Lesions Reveal Shared and Distinct Neurocognitive Bases of Oral Reading and Silent Word Recognition

Abstract

Reading is a fundamental skill for communication, learning, and daily life. Alexia, an acquired reading disorder typically caused by left-hemisphere stroke, impairs both oral and silent word reading. However, most research and clinical assessments focus on oral reading. Functional imaging studies in typical readers suggest partly shared and partly distinct neural bases of oral versus silent reading, but this has never been assessed in a lesion study. We examined oral word reading and silent word recognition, measured by a lexical decision task, in 85 chronic left-hemisphere stroke survivors (46 males, 39 females), comparing efficiency scores combining speed and accuracy to those of 69 neurologically healthy adults (36 males, 33 females). The stroke group had greater deficits in oral reading than lexical decision. Performance on the two tasks was highly correlated suggesting shared substrates, albeit with considerable unexplained variance. Support vector regression-based lesion-symptom mapping localized lesions associated with deficits in the two tasks. Oral reading deficits were associated with lesions to the superior temporal and supramarginal gyri and the rolandic operculum after controlling for lexical decision performance. Lexical decision deficits were linked to lesions in the angular gyrus, even after controlling for oral reading performance. A novel conjunction analysis found that lesions of angular and middle temporal gyri, extending into ventral occipitotemporal cortex, affected performance on both tasks, suggesting shared neural substrates. These findings suggest that neurocognitive bases of silent reading and reading aloud are partly shared and partly distinct. Alexia assessments should include both silent and oral reading tasks.

Significance Statement

Silent reading is essential for daily life but has received less attention than oral reading in research on acquired reading deficits, i.e., alexia. While functional imaging suggests partially distinct activation patterns for oral and silent reading, lesion-symptom mapping offers a critical method for identifying brain regions necessary for task performance. This is the first lesion mapping study to compare oral and silent reading within the same individuals. Using lexical decision as a proxy for silent reading, we found partly shared and partly distinct neural substrates for silent versus oral reading. These findings reveal overlapping but differentially weighted neural systems and underscore the need to assess silent reading alongside oral reading when assessing alexia.

Introduction

Reading is a critical skill in modern society. Understanding its neurocognitive mechanisms is essential for diagnosing and treating reading disorders, including alexia, an acquired reading deficit most commonly caused by left-hemisphere stroke (Friedman, 1995; Brookshire et al., 2014; Bahrami Balani and Bickerton, 2023). Alexia affects both oral and silent reading (Friedman, 1995; Leff et al., 2001). Despite the importance of silent reading in daily life, research has primarily focus on oral word reading, leaving silent reading deficits understudied.

One challenge in studying silent reading is the absence of overt behavioral output. Lexical decision, which assesses silent word recognition by asking participants whether or not a letter string forms a real word, has served as a proxy for silent reading for decades (Milberg and Blumstein, 1981; Coslett et al., 1993; Grainger and Jacobs, 1996; Joubert et al., 2004; Carreiras et al., 2007). Although lexical decision does not strictly require comprehension, lexical decision performance has long been known to rely on semantic access (James, 1975; Fischler, 1977; Milberg and Blumstein, 1981). The simplicity of the lexical decision task compared to comprehension tasks like word-to-picture matching makes it a valuable silent reading contrast to oral word reading (Balota and Chumbley, 1984; Plaut, 1997; Katz et al., 2012).

Neurocognitive models agree that reading relies on integration of orthographic, phonological, and semantic information (Seidenberg and McClelland, 1989; Plaut, 1997; Coltheart et al., 2001; Perry et al., 2007), though not all reading tasks engage these components equally. Models predict overlap in the processing demands of oral and silent reading, with task-specific weighting of phonological versus semantic pathways, depending on output requirements (Fig. 1). Reading aloud is thought to emphasize orthography-to-phonology mapping to plan for speech production (Caramazza and Hillis, 1990; Rapp and Goldrick, 2000; Hillis et al., 2004; Hillis, 2008), whereas lexical decision is thought to emphasize orthography-to-semantic mapping for word recognition (Seidenberg and McClelland, 1989; Plaut, 1997; Dilkina et al., 2010).

Figure 1.

Neurocognitive model of oral reading and lexical decision. Left, Model illustrating cognitive processes involved in oral reading and lexical decision of single written words. Both tasks begin with visual input and engage orthographic, phonologic, and semantic representations. Oral reading is thought to place greater demands on orthographic-to-phonology mapping and requires motor planning and articulation. Lexical decision is thought to place greater demands on orthographic-to-semantic mapping and requires decision-making. Right, Key brain regions thought to be associated with each process are overlaid on a left-hemisphere brain image, with colors corresponding to the cognitive components in the model. Gradient bars indicate the relative contribution of each region to oral reading versus lexical decision.

Prior functional imaging studies in healthy adults have examined the neural basis of oral reading and lexical decision separately. Single-word oral reading consistently activates regions involved in phonological processing and speech output, including the left supramarginal gyrus (SMG) and superior temporal gyrus (STG; Turkeltaub et al., 2002; Graves et al., 2010; Taylor et al., 2013). Lexical decision engages regions associated with lexical-semantic processing, particularly in the angular gyrus (AG) and middle temporal gyrus (MTG; Binder et al., 2003; Taylor et al., 2013), along with ventral prefrontal regions (Binder et al., 2003). The few fMRI studies directly comparing reading aloud and lexical decision have found overlapping activation in orthographic and semantic regions, along with task-specific engagement: premotor and sensorimotor regions for oral reading and regions involved in decision-making for lexical decision (Carreiras et al., 2007; Fawcett et al., 2022). A representational similarity analysis found that reading aloud and lexical decision relied on overlapping visual, orthographic, phonological, and articulatory representations but to different degrees: reading aloud elicited greater visual, phonological, and articulatory feature decodability, while silent reading elicited greater orthographic information decodability (Bailey et al., 2025). Overall, functional imaging findings suggest both tasks may rely on a shared reading network but differ in the degree to which they weigh different components.

Lesion-symptom mapping (LSM) provides a complementary method to functional imaging by identifying regions where damage causally impairs behavior (Bates et al., 2003; Rorden and Karnath, 2004; Wilson, 2017). Several lesion studies have examined reading aloud, together demonstrating that lesions involving posterior STG, SMG, premotor cortex, and inferior frontal cortex can impair reading (Woollams et al., 2007; Piras and Marangolo, 2009; Rapcsak et al., 2009; Ripamonti et al., 2014; Binder et al., 2016; Dickens et al., 2019; Li et al., 2021; Dyslin et al., 2025; Staples et al., 2025). To our knowledge, no LSM studies have examined written lexical decision.

To identify shared and distinct neural substrates of oral reading versus lexical decision, we apply support vector regression-based LSM (SVR-LSM) in 88 left-hemisphere stroke survivors. We hypothesized that oral word reading depends more heavily on regions associated with phonological-motor processing, while lexical decision relies more on region-associated lexical-semantics.

Materials and Methods

Participants

Eighty-five participants with chronic left-hemisphere stroke (>6 months) and 69 age- and sex-matched neurotypical controls participated in this study (Table 1). Participants for this study were recruited for a larger study investigating the brain basis of aphasia recovery (Clinicaltrials.gov ID: NCT04991519). Participants had at least 9 years of education and adequate vision and hearing. All participants were native English speakers and had no history of significant brain injury, neurological or psychiatric disorders, learning disorders, or language impairments prior to stroke. Of the 108 participants originally enrolled with left-hemisphere strokes, 7 were excluded due to missing data and 16 were excluded following post-enrollment discovery of bilateral strokes or other neurological disorders. The study was approved by the Georgetown University IRB, and all participants provided informed consent.

Table 1.

Participant demographics

	Left-hemisphere stroke (n = 85)	Controls (n = 69)
	Mean (SD)	Mean (SD)
Age	61.0 (11.6)	61.4 (11.6)
	Min: 34; max: 92	Min: 31; max: 83
Education	16.4 (3.0)	17.1 (2.5)
	Min: 9; max: 22	Min: 12; max: 21
Sex	46 M/39 F	36 M/33 F
Race	34 Black or African American, 50 White or Caucasian, 1 more than one race	22 Black or African American, 47 White or Caucasian
Ethnicity	4 Hispanic/Latino, 80 not Hispanic/Latino, 1 declined to answer	67 not Hispanic/Latino, 2 declined to answer
Handedness	75 right-handed/10 left-handed	64 right-handed/5 left-handed
Time since stroke (months)	47.9 (52.2)	–
	Min: 3.9; max: 270.7
Mean Lesion Volume (cm3)	97.5 (91.4)	–
	Min: 0.69; max: 415.2

Values are mean (standard deviation). Education reflects education by degree. Handedness was assessed using the Edinburgh Handedness Inventory. Participants with laterality index greater than +20 were classified as right-handed, and those below −20 were classified as left-handed.

Experimental design

Reading assessments

Oral reading assessment. Participants completed an oral reading assessment of real words and pseudowords. We used both stimulus types for the analyses to match the lexical decision task, which presents both real words and pseudowords. Real oral word reading stimuli were 200 monosyllabic words that varied orthogonally along dimensions in frequency, regularity, and imageability (Staples et al., 2025). Pseudoword stimuli were 60 monosyllabic pseudowords and 3–6 letters in length and were matched to words on body consistency, positional bigram frequency, and articulatory complexity (Dyslin et al., 2025). The reading tasks were programmed in E-Prime 3.0 and presented on a 17″ Dell Inspiron touch-screen laptop in tent mode. Real words were presented in blocks of 25 items with breaks in between blocks, and pseudowords were presented in blocks of 20 items. Participants were informed that stimuli were not real words for the pseudoword condition. Participants touched the screen to advance to the next item, with time limit of 10 s. Responses were video recorded and scored for accuracy offline. For real word reading, scoring was based on the first complete reading attempt containing both a consonant and a vowel. Any response that included a plausible pronunciation of the presented pseudoword or real word was scored as correct. Pseudowords were scored as incorrect if they contained spelling-to-sound mappings that do not occur in American English. Participant video and audio were recorded and scored after the session (see Dyslin et al., 2025 for details). Accuracy was based on the first complete attempt, defined as any attempt containing both a consonant and a vowel. Response times (RTs) were measured based on video and audiograms as the interval between the stimulus onset and the onset of the first complete attempt. Only RTs for accurate responses were used in the analyses.

Lexical decision task

The paradigm implemented a classic written lexical decision task. The experiment included 400 total trials. Of the 400 trials, 200 were real words used from the oral reading task. The remaining 200 stimuli were pseudowords: 80 pseudowords were the same as those used for oral reading, including pseudohomophones, 60 were created by altering regular words from the real word corpus, and 60 were created by altering irregular words from the corpus. Each trial began with a 500 ms fixation cross followed by the presentation of the stimulus word and response buttons on the bottom of the screen. Participants were asked to respond whether the stimuli were a “real word” or “not a word” by pressing the corresponding button on the screen. The task was administered in two parts, each containing 200 stimuli. The task advanced to the next item after a response or after 10 s without a response. RT was defined as the interval between stimulus onset and the button press initiated by the participant. Only RTs for accurate trials were used in the analyses. Two similar versions of the task were used across participants, which differed only in the specific stimuli presented. To ensure consistency across participants, only the 310 items (160 words and 150 pseudowords) that were common to both versions were included in the present analyses.

Statistical analysis

Behavioral analyses

Since mild deficits are reflected in slow RTs, while more severe deficits are reflected in lower accuracies, we used efficiency (accuracy/RT) as an integrated measure of performance for both oral reading and lexical decision (Sabatini, 2002; Eason et al., 2013; Laurence et al., 2018). Efficiency was defined as the accuracy across all trials divided by median RT for correct trials only. This measure allows more valid comparisons with controls, who often perform near ceiling accuracy on lexical decision, and offers a more sensitive measure of reading fluency than accuracy alone, especially when ceiling effects or compensatory slowing might mask underlying deficits.

Because lexical decision is a forced-choice task with a 50% chance baseline, lexical decision efficiency values were scaled so that chance corresponded to zero (scaled efficiency = [2 * (accuracy − 0.5)] / median RT on correct trials). The scaling provided a measure of efficiency more comparable to oral reading, where the floor accuracy is zero. Behavioral analyses were conducted using scaled lexical decision efficiency values. Although lexical decision efficiencies were scaled so that chance performance was equivalent to oral reading, caution is warranted in directly comparing the efficiency scores within groups, since one can have some accurate trials on lexical decision by guessing, while this is not likely for oral reading. Therefore, we restrict interpretation to relative differences between efficiency on the two tasks for stroke survivors as compared with the differences between tasks in controls. For completeness, we also report results from analyses using unscaled efficiency scores. Both approaches yielded the same pattern of results. The distributions of scores were checked within the task and group. Skewness values were within ±1 and kurtosis within ±0.85. Shapiro–Wilk tests were nonsignificant.

To assess group differences and task effects on reading efficiency, we conducted a 2 × 2 mixed ANCOVA with task (oral reading vs lexical decision) as a within-subject factor and the group (stroke vs control) as the between-subject factor, with age and education as covariates. To examine the relationship between the two tasks, we conducted separate linear regressions for stroke survivors and controls, with oral reading efficiency as the dependent variable and lexical decision (scaled) efficiency as the primary predictor, with age, education, and lesion volume (in stroke survivors) as additional predictors. Lesion volume was included to account for global stroke severity. We also report a parallel analysis controlling for “overall aphasia severity” as measured by the Western Aphasia Battery Aphasia Quotient (WAB-AQ).

Neuroimaging

Image acquisition

Brain images were acquired at the Center for Functional and Molecular Imaging at Georgetown University Medical Center. MRI data were acquired on a 3T Prisma scanner. For each participant, a high-resolution T1-weighted magnetization-prepared rapid gradient echo (MPRAGE) sequence was acquired with the following parameters: 176 sagittal slices; slice thickness = 1 mm; 1 mm cubic voxels; field of view (FOV) = 256; matrix = 256 × 256; repetition time (TR) = 1,900 ms; echo time (TE) = 2.98 ms; inversion time (TI) = 900 ms; flip angle = 9; and generalized autocalibrating partial parallel acquisition = 2. Additionally, to assist in manual lesion tracing, a fluid-attenuated inversion recovery (FLAIR) sequence was acquired: 192 sagittal slices, slice thickness = 1 mm; 1 mm³ voxels; flip angle = 120°; FOV = 256 mm; matrix = 256 mm × 256; TR = 5,000 ms; TE = 386 ms; and TI = 1,800 ms.

Lesion tracing and normalization

Lesion segmentation was done manually on each participant's MPRAGE and FLAIR image using the ITK-SNAP software by author P.E.T., a board-certified neurologist. Native space MPRAGEs and lesion tracings were warped to MNI space using Clinical Toolbox Older Adult Template as the target template (Rorden et al., 2012) as described below and previously (Avants et al., 2010; Martin et al., 2024).

Brain parenchyma was first extracted from each native space image using a mask designed to minimize clipping of gray matter edges. The initial mask was generated by combining binarized lesion tracing images with white and gray matter tissue probability maps applied to the original native space image, cost function masked with the lesion tracing. The mask was then blurred and inverted to remove the nonbrain tissue from the image and finally normalized using the Advanced Normalization Tools software (ANTs). Lesion masking was used at each step of the normalization process. After bias field correction was applied, normalization proceeded using a typical ANTs procedure, including rigid transform step, an affine transform step, and a nonlinear SyN step. The initial warp output was further refined by applying three additional iterations of the SyN step, aimed at improving registration accuracy. The final step involved concatenating the transformation fields from both the linear and nonlinear warps. These were used to register the original MPRAGE and lesion masks to the template space, employing BSpline interpolation for smooth resampling. This iterative nonlinear warping process was designed to improve alignment of expanded ventricles and shifted deep brain structures. To finalize the process, the normalized lesion masks were downsampled to a voxel size of 2.5 mm³.

Multivariate LSM

To localize brain areas that contribute to oral reading and silent reading, we employed SVR-LSM (Zhang et al., 2014). Nonrandom lesion distributions and voxel-level autocorrelation make univariate voxel-based LSM prone to mislocalization, particularly for functions involving multiple brain regions (Mah et al., 2014). In contrast, the multivariate SVR-LSM approach incorporates lesion status across all voxels in a single regression model, improving sensitivity to nonlinear relationships (Zhang et al., 2014). We conducted SVR-LSM in MATLAB R2022b via a graphical user interface tool developed in our laboratory (DeMarco and Turkeltaub, 2018). Only voxels lesioned in at least 5% of participants were included in each analysis. While a 10% threshold is common in LSM studies to balance coverage and statistical power (Kimberg et al., 2007; Dyslin et al., 2025), we used a 5% threshold to ensure inclusion of ventral occipitotemporal regions and to better test distributed network hypotheses involving both visual and semantic areas. Lesion volume, age, and education were regressed out of both the lesion data and behavioral outcomes of interest (DeMarco and Turkeltaub, 2018).

We estimated five SVR-LSM models using unscaled efficiency scores. First, two separate one-tailed SVR-LSM analyses were run to identify lesions associated with low oral word reading efficiency and low lexical decision efficiency. Next, to isolate distinct neural substrates supporting oral reading versus lexical decision, performance on the corresponding reading task was included as a covariate in each lesion symptom mapping analysis (i.e., oral word reading accuracy was covaried in the lexical decision analysis and vice versa). Since both tasks are thought to engage overlapping cognitive and neural processes, covarying performance on the matched task allows us to account for shared variance between the two and test whether lesions in specific brain regions are associated with disproportionately greater deficits in one task relative to the other. To provide a formal statistical test of whether there are shared substrates for both oral word reading and lexical decision, we used a novel SVR-LSM conjunction procedure. In this procedure, the minimum SVR-β value from the LSMs of the two behaviors is taken to create a new min-SVR-β map. This calculation yields a map in which lesions must result in deficits in both tasks to achieve a high voxel value. Combined with the permutation-based statistical test below, it provides a simple way to empirically identify brain structures where lesions result in deficits in both behaviors that are unlikely to have occurred by chance.

A permutation approach was used to control the cluster-level family-wise error rate. Behavioral scores were randomly reassigned to participants across 10,000 permutations generating voxelwise SVR-β-value maps, thresholded at p < 0.005. The same permutation procedure is used for the conjunction analysis, except in this case the same permutations are applied to both the oral reading and lexical decision scores and the min-SVR-β calculation is applied to the two permutation maps to create a null distribution of min-SVR-β maps. A cluster size threshold was applied to achieve a family-wise error rate of 0.05 based on the distribution of the maximum cluster size in each of the voxelwise thresholded maps. This permutation approach accounts for lesion autocorrelation effects in LSM analyses (Kimberg et al., 2007). The Montreal Neurological Institute 152 (ICBM152) was used to identify neuroanatomical regions.

Results

The two-way mixed ANCOVA showed a significant main effect of group (F_(1,146) = 130.6; p = 5.2 × 10⁻²²; Fig. 2A), reflecting greater overall efficiency in controls than in stroke survivors. A significant Task × Group interaction (F_(1,146) = 53.5; p = 1.6 × 10⁻¹¹; Fig. 2A) showed that oral reading efficiency was disproportionately reduced relative to lexical decision in the stroke survivor group compared with controls. The main effect of Task was not significant (F_(1,146) = 0.1; p = 0.74). Age and Education did not show significant main effects [(F_(1,146) = 0.6; p = 0.43), (F_(1,146) = 2.1; p = 0.14), respectively], nor did Education interact with Task (F_(1,146) = 0.9; p = 0.32). A significant Task × Age interaction (F_(1,146) = 8.6; p = 0.004) indicated that the difference between tasks varied slightly with age such that older age reduced lexical decision efficiency relatively more than it did oral reading. All participants performed above chance on the lexical decision task. Subject-level data are provided in Dataset S1.

Figure 2.

Group differences and task relationships in reading efficiency. A, Box and whisker plots showing oral word reading (purple) and lexical decision (orange) efficiency for controls and stroke survivors. Lexical decision efficiency was scaled so that chance performance corresponded to zero in order to match it better to oral reading. Gray lines connect data points form the same participant. Two-way mixed ANCOVA (covarying for age and education) revealed a significant main effect of group and a significant Task × Group interaction, reflecting greater overall efficiency in controls and disproportionately reduced oral reading efficiency in stroke survivors relative to controls. B, Relationship between oral reading and lexical efficiency in controls (cyan) and stroke survivors (navy). Regression lines and 95% confidence intervals are shown separately for controls and stroke survivors. Both groups showed a positive relationship between tasks, indicating shared cognitive resources for word recognition. The large shift left for stroke survivors and the small shift down for stroke survivors relative to controls again illustrates the disproportionate effect of strokes on oral reading relative to lexical decision. The dotted line represents the unit line, reflecting equal efficiency across tasks. Although lexical decision efficiencies were scaled so that chance performance was equivalent to oral reading, caution is warranted in directly comparing the efficiency scores within groups, since one can have some accurate trials on lexical decision by guessing, while this is not likely for oral reading.

The same pattern of results was observed when the analysis repeated using unscaled lexical decision efficiency values, with a significant main effect of Group (F_(1,146) = 130.9; p = 4.8 × 10⁻²²) and a significant Task × Group interaction (F_(1,146) = 59.1; p = 2.0 × 10⁻¹²), but no main effect of Task (F_(1,146) = 0.16; p = 0.69). Age and Education did not show significant main effects [(F_(1,146) = 0.4; p = 0.52), (F_(1,146) = 1.7; p = 0.20), respectively], nor did Task × Education (F_(1,146) = 0.3; p = 0.56), while the Task × Age interaction remained significant (F_(1,146) = 9.5; p = 0.002).

Pearson's correlations revealed a significant positive correlation between lexical decision and oral reading scores in both groups, suggesting that, on average, individuals with greater lexical decision efficiency also demonstrated greater oral reading efficiency (controls, r₍₆₈₎ = 0.54; p = 2.6 × 10⁻⁶; stroke, r₍₈₂₎ = 0.62; p = 6.2 × 10⁻¹⁰; Fig. 2B). Partial correlations remained significant after adjusting for age, education, and lesion volume (controls, r₍₆₈₎ = 0.55; p = 1.6 × 10⁻⁶; stroke, r₍₈₂₎ = 0.54; p = 3.0 × 10⁻⁷). Within the stroke group, the relationship between tasks was also preserved after adjusting for age, education, and WAB-AQ (r₍₈₂₎ = 0.52; p = 1.0 × 10⁻⁶). Notably, there was considerable unexplained variance in the relationship between the tasks, as evidenced by the scatter in Figure 2B, demonstrating that an individual stroke survivor's deficits can differ between oral reading and lexical decision.

A lesion overlap map demonstrated broad coverage of frontal, parietal, and superior temporal areas involved in reading and language processing (Fig. 3). SVR-LSM analyses yielded one significant cluster associated with lower oral reading efficiency in the STG (cluster size, 6.39 cm³; cluster p = 0.033; Fig. 4). When covarying for lexical decision efficiency, lesions in a cluster centered on SMG and extending into the STG, pre- and postcentral gyri, and the rolandic operculum were related to lower oral reading efficiency (cluster size, 14.97 cm³; cluster p = 0.0042; Fig. 4). Lesions to AG and MTG related to lower lexical decision efficiency (cluster size, 12.91 cm³; cluster p = 0.008; Fig. 5). When covarying for oral reading efficiency, lesions in a cluster centered on AG were related to lower lexical decision efficiency (cluster size, 7.58 cm³; cluster p = 0.02; Fig. 5). The conjunction analysis revealed that lesions to the AG extending into the posterior occipitotemporal sulcus (OTS), posterior SMG, the posterior STG, MTG, and underlying white matter were related to lower efficiency in both tasks (cluster size, 24.56 cm³; cluster p = 0.0005; Fig. 6). Table 2 summarizes cluster locations and sizes identified through SVR-LSM.

Figure 3.

Lesion overlap map (n = 85). Only voxels with at least five overlapping lesions, shown in the purple–yellow scale, were examined in the SVR-LSM analyses.

Figure 4.

SVR-LSM results for oral reading. Lesions associated with reduced oral word reading efficiency localized to the STG (red overlay, p = 0.0327). Lesions associated with lower oral reading efficiency after covarying for lexical decision efficiency centered on the SMG and extended into the rolandic operculum and STG (blue overlay, p = 0.0042). Purple regions represent overlap between the two maps. Results are shown at voxelwise p < 0.005 and clusterwise p < 0.05, controlling for the lesion size, age, and education.

Figure 5.

SVR-LSM results for lexical decision. Lesions associated with reduced lexical decision efficiency localized to the AG (yellow overlay, p = 0.008). Lesions associated with reduced lexical decision after covarying for oral reading efficiency centered on the AG (magenta overlay, p = 0.0245). Orange regions represent overlap between the two maps. Results are shown at voxelwise p < 0.005 and clusterwise p < 0.05, controlling for lesion size, age, and education.

Figure 6.

SVR-LSM conjunction analysis results. Lesions associated with reduced efficiency in both oral word reading and lexical decision tasks localized to the left AG extending into the posterior SMG, STG, and MTG and underlying white matter (green overlay, p = 0.0005). Results are shown at voxelwise p < 0.005 and clusterwise p < 0.05, controlling for lesion size, age, and education.

Table 2.

Cluster locations and sizes

Task	Cluster P	Cluster volume (cm3)	Center of mass (x, y, z)	Anatomical sites
Oral word reading	0.0327	6.39	MNI = −54.6, −40, 9.7	Left STG; SMG; arcuate posterior segment
Oral word reading covary for lexical decision	0.0042	14.97	MNI = −51, −19.1, 16.9	Left posterior lateral sulcus; STG, arcuate anterior segment; SMG; rolandic operculum; pre- and postcentral gyrus
Lexical decision	0.008	12.91	MNI = −42.6, −55.1, 22.7	Left MTG; inferior occipitofrontal fasciculus; AG
Lexical decision covary for oral word reading	0.0245	7.58	MNI = −38, −60, 26.5	Left AG, middle occipital gyrus
Conjunction	0.0005	24.56	MNI = −46.7, −47.9, 16.4	Left AG; SMG; intraparietal sulcus; posterior MTG; STG; posterior occipitotemporal gyrus

MNI, Montreal Neurological Institute coordinates. Anatomical sites were defined by Montreal Neurological Institute 152 (ICBM152).

Discussion

Our results suggest that oral reading and silent reading, as measured by lexical decision, rely on partially overlapping and partially distinct neural substrates. Lesions centered on the AG, posterior STG, and MTG resulted in deficits in both tasks. Lesions centered on the SMG resulted in relatively greater oral reading deficits, while lesions to the AG resulted in relatively greater lexical decision deficits.

Consistent with the idea that oral reading and silent reading rely on shared neurocognitive resources, a strong linear relationship between the two tasks was observed. The conjunction LSM analysis revealed shared lesion effects on both tasks in the AG, posterior STG, and MTG regions consistently implicated in lexical–semantic processing. The conjunction cluster also extended into the posterior OTS, near the visual word form area, and a functionally defined region supporting access to orthographic representations (Hillis et al., 2005; Dehaene and Cohen, 2011). Its involvement in both oral reading and lexical decision is expected, as both tasks require rapid visual-orthographic encoding to access semantics or phonology.

These findings align with long-standing models of reading that assign the AG a central role in mediating between visual, phonological, and semantic representations (Dejerine, 1891; Geschwind, 1965). This view of the AG was questioned after early functional imaging studies of reading did not reliably activate the AG (Fiez and Petersen, 1998; Turkeltaub et al., 2002). These early studies tended to use passive control tasks (e.g., fixation). Since the AG is part of the default mode network, it activates more for these passive tasks than active tasks, like reading (Seghier, 2013; Kim et al., 2022). A more recent parametric analysis found that the AG activates in proportion to the frequency and imageability of words during reading aloud, suggesting a role in processing semantic contributions to reading (Graves et al., 2010).

Similarly, LSM studies of oral reading have often not identified the AG as critical structure for reading (Piras and Marangolo, 2009; Dickens et al., 2019; Dyslin et al., 2025; Staples et al., 2025). Indeed, the LSM of oral reading in the present study did not identify the AG. These null results, however, may simply result from the limitations of the LSM technique. While damage to many different brain structures may result in deficits in complex tasks like reading, LSM compares lesions at each location in the brain to lesions elsewhere in the brain. This procedure tends to identify only the lesion locations that are most consistently related to the most severe deficits. Lesions that result in less consistent or less severe deficits may be missed because they are contrasted with lesions elsewhere that produce greater deficits. Multivariate LSM methods like SVR-LSM address this limitation to some degree (Mah et al., 2014), but results remain highly similar to those of univariate LSM (Ivanova et al., 2021), demonstrating that they do not completely address the issue. As a result, prior LSM studies of reading, which have focused on oral reading alone, may have been biased toward lesions that primarily affect orthography-to-phonology mapping or speech output processing, including the STG, SMG, and premotor cortex, to the exclusion of lesions affecting lexical semantic processes involved in both silent and oral reading (Woollams et al., 2007; Piras and Marangolo, 2009; Rapcsak et al., 2009; Ripamonti et al., 2014; Binder et al., 2016; Dickens et al., 2019; Li et al., 2021; Dyslin et al., 2025; Staples et al., 2025).

Recent evidence suggests the AG serves as a cross-modal hub involved in semantic integration and conceptual feature retrieval (Seghier, 2013; Xu et al., 2017; Kim et al., 2022). In addition to its semantic role, the AG has been proposed to function as an orthographic working memory system or orthographic buffer, which temporarily maintains letter identity and order during reading and spelling tasks (Rapp et al., 2016). Supporting this, a recent LSM study (Purcell et al., 2021) found that the left AG specifically supports orthographic working memory, distinguishing it from phonological working memory, which is associated with the SMG. While the AG is not always activated during simple word reading in skilled adults, it shows stronger engagement in tasks with greater semantic or working memory demands, such as lexical decision and sentence comprehension (Binder et al., 2005; Price and Mechelli, 2005), corresponding with the LSM results here demonstrating that AG lesions relate to lexical decision deficits even after controlling for oral reading.

Similarly, the MTG has been implicated in lexical–semantic access and semantic-to-phonological mapping (Graves et al., 2010), suggesting its involvement in both oral and silent reading. These overlapping effects support the idea that both tasks draw on shared ventral stream regions for semantic processing. In particular, the posterior MTG has been proposed to function as a lexical interface, which links phonological and semantic information (Hickok and Poeppel, 2007). A recent LSM study using a partly overlapping cohort to the present study to examine imageability effects in oral reading found that lesions to the posterior MTG and superior temporal sulcus specifically disrupt semantic-to-phonological mappings that support oral word reading (Staples et al., 2025). This finding suggests that this region acts as a critical computational hub for integrating semantic and phonological information. In this context, the present findings suggest that although oral and silent readings diverge in their reliance on articulatory–phonological systems, they converge on shared lexical mechanisms in the MTG, including where meaning is mapped onto phonological information.

While the above evidence clearly supports shared neurocognitive substrates of silent reading and oral reading, group-level deficits in the stroke group compared with controls were worse on oral reading than lexical decision. This suggests that oral reading builds on processes used for silent word recognition, adding processes required for speech production, which are commonly impaired by left-hemisphere strokes. Consistent with this idea, we identified unique lesion effects for oral reading when controlling for lexical decision in regions associated with speech production including the SMG and rolandic operculum (Buchsbaum et al., 2001; Blank, 2002; Dronkers and Ogar, 2004). The SMG is proposed to contribute to the phonological processing of aloud reading, mediating transformation of orthographic input into articulatory output (Price and Mechelli, 2005; Price, 2012). Consistent with this, prior LSM studies have found that SMG lesions produce relatively greater oral reading deficits for pseudowords compared with words (Dickens et al., 2019; Dyslin et al., 2025). The LSM results of the present study, in which SMG lesions were associated with oral reading deficits when covarying for lexical decision performance, add another line of evidence for the role of SMG in transforming orthographic input to articulatory output during oral reading.

These findings have potential implications for assessment of poststroke alexia. Silent word recognition cannot be reliably inferred from oral reading performance alone, as demonstrated by considerable unexplained variance observed in the relationships between the tasks. Assessments that focus solely on oral reading may bias alexia diagnosis toward reading deficits related to orthographic-to-phonological mapping and speech production processes. These assessments may miss deficits in orthographic-to-semantic mapping or lexical–semantic processing that cause more subtle deficits in oral reading but also impair silent word recognition and comprehension. Including lexical decision in clinical batteries may help differentiate levels of reading impairment in alexia, thereby informing interventions tailored to the specific component of the reading network that is impaired.

While lexical decision is a practical proxy for silent single-word reading, particularly in comparison with oral single-word reading, it also has important limitations. Most importantly, lexical decision does not strictly require comprehension, reducing its relevance to real-life silent reading. However, computational models and behavioral studies have shown that semantic variables systematically influence lexical decision performance, indicating access of semantic representations during word recognition (James, 1975; Plaut, 1997; Dilkina et al., 2010). Additionally, word-level comprehension tasks introduce additional nonreading demands. For example, word-to-picture matching requires visual search, object recognition, and semantic comparisons influenced by task strategy (e.g., process of elimination). Deficits in these abilities can affect task performance irrespective of reading comprehension ability. Although not a perfect analog for naturalistic silent reading, lexical decision provides a simple, clinically feasible silent word recognition task to compare to oral word reading. Future work should compare lexical decision to silent reading comprehension tasks using parallel written and auditory versions to control for additional task demands.

Single-word reading is less ecologically valid than sentence or passage reading comprehension. However, establishing the relationship between single-word oral and silent reading provides a foundation for future studies of sentence-level comprehension. Single-word reading ability accounts for a portion of text-level comprehension performance in alexia (Smith and Ryan, 2020). Understanding these relationships may inform alexia treatment as oral reading treatments have been shown to improve reading comprehension in some but not all individuals with alexia (Purdy et al., 2019).

Conclusion

This study provides lesion-based evidence on the neurocognitive substrates of oral and silent reading. Our findings suggest that they rely on partly shared and partly distinct neural substrates. Critically, silent reading abilities cannot be inferred from oral reading performance alone, with strokes tending to affect oral reading more than silent word recognition. This finding highlights the need for more comprehensive assessment tools that evaluate both modalities in parallel. These insights may also inform targeted interventions for individuals with alexia following stroke depending on their level of deficit.