One right can make a left: sentence processing in the right hemisphere after perinatal stroke

Abstract

When brain regions that are critical for a cognitive function in adulthood are irreversibly damaged at birth, what patterns of plasticity support the successful development of that function in an alternative location? Here we investigate the consistency of language organization in the right hemisphere (RH) after a left hemisphere (LH) perinatal stroke. We analyzed fMRI data collected during an auditory sentence comprehension task on 14 people with large cortical LH perinatal arterial ischemic strokes (left hemisphere perinatal stroke (LHPS) participants) and 11 healthy sibling controls using a “top voxel” approach that allowed us to compare the same number of active voxels across each participant and in each hemisphere for controls. We found (1) LHPS participants consistently recruited the same RH areas that were a mirror-image of typical LH areas, and (2) the RH areas recruited in LHPS participants aligned better with the strongly activated LH areas of the typically developed brains of control participants (when flipped images were compared) than the weakly activated RH areas. Our findings suggest that the successful development of language processing in the RH after a LH perinatal stroke may in part depend on recruiting an arrangement of frontotemporal areas reflective of the typical dominant LH.

Introduction

In most adults, certain critical language processes are so strongly lateralized to the left hemisphere (LH) that a stroke to left frontotemporal cortex results in a lifelong language impairment. This consistent outcome initially led to the conclusion that the left but not right perisylvian cortex is essential for language processing (Broca 1865). Strong convergent evidence has also been observed in the behavioral outcomes when only one hemisphere is available to respond during language tasks in adults, such as after transection of the corpus callosum (split-brain; Gazzaniga and Sperry 1967) or unilateral anesthetization of each cerebral hemisphere (the Wada test; Wada and Rasmussen 1960). Indeed, the common finding is that the LH supports normal verbal abilities, while the right hemisphere (RH) is either severely limited or incapable of verbal responses (Gazzaniga and Sperry 1967). In most healthy adults, neuroimaging studies have demonstrated that language activation is strongly left-lateralized regardless of handedness (Knecht et al. 2000a; Knecht et al. 2000b; Szaflarski et al. 2002) and regardless of whether the language is spoken or signed (Petitto et al. 2000; MacSweeney et al. 2008; Newman et al. 2015).

However, when a stroke damages left perisylvian regions at the beginning of life, the brain can develop to support normal language abilities. Perinatal strokes occur between the end of the third trimester and the first two weeks of life, with an incidence of approximately 1 in 3500 newborns (Ferriero et al. 2019). Over two-thirds of these strokes impact the left middle cerebral artery (MCA). This means that in many cases, left perisylvian cortex is irreversibly damaged around the time of birth, often without additional neurological or cardiovascular complications. Despite the extensive loss of LH regions that are critical for language processing in most adults, language abilities often develop to be in the normal range. An initial developmental delay has been reported after either an LH or RH perinatal stroke when language abilities are compared with typically developing peers (Bates 2004; Trauner et al. 2013), but normal performance is generally achieved when tested as adolescents or adults (Fair et al. 2010; Ilves et al. 2014; Stiles et al. 2012; Newport et al. 2017, 2022). These outcomes indicate that left perisylvian cortex is not in fact required for successful language processing in many individuals with an LH arterial ischemic perinatal stroke (see François et al. 2021 for a review of the variables that can lead to divergent outcomes in this population). Indeed, even in rare cases where the full LH is infarcted, language abilities can develop in the normal range, exclusively in RH regions (Newport et al. 2017, 2022).

In individuals who develop normal language processing in the absence of LH perisylvian regions there is an opportunity to examine the resulting alternative pattern of organization and to ask what features may be important for the implementation of a language system. Under these extreme circumstances, right perisylvian regions are commonly recruited for language processing (Martin et al. 2022a). Several neuroimaging studies, including our own recent work, have found that right frontal and temporal regions are active in people with an LH perinatal stroke during language tasks, including sentence processing (Newport et al. 2017, 2022; Ilves et al. 2014), passive story-listening (François et al. 2019), and word, verb, or rhyme generation (Staudt et al. 2002; Guzzetta et al. 2008; Tillema et al. 2008; Ilves et al. 2014). However, group-level activation maps provide a limited view of this pattern in that they only reveal the regions that are most strongly and consistently active across the group. Even in neurotypical individuals, language activation varies substantially from person to person, so the same is expected in in individuals with left hemisphere perinatal stroke (LHPS). In order to determine whether successful language processing can be implemented in a variety of arrangements, or whether there is one consistent arrangement that may be critical for such successful outcomes, we must carefully examine the spatial organization of the cortical territories recruited for language processing in individual LHPS participants. This is the focus of the present paper.

There are three potential outcomes for this investigation. One possibility is that language processing consistently recruits a “dominant hemisphere like” arrangement of perisylvian areas in the RH of individuals with an LH perinatal stroke. This finding would suggest that the RH can support normal language processing only if it can develop the same pattern of cortical recruitment for language processing that is implemented in the typical dominant LH. A second possibility is that in some or all people with an LH perinatal stroke, the RH areas recruited for language processing are the same homotopic RH areas that are weakly active during language processing in the typical brain. Indeed, the few studies that have examined the activity in homotopic RH areas during language processing in healthy participants have found that these regions respond more during language tasks relative to control tasks; the strength of the response is simply much weaker than what is measured in the LH (Just et al. 1996; Quillen et al. 2021; Martin et al. 2022b). In healthy adults, these weakly-activated RH areas appear roughly homotopic to typical LH areas, but are not entirely symmetrical (Martin et al. 2022b). This means there is a small but potentially significant margin of difference between the particular patches of perisylvian cortex strongly recruited in the typical LH and those weakly recruited in the typical RH (i.e. the distinction between the first two possible outcomes named here). A third possibility is that individual people with an LH perinatal stroke may vary in the particular areas of the RH that they recruit for language processing. Even in a participant sample with relatively homogenous stroke characteristics (large perinatal arterial ischemic strokes in the left MCA), there is still potentially relevant variability in RH language organization that might only be revealed by an individual participant analysis. Which of these three possibilities best describes the outcome after an LH perinatal stroke will be important for our understanding of the essential cortical arrangements for successful language processing.

In the current study we examined the single-subject data collected in Newport et al. 2022 to ask whether brain activity during sentence processing in individual LHPS participants was localized to consistent RH perisylvian areas that are homotopic to the LH areas active in controls. We also asked whether the spatial arrangement of RH perisylvian areas active in LHPS participants was more similar to the dominant, strongly active LH perisylvian areas of healthy controls, or rather was more similar to the non-dominant, weakly active RH perisylvian areas of controls. First, we examined whether there were differences between healthy controls and LHPS participants in the magnitude and extent of activation during sentence processing. Then, to investigate our questions regarding the spatial organization of the activation, we employed a “top voxel” approach (see also Martin et al. 2022b), which equalizes the number of active voxels compared in each hemisphere and in each participant. We hypothesized that the RH regions active during sentence processing in individual LHPS participants would be consistently homotopic to the LH regions active in controls and that the spatial arrangement of this activity would be more similar to the arrangement observed in the typical dominant LH than in the typical non-dominant RH. This result would suggest that the RH can support normal language processing only if there is sufficient plasticity at the time of injury to develop a full language system in right perisylvian regions.

Materials and methods

Participants

Fourteen perinatal stroke participants (six female) who survived a perinatal or presumed perinatal arterial ischemic stroke impacting a large portion of left frontal and/or temporal cortex were recruited and tested as part of the Pediatric Stroke Research Project at Georgetown University, along with the healthy siblings of the stroke survivors who participated as healthy controls (Newport et al. 2022). See Table 1 for participant characteristics. The study was approved by the Institutional Review Board at Georgetown University Medical Center; all participants provided consent (adults) or parental consent and child assent (children). Participants were excluded if they were diagnosed with a congenital heart disease, epilepsy, or any neurological disorder other than perinatal stroke. The 11 healthy controls (three female) were roughly age-matched to the stroke participants. We included all participants who were 10 years or older at the time of the study because this age is associated with mature language lateralization (Berl et al. 2014a). The sample in our recent publication (Newport et al. 2022) included one additional stroke participant and one additional healthy control who were just under 10 years of age at the time of testing and were therefore excluded from the current analysis. See Newport et al. (2022) for more detailed information about participant inclusion and cognitive outcomes.

Table 1.

Participant characteristics.

	N	Age (years)	Handedness	L1	Sex	Stroke
LHPS	14	17.88 +/− 4.89 (10.0–25.6)	14 L	English	6 F	13 middle cerebral artery, 1 internal carotid artery
Controls	11	15.95 +/− 5.0 (11.5–29.5)	1 L	English	3 F	N/A

Magnetic resonance imaging

Auditory Description Decision Task

The Auditory Description Decision Task (ADDT) is an fMRI task designed by Gaillard and colleagues (Gaillard et al. 2007; Berl et al. 2014a; Berl et al. 2014b) in which participants hear alternating blocks of short sentences describing target nouns (Forward Speech condition) and unintelligible sound sequences of the same auditory waveforms played backwards (Reverse Speech condition). During the Forward Speech condition, participants press a button when they judge the description to be accurate; during the Reverse Speech condition they press a button when they hear a beep at the end of the sequence. In the present study, we modified this task so that button responses were prompted on 50% of sentences and 50% of control condition sound sequences, and we also added blocks of silence that were interleaved with the Forward and Reverse conditions. Each run of the task was 5:48 in duration and included four 30-second blocks of forward and reverse speech in counterbalanced order, interleaved with 12-second periods of silent rest. Stimuli in each block were presented every 5 seconds, with a 3-second stimulation period followed by a 2-second response window. Runs were repeated if motion was excessive. Two runs were analyzed for each participant. The task was designed to be engaging but easy for different age levels and clinical populations (Berl et al. 2014a; Berl et al. 2014b). There are three levels of the task that differ only in the word frequency of the nouns described in the short sentences. It is designed to present the word frequency that is appropriate for the age of the participant. In our study, the decision about which level of the task to be used was made by the researchers who had been interacting and testing the participant, based on their estimate of their language and reading level. Most participants performed the “hard” version of the task (n = 11 LHPS and n = 10 controls) and the rest performed the “medium” version (n = 3 LHPS and n = 1 control).

For the Forward speech condition, controls had a mean accuracy of 97.5% +/− 2.4% (93.8–100.0%) and LHPS participants had a mean accuracy of 95.2% +/− 5.9% (79.2–100.0%). For the Reverse speech condition, controls had a mean accuracy of 99.2% +/− 1.4% (95.8–100.0%) and LHPS participants had a mean accuracy of 97.9% +/− 2.0% (93.8–100.0%). Independent samples t-tests did not reveal differences between participant groups in the response times for the Forward or Reverse conditions (P > 0.1).

Scanner and auditory equipment

Imaging data were acquired through a 12-channel headcoil on a 3 Tesla Siemens MAGNETOM Trio scanner at Georgetown University’s Center for Functional and Molecular Imaging. Three participants were scanned after a scanner upgrade to a Prisma model with a 20-channel head coil. Participants were fitted with Sensimetrics Model S14 insert headphones and additional Bilson ear defenders and confirmed that they could hear the auditory stimuli for the fMRI task over the scanner noise.

Scan sequences

Participants watched a movie while we collected a high-resolution anatomical image, which was visually inspected for motion and re-collected if necessary: Siemens MPRAGE, 176 sagittal slices, TR = 2.53 seconds, TE = 3.5 ms, flip angle = 7°, 1 × 1 × 1mm voxels, whole brain coverage. Functional echo-planar images were collected while participants completed the ADDT: 50 horizontal slices, descending order, TR = 3 seconds, TE = 30 ms, flip angle = 90°, 3 × 3 × 3 mm³ voxels, 116 whole brain volumes. One control participant completed an older version of the task that included 120 whole brain volumes.

fMRI processing

The preprocessing of the functional images was performed with SPM12 (Wellcome Trust Centre for Neuroimaging at University College London, https://www.fil.ion.ucl.ac.uk/spm/doc/) and included the following steps: slice-time correction, realignment to the middle volume of each run, co-registration to the native-space anatomical image, spatial normalization to the MNI-152 average template (resulting resolution 2 mm), and then smoothing (8 mm FWHM Gaussian kernel). We masked the images to exclude voxels outside the brain and used a general linear model to calculate the beta weights for the Forward and Reverse Speech conditions of interest, which were convolved with a canonical hemodynamic response function, as well as motion estimates for rotation and translation along the x-, y-, and z-axes, and a high-pass filter for the duration of the task run. Finally, we performed voxel-wise t-tests on the resulting beta maps for the Forward and Reverse Speech predictors to identify voxels that were more active during language processing.

Activation magnitude and extent analysis

We first investigated whether there were differences between LHPS participants and healthy controls in the amount of activation during sentence processing to evaluate whether the activation was perhaps more focal or diffuse in either of our participant groups. First, we applied a lenient voxelwise p-threshold (P < 0.01) with a minimal cluster size requirement (k ≥ 4) to each participant’s whole-brain Forward>Reverse Speech statistical map. We used a lenient threshold to err on the side of being more inclusive on the amount activation we would compare. We first summed the number of voxels active for each participant, and found the average size of their activation clusters. These were our two measures of activation extent, which together capture how diffuse the activation is in terms of overall quantity of recruitment (total number of active voxels) and localized expansion of recruitment (average cluster size). We then also found the mean and maximum t-value of the active voxels for each participant. These were our two measures of activation magnitude, which together capture the general spread (mean t-value) and greatest amount (maximum t-value) of language activation. We also calculated the magnitude of the activation (mean t-value) across the top voxels included in the spatial overlap comparison (described below) in the LH and RH of each control and the RH of each LHPS participant.

Flipped activation maps for LHPS participants

In the current study we sought to measure the similarity between the spatial arrangement of activity during sentence processing in the RH of LHPS participants and the LH of healthy controls. We approached this question by flipping the sentence processing activity maps for each LHPS participant across the midline, making them like LH maps, for best comparison with the unflipped activity maps for each healthy control (Fig. 1A). This required several steps to make the comparison as precise as possible. Before preprocessing we used SPM12’s reorient utility to flip the raw images (functional and structural) for each LHPS participant across the midline. We co-registered each individual’s flipped RH activity to their individual RH anatomy to spatially align the functional scans. Then, in order to address minor structural asymmetries between the two hemispheres (Amunts 2010), we spatially normalized the flipped images to the MNI template anatomy.

Fig. 1.

Top voxel analysis workflow. We (A) flipped the RH activation map for each LHPS participant for comparison to the LH and RH of each control, (B) masked activation in anatomical ROIs, (C) applied a top voxel cutoff to equalize the quantity of activation within the ROI for each hemisphere and participant, and (D) calculated the spatial overlap with a Dice Coefficient between the flipped-RH activation areas in LHPS participants and the LHs of controls, as well as the unflipped RHs of LHPS participants and the RHs of controls. LHPS: left hemisphere perinatal stroke, L: left, R: right.

Regions of interest

To investigate homotopicity within frontal and temporal regions, we used the frontal and temporal lobe masks from the WFU-Pick Atlas (Maldjian et al. 2003) in the SPM12 toolbox. We decided to use anatomical masks of each full lobe to capture any activity that may be atypically localized in people with LHPS outside of language areas that are common in healthy individuals. These region of interest (ROI) masks were resliced to match the dimensions of the MNI-space anatomical and statistical maps and were used to isolate frontal and temporal activations for comparison between participants (Fig. 1B). We also included an analysis of the whole hemisphere for each participant in the (Supplementary Materials Fig. S5).

Table 2.

Number of voxels analyzed in regions of interest.

	ROI size	Level 1 (P < 0.01)	Level 2 (P < 0.005)	Level 3 (P < 0.001)	Level 4 (P < 0.0005)
Frontal	34,933	2,962	2,427	1,556	1,303
Temporal	16,369	1,487	1,228	810	687

Top voxel approach

Conventional approaches for analyzing fMRI task data involve thresholding the statistical maps to isolate the t-values that are high enough to meet a somewhat arbitrary p-value level, and the areas with surviving t-values are considered the active regions for each participant. However, this type of approach is not ideal for comparing spatial arrangements of activity between participants because the magnitude of task-driven activity (i.e. the t-values) can be quite variable between individuals. Instead, we decided to employ a “top voxel” approach (see also Wilson et al. 2018; Martin et al. 2022b). This method includes the same number of active voxels for every participant, so the interpretation of high versus low overlap is more straightforward.

For the top voxel analysis, we first determined four cutoffs for the number of active voxels to include in our spatial overlap analysis (Fig. 1C): within the dominant hemisphere ROIs for each participant (LH for controls, RH for LHPS participants), we applied four statistical thresholds (P < 0.01, P < 0.005, P < 0.001, P < 0.0005) with minimal clustering (k = 4), and then we averaged the number of voxels that survived each threshold across all participants (listed in Supplementary Table 1 and visualized in Supplementary Fig. S1). We examined four cutoffs to limit bias due to a cutoff that was too conservative or too lenient. To apply these cutoffs, we ranked the t-values within each ROI from highest to lowest for each participant and generated a new constrained map that included only the top “N” voxels in the ROI. For control participants, we applied the cutoffs determined in the LH ROIs to their right hemisphere (RH) ROIs. From this procedure we obtained four equal size activation maps within each ROI for all participants, one at each of four increasingly strict top voxel cutoffs (examples in Fig. 1D).

Pairwise spatial overlap analysis (Dice Coefficient)

A Dice Coefficient was then calculated for each pair of participants (x and y), which represents the number of overlapping voxels between their maps as a ratio of the total number of voxels being compared:

(1)

For each participant, we then averaged the Dice Coefficients across the four top voxel cutoffs to obtain one value for spatial overlap for each region (Fig. 1D). We performed four comparisons (between controls, between LHPS participants, and between controls and LHPS participants), listed in Table 3 and described below.

Table 3.

Comparisons of interest.

Purpose of comparison	Maps compared
1. Consistency across LHPS participants	1. Unflipped RH maps for each pair of LHPS participants
2. Reference point for consistency in the typical brain	2. Unflipped LH maps for each pair of controls
3. Similarity of language activity in LHPS participants to typical dominant hemisphere	3. Flipped RH map of each LHPS participant and unflipped LH map of each control
4. Similarity of language activity in LHPS participants to typical non-dominant hemisphere	4. Unflipped RH map of each LHPS participant and unflipped RH map of each control

Statistical comparisons

Statistical analyses were performed using R through R-Studio (http://www.R-project.org/). Independent samples t-tests were performed to measure potential differences between healthy controls and LHPS participants in the extent and magnitude of whole-brain Forward>Reverse speech activation. We compared mean differences in the activation magnitude of the top voxels (level 1 top voxel cutoff) in the LH and RH of each control using paired samples t-tests, and between the LHPS RH and the control LHs and RHs using independent samples t-tests. We calculated z-scores to characterize how similar the spatial arrangement of sentence processing activity was for each LHPS participant relative to the group of healthy controls: We took each LHPS participant’s Dice Coefficient with all controls, subtracted the average within-group Dice Coefficient for controls, and divided by the standard deviation of the within-group values for controls. A z-score close to zero indicates that the spatial arrangement of sentence processing activity in the LHPS participant (flipped into the LH) was just as similar to controls as controls were to each other.

Results

Sentence processing activation is neither more focal nor more diffuse in LHPS participants

Before examining the spatial organization of the activation during sentence processing, we first investigated whether healthy controls and perinatal stroke participants differed in the focality and strength of their activation. For example, compensatory plasticity after LHPS may involve more diffuse engagement of the remaining tissue, as measured by larger clusters with perhaps weaker magnitude. LHPS participants and controls did not differ in either the number of voxels that were active or in the average size of the activation clusters (independent samples t-tests; P > 0.6). Similarly, LHPS participants and controls did not differ in either the mean or maximum t-value of the voxels active (independent samples t-tests; P > 0.1). See Fig. 2 for the group data on activation extent and magnitude at each threshold.

Fig. 2.

Activation extent and magnitude. The forward>reverse sentences statistical map for each participant was thresholded voxelwise at P < 0.01 with a lenient cluster size threshold (k ≥ 4). Controls (blue; left distribution of each plot) and LHPS participants (purple; right distribution of each plot) were compared using independent samples t-tests, and the uncorrected P-values are shown above the brackets in each plot. (A) We examined two measures of activation extent: the number of active voxels (left) and the average size of the activation clusters in each participant’s statistical map (right). (B) We examined two measures of activation magnitude: the average t-value (left) and the maximum t-value (right) across all voxels that survived the voxelwise threshold.

Sentence processing recruits consistent homotopic regions in individual LHPS participants

To gauge what degree of overlap would be considered high for our task, we first examined the consistency between control participants in the spatial arrangement of sentence processing activity in the typical LH. When the spatial arrangement of sentence processing activity was compared for each pair of healthy control participants (Fig. 3, CxC), the Dice Coefficients averaged across top voxel cutoffs had a mean in the frontal lobe of 0.245 +/− 0.042 (0.162–0.298) and a mean in the temporal lobe of 0.185 +/− 0.030 (0.122–0.224).

Fig. 3.

Sentence processing activates homotopic RH regions. For each participant, the same number of the most active voxels for the Forward>Reverse speech contrast were compared in the left frontal and temporal ROIs. For LHPS participants, functional images were flipped and normalized to standard LH anatomy so language activity in the intact RH could be spatially compared with LH activity in controls. A Dice Coefficient was used to quantify the spatial overlap between homotopic activation areas, and was calculated at four different top voxel cutoffs (corresponding to the average number of voxels across all participants that survived at P < 0.01, 0.005, 0.001, and 0.0005). (A) The y-axes reflect the average Dice Coefficient across all four top voxel levels. Gray lines indicate the overlap for individual LHPS participants in the within- and between-group comparisons. When the location of activity in stroke participants was compared with controls (CxL, purple bars; middle bar on each plot), a comparable degree of spatial overlap was observed as when controls were compared with each other (CxC, blue bars; left bar on each plot). See Supplementary Table 2 for test statistics. (B) The average spatial overlap across all four top voxel cutoffs was calculated for each LHPS participant with all controls. Z-scores reflect the difference between this average Dice value for each individual LHPS participant and the average within-group Dice value for controls, divided by the standard deviation across controls. Z-scores for the frontal ROI are in yellow, and for the temporal ROI in red. Gray shading represents one (darker) and two (lighter) standard deviations away from the controls’ normalized average.

We next examined whether the degree of overlap fell within this same range when we compared the flipped RH of each LHPS participant to the unflipped LH of each control (i.e. homotopicity). The Dice Coefficients for flipped LHPS RHs with the unflipped LHs of controls (Fig. 3A, CxL) had a mean in the frontal lobe of 0.214 +/− 0.036 (0.148–0.275) and a mean in the temporal lobe of 0.164 +/− 0.048 (0.045–0.231). An F-test did not reveal significant differences in the degree of variance between the Dice Coefficients measured between controls and those measured between LHPS participants and controls (Frontal: F(10,13) = 1.34, P = 0.61; Temporal: F(10,13) = 0.39, P = 0.14), suggesting that the amount of variance in our measure of pairwise spatial overlap was comparable between these groups. An independent samples t-test between the distribution of overlap values among controls and the distribution of overlap values for LHPS participants with controls did not reveal significant differences (Fig. 3; CxC vs. CxL in Frontal: t(19.9) = 1.91, P = 0.07, and Temporal: t(22.0) = 1.36, P = 0.19). This result indicates that sentence processing activates RH frontal and temporal regions in people with a large cortical LH perinatal stroke that are a near mirror-image of the left frontal and temporal regions active in controls. The t-test results for all of the spatial overlap comparisons are listed in Supplementary Table 2.

In addition to this comparison of the mean overlap between groups, we also characterized how homotopic the activity was for each individual LHPS participant by calculating a z-score (Fig. 3B). A z-score close to zero would indicate that activity for the individual LHPS participant is so precisely homotopic that when flipped into the LH, it is within the expected range of overlap for a given pair of control participants. In the frontal lobe, nine LHPS participants were within one standard deviation (SD) of the control mean; four were between one and two SDs, and one participant was slightly more than two SDs below (L9, z = −2.30). In the temporal lobe, seven LHPS participants were within one SD of the control mean; five were between one and two SDs, and two participants were more than two SDs below (L3, z = −2.92; L14, z = −4.65). Regarding the three participants who were more than two SDs below the control mean in either the frontal or temporal lobe, there is no apparent difference in these individuals’ age or stroke characteristics from the rest of the group that might explain why their spatial arrangement for sentence processing activity was not as precisely homotopic; and in each case the z-score was only so deviant in one of the lobes (frontal or temporal). These results suggest that for most LHPS participants in our sample, sentence processing activates areas of the RH that (when flipped) are highly overlapping with the areas of the LH active in controls. Importantly, these findings are not driven by any imbalance in the size of activation area in either group because we compared the same number of active voxels in every participant.

Finally, we measured how similar stroke participants were to one another regarding where sentence processing activated areas of the right perisylvian cortex. When the spatial arrangement of sentence processing activity in the unflipped RH was compared for each pair of LHPS participants (Fig. 3A, LxL), the Dice Coefficients averaged across top voxel cutoffs had a mean in the frontal lobe of 0.202 +/− 0.030 (0.159–0.234) and a mean in the temporal lobe of 0.168 +/− 0.043 (0.053–0.220). We ran a paired samples t-test to compare these values to the overlap between LHPS participants and controls, which did not reveal significant differences (Fig. 3A; CxL vs. LxL in frontal: t(13) = 0.77, P = 0.45, and temporal: t(13) = −0.53, P = 0.60). This result suggests that people with a large cortical LH perinatal stroke are similar to one another in where sentence processing activates perisylvian cortex, in addition to being similar to controls (in the opposite hemisphere).

LHPS RH sentence processing activity is more similar to the typical LH than the typical RH

We also asked whether the spatial arrangement of sentence processing activity in the RH of LHPS participants was more similar to the activity in the typical LH of controls or to the typical RH of controls. In contrast to the overlap with the typical dominant hemisphere (reported above), when we compared the unflipped RH of each LHPS participant to the unflipped RH of each control (i.e. comparison to the typical non-dominant hemisphere; Fig. 4, dark purple bars), the Dice Coefficients had a mean in the frontal lobe of 0.123 +/− 0.014 (0.095–0.139) and a mean in the temporal lobe of 0.104 +/− 0.021 (0.072–0.140). Independent samples t-tests revealed differences between the typical dominant and typical non-dominant comparisons in the frontal and temporal lobes (P < 0.00005 in both ROIs), with greater similarity between activity in the flipped RH of LHPS participants and the unflipped LH of controls. This result indicates that the spatial arrangement of sentence processing activity in the RH of people with a large cortical LH perinatal stroke is substantially more similar to the typical dominant LH than it is to the typical non-dominant RH. See Supplementary Materials Fig. S5 for a comparison between the full hemispheres of each pair of participants, which reproduces the relationships in the frontal and temporal lobes discussed here.

Fig. 4.

The spatial organization of RH sentence processing activity is more similar to the typical left than the typical RH. The flipped right frontal and temporal top voxel maps for each individual LHPS participant were compared with the left frontal and temporal top voxel maps of each control using a Dice Coefficient to measure the similarity to the typical dominant hemisphere. Unflipped right frontal and temporal top voxels maps for each individual LHPS participant were compared with the right frontal and temporal top voxel maps of each control to measure the similarity to the typical non-dominant hemisphere (top voxel maps were the same size for the dominant and non-dominant comparisons). The y-axis reflects the average Dice Coefficient across all four cutoffs tested. Gray lines indicate individual participants. LH maps of controls were compared with each other (light blue bars; left-most bars on each plot—approximate consistency ceiling; same as Fig. 3A), LH maps of controls were compared with the flipped RH maps of other controls (dark blue bars; second bar from the left on each plot—measure for effect of flipping), flipped RH maps of LHPS participants were compared with LH maps of controls (light purple bars; third bar from the left on each plot—typical dominant hemisphere comparison; same as Fig. 3A), and RH maps of LHPS participants were compared with RH maps of controls (dark purple bars; right-most bar on each plot—typical non-dominant hemisphere comparison). See Supplementary Table 2 for test statistics and Supplementary Fig. S2 for an analysis of these relationships at each individual top voxel level. See Supplementary Fig. S5 for a comparison in the full hemisphere.

We performed an additional validation analysis to confirm that the overlap we measured between the LH of controls and the flipped RH maps of LHPS participants was not somehow inflated by the comparison of flipped and unflipped maps. We flipped the RH map for each control participant and calculated the overlap with the LH maps for every other control participant. We found that the overlap between controls’ flipped right and unflipped left frontal and temporal regions was significantly lower than the overlap between LHPS participants’ flipped right frontal and temporal regions and controls’ unflipped left frontal and temporal regions (Fig. 4 light purple and dark blue bars; two-tailed independent samples t-tests, frontal: P < 0.005, temporal: P < 0.05). Importantly, this result indicates that flipped right perisylvian activity in LHPS participants is so highly aligned with left perisylvian activity in controls that it even surpasses the alignment of left and flipped right perisylvian activity of other controls.

Finally, we compared the magnitude of the activation in the top voxels that were spatially compared in the LH and RH of controls and in the RH of LHPS participants (Supplementary Materials Fig. S4). This comparison of activation magnitude, in contrast to the whole-brain comparisons of activation magnitude and extent discussed above, is specific to the subset of the most active voxels in the dominant and non-dominant hemispheres that were also spatially compared in the overlap analysis. Paired samples t-tests revealed that the magnitude of activation was significantly higher in the LH compared with the RH of each control participant, as expected (frontal: t(10) = 4.59, P = 0.001; temporal: t(10) = 5.91, P = 0.0001). Independent samples t-tests did not reveal significant differences in activation magnitude between the LHs in controls and the RHs in LHPS participants (frontal: t(22.8) = 0.097, P = 0.924; temporal: t(22.4) = 0.903, P = 0.376). We found significantly higher activation magnitude in the RH of LHPS participants compared with the RH of controls in the temporal ROI (t(22.1) = −2.76, P = 0.011), and there was a similar trend in the frontal ROI that did not reach significance (t(22.9) = −1.92, P = 0.068).

We created penetrance maps based on top-voxel activation maps to visualize the brain areas consistently recruited for sentence processing in the LH and RH of controls and the RH of LHPS participants (Fig. 5A; penetrance maps based on single-voxel thresholded maps can be found in Supplementary Materials Fig. S1). LHPS participants exhibited high consistency of activation in the inferior frontal cortex, superior temporal pole, and posterior superior temporal sulcus (STS) and posterior middle temporal gyrus. Control participants exhibited high consistency in these same areas in the LH. In the RH, controls exhibited less consistency in general, with some areas of high consistency in the temporal pole, superior and middle orbitofrontal cortex, and a dorsal portion of the inferior frontal gyrus (IFG) pars triangularis.

Fig. 5.

Penetrance maps and contrasts showing consistency of brain areas recruited. Penetrance maps show the percentage of the participants in each group who had activation at each voxel. (A) Maps for individual participants at the most lenient top voxel level (2962 voxels in the frontal ROI and 1,487 voxels in the temporal ROI; Table 2) were merged for all control participants in the LH and flipped RH, and for all LHPS participants in the flipped RH. Darker blue areas were less consistently recruited across the group (e.g. only one LHPS participant or one control), while white areas were recruited by more than half of the participants within the group. (B) These group penetrance maps were thresholded at 20% (i.e. ≥3 LHPS participants or ≥ 3 control participants) and binarized for comparison. Each pair of these thresholded, binarized penetrance maps was compared. Brain areas recruited in both maps for each pair are in orange.

We then contrasted these penetrance maps to examine the brain areas that were consistently shared and the areas that were consistently more active in one group or hemisphere (Fig. 5B). There is extensive overlap (orange areas) between the LH of controls and the RH of LHPS participants, and far less between the RH of controls and the RH of LHPS participants. Certain ventral portions of IFG and more posterior portions of STS and middle temporal gyrus appeared to be more consistently active in the LH of controls and the RH of LHPS participants than the RH of controls (Fig. 5B, areas that are red in the leftmost image and yellow in the rightmost image). Certain anterior portions of superior and middle orbitofrontal cortex and a more inferior portion of the temporal pole appeared to be more consistently active in the RH of controls than in the LH of controls or the RH of LHPS participants (Fig. 5B, areas that are yellow in the leftmost image and red in the rightmost image). And finally, a small dorsal portion of the IFG pars triangularis appeared to be more consistently recruited in the RH of controls and the RH of LHPS participants than in the LH of controls (0/11 Controls in the LH, 3/11 of Controls in the RH, 7/14 LHPS in the RH; Fig. 5; areas in orange in rightmost image that were yellow in the other two images). The most important evidence from these images is the high degree of overlap between the control LH and LHPS RH that is not seen when the control RH is compared.

Discussion

In our previous work we showed that individuals with an LH perinatal arterial ischemic stroke can organize a language system in RH perisylvian regions that supports proficient language abilities (Newport et al. 2022). Here we aimed to characterize the spatial arrangement of this activity as it relates to typical arrangements in the left perisylvian regions of healthy controls to understand which organizational features might be important for the successful implementation of a language system. Our analyses produced two key findings. First, we found that sentence processing recruits consistent RH perisylvian areas in individual LHPS participants that are homotopic to the left perisylvian areas recruited in healthy controls. This consistency is especially interesting, given that our LHPS participants all had large cortical strokes that varied in location and size. Second, we found that the spatial arrangement of activity was more similar to that of the typical dominant LH than the typical non-dominant RH—even when the amount of activity being spatially compared was equated. The brain areas that were especially consistent (i.e. active in more than 20% of participants in each group) in the dominant LH of controls and dominant RH of LHPS participants, relative to the non-dominant RH of controls, included the IFG and STS. The magnitude and extent of activation in the LHPS RH were also comparable to that of the dominant LH of controls.

Previous studies have reported that people with an LH perinatal stroke activate regions in the RH that appear to be a mirror-image of the LH regions activated in controls during language tasks (Staudt et al. 2002; Guzzetta et al. 2008; Tillema et al. 2008). In the current study we aimed to quantify how consistent this homotopic organization was in individual LHPS participants. Across the group of LHPS participants, we found a high degree of overlap with controls that was within the same range of how overlapping controls were to one another in frontal and temporal regions. This suggests that sentence processing is organized in consistent homotopic regions in our sample of LHPS participants. Overall, we have shown in a rigorous quantitative way what has been suggested previously based on qualitative visual comparisons of activity in the two hemispheres: that language processing in people with a large cortical LH perinatal stroke recruits consistent RH regions that are homotopic to typical LH language regions in the typical brain (Newport et al. 2017, 2022; François et al. 2019; Tillema et al. 2008; Guzzetta et al. 2008).

Some investigators have suggested that RH activity in LHPS participants relates to incorrect trials (Fair et al. 2010; Raja Beharelle et al. 2010). This hypothesis cannot explain the results here because our participants perform extremely well on our task. Furthermore, one of our participants (L4) has no available tissue in the left cerebral hemisphere and has a fully right-lateralized and functional language system, which unambiguously indicates that the RH can support language abilities. Even though our perinatal stroke participants vary in the amount of tissue loss and the particular regions damaged in the LH, they all still developed successful language processing, and it appears to recruit a consistent pattern of right perisylvian areas that are homotopic to the left perisylvian areas they might have used if they had not had a stroke early in development.

What are the areas crucial to the LH in the healthy brain that we also see activated in the RH of perinatal stroke participants? The IFG and posterior STS/middle temporal gyrus stand out as areas with high consistency in the overlap between the dominant hemispheres of LHPS participants (RH) and controls (LH), whereas the overlap was less consistent in these regions with the non-dominant RH of controls. In the healthy brain, the right IFG may (perhaps even more than temporal regions) undergo maturational changes that make it less involved in language processing in adulthood, and therefore also less similar to the typical language-dominant LH. In our recent work (Olulade et al. 2020), the magnitude of activity during sentence processing was found to significantly decline over the course of childhood in this region. This change may occur because these regions become more involved in other cognitive functions in the healthy adult brain (e.g. emotional prosody processing; Ross and Marek-Marsel 1979; Seydell-Greenwald et al. 2020). More generally, the particular perisylvian areas recruited for language processing in the dominant LH of healthy controls, which are also recruited in the right perisylvian cortex of LHPS participants, are apparently important for fully supporting a language system.

One interpretation of these findings is that the RH can support normal language processing if there is sufficient plasticity at the time of injury to organize a full language system in right perisylvian regions. A number of studies have sought to identify the features of the LH that create an advantage for language processing in the typical brain, including larger perisylvian brain structures (Foundas et al. 1994; Penhune et al. 1996; Dorsaint-Pierre et al. 2006; Powell et al. 2006; Barrick et al. 2007). Other studies have focused on the response properties of neuronal populations in the left and right auditory systems to understand why there may be a leftward bias for temporally rich information (e.g. speech processing) and a rightward bias for spectrally rich information (e.g. music; Albouy et al. 2020; Poeppel 2003; Zatorre et al. 1992; Zatorre and Belin 2001). Future studies should aim to characterize which features of the typical dominant hemisphere are present in the RH of people with an LH perinatal stroke who develop normal language processing, to further elucidate which features may be important for the implementation of a language system.

A small percentage of people are naturally right-lateralized for language processing, even with an intact LH. A few studies have investigated the organization of the language system in samples of healthy adults who exhibit atypical lateralization (Szaflarski et al. 2002; Gerrits et al. 2020; Labache et al. 2020), but none to our knowledge have directly compared the functional layout of regions recruited in the LH of left-dominant individuals and the RH of right-dominant individuals. One study by Chang et al. examined language mapping in presurgical epilepsy and tumor patients who exhibited a right-lateralized or bilateral language response during Wada testing. Using electrocorticography, they probed the loci for language disruption (speech arrest, anomia, and alexia) and found that the mapping of language processing in right-lateralized individuals mirrored the mapping in left-lateralized individuals (Chang et al. 2011). It would be important to discover whether RH activation during sentence processing in right-lateralized healthy adults aligns with our findings in the LHPS RH. If RH language recruitment in healthy individuals is also more similar to activation in the typical dominant LH than the typical non-dominant RH, then perhaps there is an optimal arrangement of regions for organizing a language network in one dominant hemisphere (left or right).

It has been suggested previously that there are symmetrical perisylvian networks in the young developing brain that may be “equipotential” for language processing (Newport et al. 2017, 2022; Lenneberg 1967, 1969; Mbwana et al. 2009; Rosenberger et al. 2009; Tivarus et al. 2012). Newport et al’.s “Developmental Origins Hypothesis” (Newport et al. 2017) suggests that cognitive functions such as language are more distributed in children, and after an injury early in life such as a stroke to left perisylvian cortex, these distributed representations (e.g. in right perisylvian cortex) are available to support developing the function in the RH. Adding empirical support to this idea, Olulade et al. found that young typically developing children activate right perisylvian cortex during language processing more than healthy older children and adults do (though even young children activate left perisylvian cortex more robustly than the right). Our current findings indicate that the RH areas recruited for sentence processing in people with a large cortical LH perinatal stroke are in a spatial arrangement that is more closely aligned with the activity in typical left perisylvian regions than the activity in typical right perisylvian regions. It is unknown whether the RH areas recruited in people with an LH perinatal stroke are a reflection of highly symmetrical left and right systems for language that may exist at the very earliest stages of development, or whether these areas become recruited as a result of injury-induced and developmental plasticity under these extreme circumstances.

By examining how language processing is successfully implemented in the RH after an LH perinatal stroke, we also hope to gain insight into how a similar outcome may be achieved in the RH after adult stroke. However, if the RH must develop a pattern of cortical recruitment during language processing that is similar to the typical LH and measurably different from the typical RH, as our current results suggest, the degree of neuroplasticity required may only be available during the developmental critical period. That is, recovering language processing in the RH after a stroke in adulthood may require different mechanisms than those involved in developing a language system in the RH after perinatal stroke. For example, language processing in the RH after stroke in adulthood may be limited to the areas that are weakly active in the typical non-dominant RH; but as we have discussed in prior work (Martin et al. 2022), increasing the strength of activity in these regions may nonetheless be important for recovering language abilities.

Conclusion

In individuals with a large cortical LH perinatal stroke, RH regions that are homotopic to typical LH language regions are recruited for language processing following a highly similar spatial layout. This result confirms an observation made by several previous studies on the same population, but, importantly, provides a rigorous, quantitative assessment of homotopicity. We additionally find that the RH recruitment for sentence processing in people with a large cortical LH perinatal stroke is more similar to the recruitment in the typical dominant LH than the typical non-dominant RH (even when the amount of activity being compared is equated). This suggests that the particular spatial arrangement of the perisylvian regions activated, especially in IFG and STS, is a potentially important feature of cerebral dominance.

Funding

This work was supported by funds from Georgetown University and MedStar Health and the Solomon James Rodan Pediatric Stroke Research Fund, the Feldstein Veron Innovation Fund, and the Bergeron Visiting Scholars Fund to the Center for Brain Plasticity and Recovery; NIH (Grants K18DC014558 to E.L.N. and R01DC016902 to E.L.N. and W.D.G.); American Heart Association (Grant 17GRNT33650054 to E.L.N.); NIH (Grant P50HD105328 to the District of Columbia-Intellectual and Developmental Disabilities Research Center at Children’s National Hospital and Georgetown University); and NIH (Grant T32NS041218 to K.C.M.) via Georgetown University's Center for Neural Injury and Recovery.

Conflict of interest statement: None declared.

CRediT statement

Kelly Martin (Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing—original draft, Writing—review and editing), Anna Seydell-Greenwald (Data curation, Investigation, Methodology, Resources, Supervision, Writing—review and editing), Peter Turkeltaub (Investigation, Methodology, Resources, Supervision, Writing—review and editing), Catherine E. Chambers (Data curation, Project administration, Writing—review and editing), Margot Giannetti (Data curation, Project administration), Alexander W Dromerick (Funding acquisition, Resources), Jessica L. Carpenter (Resources, Writing—review and editing), Madison M. Berl (Methodology, Resources, Writing—review and editing), William D. Gaillard (Funding acquisition, Methodology, Resources, Writing—review and editing), Elissa Newport (Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing—review and editing).

Data availability

For access to the unthresholded activation maps, go to https://osf.io/dp23q.