 \n 11 \nXZ - Investigation, Data Curation, Data Analysis, Visualisation, Writing – original draft 12 \nJA – Conceptualisation, Methodology, Writing – Review and editing 13 \nQB – Investigation, Data Curation, Writing – Review and editing 14 \nUG - Conceptualisation, Project Administration, Funding Acquisition, Resources, 15 \nSupervision, Writing – original draft 16 \n* Correspondence: 17 \nXuanci Zheng:  xz449@cam.ac.uk 18 \n 19 \nWord Count: 7465 20 \nFigure Count: 4 21 \nRunning title: BCI for Phonological Processing in Dyslexia 22 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n2 \n \nAbbreviations 23 \nACC  accuracy 24 \nADHD  attention deficit hyperactivity disorder 25 \nAM   amplitude modulation 26 \nART   amplitude rise time 27 \nASR  artifact subspace reconstruction 28 \nBCIs  brain-computer interfaces  29 \nCTRL  typically-developing control group 30 \nDYS  dyslexia group 31 \nEEG  electroencephalography 32 \nEI   efficiency index 33 \nEMG  electromyography 34 \nEOG  electrooculography 35 \nFDR  false discovery rate 36 \nHCPC  Health and Care Professions Council 37 \nICA   independent component analysis 38 \nPA   phonological awareness 39 \nPDE  Phonemic Decoding Efficiency 40 \nPSD  power spectral density 41 \nRAN  Rapid Automatized Naming 42 \nRT   reaction time 43 \nSpLD  Specific Learning Difficulties 44 \nSWE  Sight Word Efficiency 45 \nS.D.   standard deviation 46 \nTOWRE  Test of Word Reading Efficiency 47 \nTS   Temporal Sampling 48 \nWAIS  Wechsler Intelligence Scale for Adults 49 \nWRAT  Wide Range Achievement Test   50 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n3 \n \nAbstract 51 \nBrain-computer interfaces (BCIs) have immense potential regarding the provision of 52 \ntherapies for disorders of development, but to date have typically been created for non-53 \nlinguistic disorders such as ADHD (attention deficit hyperactivity disorder). Here we 54 \npresent a BCI that aims to improve linguistic phonological processing in developmental 55 \ndyslexia. Phonological ‘deficits’ are considered a core feature of dyslexia across 56 \nlanguages. A non-invasive EEG-BCI relying on auditory inputs and visual feedback was 57 \ndeveloped to optimise brain patterns related to phonology (speech-sound processing). 58 \nThese patterns were identified using Temporal Sampling (TS) theory, which proposes that 59 \nphonological difficulties in dyslexia are related to impaired auditory processing of 60 \namplitude envelope rise times and low-frequency speech envelope information <10 Hz. 61 \nThese impairments are thought to affect automatic features of speech processing from 62 \nbirth, impairing the development of a phonological system. Adults with and without a 63 \ndiagnosis of developmental dyslexia played the BCI for 16 sessions, and received pre- 64 \nand post-testing regarding phonological awareness and single word and nonword 65 \nreading skills. Significant associations between their BCI scores (a measure of BCI 66 \nlearning) and improvements in syllable stress discrimination, nonword reading and 67 \namplitude rise time discrimination were found. The data are interpreted with respect to TS 68 \ntheory. 69 \n 70 \nKey words: Brain–Computer Interface, Dyslexia, Phonological Processing, Temporal 71 \nSampling Theory, EEG 72 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n4 \n \n1. Introduction 73 \nTheories of developmental dyslexia attempt to provide a systematic causal 74 \nframework for understanding this specific learning difficulty (e.g., Magnocellular theory, 75 \nStein & Walsh, 1997; Visual Attention Span theory, Valdois et al., 2004; Sluggish 76 \nAttentional Shifting theory, Facoetti et al., 2010, Temporal Sampling [TS] theory, 77 \nGoswami, 2011). A focus on development is absolutely critical to identifying core factor/s 78 \nfor effective remediation, accordingly here the focus is on the phonological ‘core deficits’ 79 \nthat pre-date learning to read (Stanovich, 1998), and on TS theory. Theories focused on 80 \nthe visual system are not considered, as typically the theorised deficts can only be 81 \ndetected once reading instruction commences (see Goswami, 2022a, for a recent survey 82 \nof dyslexia theories). Regarding the phonological ‘core deficit’, studies in many 83 \nlanguages have demonstrated that a key developmental factor in the etiology of dyslexia 84 \nis phonological learning. Via the natural acquisition of spoken language, infants and 85 \nchildren implicitly learn a phonological system comprising the sounds and combinations 86 \nof sounds that are permissible in their language/s, long before reading instruction 87 \ncommences (Kuhl, 2004). In effect, their brains develop phonological representations of 88 \nthe sound structures of individual words, via automatic sensory-motor learning, and TS 89 \ntheory proposes that this automatic learning is impaired in dyslexia. The current BCI 90 \nfocuses on phonological learning at the level of speech rhythm patterns, the factor that 91 \ngoverns infant language acquisition across all languages studied to date (Mehler et al. 92 \n1988; Nazzi et al., 1998).   93 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n5 \n \nThe development of ‘phonological awareness’ (PA) in children is typically measured 94 \nby behavioural performance in PA tasks, simple oral tasks that explore a child’s ability to 95 \nconsciously detect and manipulate the component sounds in words at all linguistic levels 96 \n(speech rhythm and prosody, syllables, rhyme, phonemes, see Ziegler & Goswami, 97 \n2005). These phonological impairments persist into adulthood, although in consistent 98 \northographies like Italian, German or Spanish, in adulthood they are indexed by 99 \nsignificantly impaired speed in PA tasks (Landerl & Wimmer, 2000; Ziegler et al., 2010). 100 \nIn inconsistent orthographies like English, phonological difficulties in adulthood can be 101 \nindexed by impairments in both speed and accuracy in PA tasks (Snowling, 2000). PA 102 \nfollows a similar developmental sequence across languages, predicts reading acquisition 103 \nin all languages so far studied, and is impaired in children with dyslexia across languages 104 \n(Ziegler & Goswami, 2005). Training phonological skills, particularly in the pre-school and 105 \nearliest school years, can significantly mitigate the impact of a family risk for dyslexia 106 \n(Schneider et al., 2000). Accordingly, the current BCI for dyslexia was developed to 107 \nremediate the unconscious neural factors associated with inefficient phonological 108 \nprocessing. 109 \nAs phonological learning in infants begins with speech rhythm, recent infant EEG 110 \n(Electroencephalography) studies of neural speech processing also informed the design 111 \nof the BCI. When infants listen to sung infant-directed speech, which is highly rhythmic, 112 \ncortical tracking of low-frequency speech envelopes appears to come online first 113 \n(measurable from 2 months of age), notably in the delta and theta electrophysiological 114 \nbands (0.5 – 4 Hz, 4 – 8 Hz, see Attaheri et al., 2022; Ni Choisdealbha et al., 2023). This 115 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n6 \n \nlow-frequency cortical tracking also underpins the learning of phonetic information (Di 116 \nLiberto et al., 2023), which begins to emerge around 7 months. Individual differences in 117 \nboth delta-band cortical tracking at 11 months and in the ratio of theta-delta PSD (power 118 \nspectral density) predict individual differences in language outcomes at age 2 years 119 \n(measured by vocabulary tests and a nonword repetition task, Attaheri et al., 2024). More 120 \naccurate delta band cortical tracking and a lower theta-delta ratio predicted better 121 \nlanguage outcomes. These infant studies were informed by TS theory, an auditory theory 122 \nof dyslexia, which also informed the current study (Goswami, 2011, 2015, 2022b).  123 \nWith respect to dyslexia, TS theory proposes that sensory/neural processing 124 \ndifferences regarding speech prosody (speech rhythm patterns) lead affected children to 125 \ndevelop atypical phonological representations of spoken language, from infancy onwards 126 \n(Goswami, 2022a). Neurally, adult studies suggest that speech is encoded by 127 \nneuroelectric oscillations (rhythmic changes in electrical brain potentials in large cell 128 \nnetworks) which respond to different temporal levels of speech information (such as 129 \nphrases, syllables and phonemes, Giraud & Poeppel, 2012; Gross et al., 2013). TS 130 \ntheory suggests that in developmental dyslexia, encoding of low-frequency envelope 131 \ninformation <10 Hz (delta and theta band information) is impaired, in part because of 132 \npoorer auditory discrimination of amplitude ‘rise times’ in the speech envelope. ‘Rise 133 \ntimes’ in amplitude (the rates of change between sound onset and sound peak in a given 134 \namplitude modulation, AM) provide sensory landmarks that automatically trigger brain 135 \nrhythms and speech rhythms into temporal alignment, via phase-resetting ongoing neural 136 \nactivity (Doelling et al., 2014). This phase-resetting process is known to be impaired in 137 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n7 \n \ndyslexia (Lizarazu et al., 2021). The amplitude envelope is the slow-varying energy 138 \ncontour of speech that determines the perception of speech rhythm (Greenberg, 2006), 139 \nand it contains a range of AM patterns at different temporal rates which broadly match 140 \nEEG rates such as delta, theta and beta/low gamma. Further, speech modelling studies 141 \nof infant- and child-directed speech show that the phase relations between these different 142 \nAM rates provide systematic statistical cues to phonological units such as stressed vs 143 \nunstressed syllables, syllables, and onset-rimes (‘acoustic-emergent phonology’, Leong 144 \n& Goswami, 2015; Leong et al., 2017). Accordingly, a nascent phonological system can 145 \nbe extracted from the speech signal via the automatic alignment of neuroelectric 146 \noscillations to the AM information in speech via efficient phase-resetting driven by 147 \namplitude rise time (ART) discrimination.  148 \nChildren with dyslexia in a range of languages exhibit impaired ART discrimination 149 \ncompared to chronological age matched-controls (English, Spanish, French, Finnish, 150 \nChinese, Hungarian, and Dutch; Goswami, 2015, for review). Children with dyslexia 151 \nlearning English, Spanish and French also show impaired neural encoding of low-152 \nfrequency speech envelope information in the delta and theta neurophysiological bands 153 \nduring natural speech listening (DiLiberto et al., 2018; Molinaro et al., 2016; Destoky et 154 \nal., 2020; other languages not yet tested). A BCI for dyslexia could therefore target neural 155 \nencoding directly, for example via improving phase locking values (see Arias, Molinaro & 156 \nLizarazu, 2021). However, TS-driven studies have shown that one neural marker of 157 \nimpaired phonological processing appears to be the theta-delta oscillatory ratio during 158 \nnatural speech listening (Attaheri et al., 2024; Araújo et al., 2024). During continuous 159 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n8 \n \nspeech listening, English-speaking children with dyslexia show a higher theta-delta ratio 160 \nthan control children, which is significantly related to their offline performance in PA tasks 161 \n(a higher ratio is associated with worse performance, Araújo et al., 2024). Further, 162 \nEnglish-learning infants aged 4 – 11 months with a higher theta-delta ratio during 163 \ncontinuous speech listening go on to exhibit poorer language skills at 24 months (poorer 164 \nvocabulary and nonword repetition, see Attaheri et al., 2024). The recent developmental 165 \nresearch base thus suggests that the theta-delta ratio during natural speech listening 166 \ncould also be an effective target for a BCI for dyslexia. 167 \nThese TS-driven developmental data informed the current BCI. The aim of the BCI 168 \nwas to change the ratio of the neural oscillations that (by TS theory) underpin statistical 169 \nlearning of the AM hierarchy, thereby ameliorating the ‘phonological deficit’ in dyslexia. A 170 \nnon-invasive BCI targeting the self-regulation of low-frequency (delta and theta) neural 171 \noscillations during natural speech listening was developed by the second author as part 172 \nof his PhD and piloted with 15 adult participants, 7 of whom had a statement of dyslexia. 173 \nAraújo (2023) designed an engaging interface based on a space ship rocketing up into 174 \nspace, aimed at teaching learners of the BCI how to self-regulate their own theta-delta 175 \nratio by controlling the space ship’s position using their brains. A closed-loop operant 176 \nlearning BCI was created, in which learners aimed to make the space ship ascend as far 177 \nas possible on the gaming window in each of 16 BCI sessions (described in detail in 178 \nAraújo et al., 2023). Participants received a stronger visual reinforcement (the screen 179 \nglowed greener) the higher they made the spaceship go. No visual reinforcement was 180 \ngiven if the spaceship’s position remained below a threshold line located across the 181 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n9 \n \nmiddle of the gaming window. Listening to the audio signal of a story as input, the 182 \nparticipant was then encouraged to try out cognitive strategies focused on auditory 183 \nprocessing to modulate their oscillatory patterns that controlled the spaceship. The 184 \nspaceship’s position was estimated via real-time classification of time-series EEG data 185 \nusing a pre-trained signal processing and machine learning model (described below). 186 \nThis feedforward model shows minimal computational overhead, allowing for smooth 187 \nonline control of the BCI with minimal lags.  188 \nIn the original paradigm, successful BCI learning was indexed by whether the 189 \nspaceship position distribution of session 1 had a significantly lower mean than session 190 \n16 (using a t-test, see Araújo, 2023). The value of the t statistic became the participant’s 191 \n‘BCI Score’, the magnitude of which reflected the degree to which the participant had 192 \nreduced their theta-delta ratio. Inspection of the BCI scores suggested that 12 of the 15 193 \nparticipants had learned the BCI successfully (2 controls and 1 dyslexic did not learn). 194 \nThe baseline-normalized band frequency magnitude across the learners’ delta and theta 195 \nrhythms was then used to compare their distributions in session 1 with distributions from 196 \nsession 16. The data showed that the BCI helped participants to reduce their theta-delta 197 \nratio by significantly increasing neural signal magnitude for the slower delta rhythm and 198 \nsignificantly decreasing it for the faster theta rhythm. Further, individual BCI scores were 199 \nassociated with significant improvement in the speed of syllable stress discrimination 200 \njudgements (r= 0.59, p< .05) and showed a trend in improvement for single word reading 201 \nas measured by the TOWRE (Test of Word Reading Efficiency, Torgesen, Wagner & 202 \nRashotte, 1999; r= 0.48, p= .07). The cortical dynamics targeted by the BCI should (by 203 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n10 \n \nTS theory) only improve phonology and reading. Adult participants also received a test of 204 \narithmetic reasoning (WRAT, Wide Range Achievement Test, Snelbaker et al., 2001) 205 \nduring the study, and as expected BCI scores were not associated with changes in 206 \narithmetical reasoning from pre-test to post-test (Araújo, 2023).      207 \nThe original paradigm was developed during the Pandemic, therefore the BCI was 208 \nbased on a g-tec hardware set-up which was not suitable for taking to schools and using 209 \nwith children. Accordingly, as a further pilot, in the current study the second author 210 \nadapted the closed-loop operant learning system to work with mobile EEG headcaps 211 \nspecifically the CGX Quick-20m wireless headset. This system employs dry electrodes 212 \nrecorded positioned at 19 scalp locations following the International 10-20 system (P1, 213 \nFP2, F3, F4, Fz, F7, F8, C3, C4, T3, T4, T5, T6, P3, P4, Pz, O1, and O2), with A1 and A2 214 \nserving as linked-ear references. Signals were digitized at 24-bit resolution and sampled 215 \nat 500 Hz. The portability and ease of setup of this dry electrode system make it 216 \nparticularly well suited for use in schools with children, where the application of traditional 217 \nEEG systems would be impractical. A new group of adults with and without dyslexia were 218 \nrecruited by the first author, and received a similar protocol to that used in Araújo (2023), 219 \nwhich is described fully below. Participants were pre- and post-tested on a range of 220 \nphonological, reading and control tasks (detailed below) before and after 16 gaming 221 \nsessions with the BCI. The hypothesis was that learning the BCI would improve their 222 \nneural theta-delta ratios during natural language listening, and that this improvement 223 \n(indexed by their BCI scores based on spaceship position, the measure of real-time 224 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n11 \n \ntheta-delta learning) would be significantly associated with improvements in reading and 225 \nphonological processing. 226 \n2. Materials and Methods 227 \n2.1. Participants 228 \nTwelve control (typically-developing) adults (mean age of 24.21 ± 6.31 years; 7 female and 229 \n5 male) and twenty adults with a current or childhood diagnosis of dyslexia (mean age of 230 \n23.65 ± 5.52 years; 16 female and 4 male) participated in the study. All participants were 231 \nnative English speakers with normal or corrected-to-normal vision and no reported hearing 232 \nimpairments. Typically developing participants were included if their efficiency index (EI) 233 \non the Test of Word Reading Efficiency (TOWRE, Torgesen et al., 1999) exceeded 95 (the 234 \nEI mean is 100, S.D. 15, see  Section 2.2 for further detail) . Participants in the dyslexia 235 \ngroup were included only if they could provide formal documentation of a dyslexia 236 \ndiagnosis from a qualified professional , such as  a Health and Care Professions Council 237 \n(HCPC) registered assessor, the Accessibility and Disability Resource Centre at University 238 \nof Cambridge, or a specialist teacher with a current Specific Learning Difficulties (SpLD) 239 \nAssessment Practicing Certificate. All participants provided informed consent for the study 240 \nin accordance with the Declaration of Helsinki, and the study was reviewed by the 241 \nPsychology Research Ethics Committee of the University of Cambridge  who gave it a 242 \nfavourable opinion. 243 \n2.2. Experimental Protocol 244 \nThe experimental protocol spanned ten days and included two assessment sessions, one 245 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n12 \n \nat the beginning (Day 1) and one at the end (Day 10), to evaluate participants ’ cognitive 246 \nand linguistic profiles. These pre - and post -intervention sessions comprised tasks 247 \nmeasuring phonological and reading skills, acoustic processing, and skills that were not 248 \nexpected to be improved by the BCI (non-verbal reasoning and arithmetic ability). The eight 249 \ndays in between were dedicated to the BCI training intervention, which will be described in 250 \nSection 2.3. The experimental protocol can be found in Figure 2.1(a). The measures used 251 \nin the pre- and post-test sessions are described below. 252 \n1) Phonology and Reading Measures 253 \nPhonology and reading skill s were measured using an experimental  phoneme deletion 254 \ntask, an experimental syllable stress recognition task, an experimental Rapid Automatized 255 \nNaming (RAN), and the standardized word and nonword item lists from the TOWRE.  256 \n 257 \nThe phoneme deletion task was adapted from McDougall et al. (1994). This task required 258 \nparticipants to listen to a spoken item and delete a target phoneme  (e.g., “BICE” without 259 \nthe /b/ becomes “ICE”). The target consonant phoneme appeared in initial, medial, or final 260 \npositions, and all correct responses formed real English words. The task comprised 18 261 \ntrials (3 practice and 15 experimental items), presented through sound files recorded by a 262 \nfemale speaker of standard Southern British English. The same test was administered both 263 \nbefore and after the BCI training. 264 \n 265 \nIn the syllable stress discrimination task, participants heard pairs of different four-syllable 266 \nwords and made a same-different judgement regarding whether the pair of words shared 267 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n13 \n \nthe same stress pattern (e.g., difficulty - voluntary = yes). The items were from Leong et al. 268 \n(2011), and comprised 80 randomized pairs, some of which were deliberately mis-stressed 269 \n(e.g., di-FFI-cul-ty – VO-lun-ta-ry = NO). This task was administered before and after the 270 \nBCI training. This task was a variation to the protocol used in Araújo (2023), in which pairs 271 \nof identical words were used as stimuli (also drawn from Leong et al., 2011). Here we used 272 \ncomparisons between different words to increase  task difficulty, with the aim of reducing 273 \nthe ceiling effect observed with original design.  274 \n 275 \nIn the RAN task, participants named pictures of familiar items (e.g., cup, book, tree) aloud 276 \nas fast as possible. Four pages of pictures (two pages of target words with low phonological 277 \nneighborhoods and two with high phonological neighborhoods) were administered at both 278 \nthe pre- and post-sessions. Both the time taken to complete the task and accuracy were 279 \nrecorded. 280 \n 281 \nReading was assessed using the TOWRE. The participant received a list of single words 282 \nto read aloud in 45 seconds, and a list of nonword items to read aloud in 45 seconds. 283 \nVersion A of this task was given at the pre-training stage while version B was used at the 284 \npost-training stage. The highest available age bracket for calculating scaled scores ranges 285 \nfrom 17 years 0 months to 24 years 11 months. Since some participants in the study were 286 \nolder than this range , raw scores were used for statistical analyses. However, for 287 \nparticipant recruitment, an EI was calculated using scaled scores from 17 -24 age group, 288 \nas all participants were over 18 years of age. These scaled scores were used for typically 289 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n14 \n \ndeveloping group screening purposes and were not included in further analysis 290 \n2) Non-verbal I.Q. 291 \nAll participants completed the matrix reasoning subtest of the Wechsler Intelligence Scale 292 \nfor Adults (WAIS; Wechsler, 1955), a widely used measure of non-verbal intelligence. This 293 \nis a nonverbal reasoning task in which individuals are asked to identify patterns in designs. 294 \nThis pattern recognition task was administered at both pre- and post-intervention sessions. 295 \n3) Arithmetic Task 296 \nParticipants completed the standardized arithmetic subscale of the Wide Range 297 \nAchievement Test (WRAT) (Snelbaker et al., 2001), which includes basic math problems 298 \nrequiring written responses. Version TAN was administered before the intervention, and 299 \nVersion BLUE after.  This task was included to test whether the BCI would affect any 300 \nacademic skill, rather than specifically affect word reading. 301 \n4) Acoustic Threshold for Amplitude Rise Time (ART): 1 Rise Task 302 \nParticipants also completed a sine tone rise time task (labelled the 1 Rise task in our prior 303 \npublications with children, e.g. Flanagan et al., 202 4) to assess sensitivity to ART. Each 304 \ntrial presented three 500 -Hz tones, with one (the target) having a slower onset rise time 305 \nthan the two standard tones. Using an AXB format displayed as cartoon dinosaurs, 306 \nparticipants were asked to identify which of the first or third sounds differed from the middle 307 \ntone. The task used 39 stimuli with rise times ranging from 300 ms to 15 ms in 7.3 ms 308 \nsteps. Verbal instructions and five practice trials with feedback were provided before the 309 \nmain task. Participants performed the same test both before and after the BCI training. 310 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n15 \n \n2.3. BCI Training 311 \nBetween the pre- and post-intervention assessment sessions, participants completed eight 312 \ndays of BCI training. The protocol was structured to span ten days in total, allowing at most 313 \ntwo rest days (typically the weekend) to accommodate participant schedules. Most 314 \nparticipants completed the intervention over two consecutive weeks. Each daily session 315 \nincluded two BCI runs. Prior to each session, an EEG cap was fitted, and electrode 316 \nimpedances were checked and maintained below 100 Ω. Participants were encouraged to 317 \nlisten carefully to the words in the story and try to identify listening strategies to keep the 318 \nspaceship ascending on the screen. However, no explicit suggestions regarding how to 319 \nachieve this goal were given. 320 \n 321 \nEach BCI run consisted of two distinct phases: a baseline stage and a BCI control stage.  322 \nThe interface of the BCI and the timeline of the experiment are depicted in Figure 2.1 (a). 323 \nIn the baseline stage (lasting  four minutes), participants viewed a vertically moving 324 \nspaceship displayed on the screen. During this phase, they had no neural control over the 325 \nspaceship’s position. Instead, the spaceship moved randomly, with positions sampled from 326 \na Gaussian distribution (mean = 0.5, SD = 0.15) and mapped onto a vertical scale ranging 327 \nfrom 0 (top) to 1 (bottom). The position updated at a refresh rate of 4 Hz. This random 328 \nmovement served two purposes: it provided data to estimate individualized decoder 329 \nthresholds based on each participant’s typical neural activity, and it avoided  any neural 330 \nentrainment that might occur with fixed or repetitive visual patterns. A semi -transparent 331 \nwhite overlay and the message “Good luck! Please wait...” were displayed to indicate the 332 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n16 \n \nsystem was in passive mode. No auditory input was presented during this stage. 333 \n 334 \nAfter the baseline, participants entered the BCI stage, which lasted for the duration of a 335 \nten-minute auditory story. At this point, the semi -transparent overlay and the baseline 336 \nmessage were removed, and participants began listening to a narrated version of Winnie-337 \nthe-Pooh through headphones. Simultaneously, they gained neural control over the on -338 \nscreen spaceship, which moved vertically based on real-time EEG activity. Specifically, the 339 \nspaceship’s position was determined by the log -transformed theta/delta power ratio 340 \nmeasured from centrally located electrodes (F3, F4, C3, Cz, C4, P3, P4) . To personalize 341 \ncontrol sensitivity, decoder boundaries were set using each participant’s baseline 342 \ndistribution: the median of their log-transformed theta/delta ratio defined the vertical midline 343 \nof the screen, while the upper and lower boundaries were set at three standard deviations 344 \nabove and below the median. Participants were instructed to raise the spaceship as high 345 \nas possible and to keep it stable during the story. In terms of neural dynamics, this 346 \ncorresponded to decreasing the theta/delta ratio and reducing its variance. 347 \n 348 \nThe BCI was designed to provide feedback based on neural patterns previously associated 349 \nwith continuous speech processing and phonological awareness in children with and 350 \nwithout dyslexia. The same decoder was used across participants, with calibration derived 351 \nfrom each session's baseline. This allowed for continuous control based on dyslexia -352 \nrelevant neural dynamics, specifically those shown to relate to phonological awareness in 353 \nprevious studies. 354 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n17 \n \n 355 \nThe neurofeedback display was designed to be intuitive and engaging. To enhance 356 \nmotivation and user experience, the traditional cursor was replaced with a spaceship 357 \ngraphic, and the background featured a subtle space-themed design. A visual midline was 358 \ndrawn on the screen to indicate the target region for upward control. In addition, a five -359 \ntimestep history trace was implemented, appearing as a contrail behind the spaceship, 360 \nallowing participants to visually track their recent performance. To enhance participant 361 \nmotivation, a cumulative score related to the real -time theta/delta ratio was presented in 362 \nthe top-left corner of the screen. The score was updated at each refresh of the spaceship 363 \nposition. The instantaneous score was derived from the log transform of the ratio  364 \nstandardized to each participant's baseline, which also determined the spaceship's position. 365 \nIt was multiplied by 10 when the spaceship occupied the lower half of the screen and by 366 \n20 when it occupied the upper half to reinforce positive feedback.  Each instantaneous 367 \nscore was  continually added to the total score displayed. The story audio was not 368 \ninfluenced by task performance and remained constant throughout the session. 369 \n 370 \nTo further reinforce successful BCI control, a visual reward system was implemented. A 371 \nsemi-transparent green overlay appeared on the screen, with its intensity varying 372 \naccording to the spaceship’s vertical position. The screen was scaled from 0 (top) to 1 373 \n(bottom), and the green glow was calculated using the formula: 374 \n𝐺𝑙𝑜𝑤 𝑖𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦 = 𝑚𝑎𝑥(𝑚𝑖𝑛(255 − 510𝑥𝑡 ,255),0) 375 \nWhere 𝑥𝑡 is the scaled spaceship position. This meant that the glow reached full intensity 376 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n18 \n \nwhen the spaceship was at the top of the screen, gradually faded when it approached the 377 \nmidline, and disappeared entirely when below the midline. This continuous visual 378 \nreinforcement served as an intuitive feedback signal to encourage better control of the 379 \nspaceship.  380 \n 381 \nTo prepare the signal for use in the BCI decoder, the theta/delta ratio was log-transformed. 382 \nThis transformation was necessary because the raw ratio data exhibited a skewed, non -383 \nGaussian distribution across and within participants, along with a wide and variable 384 \ndynamic range. Applying a log transformation reduced skewness and the influence of 385 \noutliers by compressing the range of values. Crucially, because the log function is 386 \nmonotonically increasing, it preserved the relative ordering of values in the original signal. 387 \nThis ensured that the neurofeedback interface remained stable, interpretable, and 388 \nsensitive to individual neural dynamics. The BCI neural feedback is shown in Figure 2.1(b).  389 \n 390 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n19 \n \nFigure 2.1 Panel (A) shows the whole Experimental Protocol, including the time line of 391 \neach BCI session and the spaceship interface. Panel (B) depicts the decoder logic behind 392 \nthe neural feedback. 393 \n2.4. EEG preprocessing 394 \nThe EEG signal was acquired in real time using a CGX wireless headset and continuously 395 \nprocessed throughout the neurofeedback task. Raw data were initially converted from a 396 \n24-bit compressed format to microvolts and streamed at a sampling rate of 500 Hz.  EEG 397 \npreprocessing followed two distinct strategies: real-time processing  for neurofeedback 398 \ndelivery, and offline preprocessing for subsequent data analysis.  399 \n2.4.1 Real-Time Processing 400 \nDuring the BCI intervention, real -time processing prioritized low computational demand 401 \nand effective noise suppression to ensure smooth feedback. A zero -phase, fourth-order 402 \nButterworth bandpass filter (0.5–10 Hz) was applied to selected central channels to reduce 403 \nnoise and isolate relevant neural signals. A 3 -second sliding window was used for 404 \ncontinuous feature extraction. During pilot testing, we observed that the spaceship position 405 \ncould be influenced by abnormal eye movements. To manage transient artifacts (e.g., eye 406 \nmovement, muscle activity or movement), an online threshold -based artifact rejection 407 \nmethod was employed. Samples exceeding a channel -specific threshold —determined 408 \nfrom the 95th percentile of the participant’s baseline amplitude distribution—were replaced 409 \nwith random clean segments drawn from the individual’s baseline data, preserving inter -410 \nchannel relationships . This procedure ensured that the spaceship position was less 411 \naffected by abnormal EEG segments during BCI training . The resulting preprocessed 412 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n20 \n \nsignal was then used for real-time spectral analysis and neurofeedback computation. 413 \n2.4.2. Offline Preprocessing 414 \nDespite the use of an online artifact rejection method, the influence of spaceship position 415 \ncould not be completely eliminated during online processing. To address this limitation, a 416 \nmore comprehensive offline preprocessing pipeline was applied to obtain cleaner data.  417 \nPower line noise was removed using a notch filter, and the data were bandpass -filtered 418 \nbetween 0.5 and 48 Hz using an 8th -order Butterworth filter with zero -phase filtering to 419 \navoid phase distortion. The signal was then downsampled to 250 Hz to reduce 420 \ncomputational load. Given that dry EEG systems tend to produce noisier recordings than 421 \ngel-based systems, Artifact Subspace Reconstruction (ASR) was used to suppress high -422 \namplitude transients such as muscle bursts and cable movements. Channels were marked 423 \nas noisy if their voltages exceeded ±100 μV or if their power spectra deviated more than 3 424 \nstandard deviations from the mean. The EEG data were then re-referenced to the average 425 \nof all channels. Independent Component Analysis (ICA) was performed, and components 426 \nassociated with ocular, muscular, or blink artifacts (e.g., EOG, EMG) were identified and 427 \nremoved. The cleaned data were segmented into consecutive, non-overlapping 3-second 428 \nepochs. Finally, previously identified noisy channels were interpolated using a spline 429 \ninterpolation method. 430 \n2.5. Statistical Analysis 431 \nStatistical analyses were conducted to evaluate the effectiveness of the BCI 432 \nneurofeedback intervention and its relationship with behavioral performance. First, t o 433 \nassess whether participants exhibited neurophysiological changes across the training 434 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n21 \n \nperiod, we tested whether there was a significant change in participants’ neural responses 435 \nover time. Specifically, we compared the distribution of the log-transformed theta/delta ratio 436 \nbetween the first session and the final session. This was done separately for both (i) the 437 \nreal-time ratio used for intervention feedback (derived from the online preprocessing 438 \npipeline), and (ii) the ratio extracted from the offline preprocessed data. T-tests were used 439 \nto assess differences in the ratio values across sessions for each participant. The resulting 440 \nt-statistic served as a summary measure of change in BCI performance over time.  441 \nParticipants who showed a significantly lower ratio by the final session were considered to 442 \nhave demonstrated learning. 443 \nSecond, to evaluate whether participants improved on relevant behavioral skills following 444 \nthe BCI intervention, pre- and post -intervention behavioral scores (e.g. phonological 445 \nawareness, reading ability) were compared using paired -sample t -tests. A significant 446 \nincrease in post-test scores was interpreted as evidence of behavioral improvement.  447 \nThird, we investigated whether changes in neural measures were associated with changes 448 \nin behavioral performance. Pearson correlations were conducted to assess the relationship 449 \nbetween the t-statistics derived from the BCI measures (both real-time and offline) and the 450 \ndifferences in behavioral performance between pre -test and post -test. For all analyses 451 \ndescribed above, p -values were corrected for multiple comparisons using the false 452 \ndiscovery rate (FDR), and statistical significance was defined as corrected p < 0.05. 453 \n3. Results 454 \n3.1. Neurophysiological Changes Across Sessions 455 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n22 \n \nTo evaluate whether participants exhibited neurophysiological changes across the BCI 456 \ntraining period, we compared the log-transformed theta/delta power ratios between the first 457 \nand final sessions. This analysis was conducted twice, separately for data processed using 458 \nthe real-time processing pipeline (following Araújo, 2023) and for data preprocessed offline 459 \n(analysis added here). 460 \nFor each participant, a t-test was conducted to assess whether the theta/delta ratio 461 \nsignificantly decreased from the first to the last session. The resulting t-statistic served as 462 \nan individual -level summary of neural change and was used as the participant’s BCI 463 \nlearning score (hereafter BCI score) in subsequent analyses. Participants were considered 464 \nto have learned the BCI control task successfully if their t-statistic was greater than zero 465 \nand the corresponding p-value was less than 0.05. This criterion indicates a statistically 466 \nsignificant reduction in the theta/delta ratio across BCI sessions. 467 \nBased on the real-time EEG data, 9 out of 12 participants in the control group and 13 out 468 \nof 20 participants in the dyslexia group met this learning criterion. Following offline 469 \npreprocessing, which involved the removal of ocular, muscular, and movement -related 470 \nartifacts (e.g., EMG and EOG signals), the resulting theta/delta ratios exhibited generally 471 \nlower t-statistics. Under this more stringent preprocessing, 6 of 12 participants in the 472 \ncontrol group and 10 of 20 in the dyslexia group showed significant improvement according 473 \nto the same criterion. 474 \nFigure 3.1 presents a visual summary of the t-statistics for each participant under both 475 \nprocessing pipelines. The figure includes two subplots: the left subplot (in blue) represents 476 \nthe control group, and the right subplot (in red) represents the dyslexia group. In each 477 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n23 \n \nsubplot, individual participant data are shown using paired line plots that connect the t -478 \nstatistics obtained from the real -time and offline pipelines, illustrating the direction and 479 \nmagnitude of change after artifact correction. As the offline preprocessing pipeline is more 480 \nstringent, it would be expected that the BCI performance scores are lower for the offline 481 \npreprocessing, which was the case for both groups. Overall, a higher T-score indicates 482 \nbetter learning of the BCI. Overlaid on each set of lines, boxplots depict the overall 483 \ndistribution of t-statistics for each preprocessing method within each group. 484 \n 485 \nFigure 3.1. The t-statistics used as a measure of BCI performance. Data are shown  for 486 \neach participant, under real-time feedback versus offline preprocessing.  487 \n3.2. Behavioral Improvements Following BCI Intervention 488 \nTo investigate whether participants showed behavioral improvement over the course of 489 \ntraining, we examined performance across the acoustic, cognitive and linguistic tasks. As 490 \nwill be recalled, two measures were not expected a priori to show improvement following 491 \nBCI training, nonverbal IQ (WAIS Matrices) and Arithmetic. 492 \n3.2.1. Pre-Intervention Scores 493 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n24 \n \nAs a first step, we compared pre -intervention scores between the control and dyslexia 494 \ngroups to assess baseline differences in task performance (Table 3.1). As expected, there 495 \nwere no significant group differences in non-verbal cognitive tasks such as Arithmetic and 496 \nMatrices Reasoning, indicating that both groups were matched on non-reading academic 497 \nperformance and general reasoning ability. In contrast, significant group differences were 498 \nobserved in pre-test measures of reading and phonological processing. These included 499 \nsyllable stress discrimination accuracy, RAN completion time, and both word and non-word 500 \nraw scores on  the TOWRE. These results are consistent with the known language and 501 \nliteracy difficulties associated with dyslexia. An exception was the phoneme deletion task, 502 \nwhich showed no significant group difference . This was likely due to a ceiling effect. The 503 \ntask may have been too easy for both groups, limiting its sensitivity to detect individual 504 \ndifferences. Contrary to our prior adult studies, no significant group difference in sensitivity 505 \nto ART was observed, although the dyslexic group showed worse performance. 506 \n 507 \nTable 3.1. Group performance on pre-test measures, 2-tailed t-tests 508 \nBehavioral Test DYS Mean (S.D.) CTRL Mean \n(S.D.) \nt-score p-value \nWRAT Arithmetic (scaled score) 101.2 (14.0) 106.1 (15.7) 0.905  0.266  \nWAIS Matrices (T-Score) 56.6 (5.9) 57.3 (7.6) 0.307  0.423  \nPhoneme Deletion (n correct) 13.4 (2.2) 13.8 (1.7) 0.583  0.353  \nSyllable Stress Recognition (ACC) 62.6 (10.7) 76.6 (16.9) 2.868  0.009  \nSyllable Stress Recognition (RT, s) 3.4 (1.4) 2.8 (0.9) -1.457  0.145  \nRAN (mean time, s) 32.0 (4.9) 27.4 (3.2) -2.880  0.009  \nTOWRE SWE (raw score) 81.5 (6.8) 99.3 (10.4) 5.885  <0.001 \nTOWRE PDE (raw score) 45.8 (7.3) 59.9 (7.1) 5.348  <0.001 \n1 Rise Task (time threshold, ms) 136.5 (93.2) 92.2 (75.4) -1.391  0.145  \n 509 \n3.2.2. Pre-Intervention versus Post-BCI Behavioural Scores  510 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n25 \n \nWe next compared pre - and post -intervention scores within each group to examine 511 \nbehavioral changes over time (Table 3.2). In the control group, significant improvements 512 \nwere observed in Arithmetic performance, syllable stress accuracy, RAN completion time, 513 \nand TOWRE non-word reading. In the dyslexia group, significant improvements were found 514 \nin Matrices Reasoning, phoneme deletion accuracy, syllable stress discrimination accuracy, 515 \nRAN completion time, and both TOWRE word and non-word reading. With the exception 516 \nof the improvement in Arithmetic (controls only) and Matrices Reasoning (dyslexics only), 517 \nthese improvements were in line with our a priori expectations. However, both groups in 518 \nthese analyses included participants who did not meet criterion for learning the BCI. 519 \n 520 \nTable 3.2. Pre- versus post-intervention scores for the behavioral tasks, 2-tailed t-tests 521 \nBehavioral Test Group Pre-test \nmean (S.D.) \nPost-test \nmean (S.D.) t-score p-value \nWRAT Arithmetic (scaled score) \nCTRL 106.1 (15.7) 110.4 (17.0) -2.738  0.051  \nDYS 101.2 (14.0) 102.8 (14.6) -1.281  0.108  \nWAIS Matrices (T-Score) \nCTRL 57.3 (7.6) 59.2 (7.1) -1.675  0.164  \nDYS 56.6 (5.9) 59.5 (4.9) -4.681  <0.001 \nPhoneme Deletion (n correct) \nCTRL 13.8 (1.7) 14.0 (1.3) -0.561  0.521  \nDYS 13.4 (2.2) 13.9 (2.1) -2.127  0.031  \nSyllable Stress Recognition (ACC) \nCTRL 76.6 (16.9) 80.7 (18.2) -2.865  0.051  \nDYS 62.6 (10.7) 67.4 (12.6) -2.623  0.013  \nSyllable Stress Recognition (RT, s) \nCTRL 2.8 (0.9) 2.4 (0.9) 1.670  0.164  \nDYS 3.4 (1.4) 3.4 (1.9) 0.062  0.423  \nRAN (mean time, s) \nCTRL 27.4 (3.2) 25.3 (3.2) 4.891  0.004  \nDYS 32.0 (4.9) 28.0 (4.3) 6.631  <0.001 \nTOWRE SWE (raw score) \nCTRL 99.3 (10.4) 98.6 (8.9) 0.581  0.521  \nDYS 81.5 (6.8) 87.2 (7.3) -4.467  <0.001 \nTOWRE PDE (raw score) \nCTRL 59.9 (7.1) 61.7 (5.0) -2.399  0.071  \nDYS 45.8 (7.3) 48.4 (7.4) -2.931  0.009  \n1 Rise Task (time threshold, ms) \nCTRL 92.2 (75.4) 75.2 (57.2) 1.073  0.350  \nDYS 136.5 (93.2) 119.3 (93.9) 1.440  0.095  \n 522 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n26 \n \n3.2.3 Pre-Intervention versus Post-BCI Scores, BCI Learners only 523 \nWe next assessed the same relationships for the BCI learners only  (Table 3.3). In the 524 \ncontrol group, significant improvements were observed in RAN completion time  only, for 525 \nboth real-time and offline preprocessing. In the dyslexia group, significant improvements 526 \nwere found in Matrices reasoning, RAN completion time , syllable stress discrimination 527 \naccuracy and TOWRE non-word reading, for both real-time and offline preprocessing. With 528 \nthe exception of the improvement in Matrices reasoning (dyslexics only), these 529 \nimprovements were in line with our a priori expectations based on TS theory. 530 \n 531 \nTable 3.3. Pre- and post-intervention behavioral improvement in BCI learners only by group, 532 \n2-tailed t-tests 533 \nPreprocessing Behavioral Test Group \nPre-test \nmean (S.D.) \nPost-test \nmean (S.D.) \nt-score p-value \nrealtime \n(CTRL: n=9, \nDYS: n=13) \nWRAT Arithmetic \n(scaled score) \nCTRL 108.1 (15.7) 112.0 (15.8) -2.135  0.110  \nDYS 101.3 (11.4) 101.9 (13.5) -0.409  0.431  \nWAIS Matrices \n(T-Score) \nCTRL 56.7 (7.5) 59.1 (8.0) -2.137  0.110  \nDYS 56.5 (4.1) 59.9 (3.1) -4.137  0.002  \nPhoneme Deletion \n(n correct) \nCTRL 13.6 (1.9) 13.9 (1.5) -0.894  0.397  \nDYS 14.1 (1.4) 14.5 (0.7) -1.585  0.116  \nSyllable Stress \nRecognition \n(ACC) \nCTRL 77.1 (15.6) 80.8 (18.4) -2.025  0.110  \nDYS 64.7 (12.1) 70.4 (12.9) -2.312  0.039  \nSyllable Stress \nRecognition \n(RT, s) \nCTRL 2.7 (0.9) 2.2 (0.6) 1.986  0.110  \nDYS 3.3 (1.3) 3.3 (1.5) -0.121  0.503  \nRAN (mean time, s) \nCTRL 28.0 (3.3) 25.7 (3.6) 6.984  0.001  \nDYS 33.1 (5.3) 28.4 (4.7) 6.047  <0.001 \nTOWRE SWE (raw score) \nCTRL 97.3 (11.4) 97.3 (9.8) 0.000  0.889  \nDYS 80.5 (7.2) 87.2 (8.6) -4.624  0.001  \nTOWRE PDE (raw score) \nCTRL 59.0 (7.7) 61.2 (5.3) -2.443  0.110  \nDYS 46.4 (7.5) 50.0 (6.3) -3.065  0.012  \n1 Rise Task \n(time threshold, ms) \nCTRL 75.8 (57.2) 55.7 (27.0) 1.072  0.360  \nDYS 104.7 (77.0) 93.1 (72.7) 0.768  0.327  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n27 \n \noffline \n(CTRL: n=6, \nDYS: n=10) \nWRAT Arithmetic \n(scaled score) \nCTRL 110.5 (17.7) 116.3 (14.7) -2.956  0.127  \nDYS 101.1 (15.1) 104.2 (15.2) -1.835  0.083  \nWAIS Matrices \n(T-Score) \nCTRL 57.8 (8.1) 59.5 (9.7) -1.185  0.379  \nDYS 56.8 (3.6) 59.6 (3.8) -5.250  0.002  \nPhoneme Deletion \n(n correct) \nCTRL 12.8 (1.9) 13.3 (1.5) -0.889  0.415  \nDYS 13.7 (1.6) 14.0 (1.2) -0.896  0.246  \nSyllable Stress \nRecognition \n(ACC) \nCTRL 73.8 (18.4) 76.7 (21.5) -1.075  0.379  \nDYS 61.5 (9.5) 68.6 (12.0) -2.468  0.036  \nSyllable Stress \nRecognition \n(RT, s) \nCTRL 3.0 (0.9) 2.5 (0.4) 1.343  0.379  \nDYS 3.2 (1.1) 3.2 (1.4) 0.036  0.540  \nRAN (mean time, s) \nCTRL 28.1 (4.0) 25.8 (4.1) 4.834  0.038  \nDYS 33.1 (5.9) 29.5 (5.1) 4.936  0.002  \nTOWRE SWE (raw score) \nCTRL 98.2 (10.8) 96.2 (10.3) 1.369  0.379  \nDYS 80.1 (6.0) 83.5 (6.5) -3.511  0.011  \nTOWRE PDE (raw score) \nCTRL 56.8 (8.8) 59.8 (6.1) -2.423  0.160  \nDYS 44.6 (5.2) 48.5 (5.0) -2.830  0.025  \n1 Rise Task \n(time threshold, ms) \nCTRL 80.1 (70.3) 58.0 (31.7) 0.769  0.424  \nDYS 131.7 (78.5) 107.7 (71.1) 1.055  0.228  \n 534 \n3.3. Associations Between Neural and Behavioral Changes 535 \nTo explore whether the observed improvements in performance were systematically 536 \nrelated to learning the BCI, we calculated Pearson correlations between participants’ BCI 537 \nlearning scores and their changes in the behavioral tasks (computed as post-intervention 538 \nbehavioural score minus pre-intervention score in each case ). The results are shown in 539 \nTable 3.4. BCI learning scores were quantified using the t-statistics from session -wise 540 \ncomparisons of the theta/delta ratio. T-scores were considered separately for the real-time 541 \nand offline EEG pipelines. Higher t-statistics indicated greater success in reducing 542 \ntheta/delta ratios across training, and hence should be positively related to improvements 543 \nin accuracy in the behavioural tasks, and negatively related to improvements in processing 544 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n28 \n \ntime. Given that TS theory would predict improvements in the phonological and reading 545 \ntasks, we tested these correlations using one-tailed tests.  546 \n3.3.1.  Real-Time BCI Improvement; Correlations by Group 547 \nIn the real -time pipeline, no significant correlations were found between BCI learning 548 \nscores and improvements in Arithmetic or Matrices Reasoning tasks  (Table 3.4). This 549 \naligns with expectations, as these tasks reflect non-verbal cognitive abilities that were not 550 \ntargeted by the BCI. In contrast, significant associations were observed between neural 551 \nimprovement and phonological recoding for both groups . Specifically, participants with 552 \nhigher BCI learning scores showed greater gains in TOWRE non -word reading, r= .61, 553 \np< .05 (DYS) and r= .72, p< .05 (CTRL). In the control group, better BCI performance was 554 \nalso associated with a greater decrease in response time on the syllable stress 555 \ndiscrimination task (r= .79, p< .01), suggesting faster phonological processing. 556 \n3.3.2.  Offline Preprocessing BCI Improvement; Correlations by Group 557 \nIn the offline preprocessing pipeline, only the participants with dyslexia showed significant 558 \nchanges in phonology and reading. As with the real-time data, no significant relationships 559 \nwere detected between BCI scores , Arithmetic  and Matrices reasoning tasks. In the 560 \ndyslexic group, positive correlations were found between BCI learning scores and 561 \nimprovements in syllable stress discrimination accuracy (r= .51, p< .05), and TOWRE non-562 \nword reading (r= .56, p< .05). Additionally, a significant negative correlation was observed 563 \nbetween BCI performance and rise time threshold  (r= -.50, p< .05) , indicating that 564 \nparticipants with greater neural adaptation were more sensitive to amplitude rise time 565 \nfollowing training. 566 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n29 \n \n 567 \nIn each case, Figure 3.2 (real-time processing) and Figure 3.3 (offline preprocessing) 568 \ndisplay the relevant scatter plots and regression lines for the behavioural tasks with 569 \nsignificant correlations. All correlation coefficients for each behavio ural task, across both 570 \npreprocessing strategies and participant groups, are provided in Table 3.4. 571 \n 572 \nFigure 3.2. The correlation between BCI intervention training improvement and behavioral 573 \ntask improvement (real-time preprocessing). 574 \n 575 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n30 \n \nFigure 3.3. The correlation between BCI intervention training improvement and behavioral 576 \ntask improvement (offline preprocessing). 577 \nTable 3.4 Pearson correlations between BCI score and behavioral tests by group, 1-tailed 578 \ntests 579 \nBCI score \n(t-value \npre/post) \nbehavioral test \n(Post - Pre) \nControl Dyslexia \nr p-value r p-value \nrealtime \n(CTRL: n=12, \nDYS: n=20) \nWRAT Arithmetic (scaled score) 0.133  0.628  -0.208  0.432  \nWAIS Matrices (T-Score) 0.348  0.375  0.300  0.365  \nPhoneme Deletion (n correct) 0.247  0.375  -0.134  0.634  \nSyllable Stress Recognition (ACC) -0.431  0.715  0.339  0.286  \nSyllable Stress Recognition (RT, s) -0.786  0.008  -0.013  0.478  \nRAN (mean time, s) 0.038  0.628  -0.224  0.365  \nTOWRE SWE (raw score) -0.185  0.628  0.143  0.365  \nTOWRE PDE (raw score) 0.722  0.014  0.611  0.017  \n1 Rise Task (time threshold, ms) -0.231  0.375  -0.150  0.365  \noffline \n(CTRL: n=12, \nDYS: n=20) \nWRAT Arithmetic (scaled score) 0.176  0.681  -0.037  0.584  \nWAIS Matrices (T-Score) 0.278  0.681  -0.124  0.568  \nPhoneme Deletion (n correct) 0.180  0.681  -0.191  0.584  \nSyllable Stress Recognition (ACC) -0.411  0.908  0.513  0.024  \nSyllable Stress Recognition (RT, s) -0.436  0.440  -0.094  0.443  \nRAN (mean time, s) 0.078  0.681  -0.079  0.443  \nTOWRE SWE (raw score) -0.086  0.681  -0.100  0.568  \nTOWRE PDE (raw score) 0.402  0.440  0.560  0.024  \n1 Rise Task (time threshold, ms) -0.101  0.681  -0.502  0.024  \n 580 \n3.3.3.  All BCI Learners: Correlations 581 \nFinally, given that many control participants also demonstrated learning of the BCI, we 582 \nconsidered the BCI learners only as a single pooled group. We computed Pearson 583 \ncorrelations between participants’ BCI learning scores and their changes in the 584 \nbehavioral tasks, adding the groups to achieve reasonable power (N = 22 for real time 585 \ndata, 13 dyslexics and 9 controls; N = 16 for offline data, 6 dyslexics and 6 controls). The 586 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n31 \n \nresults are shown in Table 3.5. In both the real time data and the offline data, BCI 587 \nlearners showed significant improvements in nonword reading (r= .70, p< .01; r= .66, 588 \np< .01, respectively). For the real time data (N = 22), BCI learners also showed 589 \nsignificant improvement in amplitude rise time discrimination (r= -.53, p< .05) and in the 590 \nspeed of making syllable stress pattern judgements (r= -.47, p< .05). 591 \n 592 \nTable 3.5 Pearson correlations between BCI score and behavioral test in all BCI learners, 593 \n1-tailed tests 594 \nBCI score (t-value pre/post) behavioral test (Post - Pre) r p-value \nrealtime \n(n=22) \nWRAT Arithmetic (scaled score) 0.138  0.463  \nWAIS Matrices (T-Score) 0.128  0.463  \nPhoneme Deletion (n correct) -0.067  0.463  \nSyllable Stress Recognition (ACC) 0.167  0.343  \nSyllable Stress Recognition (RT, s) -0.467  0.029  \nRAN (mean time, s) 0.067  0.463  \nTOWRE SWE (raw score) -0.239  0.572  \nTOWRE PDE (raw score) 0.696  0.001  \n1 Rise Task (time threshold, ms) -0.527  0.018  \noffline \n(n=16) \nWRAT Arithmetic (scaled score) 0.006  0.869  \nWAIS Matrices (T-Score) -0.112  0.790  \nPhoneme Deletion (n correct) -0.159  0.790  \nSyllable Stress Recognition (ACC) 0.242  0.371  \nSyllable Stress Recognition (RT, s) -0.039  0.691  \nRAN (mean time, s) -0.073  0.691  \nTOWRE SWE (raw score) -0.181  0.790  \nTOWRE PDE (raw score) 0.661  0.003  \n1 Rise Task (time threshold, ms) -0.271  0.371  \n 595 \n4. Discussion. 596 \nThe current report suggests that a BCI aimed at normalising the low-frequency 597 \noscillatory neural patterns associated with continuous speech processing in 598 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n32 \n \ndevelopmental dyslexia can improve linguistic phonological processing of syllable stress 599 \npatterns and phonological recoding of print to sound (nonword reading) for adults both 600 \nwith and without dyslexia. The BCI developed here also improved ART sensitivity for all 601 \nBCI learners. ART is an important acoustic cue used for automatic oscillatory phase-602 \nresetting during speech-brain alignment (Doelling et al., 2014). These improvements are 603 \nin line with TS theory.  604 \nTS theory is based on atypical encoding of the low-frequency envelope information 605 \nthought to govern prosodic perception in dyslexia (Goswami, 2011). Prior neuroimaging 606 \nstudies have shown that children with dyslexia learning English, Spanish and French 607 \nshow impaired neural encoding of low-frequency speech envelope information <10 Hz 608 \nduring natural speech listening (DiLiberto et al., 2018; Molinaro et al., 2016; Destoky et 609 \nal., 2020), and that English-speaking children with dyslexia show a higher theta-delta 610 \nratio during natural speech listening, which is significantly related to their performance in 611 \nphonological awareness tasks (a higher ratio is associated with worse performance, 612 \nAraújo et al., 2024). Accordingly, the theta-delta ratio was targeted by the current BCI. 613 \nImprovements in syllable stress processing following BCI training were expected on the 614 \nbasis of related TS-driven speech modelling work, which indicated that sensory 615 \ndiscrimination of the phase relations between AMs at the delta (0.5 – 4 Hz) and theta (4 – 616 \n8 Hz) rates govern whether a strong or a weak syllable is perceived (Leong et al., 2014; 617 \nLeong & Goswami, 2015). Here, significant correlations between participants’ BCI scores 618 \nand syllable stress processing were demonstrated for both control adults (for real-time 619 \nprocessing and response time) and for adults with dyslexia (for offline processing and 620 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n33 \n \nresponse accuracy). Significantly faster syllable stress processing was also exhibited by 621 \nthe pooled group of all BCI learners (Table 3.5, real time data). No associations were 622 \nfound for phoneme-level linguistic tasks in any analyses, however this could reflect 623 \nceiling effects on the phoneme deletion task that was selected for this study.  624 \nFor both adults with dyslexia (Table 3.4) and all pooled BCI learners (Table 3.5), 625 \nthere was also a significant correlation between BCI scores and enhanced ART 626 \ndiscrimination. BCI learners showed better discrimination of ART following BCI training. 627 \nThis could be promising therapeutically for children, as by TS theory it is impaired ART 628 \ndiscrimination which affects neural speech encoding via oscillatory speech-brain 629 \nalignment. Both impaired ART discrimination and associated impaired neural speech 630 \nencoding of low-frequency speech information compromise the efficient development of a 631 \nphonological lexicon. Indeed, experimental work with dyslexic adults has demonstrated 632 \nsuch a relationship regarding impaired ART discrimination and impaired speech encoding 633 \n(Lizarazu et al., 2021), while a series of studies across languages (summarized in 634 \nGoswami, 2015) demonstrate that impaired ART discrimination is significantly related to 635 \nimpairments in phonological awareness at many linguistic levels. Accordingly, if learning 636 \nthe BCI leads to enhanced ART discrimination, this should have positive effects on 637 \ndevelopmental trajectories for phonological development.  638 \nMost promising of all regarding the compromised reading skills that ensue from the 639 \nphonological processing difficulties that characterize developmental dyslexia, control 640 \nadults, dyslexic adults and all BCI learners showed enhanced nonword reading after 641 \nlearning the BCI. For dyslexic participants, both real-time BCI scores and offline BCI 642 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n34 \n \nscores showed significant correlations with nonword reading (Table 3.4), while real-time 643 \nBCI scores showed a significant correlation with nonword reading for control adults 644 \n(Table 3.4). When all BCI learners were considered as a pooled group, both real-time and 645 \noffline BCI scores were significantly correlated with improvement in nonword reading 646 \n(Table 3.5). As impaired nonword reading is a hallmark of childhood dyslexia across 647 \nlanguages, further development of the current BCI for children may thus offer significant 648 \ntherapeutic benefits.    649 \nTo our knowledge, this is the first BCI for dyslexia that targets the pre-reading 650 \n‘phonological deficit’ (see Christodoulides et al., 2022, for an EEG classifier study 651 \nintended to inform a dyslexia BCI based on magnocellular theory, Ortiz et al., 2020, for 652 \nEEG classifiers for dyslexia based on AM-noise; Arias et al., 2021, for an in-principle BCI 653 \nto enhance neural entrainment; and Günet, 2020, for a dyslexia BCI based on 654 \nmultisensory training). The BCI developed here was informed by the TS theory of 655 \ndyslexia, an auditory theory that proposes that the auditory organization of speech 656 \ninformation by a child (assigning acoustic elements of speech perception to the 657 \ngroupings comprising words in a particular language) is impaired at the prosodic level, 658 \nleading to developmental differences in the accuracy of phonological representations at 659 \nthe level of syllable stress patterning. As the prosodic or rhythmic level is the foundational 660 \nperceptual (AM) level regarding the rest of the linguistic hierarchy (syllables, onset-rimes 661 \nand phonemes, see Leong & Goswami, 2015), these inaccurate prosodic representations 662 \naffect all levels of phonological representation for affected children, making learning to 663 \nread difficult and effortful in every language (see Goswami, 2022a, for a detailed 664 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n35 \n \nexplanation). Accordingly, if the current BCI is able to improve syllable stress processing, 665 \nchildren’s access to all levels of phonology in the linguistic hierarchy should improve. 666 \nIt is important to note that neural data suggest that the phonological representations 667 \ndeveloped by individuals with dyslexia are not noisy, as previously believed, rather they 668 \nare subtly different in organization from those developed by non-dyslexic individuals 669 \n(Keshavarzi et al., 2022, children; Tan et al., 2022, adults). According to TS theory, the 670 \nmain difference regarding phonological representations lies in encoding accurately the 671 \nlow-frequency amplitude envelope information (see Keshavarzi et al., 2023, for 672 \nexperimental evidence that amplitude envelopes for multi-syllabic words are also 673 \nproduced inaccurately by children with dyslexia). This difference in phonological 674 \nrepresentations for words means that when print is encountered and visual codes for 675 \nrepresenting spoken language are acquired (culturally-specific codes that are taught and 676 \nlearned using symbol-sound correspondences), the dyslexic child is at a disadvantage 677 \nfrom the outset. If the current BCI can be applied with children prior to learning to read, 678 \nthis disadvantage could potentially be eliminated before school entry. Indeed, brain 679 \nimaging studies across languages show that visual symbol learning, whether of the 680 \nalphabet or of characters such as Kanji, is linked to sound from the very beginning of 681 \nacquiring reading (Blau et al., 2010; Froyen et al., 2009; Maurer et al., 2005, 2011; Yang 682 \net al. 2020). Accordingly, by targeting neural features of the dyslexic brain’s response to 683 \nacoustic linguistic input (natural speech) before reading instruction commences, the 684 \ncurrent BCI may be able to facilitate visual symbol learning in any language. 685 \nThe current study has a number of limitations. Firstly, the training sessions were given 686 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n36 \n \nover a relatively short period of time, and some participants did not learn the BCI (real-time 687 \nprocessing, 7/20 dyslexics, 3/12 controls; offline preprocessing, 10/20 dyslexics, 6/12 688 \ncontrols). One explanation could be insufficient gaming experience, a ccordingly a longer 689 \ngaming period may be beneficial in studies which involve child participants. Secondly, the 690 \nsample size was relatively small. However, it is comparable to prior studies attempting to 691 \ncreate BCIs for dyslexia (Günet, 2020; Christodoulides et al., 2022). Thirdly, EEG data is 692 \nprone to exhibiting highly variable day-to-day variations. To mitigate this problem, baseline 693 \ndata was collected before each BCI run and these recorded EEG patterns were used to 694 \ndefine the upper and lower limits of the spaceship on the screen on each run.  Fourth, a 695 \nsingle story (Winnie the Pooh) was used throughout the whole training protocol. While this 696 \nwas helpful in allowing direct comparison of performance across sessions, it also made the 697 \nprotocol quite tedious, which may have led to de -motivation – an especially pertinent 698 \nlimitation if the participants were to be children. Accordingly, it would be best if future work 699 \ncould devise an operant learning protocol that could handle any story input in any language. 700 \nFinally, while the provided instructions were quite clear regarding the gaming objective (i.e. 701 \nmaking the spaceship go upwards as consistently as possible on the screen while listening 702 \nto the words in the story carefully), the instructions were also kept purposefully vague so 703 \nthat participants could decide by themselves which strategy to employ. Some participants 704 \nspontaneously made remarks indicating their chosen strategies, for example “letting your 705 \nbrain flow up and down with the syllables in speech in a new way”. To optimize children’s 706 \nBCI learning, it may be useful to give them explicit suggestions about utilizing strategies of 707 \nthis nature.   708 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n37 \n \nIn conclusion, the exploratory data presented here suggest that a simple and engaging 709 \nBCI for improving phonological processing during natural speech listening can be created 710 \nusing EEG data informed by the TS theory of developmental dyslexia. Participants who 711 \nlearned the BCI showed improved processing of syllable stress patterns in words, 712 \nimproved phonological recoding skills (nonword reading), and improved ART discrimination. 713 \nThese improvements occurred even though no direct training of phonology, nonword 714 \nreading nor ART discrimination occurred during the study. This is particularly interesting 715 \ntheoretically, as it suggests that the therapeutic benefits resulted from improving the neural 716 \ntheta-delta ratio during natural speech listening. Therapeutic interventions which filter 717 \nspeech to enhance ARTs have also been shown to improve speech processing in 718 \nparticipants with dyslexia via changing the theta-delta ratio (Mandke et al., 2023; see also 719 \nVan Herck et al., 2022, for a related envelope -enhanced method that did not explore the 720 \ntheta-delta ratio). Accordingly, further investigation of neural speech processing in dyslexia 721 \nguided by TS theory may identify other, possibly more effective, neural targets for BCI 722 \ndevelopment. 723 \n  724 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n38 \n \nACKNOWLEDGEMENTS 725 \nThe authors would like to thank all the participants who volunteered for  the study. This 726 \nresearch was funded by a donation to U.G. from the Yidan Prize Foundation. The sponsor 727 \nplayed no role in the study design, data interpretation, nor writing of the report. 728 \n 729 \nDATA AND CODE AVAILABILITY 730 \nData and code will be made available on request. 731 \n 732 \nDECLARATION OF COMPETING INTEREST 733 \nThe authors declare no conflicts of interest.  734 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n39 \n \nArias, F.J.C., Molinaro, N. & Lizarazu, M. (2021). Real time EEG neurofeedback as a tool 735 \nto improve neural entrainment to speech. 736 \nhttps://www.biorxiv.org/content/10.1101/2021.04.19.440176v1.full.pdf 737 \nAraújo, J. (2023). Computational framework enabling an EEG-based BCI for 738 \nneurofeedback in language disorders: The case of dyslexia. Ph.D. dissertation, University 739 \nof Cambridge. 740 \nAraújo, J., Simons, B.D. & Goswami, U. (2023). Remediating phonological deficits in 741 \ndyslexia with brain-computer interfaces. In (Ed.) C. Guger, Brain-Computer Interface 742 \nResearch, A State of the Art Summary 12. New York: Springer. 743 \nAraújo, J., Simons, B.D., Peter, V., Mandke, K., Kalashnikova, M., Macfarlane, A., 744 \nGabrielczyk, F., Wilson, A.M., Di Liberto, G.M., Burnham, D., & Goswami, U. (2024). 745 \nAtypical low-frequency cortical encoding of speech identifies children with developmental 746 \ndyslexia. Frontiers in Human Neuroscience, 18, 1403667.  747 \nAttaheri, A., Ní Choisdealbha, A., Di Liberto, G., Rocha, S., Brusini, P ., Mead, 748 \nN., Olawole-Scott, H., Boutris, P ., Gibbon, S., Williams, I., Grey, C., Flanagan, S., M., 749 \nGoswami, U. (2022). Delta- and theta-band cortical tracking and phase-amplitude 750 \ncoupling to sung speech by infants. Neuroimage, 247, 751 \n118698. https://doi.org/10.1016/j.neuroimage.2021.118698 752 \nAttaheri, A., Ní Choisdealbha, A., Rocha, S., Brusini, P ., Di Liberto, G., Mead, N., 753 \nOlawole-Scott, H., Boutris, P ., Gibbon, S., Williams, I., Grey, C., Alfaro e Oliveira, M., 754 \nBrough, C., Flanagan, S., and Goswami, U. (2024). Infant low-frequency EEG, cortical 755 \npower, cortical tracking and phase-amplitude coupling predicts language a year 756 \nlater. PLoS ONE 19 (12): e0313274. https://doi.org/10.1371/journal.pone.0313274 757 \nBlau, V., Reithler, J., van Atteveldt, N., Seitz, J., Gerretsen, P ., Goebel, R., Blomert, L. 758 \n(2010). Deviant processing of letters and speech sounds as proximate cause of reading 759 \nfailure: a functional magnetic resonance imaging study of dyslexic children. Brain, 133(3), 760 \n868–879. DOI: 10.1093/brain/awp308 761 \nChristodoulides, P ., Miltiadous, A., Tzimourta, K. D., Peschos, D., Ntritsos, G., 762 \nZakopoulou, V., Giannakeas, N., Astrakas, L. G., Tsipouras, M. G., Tsamis, K. I., Glavas, 763 \nE., & Tzallas, A. T. (2022). Classification of EEG signals from young adults with dyslexia 764 \ncombining a Brain Computer Interface device and an Interactive Linguistic Software Tool. 765 \nBiomedical Signal Processing & Control, 76, 1036-46. 766 \nDestoky F, Bertels J, Niesen M, Wens V, Vander Ghinst M, Leybaert J, et al. (2020) 767 \nCortical tracking of speech in noise accounts for reading strategies in children. PLoS Biol 768 \n18(8): e3000840. https://doi.org/10.1371/journal.pbio.3000840 769 \nDi Liberto, G. M., Peter, V., Kalashnikova, M., Goswami, U., Burnham, D., & Lalor, E. C., 770 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n40 \n \n2018. Atypical cortical entrainment to speech in the right hemisphere underpins 771 \nphonemic deficits in dyslexia. NeuroImage, 175, 70-79. 772 \nhttps://doi.org/10.1016/j.neuroimage.2018.03.072 773 \nDiLiberto, G.M., Attaheri, A., Cantisani, G., Reilly, R.B., Ní Choisdealbha, A., Rocha, S., 774 \nBrusini, P ., & Goswami, U. (2023). Emergence of the cortical encoding of phonetic 775 \nfeatures in the first year of life. Nature Communications, 14, 7789. 776 \nhttps://doi.org/10.1038/s41467-023-43490-x 777 \nDoelling, K. B., Arnal, L. H., Ghitza, O., Poeppel, D. (2014). Acoustic landmarks drive 778 \ndelta-theta oscillations to enable speech comprehension by facilitating perceptual 779 \nparsing. Neuroimage, 85, 761-768. DOI: 10.1016/j.neuroimage.2013.06.035 780 \nFacoetti, A., Trussardi, A. N., Ruffino, M., Lorusso, M. L., Cattaneo, C., Galli, R., Molteni, 781 \nM., & Zorzi, M. (2010). Multisensory spatial attention deficits are predictive of 782 \nphonological decoding skills in developmental dyslexia. Journal of Cognitive 783 \nNeuroscience, 22, 1011-1025. https://doi.org/10.1162/jocn.2009.21232.   784 \nFlanagan, S., Wilson, A. M., Gabrielczyk, F. C., Macfarlane, A., Mandke, K., Goswami, U. 785 \n(2024). Amplitude rise time sensitivity in children with and without dyslexia: differential task 786 \neffects and longitudinal relations to phonology and literacy.  Frontiers in Psychology, 787 \n15. https://doi.org/10.3389/fpsyg.2024.1245589 788 \nFroyen, D., Bonte, M.L., van Atteveldt, N., Blomert, L., (2009). The long road to 789 \nautomation: neurocognitive development of letter-speech sound processing. J. Cogn. 790 \nNeurosci. 21, 567–580. https://doi.org/10.1162/jocn.2009.21061 791 \nGiraud, A.L., and Poeppel, D. (2012). Cortical oscillations and speech processing: 792 \nemerging computational principles and operations. Nature Neuroscience, 15, 511-517. 793 \ndoi:10.1038/nn.3063 794 \nGoswami, U. (2011). A temporal sampling framework for developmental dyslexia. Trends 795 \nin Cognitive Sciences, 15, 3-10. https://doi.org/10.1016/j.tics.2010.10.001 796 \nGoswami, U. (2015). Sensory theories of developmental dyslexia: Three challenges for 797 \nresearch. Nature Reviews Neuroscience, 16, 43-54. http://dx.doi.org/10.1038/nrn3836    798 \nGoswami, U. (2022a). Theories of Dyslexia. In M. Skeide (Ed.), The Cambridge 799 \nHandbook of Dyslexia and Dyscalculia (Cambridge Handbooks in Psychology, pp. 5-24). 800 \nCambridge: Cambridge University Press. doi:10.1017/9781108973595.002 801 \nGoswami, U. (2022b). Language acquisition and speech rhythm patterns: an auditory 802 \nneuroscience perspective. R. Soc. Open Sci. 9: 803 \n211855. https://doi.org/10.1098/rsos.211855 804 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n41 \n \nGross J., Hoogenboom N., Thut G., Schyns P., Panzeri S., Belin P., & Garrod, S. (2013). 805 \nSpeech rhythms and multiplexed oscillatory sensory coding in the human brain. PLoS 806 \nBiology, 11(12), e1001752. https://doi.org/10.1371/journal.pbio.1001752. 807 \nGünet, E. (2020). Improving reading abilities in dyslexia with neurofeedback and 808 \nmultisensory learning. Unpublished PhD thesis, Sabanci University, Turkey. 809 \nhttps://research.sabanciuniv.edu/id/eprint/41210/ 810 \nKalashnikova, M., Goswami, U., & Burnham, D. (2020). Novel word learning deficits in 811 \ninfants at family risk for dyslexia. Dyslexia, 26(1), 3-17. https://doi.org/10.1002/dys.1649 812 \nKeshavarzi, M., Mandke, K., Macfarlane, A., Parvez, L., Gabrielczyk, F.C., Wilson, A.M., 813 \nFlanagan, S. & Goswami, U. (2022). Decoding of speech information using EEG in children 814 \nwith dyslexia: Less accurate low -frequency representations of speech, not ‘noisy’ 815 \nrepresentations. Brain & Language, 235, 105198. 816 \n 817 \nKeshavarzi, M., Di Liberto, G., Gabrielczyk, F.C., Wilson, A.M., Macfarlane, A. & Goswami, 818 \nU. (2023). Atypical speech production of multi-syllabic words and phrases by children with 819 \ndevelopmental dyslexia. Developmental Science, e13428. 820 \nKuhl, P . K. (2004). Early language acquisition: Cracking the speech code. Nature 821 \nReviews Neuroscience, 5, 831-843. https://doi.org/10.1038/nrn1533 822 \nLanderl, K., Wimmer, H. (2000). Deficits in phoneme segmentation are not the core 823 \nproblem of dyslexia: Evidence from German and English children. Applied 824 \nPsycholinguistics, 21, 2 , 243 – 262. DOI: https://doi.org/10.1017/S0142716400002058 825 \nLeong, V. & Goswami, U. (2015). Acoustic-emergent phonology in the amplitude 826 \nenvelope of child-directed speech. PLoS One, 10(12), e0144411. 827 \nhttps://doi.org/10.1371/journal .pone.0144411 828 \nLeong, V., Hämäläinen, J., Soltész, F., & Goswami, U. (2011). Rise Time Perception and 829 \nDetection of Syllable Stress in Adults with Developmental Dyslexia. Journal of Memory 830 \nand Language, 64, 59-73. https://doi.org/10.1016/j.jml.2010.09.003 831 \nLeong, V., Stone, M., Turner, R.E., & Goswami, U. (2014). A role for amplitude 832 \nmodulation phase relationships in speech rhythm perception. Journal of the Acoustical 833 \nSociety of America, 136, 366-381. https://doi.org/10.1121/1.4883366 834 \nLeong, V., Kalashnikova, M., Burnham, D. & Goswami, U. (2017). The temporal 835 \nmodulation structure of infant-directed speech. Open Mind, 1(2), 78-90. 836 \nhttps://doi.org/10.1162/OPMI_a_00008 837 \nLizarazu, M., Lallier, M., Bourguignon, M., Carreiras, M., Molinaro, N. (2021). Impaired 838 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n42 \n \nneural response to speech edges in dyslexia. Cortex, 135, 207-218. 839 \nhttps://doi.org/10.1016/j.cortex.2020.09.033 840 \nMandke, K.M., Flanagan, S.A., Macfarlane, A., Feltham, G., Gabrielczyk, Wilson, A.M., 841 \nGross, J., & Goswami, U. (2023). Neural responses to natural and enhanced speech edges 842 \nin children with and without dyslexia. Frontiers in Human Neuroscience, 17, 1200950. 843 \nMaurer U., Brem S., Bucher K., & Brandeis D. (2005). Emerging neurophysiological 844 \nspecialization for letter strings. Journal of Cognitive Neuroscience, 17(10), 1532-1552. 845 \nDOI: 10.1162/089892905774597218 846 \nMaurer, U., Schulz, E., Brem, S., der Mark, S. van, Bucher, K., Martin, E., Brandeis, D. 847 \n(2011). The development of print tuning in children with dyslexia: Evidence from 848 \nlongitudinal ERP data supported by fMRI. Neuroimage 57, 714–722. 849 \nhttps://doi.org/10.1016/j.neuroimage.2010.10.05 850 \nMolinaro, N., Lizarazu, M., Lallier, M., Bourguignon, M., & Carreiras, M. (2016). Out-of-851 \nsynchrony speech entrainment in developmental dyslexia. Human Brain Mapping, 37, 852 \n2767–2783. https://doi.org/10.1002/hbm.23206   853 \nNí Choisdealbha, A., Attaheri, A., Rocha, S., Mead, N., Olawole -Scott, H., Brusini, P ., 854 \nGibbon, S., Boutris, P ., Grey, C., Hines, D., Williams, I., Flanagan, S.A., & Goswami, U. 855 \n(2023). Neural phase angle from 2 months when tracking speech and non-speech rhythm 856 \nlinked to language performance from 12 to 24 months. Brain & Language, 243, 105301. 857 \nOrtiz, A., Martinez-Murcia, F. J., Luque, J. L., Giménez, A., Morales-Ortega, R., & 858 \nOrtega, J. (2020). Dyslexia diagnosis by EEG temporal and spectral descriptors: An 859 \nanomaly detection approach. International Journal of Neural Systems, 30 (07), 2050029. 860 \nhttps://doi.org/10.1142/S012906572050029X 861 \nSchneider, W., Roth, E., & Ennemoser, M. (2000). Training phonological skills and letter 862 \nknowledge in children at risk for dyslexia: A comparison of three kindergarten intervention 863 \nprograms. Journal of Educational Psychology, 92, 284–295. 864 \nhttp://dx.doi.org/10.1037/0022-0663.92.2 .284 865 \nSnowling, M. J. (2000). Dyslexia (2nd ed.). Oxford: Blackwell Publishers.. 866 \nStanovich, K. E. (1998). Refining the Phonological Core Deficit Model. Child Psychology 867 \nand Psychiatry Review, 3, 17-21. https://doi.org/10.1111/1475-3588.00203  868 \nStein, J., & Walsh, V. (1997). To see but not to read: The magnocellular theory of 869 \ndyslexia. Trends in Neuroscience, 20, 147–152. https://doi.org/10.1016/S0166-870 \n2236(96)01005-3 871 \nTan Y , Chanoine V, Cavalli E, Anton J-L and Ziegler JC (2022) Is there evidence for a noisy 872 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n43 \n \ncomputation deficit in developmental dyslexia? Front. Hum. Neurosci. 16:919465. doi: 873 \n10.3389/fnhum.2022.919465    874 \nValdois, S., Bosse, M. L., & Tainturier, M. J. (2004). The cognitive deficits responsible for 875 \ndevelopmental dyslexia: Review of evidence for a selective visual attentional disorder. 876 \nDyslexia, 10, 4, 339-363. https://doi.org/10.1002/dys.284. 877 \nVan Herck, S., Vanden Bempt, F., Economou, M., Vanderauwera, J., Glatz, T., Dieudonné, 878 \nB., Vandermosten, M., Ghesquière, P . and Wouters, J., (2022). Ahead of maturation: 879 \nEnhanced speech envelope training boosts rise time discrimination in pre ‐readers at 880 \ncognitive risk for dyslexia.  Developmental Science , 25(3), p.e13186. DOI: 881 \n10.1111/desc.13186 882 \nYang, Y ., Yang, Y . H., Li, J. J., Xu, M., & Bi, H. Y . (2020). An audiovisual integration deficit 883 \nunderlies reading failure in nontransparent writing systems: An fMRI study of Chinese 884 \nchildren with dyslexia. Journal of Neurolinguistics, 54, 100884. 885 \nhttps://doi.org/10.1016/j.jneuroling.2019.100884 886 \nZiegler, J. C., & Goswami, U. (2005). Reading acquisition, developmental dyslexia, and 887 \nskilled reading across languages: A psycholinguistic grain size theory. Psychological 888 \nBulletin, 131, 3-29. https://doi.org/10.1037/0033-2909.131.1.3   889 \nZiegler, J. C., Bertrand, D., Tóth, D., Csépe, V., Reis, A., Faísca, L., Saine, N., Lyytinen, 890 \nH., Vaessen, A., Blomert, L. (2010). Orthographic Depth and Its Impact on Universal 891 \nPredictors of Reading: A Cross-Language Investigation. Psychological Science. 21 (4), 892 \n551-559. https://doi.org/10.1177/0956797610363406 893 \n  894 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n44 \n \nFIGURE LEGENDS 895 \nFigure 2.1 Panel (A) shows the whole Experimental Protocol, including the time line of each 896 \nBCI session and the spaceship interface. Panel (B) depicts the decoder logic behind the neural 897 \nfeedback. 898 \nFigure 3.1. The t-statistics used as a measure of BCI performance. Data are shown for each 899 \nparticipant, under real-time feedback versus offline preprocessing. 900 \nFigure 3.2. The correlation between BCI intervention training improvement and behavioral task 901 \nimprovement (real-time preprocessing). 902 \nFigure 3.3. The correlation between BCI intervention training improvement and behavioral task 903 \nimprovement (offline preprocessing). 904 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\nDaily Session:\nBCI task: 2 runs\nRest: ~ 5mins\nExperimental Protocol:\nBCI training: 8 daily sessions\nCognitive testing: 2 days (pre, post)\nRest: 2 days\nBaseline Stage: 4 mins BCI Stage: 10 mins\n𝑥 𝑡 = 𝜃(𝑡)\n𝛿(𝑡) log[𝑥(𝑡)]\nRatio distribution log distribution BCI task\n(a)\n(b) EEG data\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\nSyllable Stress RT\nTOWRE PDE\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint \n\nTOWRE PDE\nRise time threshold\nStress syllable accuracy\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700941doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}