Isolating the white matter circuitry of the dorsal language stream: Connectome‐Symptom Mapping in stroke induced aphasia

Abstract The application of ℓ1‐regularized machine learning models to high‐dimensional connectomes offers a promising methodology to assess clinical‐anatomical correlations in humans. Here, we integrate the connectome‐based lesion‐symptom mapping framework with sparse partial least squares regression (sPLS‐R) to isolate elements of the connectome associated with speech repetition deficits. By mapping over 2,500 connections of the structural connectome in a cohort of 71 stroke‐induced cases of aphasia presenting with varying left‐hemisphere lesions and repetition impairment, sPLS‐R was trained on 50 subjects to algorithmically identify connectomic features on the basis of their predictive value. The highest ranking features were subsequently used to generate a parsimonious predictive model for speech repetition whose predictions were evaluated on a held‐out set of 21 subjects. A set of 10 short‐ and long‐range parieto‐temporal connections were identified, collectively delineating the broader circuitry of the dorsal white matter network of the language system. The strongest contributing feature was a short‐range connection in the supramarginal gyrus, approximating the cortical localization of area Spt, with parallel long‐range pathways interconnecting posterior nodes in supramarginal and superior temporal cortex with anterior nodes in both ventral and—notably—in dorsal premotor cortex, respectively. The collective disruption of these pathways indexed repetition performance in the held‐out set of participants, suggesting that these impairments might be characterized as a parietotemporal disconnection syndrome impacting cortical area Spt and its associated white matter circuits of the frontal lobe as opposed to being purely a disconnection of the arcuate fasciculus.

The pioneering work of 19th and 20th century aphasiologists have shown us that cerebrovascular lesions may impair this auditorymotor circuit leading to the syndrome of conduction aphasia (Benson et al., 1973;Geschwind, 1965;Wernicke, 1874Wernicke, /1977. The symptom complex of conduction aphasia includes frequent phonological errors during production and difficulty with verbatim repetition, while preserving the comprehension of speech itself (Ardila, 2010;Buchsbaum et al., 2011;Catani & Mesulam, 2008;Goodglass, 1992). Classically, the critical lesion causing conduction aphasia was thought to be the arcuate fasciculus (Geschwind, 1965), however, more recent work has implicated cortical disruption in the posterior Sylvian region as a major source of the deficits (Anderson et al., 1999;Buchsbaum et al., 2011;Damasio & Damasio, 1980;Quigg & Fountain, 1999). At the same time, it is undeniable that fluent speech production relies, at least in some part, on the integrity of subjacent white matter connections (Benson et al., 1973;Damasio & Damasio, 1980). Cerebrovascular lesions that produce conduction aphasia nearly always extend into white matter. More recently, intraoperative direct electrical stimulation studies have shown that transient phonological paraphasias can occur upon stimulation of perisylvian white matter (Duffau, 2015; Moritz-Gasser & Duffau, 2013 for a review). Therefore, white matter involvement appears likely and mapping their organization will have important implications for resections involving white matter beneath eloquent cortical areas, which may induce undesirable speech deficits postoperatively (Ellmore, Beauchamp, O'Neill, Dreyer, & Tandon, 2009).
Although tractography studies have awarded unprecedented access to in vivo connectional neuroanatomy, the methodologies being used to make brain-behavior correlations have lagged behind considerably. Preoccupied with validating classical assumptions, previous studies have investigated the white matter correlates of speech repetition, a core ability supported by the dorsal stream network, by restricting analyses solely to the arcuate fasciculus (Berthier, Ralph, Pujol, & Green, 2012;Forkel et al., 2020;Sierpowska et al., 2017;Yeatman et al., 2011), resulting in a spatial bias which underestimates the complexity of the underlying anatomy. Indeed, the white matter configuration of the inferior parietal lobule alone is not only a passageway for the classical AF, but is also a convergence zone for several noncanonical association fibers. Among them, are the middle longitudinal fasciculus (MdLF), the so-called fronto-parietal (SLFIII) and temporo-parietal segments of the AF (SLFtp), the vertical occipital fasciculus (VOF), and the second branch of the superior longitudinal fasciculus (SLFII). Disentangling these pathways using traditional userdependent tractography methods will likely fall victim to methodological idiosyncrasies that is further compounded by the disagreements in pathway terminations (Mesulam, Thompson, Weintraub, & Rogalski, 2015) as well as in their proposed segmentations (Glasser & Rilling, 2008;Makris et al., 2013).
By applying regularization penalties on high-density connectomes, elements of the feature space (i.e., connections) can be identified on the basis of their predictive value with a response variable (Hastie, Tibshirani, & Wainwright, 2015) rather than through conventional univariate association methods , which might not generalize to independent datasets. Generally, aphasia studies have been limited in sample size and given that conduction aphasia is relatively rare among the aphasia taxonomies, this limitation has been especially pronounced when investigating its underlying biological mechanisms. To approach this problem, we measured repetition performance in a relatively large sample of subjects with stroke induced aphasia. By comprising a multitude of aphasia classifications, this cohort presents a unique opportunity to identify associations between white matter disconnection and speech repetition impairment, as each of the subtypes vary considerably with respect to their cerebrovascular lesions as well as in the extent to which repetition is disrupted. This association was established using a regularized latent projectionbased algorithm-sparse Partial Least Squares-Regression (sPLS-R)-in order to select a subset of white matter connections most predictive of repetition performance; sPLS aims to generate a parsimonious model by discarding non-informative features when optimizing prediction error (Lê Cao et al., 2009). In contrast to conventional univariate methods, sPLS uses dimensionality reduction via multivariate latent projections which accommodates both high-dimensionality and collinearity of the feature space (Rohart, Gautier, Singh, & Lê Cao, 2017).
Multicollinearity is a pertinent issue in voxel-based (VLSM) and connectome-based (CLSM) lesion studies wherein elements of the feature space are highly interdependent, as lesions tend to impact adjoining voxels-likely inflating the type I error rates (Arnoux et al., 2018). Although PLS is a well-established method in neuroimaging analysis (McIntosh, Bookstein, Haxby, & Grady, 1996), to our knowledge, it is regularized counterpart-used in high-throughput "Omics" research from computational biology (Lê Cao, Boitard, & Besse, 2011)-has not yet been applied to study the connectOmics of speech and language (Sporns, 2013). The objective of the present study is to implement sPLS on the high-dimensional connectome to first generate a parsimonious model of white matter features which will then, in turn, be used to generate predictions for speech repetition on an independent test set. By algorithmically ranking the most informative features based on their predictive value, we believe this approach will isolate the white-matter correlates of the speech repetition circuit in a data-driven, spatially unbiased manner.
2 | METHODS 2.1 | Participants, behavioral evaluations, and outcome measure Data from seventy-one participants (27 females, mean age = 60.37 ± 11.3, 7 with atypical handedness) with chronic aphasia (≥12 months post-stroke) were analyzed here. Participants were recruited as part of a larger aphasia treatment study conducted at the University of South Carolina and Medical University of South Carolina. Only individuals with aphasia resulting from ischemic or hemorrhagic stroke to the left hemisphere were included. Participants with lacunar infarcts or with damage that involved the brainstem or cerebellum were excluded. All study procedures were approved by Institutional Review Boards at both institutions.
As part of this trial, individuals completed an extensive battery of cognitive-linguistic testing and neuroimaging at baseline and at various post-treatment time points (Kristinsson et al., 2021). Details about this trial can be found in Spell et al. (2020). All data included here were obtained at baseline, before initiating treatment. To test for repetition ability, the Philadelphia Repetition Test (Dell, Martin, & Schwartz, 2007) was used which evaluates repetition at a single-word level. The PRT is a modification of the Philadelphia Naming Test (PNT) in which participants hear a pre-recorded audio file and are asked to repeat exactly what they hear. In addition, the repetition subtest of the revised Western Aphasia Battery-Revised (WAB-R) was used to evaluate repetition at both the single-word and phrase level (Kertesz, 2007). The scores for the PRT and WAB-R were then normalized into percentages by dividing the respective columns by their maximum value. These measures were strongly correlated (r = .84), and were thus averaged to form a single composite score representing the participant's ability to perform repetition tasks at both the single word and phrasal levels. Lastly, subsequent analyses were performed on the composite repetition scores after correcting for the effects of age and months post-stroke. Of the 71 participants enrolled in this study, the following seven aphasia subtypes were observed: Broca's (34 participants), anomic (15 participants), conduction (10 participants), Wernicke's (4 participants), global (4 participants), Transcortical-Motor (1 participant), and no aphasia (3 participants).

| Image acquisition, lesion mapping, and preprocessing
Imaging was acquired on a Siemens Prisma 3 T scanner equipped with a 20-element head/neck (16/4) coil at the University of South both were blinded to behavioral scores at time of lesion drawing.
Using SPM12 and MATLAB scripts developed in-house, the binary stroke lesion maps were spatially normalized to standard space through the following steps: (i) The T2 scan was co-registered with the individual's T1 scan with the transforms used to resliced the lesion into native T1 space; (ii) the resliced lesion maps were smoothed with a 3 mm full-width at half-maximum Gaussian kernel to remove jagged edges associated with manual drawing; (iii) an enantiomorphic normalization (Nachev, Coulthard, Jäger, Kennard, & Husain, 2008)

| Structural connectome analysis
Diffusion images were undistorted using TOPUP (Andersson, Skare, & Ashburner, 2003) and Eddy (Andersson & Sotiropoulos, 2016). . sPLS-R was implemented using the "mixomics" package in R, dedicated to the multivariate analysis of biological "omics" datasets (Rohart et al., 2017). First, a 70/30 split was performed on the entire dataset such that 50 subjects (70% of the dataset) would be used to build a model whose predictions could then be tested on the remaining 21 subjects (30% of the dataset). The hyperparameters for sPLS are the number of components to retain (H) and the ℓ1-regularization penalty (keepX), selected on the basis of optimizing prediction accuracy using mean absolute error). Here, we chose to fix the H at five while evaluating a list of different penalties with a maximum keepX of 100, tuned using repeated k-fold cross validation (k = 5, repeats = 50).

| Variable selection using the Bootstrapvariable importance in projection (VIP) approach
In order to rank the retained features on the basis of their predictive value, the variable importance in projection (VIP) method was used (Chong & Jun, 2005;Farrés, Platikanov, Tsakovski, & Tauler, 2015).
VIP is regarded as the impact of a given variable into the construction of the PLS components, weighted by the explained variance across the components. Features with a large VIP score, larger than one, have been shown to indicate relevance for explaining the outcome measure and this cutoff is widely used as a criterion for variable selection (Chong & Jun, 2005;Colombani et al., 2012;Farrés et al., 2015).
In order to establish distributions around the VIP estimates, the bootstrap-VIP method (Gosselin, Rodrigue, & Duchesne, 2010) was implemented wherein the sPLS model tuning procedure was replicated 8,000 times using bootstrap resamples of the training set and the VIP scores for the principal PLS component at each iteration was recorded. If a variable is truly important in predicting the outcome measure, we may expect it to not only consistently survive the regularization penalty but also maintain a relatively strong VIP score across the different pseudo-independent datasets (Afanador, Tran, & Buydens, 2013;Colombani et al., 2012). For this reason, the candidate pathways for subsequent analysis were identified using the aforementioned "greater than one" threshold on the mean of the bootstrap distribution (Chong & Jun, 2005;Farrés et al., 2015), while imposing that these features survived the regularization penalty on the majority of resampling iterations (selection stability frequency greater than 50%) (Lê Cao et al., 2011;Tillisch et al., 2017). Lastly, the resulting features were used to fit a PLS model on the training set and its predictions were then evaluated for statistical significance using their correlations with repetition scores from the unseen test set.

| RESULTS
The complete feature space of 2,591 connections is displayed as a chord diagram in Figure 1 using the circlize package in R (Gu, Gu, Eils, Schlesner, & Brors, 2014), whose arcs represent each of the distinct connections between brain regions defined by the AICHA atlas

| DISCUSSION
Using a high-dimensional connectomics approach combined with algorithmic feature selection, the present study isolated the white matter substrates for speech repetition among the broader set of cerebral white matter. By obviating the need to restrict analysis to a priori selected pathways, we show how modern machine learning algorithms could leverage connectomics data to generate novel hypotheses for uncovering brain-behavior relationships when studying the neurobiology of language. It is worth emphasizing that most- if not all-studies on the white matter correlates of speech repetition have focused exclusively on the classical arcuate fasciculus (AF) given its implication from classical neurobiological models (Berthier et al., 2012;Forkel et al., 2020;Yeatman et al., 2011). The connectomic approach implemented here identified short-and longrange connections emanating from the superior temporal cortex and parieto-temporal junction as essential for repetition function ( Figure 3). Of the connections identified, an optimal subset of 10 white matter features were capable of making significantly accurate predictions on the unseen test set-thereby attesting to the generalizability of our findings ( Figure 5, panel 4). Fifty percent of the optimal features were short-range connections in the inferior parietal lobe, with preferential contributions from short supramarginal white matter closely delineating cortical area Spt (Table 2) Consistent with modern imaging studies, our results underscore the importance of the parieto-temporal cortex for the repetition of speech (Buchsbaum et al., 2011;Fridriksson et al., 2010;Hickok & Poeppel, 2007;Yourganov, Fridriksson, Rorden, Gleichgerrcht, & Bonilha, 2016). Lukic et al. (2019) have recently shown a relationship between the cortical thickness in parieto-temporal cortex and repetition performance-a finding that has since been replicated by Forkel and colleagues (Forkel et al., 2020;Lukic et al., 2019). Similarly, Rogalsky and colleagues-using a VLSM approach-localized repetition impairments to focal brain damage in the same vicinity (Rogalsky et al., 2015). Combined with these studies and others (Buchsbaum et al., 2011;Hickok & Poeppel, 2007;Isenberg, Vaden Jr, Saberi, Muftuler, & Hickok, 2012), the present work highlights the significance of parieto-temporal cortical area Spt (Hickok et al., 2003) and of its subjacent subcortical connections for auditory-motor integration F I G U R E 5 PLS model selection.
(1) Three connectomic PLS models were evaluated by applying progressively higher feature selection thresholds on the 16 candidate features: a model containing the complete set of 16 features (dark gray), a model containing the top 10 features (orange), and a model containing the top two features (red).
(2) The three models were trained on the training set (n = 50) (2a-c) and their performances were evaluated on the held out test set (n = 21) using correlations computed between fitted and actual repetition scores for each respective model (3a-c). (4) On the basis of its predictive performance on the test set (R = 0.45, p = .02), the connectomic PLS model containing the top 10 features was chosen as the optimal model (4a). The actual versus fitted scores of the optimal model are plotted with respect to the aphasia subtypes present in the test set (4b) (Figure 3c). This claim was justified by the feature selection procedure, as nearly 70% of the candidate features had terminations in either superior temporal or supramarginal nodes (Figure 3). Furthermore, when fitting the PLS model using the optimal 10 features across the entire dataset (i.e., on the combined training and test set), the feature with the strongest VIP and loading coefficient was a short-range connection within the supramarginal gyrus (G_Supramarginal-4$G_SupraMarginal-2) which precisely overlaps with area Spt ( Figure 5, Table 2) (Isenberg et al., 2012).
One of the distinct feature clusters sharing strongly correlated connectivity patterns was localized to the inferior parietal lobule ( Figure 4, left), with connections spanning the intraparietal sulcus and S_Intraparietal-2$G_Angular-1 0.29 0.9 5.35 (6.55) 28.08 Note: PLS component coefficients and feature statistics. Loadings and variable importance in projection (VIP) coefficients of the principal component resulting from the partial least squares (PLS) model fit on the optimal 10 features across the entire dataset (n = 71). Means, standard deviations, and ranges are also reported for each individual feature.
(a) (b) F I G U R E 6 Connectome-Symptom Mapping. Using the same descending order shown and described in Figure 2, barplots of the normalized probabilistic streamline counts are shown across the entire group (N = 71) for each of the optimal 10 features identified. Both the behavioral patterns and structural connectivity measures show similar descending patterns across the aphasia taxonomies angular gyrus. Indeed, several such connections of the intraparietal sulcus persisted when selecting the predictive model that generalized to the unseen test set (Figures 5 and 6). Interestingly, Geschwind had originally speculated that association cortex of the angular gyrus was likely involved in Wernicke's aphasia-later called Geschwind's territory (Catani & Mesulam, 2008)  In addition to the short-range features identified in the parietal and temporal cortex, the present work stresses the importance of several long-range connections originating from these areas and terminating in frontal premotor cortex (Figures 3 and 4). To our knowledge, this is the first large-scale study to implicate a structural connection to dorsal premotor cortex as being involved in repetition of speech. Classical neurobiological models of language make no mention of an area yet contemporary models appear to suggest a role for a dorsal premotor area in laryngeal motor control (Dichter, Breshears, Leonard, & Chang, 2018;Hickok, 2017). The dual stream theoretical model proposed by Hickok and Poeppel (2007) noted a close coupling between sensory-motor area Spt and a dorsal premotor area involved in articulation (Hickok & Poeppel, 2007).
Recently, functional imaging studies have localized an area in close proximity to this region-area 55b-which appears to activate during language tasks (Glasser et al., 2016). Similarly, recent direct electrical stimulation studies have shown that stimulation of this dorsal premotor region causes speech disturbances intraoperatively (Hazem et al., 2021;Rech et al., 2019) and this area is both functionally and structurally connected to the superior temporal cortex (Barbeau, Descoteaux, & Petrides, 2020;Rech et al., 2019). Moreover, neurosurgical evidence suggests that tumor resection within the dorsal premotor cortex can produce long-term speech production deficits characterized by apraxia and impairments with repetition (Chang et al., 2020).
These authors also demonstrated white matter connections to exist between this dorsal frontal area and the temporal lobe via the superior longitudinal fasciculus. Indeed, a recent (anatomical study) from Barbeau and colleagues have published evidence that a dorsal branch of the arcuate fasciculus exists in humans, which terminates in the posterior dorsolateral frontal region anteriorly and the superior temporal cortex posteriorly (Barbeau et al., 2020)-precisely the connectivity pattern identified in the present behavior-connectomic analysis.
Thus, we provide novel evidence of a structure-function relationship between this pathway and a prominent language function.
The association between conduction aphasia and apraxia has been shown particularly in the presence of supra-sylvian lesions located deep to the inferior parietal lobe (IPL); specifically near the supramarginal gyrus (Basilakos, Rorden, Bonilha, Moser, & Fridriksson, 2015;Benson et al., 1973;Geschwind, 1965;Mendez & Benson, 1985;Poncet, Habib, & Robillard, 1987). This association may potentially be mediated by connectivity between area Spt and its communication with this dorsal premotor speech area via the 55b-Spt pathway critical for laryngeal motor control. These data suggest that lesions to the IPL may disrupt communication between parietotemporal sensory-motor cortex and dorsal premotor areas critical for buccofacial (Benson et al., 1973) and laryngeal motor control (Hickok, 2017), as these areas are functionally (Glasser et al., 2016;Rech et al., 2019) and structurally (Chang et al., 2020;Rech et al., 2019) interconnected. Among the 16 features identified in this study, the  (Matchin & Hickok, 2020), this characterization appears more likely than it playing the major role in phonological aspects of speech. If the AF were indeed involved in repetition, we would expect to find that damage to its anterior termination in Broca's area would emerge as a strong cortical correlate in VLSM studies-a finding not supported by recent work (Rogalsky et al., 2015). Parietotemporal regions superior to Wernicke's area increasingly appear critical for auditory-verbal working memory storage and rehearsal processes (Lukic et al., 2019;Rogalsky et al., 2015) in order to maintain complex stimuli online while simultaneously integrating with articulatory-phonological areas that are coupled with laryngeal motor cortex via the Spt-55b and Spt-vPMC pathways identified here. Taken together, the literature and the present work support the notion that the posterior STG (Basilakos, Smith, Fillmore, Fridriksson, & Fedorenko, 2018) and its underlying connections, are involved in the sensory guidance of speech whose disruption drastically impairs both production and repetition ( Figure 6).

| CONCLUSION
The purpose of the present study was to isolate pathways critical for speech repetition using a data-driven feature selection algorithm applied on the structural connectome. By evaluating this relationship in stroke patients presenting with varying lesions and varying degrees of impairment, this connectome-based lesion symptom mapping approach successfully highlighted a focal set of superficial parietotemporal connections as being essential for the prediction of repetition performance. This finding corroborates classical and contemporary models by indicating that the repetition impairments observed in conduction aphasia might rather be characterized as a parietotemporal disconnection syndrome impacting cortical area Spt and associated frontal circuits as opposed to being explained as purely a disconnection of the classical arcuate fasciculus. The study also identified an additional important circuit involving superior temporal connectivity to a dorsal premotor site that calls for further investigation in the future. In conclusion, machine learning analyses frameworks offer the unique capacity to assess clinico-anatomical correlations in a theoretically unbiased manner, leading to novel insights into the neurobiology of language without having to reduce the complexity of the underlying anatomical feature space that is often characteristic of high-dimensional neuroimaging studies.

DATA AVAILABILITY STATEMENT
The data used in this study are available to researchers upon qualified request to the corresponding author.