Abstract
Reading words aloud is a fundamental aspect of literacy. The rapid rate at which multiple distributed neural substrates are engaged in this process can only be probed via techniques with high spatiotemporal resolution. We probed this with direct intracranial recordings covering most of the left hemisphere in 46 humans (26 male, 20 female) as they read aloud regular, exception and pseudo-words. We used this to create a spatiotemporal map of word processing and to derive how broadband γ activity varies with multiple word attributes critical to reading speed: lexicality, word frequency, and orthographic neighborhood. We found that lexicality is encoded earliest in mid-fusiform (mFus) cortex, and precentral sulcus, and is represented reliably enough to allow single-trial lexicality decoding. Word frequency is first represented in mFus and later in the inferior frontal gyrus (IFG) and inferior parietal sulcus (IPS), while orthographic neighborhood sensitivity resides solely in IPS. We thus isolate the neural correlates of the distributed reading network involving mFus, IFG, IPS, precentral sulcus, and motor cortex and provide direct evidence for parallel processes via the lexical route from mFus to IFG, and the sublexical route from IPS and precentral sulcus to anterior IFG.
SIGNIFICANCE STATEMENT Reading aloud depends on multiple complex cerebral computations: mapping from a written letter string on a page to a sequence of spoken sound representations. Here, we used direct intracranial recordings in a large cohort while they read aloud known and novel words, to track, across space and time, the progression of neural representations of behaviorally relevant factors that govern reading speed. We find, concordant with cognitive models of reading, that known and novel words are differentially processed through a lexical route, sensitive to frequency of occurrence of known words in natural language, and a sublexical route, performing letter-by-letter construction of novel words.
Introduction
Reading a word aloud depends on multiple complex cerebral computations: mapping the visual input of a letter string to an internal sequence of sound representations, and then their expression through orofacial motor articulations. Models of how this mapping occurs have postulated a dual-route architecture (Coltheart et al., 2001; Perry et al., 2007, 2010, 2019; Taylor et al., 2013), with a lexico-semantic route for rapidly reading known words and a sublexical route for constructing phonology of novel words. Contrasts between phonological exception words and pseudowords (Fiebach et al., 2002; Shim et al., 2012; Taylor et al., 2013; Sebastian et al., 2014) are often used to maximally separate these two routes. Exception words contain irregular grapheme-phoneme associations (e.g., yacht, sew), and their pronunciations must therefore be retrieved from internal lexical representations, as they cannot be accurately constructed de novo. In contrast, pseudowords have no stored representations, and their phonology must be constructed rather than retrieved. Finally, regular words may be read correctly by either route, with the relative speeds of the routes depending on the level of automatization of reading as well as the size of the reader's vocabulary.
At the brain level, the two routes are thought to share an initial stage of visual word form processing in left occipitotemporal cortex (Dehaene and Cohen, 2011), with only small topological changes depending on whether the task favors lexical or phonological processing (Bouhali et al., 2019). More anteriorly, ventral temporal cortex is strongly implicated as mediating the lexical route, with mid-fusiform cortex (mFus) likely functioning as the orthographic lexicon, a region where familiar letter strings are mapped onto known words (Nobre et al., 1994; Kronbichler et al., 2004; Glezer et al., 2015; Hirshorn et al., 2016; Lochy et al., 2018; White et al., 2019; Liu et al., 2021; Woolnough et al., 2021, 2022). mFus is sensitive to lexicality and word frequency (Kronbichler et al., 2004; White et al., 2019; Woolnough et al., 2021), and its activity is modulated by visual word learning (Taylor et al., 2014a; Glezer et al., 2015). Conversely, the sublexical route, essential for articulating novel words, is thought to engage the inferior parietal lobe (IPL). The dysfunction of subregions of the IPL is associated with specific forms of reading deficits, most prominently impairing pseudoword reading (Temple et al., 2003; Raschle et al., 2011; Rapp et al., 2016; Dickens et al., 2019; Tomasino et al., 2020), in addition to broader phonological and semantic deficits (Binder et al., 2009; Hula et al., 2020; Numssen et al., 2021). The two routes of reading are presumed to be active in parallel (Simos et al., 2002) and converge in the inferior frontal gyrus (IFG) (Taylor et al., 2013).
Cognitive models of reading result in predictions for the overlap or separation between the regions activated by different word classes at each stage of processing. For example, the orthographic lexicon is predicted to be sensitive to word frequency, whereas the phoneme system is predicted to be sensitive to both lexicality and phonological regularity (Taylor et al., 2013). The theoretical embedding of the routes of reading in neural architecture has primarily been derived from lesion data and fMRI (Jobard et al., 2003; Ripamonti et al., 2014; Bouhali et al., 2019; Tomasino et al., 2020), but these modalities lack the spatiotemporal resolution to delineate the dynamics of these networks and their real-time interactions that allow us to rapidly read. For example, is word recognition primarily a serial, feedforward process, or can lexical effects also propagate in a top-down manner all the way back to visual cortices (Carreiras et al., 2014; Woolnough et al., 2021)? To attain the high temporal resolution needed to resolve such issues, we used intracranial recordings in a large cohort of patients (46 patients, 3846 electrodes), with medically intractable epilepsy, while they read aloud known and novel words, creating a comprehensive 4D map of the spatiotemporal dynamics of behaviorally important lexical and sublexical processes throughout the reading system.
Materials and Methods
Data and code availability
The raw datasets generated from this research are not publicly available due to their containing information noncompliant with HIPAA, and the human participants from whom the data were collected have not consented to their public release. We have released summary statistics at https://osf.io/e6zd9/.
Participants
Forty-six patients (26 male, 19-60 years, 5 left-handed, IQ 94 ± 14, age of epilepsy onset 18 ± 10 years) participated in the experiments after giving written informed consent. All participants were semi-chronically implanted with intracranial electrodes for seizure localization of pharmaco-resistant epilepsy. Participants were excluded if they had confirmed right-hemisphere language dominance or a significant additional neurologic history (e.g., previous resections, MRI abnormalities such as malformations or hypoplasia). One additional participant was tested but excluded from analysis as their response times (RTs) were considered outliers for all tested word classes. All experimental procedures were reviewed and approved by the Committee for the Protection of Human Subjects of the University of Texas Health Science Center at Houston as Protocol #HSC-MS-06-0385.
Electrode implantation and data recording
Data were acquired from either subdural grid electrodes (SDEs; 4 patients) or stereotactically placed depth electrodes (sEEGs; 42 patients). SDEs were subdural platinum-iridium electrodes embedded in a silicone elastomer sheet (PMT; top-hat design; 3-mm-diameter cortical contact), and were surgically implanted via a craniotomy (Tandon, 2012; Pieters et al., 2013; Tong et al., 2020). sEEGs were implanted using a Robotic Surgical Assistant (Medtech) (Tandon et al., 2019; Rollo et al., 2020). Each sEEG probe (PMT) was 0.8 mm in diameter and had 8-16 electrode contacts. Each contact was a platinum-iridium cylinder, 2.0 mm in length and separated from the adjacent contact by 1.5-2.43 mm. Each patient had 12-20 such probes implanted. Following surgical implantation, electrodes were localized by coregistration of preoperative anatomic 3T MRI and postoperative CT scans in AFNI (Cox, 1996). Electrode positions were projected onto a cortical surface model generated in FreeSurfer (Dale et al., 1999), and displayed on the cortical surface model for visualization (Pieters et al., 2013). Intracranial data were collected during research experiments starting on the first day after electrode implantation for sEEGs and 2 d after implantation for SDEs. Data were digitized at 2 kHz using the NeuroPort recording system (Blackrock Microsystems), imported into MATLAB, initially referenced to the white matter channel used as a reference for the clinical acquisition system, and visually inspected for line noise, artifacts, and epileptic activity. Electrodes with excessive line noise or localized to sites of seizure onset were excluded. Each electrode was rereferenced to the common average of the remaining channels. Trials contaminated by interictal epileptic spikes were discarded.
Stimuli and experimental design
All patients undertook a task reading aloud single words and pseudowords. Stimuli were presented on a 2880 × 1800 pixel, 15.4 inch LCD screen positioned at eye-level, 2-3 feet from the patient. Participants were presented with 80 each of monosyllabic (1) phonologically regular words, (2) phonologically irregular exception words, and (3) novel pseudowords and asked to read them aloud. Stimuli were presented using Psychophysics Toolbox (Kleiner et al., 2007) in MATLAB, in all lowercase letters, in Arial font with a height of 150 pixels (∼2.2° visual angle). Each stimulus was displayed for 1500 ms with an interstimulus interval of 2000 ms. Stimuli were presented in two recording sessions, each containing presentation of 120 stimuli in a pseudorandom order with no repeats.
Signal analysis
Analyses were performed by first bandpass filtering raw data of each electrode into broadband γ activity (BGA; 70-150 Hz) following removal of line noise (zero-phase second order Butterworth bandstop filters). A frequency domain bandpass Hilbert transform (paired sigmoid flanks with half-width 1.5 Hz) was applied, and the analytic amplitude was smoothed (Savitzky-Golay finite impulse response, third order, frame length of 201 ms). BGA is presented here as percentage change from baseline level, defined as the period −500 to −100 ms before each word presentation.
Electrodes were tested to see whether they were responsive during the task. Responsiveness was defined as displaying >20% average BGA over baseline for at least one of the three following windows: 100-500 ms following stimulus onset, −500 to −100 ms before articulation onset, or 100-500 ms following articulation onset. Of the 3846 useable electrodes, 1248 electrodes were designated responsive based on these criteria.
ROIs were selected in canonical language areas, areas identified in previous intracranial studies of reading (Hirshorn et al., 2016; Lochy et al., 2018; Woolnough et al., 2019, 2021), and areas of high peak activation in the word-type independent mixed-effects, multilevel analysis (MEMA) movie. ROI centers were defined on the cortical surface, and all responsive electrodes within a set geodesic radius of this point were included (Kadipasaoglu et al., 2015). This method was selected as currently available cortical surface parcellations have been shown to be inadequate at predicting functional boundaries of task-evoked activity (Zhi et al., 2021). An exception was made for the superior temporal gyrus (STG) ROI, which was defined using a Human Connectome Project derived parcellation (A1, A4, PBelt, MBelt, LBelt) (Glasser et al., 2016). Centers of mass for each of the left hemispheric ROIs, in Talairach space, were as follows: lateral occipitotemporal cortex (LOT), −41, −61, −13; mFus, −31, −36, −18; inferior parietal sulcus (IPS), −28, −62, 34; precentral sulcus (pCS), −37, 0, 39; posterior inferior frontal gyrus (pIFG), −42, 17, 26; anterior IFG (aIFG), −39, 37, 15; frontal operculum (FO), −35, 17, 9; ventral motor cortex (vMC), −56, −10, 21; supramarginal gyrus (SMG), −37, −29, 35; supplementary motor area (SMA), −5, −3, 46; posterior insula (PI), −36, −14, 10; STG, −53, −20, 5.
For ROI-based analyses, we averaged the BGA responses of all responsive electrodes within-ROI, within-patient before calculating the across-patient mean. This method minimizes bias related to between-patient variations in electrode coverage within each ROI. Frequentist statistical methods were corrected for multiple comparisons using a Benjamini–Hochberg false detection rate threshold of q < 0.05.
Neural decoding
Decoding analyses were performed using logistic regression classifiers, using fivefold cross validation, implemented within MNE-Python (Gramfort, 2013; Gramfort et al., 2014). For each patient, decoding performance was summarized with an area under the curve and a set of classifier weights for each electrode. Temporal decoding was performed on BGA using a sliding estimator at each time point, using all available electrodes. Spatial distribution of classifier weights was reconstructed by a cortical surface transform onto a standardized brain surface using each electrode's presumed “recording zone,” an exponentially decaying geodesic radius (Kadipasaoglu et al., 2014; McCarty et al., 2022). Cortical surface maps were amplitude normalized within patient, then averaged across patient to create a population weighting map.
Linguistic analysis
We quantified word frequency as the base-10 log of the SUBTLEXus frequency (Brysbaert and New, 2009). This resulted in a frequency of 1 meaning 10 instances per million words and 4 meaning 10,000 instances per million words. There was no significant difference between word frequency of regular (1.5 ± 0.35; mean ± SD) and exception (1.7 ± 1.0) words (Wilcoxon rank sum, p = 0.36). Positional letter frequency was calculated as the base-10 log of the sum of the SUBTLEXus frequencies of all words with a given letter in a specific ordinal position. Orthographic neighborhood was quantified as the orthographic Levenshtein distance (OLD20); the mean number of single character edits required to convert the word into its 20 nearest neighbors with a log frequency >0 (Yarkoni et al., 2008). Phonological neighborhood densities were obtained from the Irvine Phonotactic Online Dictionary (Vaden et al., 2009). Pseudowords were phonemically transcribed using the most common pronunciation.
Results
Participants read aloud phonologically regular words, exception words, and novel pseudowords (Fig. 1A) while concurrent recordings were performed using 3846 separate intracranial electrodes placed for the localization of intractable epilepsy (Fig. 1B,C). Forty-two participants had depth recordings using sEEGs and 4 had SDEs.
Experimental design and electrode coverage. A, Schematic representation of the reading task. B, Representative coverage map (46 patients) and (C) individual electrode locations (3846 electrodes) for the left hemisphere, highlighting responsive electrodes (1248 electrodes; >20% activation above baseline).
Behavioral analysis
Mean (±SD) RTs were as follows: regular words (734 ± 113 ms), exception words (738 ± 113 ms), and pseudowords (911 ± 162 ms) (Fig. 2A). Regular and exception words showed no difference in RT (Wilcoxon sign rank, p = 0.90; ln(Bayes Factor (BF10)) = −1.6), although pseudoword RT was slower than for exception words (p < 10−8, ln(BF10) = 32); 95 ± 4% of trials were correctly articulated. The most common errors were regularization of exception words (e.g., sew as sue, soot as sute) or lexicalization of pseudowords (e.g., shret as shirt, jinje as jingle). Error trials were excluded from analysis.
Population word RTs. A, RT distribution for each of the three word classes, averaged within participant. B, Sorted mean (±SE) RTs for each item within the three word classes, averaged across participants, highlighting some exemplar words.
To determine the word features that modulate RT, we performed linear mixed effects (LME) and BF analyses on each word class with fixed effects modeling linguistic factors commonly linked to word identification and articulation (Table 1). Regular and exception word RTs showed greatest modulation by word frequency. Exception words also displayed modulation by orthographic neighborhood, and we observed a significant interaction between regularity and orthographic neighborhood (LME: t(6589) = −4.2, β = −62.5, p < 10−4, 95% CI −91.7 to −33.3). Pseudoword RT was most strongly associated with orthographic neighborhood, meaning pseudowords with many known word neighbors were articulated faster.
Statistical modeling of RTa
Spatiotemporal mapping of single word reading
We used an MEMA of BGA (70-150 Hz) in group surface normalized space to create a population-level map of cortical activation across the population. This analysis is specifically designed to account for sampling variations and to minimize effects of outliers (Fischl et al., 1999; Argall et al., 2006; Saad and Reynolds, 2012; Esposito et al., 2013; Conner et al., 2014; Kadipasaoglu et al., 2014). All correctly articulated trials were used. A 4D representation of activation on the cortical surface was generated by collating MEMA on short, overlapping time windows (150 ms width, 10 ms spacing) to generate successive images of cortical activity, time-locked to stimulus onset (Movie 1; Fig. 3A) or the onset of articulation (Movie 2). The spatial distribution of activations was highly consistent across word classes (Fig. 4).
Spatiotemporal profile of cortical activations. A, Collapsed articulation-locked activation movie (Movie 2) highlighting the amplitude of peak BGA. B, Representative ROIs in 12 anatomically and functionally distinct regions, showing all responsive electrodes. C, Mean activation during word reading of each ROI, averaged within patient, time-locked to stimulus onset (left) and articulation onset (right). SEs omitted for visual clarity. Colored bars represent regions of activation greater than baseline (Wilcoxon signed-rank, q < 0.05).
Conjunction map of word class activations. MEMA conjunction maps showing overlap of binarized activation maps of each of the three word classes tested (%BGA > 5%, t > 2.58, patients ≥ 3), over three time windows locked to stimulus onset. Across all time windows, all three word classes demonstrate a gross overlap of activation (white). In the later time window, areas associated with postarticulatory processes (e.g., auditory cortex) show selective activation for known words, reflecting differences in RT between known words and novel pseudowords. Regions in black did not have consistent coverage for reliable MEMA results.
Spread of stimulus-locked activity across the cortical surface. MEMA movie of the time course of BGA across the cortical surface with trials time-locked to onset of the visual stimulus. Regions in black did not have consistent coverage for reliable MEMA results.
Spread of articulation-locked activity across the cortical surface. MEMA movie of the time course of BGA across the cortical surface with trials time-locked to the onset of articulation. Regions in black did not have consistent coverage for reliable MEMA results.
By collapsing across these frames, we visualized peak activations at each point on the cortical surface (Fig. 3A). Further, to create a more focused visualization of the spatiotemporal progression across reading-sensitive cortex, we selected ROIs at sites with maximal activation that also corresponded to sites believed to be important for written word processing, speech production, and speech monitoring (Fig. 3B,C).
Spatiotemporal representation of lexical factors
To distinguish activity patterns across word classes, we contrasted grouped γ power activations between exception versus pseudowords (lexicality) and exception versus regular words (regularity), using whole-brain MEMAs. The lexicality contrasts demonstrated clusters in early visual cortex, mFus, pCS, IPS, and aIFG (Fig. 5A). The regularity contrast only demonstrated a small cluster in pCS (Fig. 5B).
Contrasting word classes. A, B, MEMA contrasts of (A) exception versus pseudoword, and (B) exception versus regular, revealing regions of significantly different mean BGA between conditions (p < 0.01 corrected) within each time window. Regions in black did not have consistent coverage for reliable MEMA results. C, Decoding accuracies of the logistic regression decoders trained to distinguish exception word versus pseudoword trials (left) and exception word versus regular word trials (right). Gray lines indicate individual patient decoding accuracies. Colored line indicates median accuracy. Colored bars represent time periods significantly greater than chance (Wilcoxon signed-rank, q < 0.05). D, Cortical surface representation of population average electrode weightings of the exception versus pseudoword decoder between 300 and 500 ms.
To distinguish whole-brain activity patterns for each of these factors, within-individuals at a single-trial level, we used a logistic regression decoder. Decoders trained to distinguish between exception word and pseudoword trials demonstrated high decoding accuracy (Peak median accuracy: 63%), with 5 patients showing >80% decoding accuracy (peak individual accuracy: 85%) (Fig. 5C). These lexicality decoders displayed high electrode weightings in the same regions as the lexicality contrasts listed above with the addition of the superior frontal sulcus and the ventral visual stream; thus, an independent analysis corroborates the critical role of the LOT, mFus, pCS, IPS, and aIFG in lexicality processing (Fig. 5D). Decoders trained to distinguish exception and regular words did not show significantly greater decoding accuracy than in the baseline period.
To characterize the timing of lexicality distinctions between known words (regular and exception) and novel pseudowords broadly across ROIs (Fig. 6A,B), we compared their activity by word class. Lexicality distinctions were observed earliest in mFus and subsequently in IFG and IPS, with stronger activity for pseudowords in all cases. Distinctions were also observed in the opposite direction in postarticulatory auditory regions (PI and STG). In those areas, the faster RT for known words compared with novel words led to an earlier activation for known words as the patients heard themselves earlier.
Spatiotemporal activation profiles of known and novel words. A, Mean activation (±SE) for each word class, within each ROI, during word reading, averaged within patient, time-locked to stimulus onset. Number of electrodes and patients, per ROI, is indicated. Colored bars represent regions of significant difference from exception words (Wilcoxon signed-rank, p < 0.05 for >100 ms). B, Mean activation per unique word, within each ROI. Trials separated by word class and sorted by the word's mean RT for the patients included within the given ROI. Mean RT for each word is highlighted. C, Latency of first onset of activation (first derivative of BGA >3.5 SD above baseline) for each electrode within each ROI. D, E, Network representations demonstrating maximum likelihood latency differences between ROIs, within patients with simultaneous coverage, for (D) initial activation latency and (E) initial lexicality distinction latency (first derivative of pseudowords, known words >3.5 SD above baseline). Lines excluded for ROI pairs where there was no simultaneous coverage of significant electrodes. Arrowheads not shown for differences <10 ms.
As all the previous analyses had demonstrated minimal differences between exception and regular words within the ROIs tested, from this point, we combined responses across all known words to increase power in these more fine-grained analyses.
The activation evoked by words initially spread anteriorly across ventral temporal cortex before progressing to parietal and frontal cortices (Fig. 6C; Movie 1). Within patients who had simultaneous electrode coverage within multiple ROIs, we compared the onset latencies for activation relative to the prestimulus baseline, combining across all word classes (Fig. 6D), versus for the lexicality distinction (Fig. 6E). This analysis demonstrated a spread of the activation from LOT to mFus and IPS before spreading to pIFG, then aIFG. In contrast, the earliest lexicality distinctions were observed in mFus, with a subsequent spread backward to LOT (Fig. 7) and forward to aIFG. None of the individual tested pIFG electrodes showed a significant lexicality distinction.
Anterior-to-posterior spread of lexicality in ventral temporal cortex. A, Ventral temporal ROI electrodes on an N27 pial surface. B, Mean activation (±SE) for known words (regular and exception) and pseudowords, within each ROI, during word reading, averaged within patient, time-locked to stimulus onset. Number of electrodes and patients, per ROI, is indicated. Colored bars represent regions of a significant effect of lexicality (LME, q < 0.05, effects of word length regressed out). EVC, Early visual cortex (Talairach center of mass −17, −81, −15).
For the six ROIs that showed a clear prearticulatory peak in activation, we analyzed their activity for sensitivity to the main drivers of RT seen in the behavioral analysis: lexicality, word frequency of known words, and orthographic neighborhood for pseudowords. mFus showed the earliest sensitivity to lexicality, followed by LOT and pCS, and then broad sensitivity across multiple regions (Fig. 8A). All these regions demonstrated greater activity for pseudowords over known words. mFus showed an early and long-lasting word frequency sensitivity, with IPS and aIFG becoming sensitive later (500-700 ms). Frequency sensitivity manifested as greater activation for low-frequency words. Sensitivity to orthographic neighborhood of pseudowords was only seen in IPS (500-700 ms), with pseudowords that had fewer close neighbors showing greater activity. In the right hemisphere, we observed effects of lexicality only in LOT and IPS (with no involvement of right mFus) and no effect of word frequency or orthographic neighborhood. (Fig. 9).
Regression of lexical factors. A, BF analysis of lexicality, word frequency, and orthographic neighborhood effects in the six prearticulatory ROIs, for three time windows. BF analysis was used here because of the variability in coverage between ROIs. Frequentist methods can lack statistical significance either because of a lack of effect or lack of power. BF analysis allowed us to disambiguate and show the strength of evidence for or against the presence of an effect. Lexicality tested all known words against pseudowords. Word frequency was regressed across all known words. Orthographic neighborhood was regressed across all pseudowords. ln(BF10) shown for each contrast, and values >2.3 are highlighted. B, C, LME model regression of (B) word frequency in known words and (C) orthographic neighborhood in pseudowords, in three ROIs (β ± SE; mFus, 62 electrodes, 21 patients; IPS, 21 electrodes, 9 patients; aIFG, 35 electrodes, 9 patients). Colored bars represent regions of significance (q < 0.05).
Spatiotemporal activation profiles in right hemisphere. A, Right hemisphere analogs of the main analysis ROIs, highlighting all 588 responsive electrodes (of 1712 implanted in right hemisphere). B, Mean activation (±SE) for each word class, within each ROI, during word reading, averaged within patient, time-locked to stimulus onset. Number of electrodes and patients, per ROI, is indicated. Colored bars represent regions of significant difference from exception words (Wilcoxon signed-rank, p < 0.05 for >100 ms). C, BF analysis of lexicality, word frequency, and orthographic neighborhood effects in the four prearticulatory ROIs, for three time windows. Lexicality tested all known words against pseudowords. Word frequency was regressed across all known words. Orthographic neighborhood was regressed across all pseudowords. ln(BF10) shown for each contrast and values >2.3 are highlighted. Centers of mass for each of the right hemispheric ROIs, in Talairach space, were as follows: LOT, 32, −61, −15; mFus, 39, −33, −21; IPS, 30, −57, 43; aIFG, 43, 30, 9; vMC, 50, −14, 32; SMG, 36, −30, 39; SMA, 5, −9, 48; PI, 38, −16, 12; STG, 56, −31, −4.
For the three regions with evidence of word frequency or orthographic neighborhood effects (mFus, IPS, and aIFG), we generated LME models at a higher time resolution. Sensitivity to word frequency was observed earliest in mFus (200 ms) followed by IPS and aIFG (425 ms) (Fig. 8B). In IPS, we observed a period of elevated orthographic neighborhood sensitivity, but this did not show significance at this time resolution (Fig. 8C). Evidence from fMRI suggests differential reading-related function within medial and lateral aspects of fusiform cortex (Bouhali et al., 2019; Nordt et al., 2021; Rosenke et al., 2021), but we found no strong evidence of topographic differences in degree of activation or sensitivity to word frequency across the medial-to-lateral extent of mFus (Fig. 10).
Functional topography of mFus. A, Ventral view of the pial surface, highlighting the medial (green) and lateral (pink) mFus sub-ROIs. B, Mean (±SE) BGA within each sub-ROI. Number of electrodes and patients, per sub-ROI, is indicated. C, β ± SE of the LME-derived effect of word frequency for each sub-ROI. Colored bars represent regions of significance (q < 0.05).
Discussion
This work comprehensively maps the spatiotemporal spread of cortical activation across the left hemisphere during word reading to derive the dynamics of cortical networks underlying literacy, using intracranial recordings in a large population. We find two regions demonstrating early selectivity to lexicality-mFus and pCS, with subsequent engagement of the IPS and IFG. This lexicality network is broadly comparable with the spatial substrates derived from fMRI (Heim et al., 2013; Taylor et al., 2013, 2014b), with an added benefit of millisecond temporal resolution. Lexicality-sensitive regions display maximal distinction between known words versus pseudowords 300-500 ms after stimulus onset, in a manner reliable enough to enable above-chance single-trial decoding of lexicality.
Three of the most prominent cognitive models of reading are the DRC (Coltheart et al., 2001), CDP+ (Perry et al., 2007, 2010, 2019), and Triangle (Seidenberg and McClelland, 1989), models that each result in predictions for the overlap or separation between the activation induced by different word classes at each stage of processing (Taylor et al., 2013). These models are designed to simulate human behavioral performance rather than neural activity but allow a theoretical decomposition of reading into multiple cognitive stages that may correlate with neural computations. Concordant with all three cognitive models and with previous lesion data, we find evidence of two parallel pathways: (1) a lexical route via mFus to aIFG that shows sensitivity to word frequency, with greater activity for low-frequency words, as predicted by the DRC and CDP+ models; and (2) a sublexical route, via IPS and pCS, that shows sensitivity to lexicality, with greater activity for pseudowords, as predicted by all three models, and sensitivity to orthographic neighborhood of pseudowords observed in IPS.
We also observed a greater activation induced by pseudowords relative to known words in all our ROIs. This observation diverges from the predictions of all three models, which predict that regions associated with the orthographic lexicon and semantic processing should display greater engagement for words than for pseudowords (Taylor et al., 2013). We have previously documented a task-based inversion of lexical selectivity within mFus, with words > pseudowords while passively viewing but pseudowords > words when actively reading (Woolnough et al., 2021). IPS has previously been implicated in providing top-down modulation of ventral temporal responses, enhancing representation of task-relevant features (Kayser et al., 2010; Kay and Yeatman, 2017). This suggests that resources used by the lexical route for pseudowords are dependent on the specific type of pseudoword processing required by a given task and may be influenced by the extent of engagement of the sublexical route, a task-dependency that is presently not captured by any of the dominant models.
We did not observe a reliable distinction between regular and exception word activations. However, the lack of effect could be because of the present stimulus set being comprised of short, monosyllabic words, as cognitive models would suggest that any differences may be more apparent when using longer, low-frequency words with irregularities early in the word (Rastle and Coltheart, 1999).
We have previously demonstrated that mFus is the earliest region in ventral temporal cortex to show sensitivity to word frequency while reading (Woolnough et al., 2021). It is often assumed that sensitivities to statistical properties of language, such as word frequency, seen in ventral temporal cortex, arise solely from a top-down modulation from IFG (Price and Devlin, 2011; Heim et al., 2013; Carreiras et al., 2014; Woodhead et al., 2014; Whaley et al., 2016; Liu et al., 2021). Here, we demonstrate again the primacy of mFus in coding both word frequency and lexicality: its sensitivity precedes the emergence of lexicality effects, both posteriorly in occipitotemporal and early visual cortices and anteriorly in aIFG and IPS, where these effects start ∼200 ms later. These findings suggest that the directional connectivity metrics derived in earlier studies may not necessarily detect the transmission of orthographic lexical information, but perhaps higher-order semantic or phonological information. Additionally, we observe that right LOT demonstrates lexical sensitivity at a comparable time to left LOT, subsequent to the lexicality distinction in left mFus. This likely indicates that left mFus provides top-down influences to both ipsilateral and contralateral visual areas. These results fit with the putative role of mFus as the locus of the orthographic lexicon, organized based on statistical regularities of individual words in natural language (Woolnough et al., 2021). Additionally, there is evidence of multimodal semantic processing within the vicinity of mFus (Forseth et al., 2018; Binder et al., 2020). It is possible that these functions are spatially dissociable as there is separation within mFus, within individuals, between cortical stimulation sites that lead to reading specific deficits and those that result in more general semantic deficits (Mani et al., 2008; Hirshorn et al., 2016; Woolnough et al., 2021). Future research should clarify whether orthographic and multimodal effects occur at the same mFus sites within individuals, thus determining whether this area may already implement abstract lexical semantics, or whether a purely orthographic stage of lexical processing can be isolated in mFus.
The IPS was the only region with sensitivity to the orthographic neighborhood of novel pseudowords. IPS has previously been implicated in the grapheme-phoneme conversion process (Dehaene-Lambertz et al., 2018; Xu et al., 2020). Given that IPS shows both word frequency and lexicality sensitivity, its role in sublexical processing might appear to be questionable. However, it should be remembered that both routes are thought to activate in parallel and compete for speed (Simos et al., 2002). For known words of high enough frequency, the lexical route is faster and more accurate than the sublexical route; thus, once a letter string is identified as a known lexical object, sublexical processes are no longer required (unless the word is visually degraded, thus promoting a return to slow left-to-right analysis) (Vinckier et al., 2006; Cohen et al., 2008). Given that the speed of lexical identification varies with word frequency, the timing of the cessation of sublexical processes should also be frequency dependent. This interpretation is entirely consistent with our data as IPS shows more sustained activity, but not higher peak activity for pseudowords compared with words. We also observed lexicality effects in right IPS, consistent with TMS studies demonstrating the importance of right, nondominant parietal cortex in phonological (Hartwigsen et al., 2010) and visuospatial processing (Cazzoli et al., 2015) during reading.
It is theorized that pCS is involved in articulatory phonological processing, specifically feedforward control of articulator velocity (Tourville and Guenther, 2011; Matchin and Hickok, 2020). Through lesion studies, pCS has also been linked to phonological dyslexia (Rapcsak et al., 2009; Tomasino et al., 2020). Our data demonstrate that pCS activation begins early, preceding the IFG, suggesting a role in early linguistic or phonological processing, as part of the sublexical route. This is concordant with recent findings suggesting that pCS is involved in the grapheme-to-phoneme conversion process (Kaestner et al., 2021). pCS demonstrates lexical sensitivity but no effect of word frequency or orthographic neighborhood. Given the association of pCS with articulation phonology and phonological dyslexia, pCS may contribute to the process of constructing articulatory representations for both words and pseudowords, likely in a manner distinct from the processing in IPS.
This study provides further evidence that medial frontal operculum is involved in prearticulatory, preparatory processes, distinct from those of the lateral IFG (Mălîia et al., 2018; Woolnough et al., 2019). Lesions involving this region have been linked to impairment of complex articulation (Baldo et al., 2011), which may explain the greater engagement during pseudoword articulation.
We observed no substantial prearticulatory activity in peri-sylvian angular gyrus, instead peaking prominently after the initiation of articulation. This region has previously been linked to semantic and phonological processes during word processing (Stoeckel et al., 2009; Graves et al., 2010; Hartwigsen et al., 2010; Sliwinska et al., 2015), with fMRI deactivations observed when reading pseudowords (Taylor et al., 2014b). It is active during reading in children but may not be recruited in adults for simple reading tasks (Martin et al., 2015), with activation observed when comprehending multiword phrases (Dronkers et al., 2004; Wilson et al., 2014; Matchin et al., 2017; Fridriksson et al., 2018). It is also possible that fMRI deactivations are not strongly represented in the intracranial signal, as task-induced BGA decreases are not commonly observed.
In conclusion, this dataset comprehensively maps the dynamic functional characteristics of the two reading routes. Extant models suggest that all words should be initially processed through both reading pathways, with areas responsible for lexical access being activated inversely proportionally to lexical frequency, as demonstrated here in mFus and aIFG. Once a known word is detected, at a latency which depends on frequency, the sublexical route appears to be interrupted, and the word is named through the lexical route. However, if the word is novel, the lexical route continues to attempt to identify the pseudoword while the sublexical route, involving the IPS, constructs the phonology of these words, with activity driven by pseudoword orthographic complexity. Last, the early, lexically sensitive activation of ventral premotor cortex (pCS) implies a role in early grapheme-to-phoneme conversion, as part of the sublexical route, distinct from the IPS.
Footnotes
This work was supported by National Institute of Neurological Disorders and Stroke NS098981. We thank all the patients who participated in this study; the neurologists at the Texas Comprehensive Epilepsy Program who participated in the care of these patients; and the nurses and technicians in the Epilepsy Monitoring Unit at Memorial Hermann Hospital who helped make this research possible.
The authors declare no competing financial interests.
- Correspondence should be addressed to Nitin Tandon at nitin.tandon{at}uth.tmc.edu
