Introduction

Skilled reading involves a series of processes that enable accurate decoding of words and phrases to comprehend their meaning. Often, decoding and comprehension skills develop in tandem1. However, for individuals on the autism spectrum, dissociations have been reported between decoding and comprehension2. Specifically, autistic children without intellectual disabilities are often reported to have word and pseudoword (pronounceable nonsense word) decoding skills comparable to their non-autistic counterparts, but with reading comprehension skills lower than that of their age or grade level3,4,5,6,7,8,9,10. The cognitive and neural source of these differences, however, remains unclear.

Much of the literature examining reading ability in autistic individuals has defined reading and its components through the lens of the Simple View of Reading (SVR)11,12. The SVR posits that two independent components, decoding and oral language comprehension, are required for reading. Here, decoding refers to the ability to recognize isolated words quickly and accurately through learned spelling-sound correspondences, while oral language comprehension refers to the ability to process and construct literal and inferred meanings from learned linguistic components associated with speech (i.e., sound-to-meaning)11,12. The SVR assumes that once a series of words is decoded, the reader applies the same mechanisms used for understanding the spoken equivalent to the written text11,12. The assumption that language comprehension mechanisms are applied after decoding suggests that processing word meaning (i.e., semantics) is not needed at the decoding stage. This assumption is in direct contrast to the Primary Systems Hypothesis (PSH), which posits that the recognition and pronunciation of single words involves processing three independent components: orthography (written form), phonology (sound form), and semantics13,14,15. The PSH specifies that reading occurs through multiple pathways. For example, pronouncing a word can be accomplished through spelling-to-sound correspondences (orthography → phonology), or through spelling-to-meaning-to-sound correspondences (orthography → semantics → phonology). This suggests that under certain conditions, decoding a word through spelling-sound correspondences alone is not enough and may need input from semantic processes in order to be pronounced correctly.

The specificity of the PSH allows researchers to examine questions that are difficult to examine under the SVR. For example, deep orthographies such as English often have words that do not have one-to-one spelling and sound correspondences (e.g., yacht or colonel) or may have words that are similar in their written form but differing in their spoken form (e.g., hint – pint)16. These irregular words may thus require input from semantically mediated pathways (e.g., spelling-to-meaning-sound) to be correctly pronounced14,17,18,19. Additionally, the PSH provides a framework for examining how reading behavior or neural responses may be individually or simultaneously influenced by specific psycholinguistic factors related to components of reading such as word frequency (the degree to which a word appears within a given language; e.g., work vs. wilt), consistency (the degree to which a word’s written form matches its sound form; e.g., hint vs. pint) and imageability (the degree to which a word evokes of a mental image; e.g., moth vs. truth). These relationships are best captured during reading aloud paradigms, which allows for the recording of overt reading responses and, thus, a more direct examination of the relation between psycholinguistic factors and reading behavior20,21.

Previous neuroimaging studies have reported differences between autistics and non-autistics in silent reading and semantic processing tasks. Specifically, autistic adults have less neural activity (i.e., decreased BOLD response) during reading in regions typically associated with reading22,23,24, as well as the involvement of visual association regions when processing word meanings25. However, to date, no studies have directly examined correspondences between reading behavior and neural activation in autism, particularly how both behavior and brain responses are modulated by psycholinguistic features in autism using the PSH framework. We address this limitation by using a reading aloud paradigm, allowing us to directly test processes that silent reading cannot, specifically the production of phonological output from orthographic input. Therefore, in this study, we sought to better understand the neural mechanisms of reading differences in autism and determine if there are distinct neural mechanisms behind processing both the form and meaning of words.

Word processing in non-autistic individuals

Reading single words involves several primarily left-lateralized visual and language-associated regions for processing phonologic, orthographic, and semantic information. Prior neuroimaging studies of reading have identified a posterior-to-anterior network that begins within the left ventral occipitotemporal cortex (OTC), specifically the visual word form area (VWFA), a subregion functionally specialized for orthographic processing26,27,28,29. The middle temporal gyrus (MTG) is also activated during reading, and has been specifically associated with phonological and semantic word processing30,31,32,33. Unfamiliar words and pseudowords have been proposed to be processed within primarily left temporoparietal regions such as the posterior superior temporal gyrus (STG), primarily associated processing phonology32,33,34 as well as within the supramarginal gyrus21,35,36. Finally, phonological and semantic information are thought to converge in the inferior frontal gyrus (IFG), possibly to be prepared for verbal output; however, the exact role of this region is unclear37,38.

To probe the role of each reading-related region in its hypothesized processes, behavioral and neuroimaging studies have examined how various psycholinguistic features, such as word frequency, consistency, and imageability, affect reading performance and brain responses. At the behavioral level, low-frequency words (e.g., wilt) often take longer to process, have larger error rates compared to high-frequency words (e.g., work), and have been shown to modulate performance in a variety of tasks such as reading aloud39, lexical decision40, and word-association tasks41. At the neural level, word frequency effects in reading have been used in conjunction with other psycholinguistic factors to elicit activation within the lexical (whole word) system20,42,43,44.

Consistency refers to how systematically a word’s spelling maps onto its pronunciation. For example, the words “cash” and “hash” are very similar at both the phonological and orthographic levels – the “ash” rime in each is pronounced the same way. However, the rime in the word “wash” is pronounced differently despite being spelled the same as “cash” and “hash”, and is therefore considered to be spelling-sound inconsistent. Reading lower consistency words (e.g., wash) has been associated with higher activity in the IFG for both adults and children16,20,42,45,46.

Finally, imageability refers to the degree to which a word evokes a mental image. Imageability is often considered a semantic factor as highly imageable words are associated with richer semantic representations18,47,48. Imageability is typically highly correlated with word concreteness/abstractness; that is, more concrete words such as “farm” also have higher imageability, while more abstract words such as “truth” have lower imageability. Imageability is thought to play a role in word recognition, particularly for infrequent and inconsistent words in which spelling-sound associated computations would not be efficient, and thus rely on a semantically mediated pathway14. Aligning with this theory, words that are low in both frequency and consistency but high in imageability (e.g., moth) are processed faster than words that are low in all three lexical factors (e.g., dour)18,19,49. Neurally, positive imageability effects (i.e., higher activation for more imageable words) have been reported in areas associated with the semantic network including the bilateral AG, precuneus, and posterior cingulate cortex (pCC)20,30,50. Lower imageability words are associated with increased activation within frontal regions such as the IFG50,51,52, likely indicating more domain-general processing for those more difficult items20. To date, there are no studies that examine how brain responses to single-word reading are modulated by these psycholinguistic features in autism.

Word processing in autistic individuals

Compared to both typically developing children and adults, autistic individuals without intellectual disabilities appear to have an intact, and sometimes precocious, ability to read. Behavioral studies report that autistic individuals performed equally well or better than age-matched non-autistic counterparts in both word and pseudoword reading tasks4,5,7,8,53. Furthermore, studies have found a higher prevalence of hyperlexia, a phenomenon in which an individual demonstrates word-decoding abilities beyond what is expected for their age, amongst autistic children54,55,56. However, within those same studies, it was also found that autistics tend to perform more poorly in comprehension tasks compared to age-matched non-autistics4,5,6,7,53,55. This pattern of results fits with research examining semantic processing in memory tasks. For example, Toichi & Kamio57,58 revealed that autistic adults do not show the expected recall advantage for concrete words compared to abstract words. Taken together, these results suggest that autistic adults are generally proficient in their decoding ability but may struggle to apply or integrate the semantic information necessary to extract meaning from text.

To better understand the neural mechanisms underlying semantic processing in autism, several fMRI studies have used semantic decision or judgment tasks22,24,25. In Harris et al.22, age-matched autistic and non-autistic adults completed a semantic judgment task in which they were asked to identify via button press if the word had positive or negative valence. At the behavioral level, both groups performed similarly in both valence attribution and reaction times in response to the words. At the neural level, the autistic group showed significantly less activation in the IFG compared to non-autistics. However, the autistic group had greater activation in the MTG than non-autistics. In a similar study, Gaffrey and colleagues25 examined the processing of word meanings using a semantic decision task that involved indicating if a word belonged to a given category. They did not find differences in IFG activity; instead, autistic individuals showed greater activation in bilateral extrastriate visual cortices (Brodmann areas 18 and 19) compared to controls. Furthermore, this greater activation was correlated with the number of errors produced during the task25. Overall, these studies suggest an atypical neural organization of semantic processing in autism.

In addition to evidence of reduced activation of the reading network and atypical involvement of sensory processes, evidence of hemispheric differences in lexical processing has also been found in autism. A meta-analysis conducted by Herringshaw and colleagues23 revealed that autistic individuals showed decreased left IFG and STG activation and increased activation in right IFG and STG, particularly in lexical comprehension and tasks involving the visual presentation of stimuli. In a study by Knaus and colleagues24, adolescent autistics and age-matched non-autistics completed a semantic similarity task during fMRI to examine the neural bases of processing action words (e.g., verbs). They found that even though both groups had activation in left hemisphere language regions, the autistic group did not show expected activation in the left pre-supplementary motor area (pre-SMA) that is typically associated with processing action words. Furthermore, while there were no hemispheric differences in activation between groups, faster reaction times within the autistic group were specifically associated with less left-lateralized activation, particularly within frontal regions24. Overall, these results suggest an atypical lateralization of language processing in autism in which left frontal language regions are less efficient, and therefore may invoke right hemisphere homologues as compensatory mechanisms.

Current study

To date, there have been no studies directly examining the correspondences between reading behavior and neural activation in autism, particularly how both behavior and brain responses are modulated by psycholinguistic features in autism using the PSH framework. Therefore, as a first step, we focus on examining responses to reading single words and pseudowords aloud using whole-brain analyses and hypothesis-driven regions of interest (ROI) in autistic and non-autistic comparison groups.

By using single words and pseudowords, we can exert tight control over psycholinguistic features to precisely probe the reading system20. One feature that shows widespread effects is word frequency. It is thought to reflect overall familiarity with word forms, rather than more specific aspects such as spelling-sound mapping or semantics15,39,59. Therefore, we selected stimuli to ensure that spelling-sound consistency and imageability were not correlated with word frequency. When considering consistency, we focused on brain regions such as the OTC, STG, and MTG, which are thought to be involved in mapping from orthography to phonology with varying contributions from semantics38. Based on prior work showing that autistic individuals have equal or potentially greater word and pseudoword decoding abilities3,4,5,6,7,8, we predict that the autistic group will show greater consistency effects (i.e., greater associations between neural activity and consistency) in regions such as the OTC, STG, and MTG compared to the non-autistic group. Specifically, we predict greater decreases in activity as consistency increases in the autistic group compared to the non-autistic group. Regarding imageability, we focused on brain regions thought to be associated with semantic processing, such as the AG and MTG20,30. Based on behavioral and brain imaging results suggesting less efficient semantic processing in autism and greater reliance on regions associated with perceptual processing, we predict that imageability effects will be weaker in the autistic group in regions such as the AG and MTG, but may be present in visual processing regions. Additionally, due to evidence of hemispheric differences in activation during semantic tasks23, we can also expect greater imageability effects in the right hemisphere homologues of semantic processing regions in the autistic group.

Methods

Participants

Participants included 40 individuals in either the autistic (N = 19; mean age = 20.3) or non-autistic group (N = 21; mean age = 21.7). The data reported here have not been previously reported but were acquired during the same overall scanning sessions as other data previously reported by Graves et al.60. Participants were compensated $15 per hour for behavioral sessions and $30 per hour for MRI sessions. All participants met the following overall enrollment criteria: 1) no reported history of psychiatric or neurologic disorders (other than autism in the autistic group); 2) right-handed, 3) English learned as primary language at no later than five years old; 4) no metal embedded in soft tissue or other contraindications to MRI; and 5) minimum full-scale IQ of 85 on the Wechsler Abbreviated Scale of Intelligence – II61. Participants in the autistic group additionally required a primary diagnosis of autism spectrum disorder as well as a score that was above the cut-off point of 6 on the shortened version of the Autism Spectrum Quotient (AQ-10) questionnaire62. Autism diagnosis was confirmed through the Autism Diagnostic Observation Schedule – Second Edition (ADOS-2) Module 4, administered by a licensed clinical psychologist who was research-reliable or a doctoral student supervised by such a clinician60. Participants within the non-autistic group were not administered the ADOS-2, as is often the case for adult participants who are assumed to reliably self-report63. To quantify potential autistic traits across all participants, we administered the Social Responsiveness Scale, Second Edition (SRS-2)64, a validated tool for assessing autistic tendencies in adults65. Independent t-tests for group differences confirmed that there were significant differences between autistic and non-autistic groups in their SRS scores (autistic mean = 88.80, SD = 20.16; non-autistic mean = 56.71, SD = 16.43; t(34) = 4.60 , p < 0.001).

Groups were matched on relevant factors (Table 1, all p > 0.1) such as age (t (37) = 1.08), full-scale IQ (t (37) = -1.43), verbal IQ (t (32) = -1.22), and performance IQ (t (36) = -1.15). The autistic group included 1 transgender female (MtF), 1 transgender male (FtM), and 1 gender non-conforming (NC) individual. We were not able to individually match autistic participants by gender; however, we tested for independence between group and gender. Chi-square tests for independence showed no differences in gender identity ratios between groups (X2 = 1.54, p = 0.47). Participants also completed a battery of reading-related tests from the Wechsler Individual Achievement Test – Third Edition66. Independent t-tests for group differences found there were no significant differences between groups (Table 1, all p > 0.1) in reading comprehension (t (19) = -0.07), word reading (t (37) = -0.74), and pseudoword decoding (t (34) = -1.64).

Table 1 Descriptive information regarding non-autistic and autistic groups.
Full size table

The study was conducted in accordance with the Declaration of Helsinki, and the study protocol was approved by the Rutgers University Institutional Review Board (IRB). All participants provided written informed consent as approved by the Rutgers University IRB.

Stimuli

Stimuli consisted of 220 monosyllabic English words and pseudowords (110 each). Words were a subset of stimuli from the Graves et al.20 465-word list and were uncorrelated in frequency, spelling-sound consistency, imageability, and length (Table 2; p = 0.1). Word frequency values were obtained from CELEX67 using a log-transformed occurrence per million estimate. Word consistency values were defined by the number of friends (words with a rime that is spelled the same and pronounced the same as the target word) minus the number of enemies (words with a rime that is spelled the same but pronounced differently from the target word). Finally, mean imageability ratings (on a scale of 1–7) were obtained from a database comprised of 6 sources: Bird et al.68, Clark & Paivio69, Cortese & Fugett70, Gilhooly & Logie71, Paivio et al.,72, Toglia & Battig73.

Table 2 Correlation matrix (r-values) for word psycholinguistic features of interest.
Full size table

The pseudowords were generated by entering each of the word stimuli into Wuggy74 to generate 10 pseudoword variations for each word. Generated pseudowords were then entered into MCWord75 to check if words included a trigram necessary for pronunciation. Pseudohomophones (or nonwords that sound like real words, e.g., work vs werk) were then manually eliminated. Consistency values for each pseudoword on the selected list were defined by calculating the number of friends minus the number of enemies, in essentially the same way as calculated for words. Specifically, the rime of each selected pseudoword was manually identified. All monosyllabic words with a rime that was an orthographic match to the pseudoword rime were identified from CELEX. These were then indexed to their corresponding phonetic transcriptions. Phonetic transcriptions matching the pronunciation of the target pseudoword rime were considered friends (e.g., wint – tint), while any not matching the target were considered enemies (e.g., wint – pint). Spelling-sound consistency was then calculated for each pseudoword the same way as for words: number of friends minus the number of enemies.

In-scanner task

Participants were asked to read both words and pseudowords aloud as they appeared in the scanner (Fig. 1). Their spoken responses were recorded using an MRI-compatible microphone. Participants viewed the stimuli from a projection screen through a mirror mounted on the head coil. Word and pseudoword stimuli were presented in a randomized order over two 9-min runs (110 stimuli per run) using PsychoPy software76. Each trial consisted of the following: 1) an initial fixation cross displayed centrally on the screen for 1000 ms; 2) the stimulus (word or pseudoword) displayed for 1000 ms; 3) a fixation cross displayed centrally for a random jittered inter-trial interval (minimum = 1000 ms; maximum = 3000 ms) before returning to the initial fixation cross for 1000 ms. Verbal responses during reading aloud were recorded using Audacity software to determine both trial-level accuracy and reaction time (RT) measures.

Fig. 1
figure 1

Task Design.

Full size image

Image acquisition

Structural and functional data were collected using a Siemens Trio 3 T scanner with a 12-channel head coil at the Rutgers University Brain Imaging Center. T1-weighted (1 mm isotropic resolution) structural images were collected with an MPRAGE sequence using the following parameters: TR = 1900 ms, TE = 2.52 ms, flip angle = 9°, image matrix = 256 × 256, field of view = 256 mm, number of slices = 176. Functional images were collected as T2*-weighted images (3 mm isotropic resolution) using a gradient-echo single-shot echo planar image sequence (EPI) with the following parameters: TR = 2000 ms, TE = 25 ms, flip angle = 77°, image matrix = 64 × 64, field of view = 195 mm, number of slices = 35.

Image preprocessing

Neuroimaging data were preprocessed using AFNI software77 within custom Linux shell scripts. Functional data were first converted from DICOM to AFNI HEAD/BRIK format, after which motion correction and anatomical alignment were conducted using the AFNI script, align_epi_anat.py78. 3dToutcount was used to identify TRs with motion displacement greater than 10% of mean displacement and censored from the analysis. Average TR’s censored from the analysis was less than 1% in each group (autistic = 0.62%; non-autistic = 0.67%), with no significant differences in the average TRs censored between groups (t (36) = 0.17; p = 0.86). Finally, a nuisance regressor was created to model non-physiological signal fluctuation based on the averaged signal from the lateral ventricles using 3dmaskave.

Data analyses

Behavioral

Trial-level accuracy and RT were analyzed using R 4.1 software79. For all RT analyses, only the correct trials were included. For words, correct trials were those that matched a pronunciation listed in a standard American English dictionary such as Merriam-Webster (https://www.merriam-webster.com/). For pseudowords, all authors met and arrived at a joint consensus for preferred pronunciations and reasonable alternatives based on standard readings by first-language speakers of American English. Participant pronunciations matching those on the list were considered correct, otherwise incorrect. Production of plausible alternatives not on the list was rare and scored based on consensus rating. This procedure was followed in lieu of formal tests of inter-rater reliability. T-tests for group differences were first conducted for both accuracy and RT. Mixed-effect linear regressions for accuracy and RT respectively for words were then conducted using the lme4 package80 to test for both group and individual level effects using the following terms of interest as fixed effects: group, and continuous measures of consistency, frequency, and imageability. Fixed effect interaction terms were included between group status and consistency, frequency, and imageability respectively, as well as between consistency and frequency, consistency and imageability, imageability and frequency, and finally, a three-way interaction between all three psycholinguistic features. A random effect interaction term was also included for group status and individual participants. Similar mixed-effects linear regression models were conducted for accuracy and RT for pseudowords using all of the above terms except for frequency and imageability, and associated interaction terms, as these are not defined for pseudowords.

Neuroimaging

To test for potential differences in overall activity for words and pseudowords and how this activity was modulated by psycholinguistic factors, we used both whole-brain and ROI-based approaches.

Whole-brain

Neuroimaging data were analyzed using AFNI software77. First-level analyses of individual participant imaging data were performed using 3dDeconvolve with seven nuisance covariates derived from preprocessing, and eight explanatory variables of interest. Six of the nuisance covariates consisted of motion parameters, 3 for rotation in the x (pitch; side-to-side), y (roll; front-to-back) and z (yaw; vertical axis rotation) directions, and 3 for displacement in those directions. Finally, the last nuisance covariate was a physiological noise regressor from the lateral ventricles. Of the explanatory variables of interest, there was a separate term associated with correct responses to words, pseudowords, and z-score normalized measures of each of the psycholinguistic features of interest: imageability, consistency, frequency, and pseudoword consistency. There were additional separate terms for a z-score normalized regressor for trial RT for both words and pseudowords and finally a term modeling erroneous responses, for a total of nine explanatory terms in the full model. The output images were then nonlinearly warped to Talairach (TT_N27) space81 using 3dQWarp and smoothed with a 5 mm FWHM kernel using 3dmerge in AFNI. Second-level analysis of groups was performed in AFNI using the 3dtest +  + program to compare first-level contrasts between autistic and non-autistic groups. In all analyses, the “Clustsim” option was used to implement map-wise cluster correction for multiple comparisons82. Thresholds for all reported contrasts were set at a voxel-level of p < 0.005 and map-wise cluster corrected to p < 0.05.

Regions of interest

To probe for potential group differences in how consistency and imageability modulate brain activity, five non-overlapping left hemisphere ROIs were defined based on previous functional neuroimaging results20,32 using AFNI software. The ROIs consisted of five areas that have been associated with orthographic and phonological processing: the ventral OTC, the posterior portions of the STG (pSTG) and MTG (pMTG), the AG, and the IFG. The specific contours were the same as those tested in Graves et al. (2014), where each ROI mask was restricted so that they lay within anatomical boundaries as defined within the Talairach atlas file (TT_N27_EZ_ML) included in AFNI83. Due to previous reports of less left-lateralized activation during language processing in autism23, right hemisphere masks were also created by flipping the orientation of each ROI from left to right, using the AFNI program 3dLRflip. For each individual, activation estimate statistics for each ROI were extracted from the results of the first-level analyses for both words and pseudowords, as well as word and pseudoword consistency and word imageability using 3dROIstats. Finally, two-way or three-way ANOVAs were conducted to test for differences in activity associated with psycholinguistic features between and within groups while also factoring in hemisphere and stimulus type (words vs pseudowords) using R 4.1 software79.

Results

Behavioral

Participants in both groups had very high reading accuracy scores (Fig. 2A and 2C) for both words (autistic: 97%; non-autistic: 96%) and pseudowords (autistic: 91%; non-autistic: 89%) and were not reliably different from each other (words: t = -0.61, p = 0.5; pseudowords: t = -1.01, p = 0.3). Mixed-effect linear regression models for word accuracy (all statistics reported in Table 2) revealed main effects for consistency, imageability, and frequency. As predicted, there was also a significant interaction between consistency and imageability in which imageability had a greater impact on accuracy for lower-consistency words. There was no main effect of group status in the model. However, there was an interaction between group status and imageability, in which accuracy scores increased as word imageability increased for the non-autistic group but not the autistic group (Table 2; Fig. 3A). Regression models for pseudoword accuracy did not reveal any significant effects for either group, consistency, or for their interaction (all p > 0.1; see Tables 3 and 4 for full statistics).

Fig. 2
figure 2

Accuracy and RT responses to words (A and B) and pseudowords (C and D) between groups. Groups were not significantly different in either accuracy or RT measures. Dots represent individual participant data points.

Full size image
Fig. 3
figure 3

Mixed-effect linear regressions depicting interaction effects between group status and imageability (A and B), consistency (C and D), and pseudoword consistency (E and F) on accuracy and reaction times responses. Dots represent individual word data points. Shading around the lines represents 95% confidence intervals. When the difference of slopes is significant at the p < .05 threshold, the z-scores in black appear in bold, while the z-score for the corresponding line also appears in bold.

Full size image
Table 3 Models of word accuracy and RT.
Full size table
Table 4 Models of pseudoword accuracy and RT.
Full size table

Considering RT as the dependent variable when reading words (Fig. 2B) or pseudowords (Fig. 2D), t-tests for group differences revealed no significant differences between the autistic and non-autistic group for either words (autistic group mean = 791 ms, non-autistic group mean = 761 ms; t = -1.01, p = 0.3) or pseudowords (autistic mean = 842 ms, non-autistic mean = 808 ms; t = -1.05, p = 0.3). Mixed-effect linear regressions for word RT (all statistics reported in Table 2) revealed significant main effects of frequency and imageability. There was an interaction effect between imageability and frequency, in which imageability had a greater effect on lower frequency words. Consistency did not significantly predict RT. There was no main effect of group status; however, there was an interaction of group status and imageability in which RT decreased as word imageability increased for the autistic group but not the non-autistic group (Fig. 3B). Additionally, there was an interaction between group status and consistency in which RT decreased as word consistency increased for the non-autistic group but not the autistic group (Table 2; Fig. 3D). Regression models for RT to pseudowords (see Tables 3 and 4 for full statistics) did not reveal any significant effects of group but did reveal a main effect of consistency in which reaction time decreased as pseudoword consistency increased (Fig. 3F).

Functional neuroimaging

Whole-brain

In the non-autistic group, compared to baseline fixation, reading words revealed robust activation of regions associated with the reading network with significant clusters of activation in bilateral STG, precentral gyrus (preCG), postcentral gyrus (postCG), and lateral occipitotemporal regions (OTC), with the largest cluster of activation having its maximum point within the left posterior STG (Table 5; Fig. 4A). For the autistic group, analyses revealed similar areas of activation, albeit to a lesser degree, with significant clusters of activation in the bilateral STG, preCG, and postCG, but not the OTC, with peak activation in more dorsal parts of the preCG (Table 6; Fig. 4B). While the extent of activation was much smaller in the autistic group relative to the non-autistic group, direct comparisons between the groups did not reveal any statistically significant group differences in activation for words compared to baseline at the whole-brain level.

Table 5 Location of maximum activation of clusters within the non-autism group.
Full size table
Fig. 4
figure 4

Words versus baseline contrast for the non-autistic group (A) and autistic group (B).

Full size image
Table 6 Location of activation peaks of clusters for the autism group.
Full size table

Compared to baseline fixation, reading pseudowords revealed clusters of activation within the left preCG, with additional clusters of activation within the right postCG for both the non-autistic (Fig. 5A) and autistic (Fig. 5B) groups, with no significant differences between the groups at the whole-brain level. Finally, to test for differences in lexicality, we contrasted words and pseudowords. However, neither within-group nor between-group contrasts revealed any significant differences in activation.

Fig. 5
figure 5

Pseudowords versus baseline contrast for non-autistic group (A) and autistic group (B).

Full size image

Analyses testing for modulation of activation by word and pseudoword characteristics revealed significant effects of word frequency and pseudoword consistency. Effects of word frequency were found in the non-autistic group only. Specifically, increasing word frequency correlated with decreased activation in the left middle frontal gyrus (MFG), and increased activation in the right anterior temporal lobe (ATL) and AG (Supplementary Fig. S2). Effects of pseudoword consistency were found in the autistic group only. Specifically, increasing pseudoword consistency corresponded with decreased activation (or equivalently, decreasing pseudoword consistency corresponded with increased activation) in bilateral clusters surrounding the intraparietal sulcus (IPS) with extension to the superior and inferior parietal lobules (Fig. 6). Between-group contrasts for word frequency, word and pseudoword consistency, and word imageability effects did not reveal any significant differences at the whole-brain level.

Fig. 6
figure 6

Brain areas where activation is modulated by pseudoword consistency in the autistic group only. Cooler colors represent decreasing activity as pseudoword consistency increases.

Full size image

Regions of interest

To test for hemispheric differences in the activation of words and pseudowords, two-way ANOVAs were conducted for each ROI. Additionally, to test for hemispheric differences as well as differences in the effects of word and pseudoword consistency between the autistic and non-autistic groups, we conducted three-way ANOVAs for each ROI. Finally, two-way ANOVAs were conducted to test for hemispheric differences as well as differences in the effects for word frequency and word imageability. Uncorrected p-values are provided for these ANOVAs. For words, no effects of group or hemisphere were found (all p > 0.1; Supplementary Fig. S3). However, for pseudowords, a main effect of hemisphere and an interaction effect between hemisphere and group were found within the IFG and pSTG. Follow-up t-test analyses revealed significantly greater activation in the left IFG (M = 1.08) compared to the right IFG (M = -0.06, p = 0.005) in the non-autistic group only. Additionally, there was a non-significant trend of greater activation in the left pSTG (M = 1.26) compared to their right pSTG (M = 1.09, p = 0.06) in the non-autistic group only (Supplementary Fig S3).

Regarding the effects of psycholinguistic factors, three-way ANOVAs did not reveal a main effect group but did reveal an interaction effect of group and stimuli type in the pSTG and OTC (Table 7; Fig. 7 and 8). Follow-up t-tests revealed the autistic group had greater positive word consistency effects in both the left (autistic group M = 0.10, non-autistic group M = -0.12; p = 0.006) and right pSTG (autistic group M = 0.09, non-autistic group M = -0.08; p = 0.03). This same effect was also observed in both the left (autistic group M = 0.13, non-autistic group M = -0.19; p = 0.005) and right OTC (autistic group M = 0.18, non-autistic group M = -0.07; p = 0.02). Additionally, follow-up t-tests (Table 8) revealed the autistic group showed greater negative pseudoword consistency effects in the left OTC (M = -0.13) compared to the non-autistic group (M = 0.10; p = 0.02). No differences were found in the right hemisphere. Regarding word imageability, two-way ANOVAs revealed a main effect of hemisphere in the IFG. Follow-up t-tests revealed that, regardless of group, there was greater effect of imageability in the left IFG (autistic group M = -0.10, non-autistic group M = -0.11) compared to the right IFG (autistic group M = -0.001, non-autistic group M = -0.03; Table 9). Finally, two-way ANOVAs for word frequency-related activation in the IFG revealed a main effect of hemisphere but no effect of group.

Table 7 Results of t-tests for group differences in effects of word consistency in ROIs between non-autistic and autistic groups.
Full size table
Fig. 7
figure 7

Effects of word and pseudoword consistency in the left and right pSTG (orange) for the autistic and non-autistic groups. Compared to the non-autistic group, the autistic group had significantly greater effects of word consistency in both hemispheres. Stars indicate differences at p < .05 threshold with top lines indicating between-group differences and bottom lines indicating within-group differences. Dots represent individual participant data points.

Full size image
Fig. 8
figure 8

Effects of word and pseudoword consistency in the left and right OTC (green) for the autistic and non-autistic groups. Compared to the non-autistic group, the autistic group had significantly greater effects of word consistency in both hemispheres and greater pseudoword consistency in the left hemisphere only. Stars indicate differences at p < .05 threshold with top lines indicating between group differences and bottom lines indicating within group differences. Dots represent individual participant data points.

Full size image
Table 8 Results of t-tests for group differences in effects of pseudoword consistency in ROIs between non-autistic and autistic groups.
Full size table
Table 9 Results of t-tests for group differences in effects of imageability in ROIs between non-autistic and autistic groups.
Full size table

Discussion

In this study, we sought to better understand the neural bases of reading differences in autism and determine if there are distinct neural mechanisms behind processing both the form and meaning of single words. To date, we are aware of no studies that have directly examined the neural correlates of reading aloud in autism. Using a reading aloud paradigm with overt responses is important because it enables us to directly test for correspondences between reading behavior and neural activation. Our design also enabled examining how psycholinguistic features such as word frequency, imageability, consistency, and pseudoword consistency modulate brain and behavior responses in autistic and non-autistic individuals. At the behavioral level, we found that while there were no differences in overall RT and accuracy between groups, more fine-grained analyses revealed that imageability and word consistency modulate RT differently between groups. At the whole-brain level, we found no significant activation differences between groups. Within groups, however, activation was significantly modulated by psycholinguistic factors such as word frequency and pseudoword consistency. Specifically, in the autistic group, we found that increasing pseudoword consistency corresponded with decreased activity in bilateral IPS and the surrounding cortex. ROI analyses also revealed significant differences in activation to word and pseudoword consistency. Specifically, the autistic group showed greater positive effects of word consistency in bilateral pSTG and OTC compared to the non-autistic group, while the autistic group showed negative effects of pseudoword consistency compared to the non-autistic group in the left OTC. Our findings provide insights into how these factors are related to the main components of reading, particularly phonology and orthography, as well as how the processing of these components may differ in autism.

Differences in reading behavior

Consistent with the well-matched characteristics of the samples (Table 1), there were no significant differences between groups in overall accuracy and RT for reading words or pseudowords. Regarding accuracy, positive imageability effects were found in the non-autistic group only, where increasing imageability was associated with higher accuracy. However, we found there were very high accuracy scores for words in both groups (autistic = 97%; non-autistic = 96%). Further inspection of the data found that the accuracy scores across groups were not normally distributed. Therefore, due to these concerns, we are cautious about interpreting these results.

Regarding RT, surprisingly, we found that increasing word imageability corresponded with reduced RT in the autistic group only. One possible explanation for this finding relates to how imageability may be processed. Previous behavioral research has suggested that autistic individuals may struggle to integrate semantic information during reading5,8 and non-reading memory tasks57,58. While imageability is considered a semantic feature and has been associated with regions within the semantic network in the past20,30,50, imageability may also index components of perceptual processing. That is, words that are high in imageability have a direct, tangible reference that can be mentally visualized (e.g., farm). Current neurocognitive theories of autism such as Enhanced Perceptual Functioning (EPF)84 and Weak Central Coherence (WCC)85 have suggested that people with autism have either a bias or higher proficiency in perceptual processing compared to their non-autistic peers. Previous neuroimaging research has found the involvement of parietal and occipital regions associated with mental imagery even when processing sentences that are low in imagery in autistic individuals86 as well as the involvement of visual associations regions when processing word meaning25. If autistic individuals do have a bias or higher proficiency in perceptual processes, then they may be able to utilize specific semantic features such as imageability that also index perceptual processing. This would also suggest that other semantic features that may not index perceptual processes, such as semantic diversity, a measure of semantic ambiguity87, would have minimal impact on reading behavior in autism. Future studies should consider increasing the number of stimuli and using a wider range of psycholinguistic features, particularly semantic features, to not only test this possibility but to also allow the use of fine-grained pattern-based analyses. Such approaches could be useful in revealing potential differences in how semantic features are neurally represented in autistic individuals.

Neural word processing

Although there were no statistically significant whole-brain activation differences between groups, visual inspection of the thresholded activation maps within each group shows apparent qualitative differences. Activation for words and pseudowords, each compared to fixation baseline, was more widespread in language-related areas for the non-autistic group, and conversely more spatially restricted within the autistic group. Specifically, reading words for both groups showed activation within the STG, preCG, and postCG. However, the largest cluster of activation for the non-autistic group was primarily within the left pSTG, while the largest cluster of activation for the autistic group was within the left preCG. The pSTG has been associated with phonological processing34,88,89. The preCG has also been associated with phonological processing, particularly with phonological decoding90 and phonological rehearsal91. However, the preCG has also been associated with the processing of speech and the auditory-motor coordination network35 and may be involved in processing auditory error signals associated with maintaining fluency during speech92. While the groups tested here were matched on measures of language and reading ability, these trends toward differences in phonology-related neural regions point to the possibility that autistic individuals have a different functional neural organization of reading, specifically when processing phonology.

Processing of psycholinguistic features

Regarding the effects of psycholinguistic features, analyses did not reveal any neural effects of imageability for either whole-brain or ROI-based analyses. Word frequency effects were observed at the whole-brain level for the non-autistic group only, where increased word frequency was associated with decreased activation in the left MFG, and increased activation in the right MTG and AG. Follow-up ROI analysis of the left IFG showed no differences between groups. Frequency-related activations in the left greater than right IFG, regardless of group, were found, possibly consistent with previous studies showing both language-related and more domain-general effects in this area20,93. The MTG and AG have previously been associated with semantic processing somewhat bilaterally, though primarily in the left hemisphere20,30,35,94. The lack of word frequency effects specific to the autistic group, including within semantic regions such as the AG, may point to less emphasis on processing meaning during word reading, particularly for higher-frequency words.

Turning to the effects of word consistency, whole-brain analyses did not reveal effects in either group. However, ROI analyses showed significantly greater positive word consistency effects for the autistic group in bilateral pSTG and OTC compared to the non-autistic group. The left OTC, which includes the VWFA, has been shown in previous neuroimaging studies to be positively correlated with orthographic probability26,29 and may support the mapping of orthography to phonology20,95. Regarding the left pSTG, lesions to this region have been associated with deficits in phonological processing96,97,98,99. Furthermore, previous neuroimaging studies have found that activation in this region is associated with accessing and extracting phonological components, particularly during speech32,34. Combined with the lack of neural effects for frequency and imageability in the autistic group, these consistency effects suggest that the autistic group may be placing more emphasis on orthography-phonology mapping. Additionally, greater activation for consistency in the right hemisphere pSTG and OTC in the autistic group suggests that processing words in the left hemisphere may be less efficient for the autistic group, possibly due to being unable to take advantage of semantic features. Alternatively, greater activation in the right hemisphere could suggest an inability to take advantage of long-range connections between semantics and phonological resources in the left hemisphere, leading to increased cognitive effort in the form of recruitment of additional right hemisphere resources. Both of these suggestions fit with previous neuroimaging literature on language processing in autism that found less left-lateralized activation for language comprehension tasks, particularly in the IFG and STG23,24, as well as evidence of differences in frontal-posterior connectivity in autism86,100,101.

Whole-brain effects for pseudoword consistency were only observed in the autistic group, in which increased pseudoword consistency was associated with decreased activity in bilateral IPS spreading to adjacent cortices in the superior and inferior parietal lobules. Similarly, ROI analyses revealed that the autistic group had significantly more negative effects of pseudoword consistency in the left OTC compared to the non-autistic group. While we did not find performance differences between groups for in-scanner pseudoword reading, there was a trend toward better Pseudoword Decoding ability in the autistic group (Table 1, autistic = 110.8 vs. non-autistic = 104.9), consistent with previous studies showing that autistic individuals tend to read pseudowords better than their counterparts without autism4,7. These results suggest that autistic individuals may be better at applying rule-like spelling-sound mappings. Previous neuroimaging research has shown the IPS to be associated with the task-positive network102,103. Furthermore, recent studies of typical reading aloud have found that orthographic and phonological similarities among stimulus words are associated with neural patterns in the IPS and adjacent cortices21. Here we found that the more consistent the spelling-sound mappings were for pseudowords, the faster participants were to initiate reading them aloud, in both groups. With very high accuracies suggesting a lack of any speed-accuracy trade-off, faster RTs for higher-consistency pseudowords can be interpreted as those pseudowords being “easier” or less effortful to pronounce. Thus, decreasing activity in the bilateral IPS and left OTC with increasing consistency found only in the autistic group may reflect less effort required to read pseudowords for which decoding skills straightforwardly apply.

Put another way, the defining difference between words and pseudowords is that words have meaning and pseudowords do not. The OTC is both part of the ventral attention system and houses the VWFA104. Therefore, it can reasonably be expected to respond to factors that contribute to the ease or difficulty of reading. Previous studies have reported greater activation for pseudowords than words for non-autistic participants in the left OTC16. Here that group activated left OTC more for spelling-sound consistent pseudowords, and more for inconsistent words. The PSH proposes that information about semantics is important for correctly pronouncing inconsistent words. For autistic participants, we expect reading to be less efficient for such words because of potential difficulties recruiting the semantic information that would typically help with pronouncing them. This may be what is driving the opposite effect for word consistency in the OTC between groups. For pseudowords, there is no semantics to appeal to. Instead, reading higher consistency pseudowords corresponds with lower activation in both the OTC and IPS for the autistic participants only. This suggests particular sensitivity to spelling-sound consistency among autistic participants when such information is sufficient for pronunciation. Considered alongside the frequent co-occurrence of hyperlexia with autism54,55, this suggests autistic individuals may be sensitive to statistical regularities for spelling-sound correspondences, especially for pseudoword stimuli for which semantics would not help guide the pronunciation. Overall, results from both word and pseudoword consistency point to a greater emphasis during reading on orthography-phonology mappings, and correspondingly less emphasis on dynamic recruitment of semantics, in participants with autism. A potential implication is that, when learning new words, rather than initially focusing on their meanings, it might be more helpful to encourage autistic individuals to focus on their phonological similarity with known words.

Potential limitations

One potential concern to consider when interpreting the current results is that the stimuli consisted of only 110 words and 110 pseudowords. Therefore, it is possible that the lack of significant differences in the word vs. pseudoword contrast between groups as well as the lack of significant group-level whole-brain differences with parametric modulation of word and pseudoword features could be attributed to not having sufficient trials to robustly elicit these effects. A second possible concern is the sample size. While our current sample of participants with autism (N = 19) is higher than that of similar studies in autism such as Gaffrey et al. (N = 12)25, Harris et al. (N = 14) 22 , and Knaus et al. (N = 15)24, it can still be considered rather limited. Additionally, to address concerns about potential false-positive results, we chose to use the current best practices that account for the actual smoothness of the data and temporal auto-correlation when calculating cluster-size thresholds to minimize potential false positives in our whole-brain analyses82. Therefore, it is possible that in using these practices to minimize false positives, we may have limited the sensitivity to detect more subtle effects. Finally, the participant sample was restricted to individuals with a full-scale IQ of 85 or greater. This ultimately limits the generalizability of our findings to autistic individuals without intellectual disabilities. For the non-autistic comparison participants, inclusion criteria included having an AQ-10 score of less than 6. We also collected SRS-2 responses, including for the non-autistic sample, to supplement self-reports of no autistic diagnosis. There were 3 participants included in the sample with high SRS-2 scores (> = 80). Removing the behavioral data from those three participants and re-running the analyses revealed no differences from the patterns reported here. The presence of autism-related traits, as measured by the SRS-2, has also been previously reported to occur among people without autism105.

Conclusions

This study directly examined the neural correlates of word and pseudoword reading aloud in younger adults with and without autism. Our results provide evidence of differences in the functional organization of reading in autism, particularly when processing elements of orthography and phonology. Specifically, we show neural evidence of more bilateral engagement during reading in autism, particularly when reading pseudowords that depend on spelling-sound correspondences with no or minimal influence from semantics. Overall, the results of our study suggest a focus on spelling-sound mapping abilities rather than semantics when reading aloud in autism.