Altered efficiency of white matter connections for language function in children with language disorder

Abstract

To characterize structural white matter substrates associated with language functions in children with language disorders (LD), a psychometry-driven diffusion tractography network was investigated with canonical correlation analysis (CCA), which can reliably predict expressive and receptive language scores from the nodal efficiency (NE) of the obtained network. The CCA found that the NE values of six regions: left inferior-frontal-opercular, left insular, left angular gyrus, left superior-temporal-gyrus, right hippocampus, and right cerebellar-lobule were highly correlated with language scores (ρ_expressive/ρ_receptive=0.609/0.528), yielding significant differentiation of LD from controls using new imaging predictors u_expressive (F=15.024, p=.0003) and u_receptive (F=7.421, p=.009). This study demonstrates the utility of intrinsic language network analyses in distinguishing and potentially subtyping the type and severity of language deficit, especially in very young children (≤3 years) with LD. The use of structural imaging to identify children with persisting language disorder could prove useful in understanding the etiology of language disorder.

1. INTRODUCTION

Language disorder (LD) is a condition in which children evidence difficulties in the acquisition and use of language due to deficits in the comprehension or production of vocabulary, sentence structure, and/or discourse (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, DSM-5, American Psychiatric Association, 2013). About 5-8% of U.S. children are affected (Burden, Stott, Forge, & Goodyer, 1996; Nelson, Nygren, Walker, & Panoscha, 2006), and impairments are by definition age- and language specific, not directly attributable to other sensorimotor, neurological, cognitive, or intellectual disability. Assessment and treatment of children with a potential LD is challenging for several reasons: LD encompasses a heterogeneous group of impairments with varied etiologies, it is difficult to estimate the severity of language dysfunction in very young children (i.e. distinguishing “late talkers,” whose language ability may later normalize by age 4 or 5, from more profound LD), and small children with language problems also frequently demonstrate comorbid behavioral concerns, including inattention, overactivity, and noncompliance, that renders assessment of their language functions unreliable.

Despite some nomenclatural ambiguity in the diagnosis of LD compared to more developmental causes of language disruption (e.g. DSM-5’s “LD” vs. the International Classification of Diseases 10^th edition’s “Specific developmental disorders of speech and language”), such complications do present at a developmentally-sensitive time and therefore motivate treatment with a sense of urgency: while intervention is delayed, language difficulties may negatively affect school performance and social function (see Girbau-Massana, Garcian-Marti, Marti-Bonmati, & Schwartz, 2014; Shevell et al., 2003; Vydrova et al., 2015), and the window of opportunity for early interventions, which bring greater likelihood of more positive outcomes (Bashir & Scavuzzo, 1992; Hustad, Sakash, Broman, & Rathouz, 2018), is closing. There is a critical need for strategies, and especially for objective tools circumventing the aforementioned diagnostic challenges, that help identify children with LD as characterized in the 2017 paper by Bishop et al.; those having language difficulties that are likely to be enduring and have impact on everyday life.

Neuroimaging may provide some such objective insights into LD patients by elucidating alterations in the brain structure that normally facilitates development and execution of language abilities. Previous T1-weighted MRI studies, for example, have found that children with LD show significant gray matter differences when compared with typically-developing children. Decreases in right postcentral parietal gyrus, bilateral medial occipital gyri, and right superior occipital gyrus gray matter volume have been associated with limited language understanding and knowledge representations (Girbau-Massana, Garcian-Marti, Marti-Bonmati, & Schwartz, 2014), while various differences in left inferior frontal gyrus, right caudate nucleus, and bilateral superior temporal gray matter volume have been associated with decreased language production and comprehension (Badcock, Bishop, Hardiman, Barry, & Watkins, 2012). Past research has shown that changes in white matter volumes are also associated with developmental LD, with both whole-brain increases (Herbert et al. 2003) and language- and motor-region specific decreases (Jäncke, Siegenthaler, Preis, & Steinmetz, 2007) observed in this population. Given the developing population studied in this research, several physiological processes could contribute to increased or decreased white matter volume, including changes to the density of neurons and axons or changes related to axonal development. However, it is unclear how regions outside the language network or across the whole brain contribute to relatively specific cognitive deficits in language development, and these findings highlight the need to address LD as a network disorder underpinned by brain structure connecting and facilitating heterogeneous regional function.

To address this issue, further research has utilized diffusion weighted imaging (DWI) tractography, which can help elucidate the network of white matter fiber tracts connecting various language-relevant regions in the brain. Such studies have reported LD-associated diffusion anisotropy abnormalities within this white matter network, involving regions such as the arcuate fasciculus, inferior fronto-occipital fasciculus, and inferior longitudinal fasciculus (Vydrova et al., 2015). Utilization of similar diffusion abnormalities to distinguish between multiple abnormal-language phenotypes, such as the presence or absence of a comorbid speech disorder, reveals that DWI signal in these patients differs more reliably in the motor-associated corticobulbar tracts, as opposed to strictly language-network tracts (Morgan et al., 2018). Complicating the interpretation of these markers further, other recent studies have shown that diffusion properties from language network pathways may distinguish developmental LD from healthy controls, even in the presence of co-morbid conditions like Dyslexia or Autism, but can only stratify measured language performance in healthy children (Verly et al., 2019). The structural brain network disrupted in LD is clearly heterogeneous and difficult to characterize. Employing a whole-brain connectome analysis to move beyond the known major language network, another recent study demonstrated that, compared with age and gender matched controls, children with LD showed significantly decreased connectivity between left fronto-temporal language regions, especially in the mid frontal and superior temporal gyrus (Jeong, Sundaram, Behen, & Chugani, 2016).

In the present study we attempt to combine the strengths of this past work, using DWI tractography across the whole brain to define a cohort-specific modular language network (Chai, Mattar, Blank, Fedorenko, & Bassett, 2016; Del Gaizo et al., 2017), identifying structural white matter pathways with values that correlate with Clinical Evaluation of Language Function (CELF-4) language scores. In this way, our approach can provide new insights toward identifying language networks that account for structural brain variation in our patient cohort, while still including known language-related brain regions. Cortical and subcortical regions are represented as nodes in the brain network, linked by measurements of DWI pathways connecting different nodes. Parametric correlation analyses can then be used to assess the relationship between neuropsychological measures of language and pair-wise connection strengths, helping identify the key nodes that constitute the language modular network of interest. The proposed modular network analysis takes advantage of quantitative graphic theory to fully describe whole-brain white matter alterations, innovatively identifying subtle but precise white matter abnormalities associated with LD phenotypes.

We hypothesize that 1) compared with typically developing children, children with LD will have significantly altered efficiency in both long and short range axonal connections within the modular language network, which will be centered on anatomically well-established language areas such as fronto-temporal cortical regions (Glasser & Rilling, 2008), hippocampus (Kepinska, de Rover, Caspers, & Schiller, 2018), and cerebellum (Mariën et al., 2014); 2) the combination of imaging variables in the modular language network (i.e., values of nodal efficiency) will accurately distinguish subtypes of LD children from healthy controls. To investigate these hypotheses, the present study employs canonical correlation analysis (CCA) to predict language scores of children with LD using measures of altered nodal efficiency in the modular language network.

2. MATERIAL AND METHODS

2.1. Subjects

This study recruited thirty-one children clinically diagnosed with LD (age: 4.25±2.38 years, 20 male) and sixteen controls with typical development (TD, age: 7.26±3.42 years, 9 male; age range 2-13 years) through the pediatric neurology clinic at Children’s Hospital of Michigan (CHM) in Detroit, MI USA. Intellectual and language functions were assessed using a battery of age- and developmentally-appropriate measures, which included age-specific choice of the Mullen Scales of Early Learning (Mullen, 1995), the Wechsler Preschool and Primary Scales of Intelligence (for children aged 2-6 years, Wechsler, 2012) and/or the Wechsler Intelligence Scale for Children (for children older than 6 years, Wechsler, 2014) to measure intellect, the Comprehensive Evaluations of Language Functions (CELF-P, Semel, Wiig, & Secord, 2006) (to measure expressive and receptive language functions in children aged 2-5 years), and CELF-4 (Semel, Wiig, & Secord, 2003), and the Token Test for Children (Gallardo, Guardia, Villasenor, & McNeil, 2011) to measure expressive and receptive language functions, respectively in children over 5 years of age, and the Vineland Adaptive Behavior Scales to measure communication, daily living, and socialization skills in all children. The assessment also included measures of fine motor dexterity (using the age-appropriate pegboard, Purdue Pegboard for children <5 years, Tiffin, 1968; Grooved Pegboard for children >5 years, Trites, 1989), internalizing and externalizing behavioral problems (CBCL, Achenbach & Rescorla 2001), and caregiver report and direct observations of autism spectrum problems (Autism Diagnostic Observation Schedule, 2^nd Edition, ADOS-2, appropriate module based on child’s age and verbal level, Lord et al., 2012). All of the measures included in the study have well-established psychometric properties and are very widely used in clinical and research populations as measures of intellect and language functions, respectively.

All psychometric indices were converted to t-scores (mean = 50; SD = 10). Mean verbal and nonverbal intellectual indices for the LD group were (41.4+5.8); mean expressive and receptive language scores for the group were 33.04±10.09 and 29.16±7.29 in expressive and receptive functions, respectively. These scores were used as ground truth data to define the modular language network and generate our predictive model using CCA. The time interval between assessment of intellect and language functions and MRI (Δt = date of assessment – date of MRI) was 2.65±1.84 years for the children in the LD group; children in the TD group completed the MRI on the same visit as their psychological evaluation.

All children in the LD group were seen by their neurologist at CHM Pediatric Neurology – potential participants included children for whom a clinical MRI was ordered; children then underwent psychological evaluations, including assessment of language functions. Children in the typically developing group were recruited as part of a study aimed at acquiring imaging data on a sample of typically developing children that might be used as controls in many of our studies of children with various developmental and psychiatric conditions.

2.2. Inclusion/exclusion criteria

All participants were monolingual English and were right-handed, and all of the children had hearing measured within normal limits. All of the children in the LD group had clinical diagnoses of language disorder as defined by DSM-5 criteria and consistent with the population described in the Bishop et al., 2017 paper (Statement 2), and more specifically, would also meet the criteria for the group of children having developmental LD as defined in the Bishop et al., 2017 paper (Statement 7), based on clinical psychological evaluations. All of the children with language disorders had nonverbal intellect measured within 1.5 standard deviations of the normative mean (t-score>35), and also at least one measure of language function measured below 1.5 SDs of the normative mean (SS<36) – see Table 1 for a full characterization of the neurocognitive and behavioral characteristics of the children in the language disorder group, including intellectual, language, motor functions, and internalizing and externalizing behavioral problems for each of the participants. All of the children in the LD group also were measured below the clinical cutoff (on caregiver report and direct observations) for an autism spectrum disorder. Children with any medical conditions, including seizure disorder, were excluded from the study, as were children with gross abnormality (i.e., tumor, heterotopia) on T1 weighted MRI. All of the children in the TD group had measured intellect SS>84 (mean = 100, SD=15) and none of the children in this group had any clinical elevations t-scores<66 (mean=50, SD=10) on any scales on a caregiver reported behavioral checklist.

Table 1.

Neurocognitive characteristics of participants in the language disorder group.

Participant	Age at assessment (years)	Age at MRI (years)	NVI Q*	Expressive Language*	Receptive Language*	Motor Dominant Hand	Externalizing Behavior**	Internalizing Behavior**
LD1	5.7	1.0	40	28.0	28.0	wnl	60	51
LD2	6.1	1.3	47	45.3	25.3	wnl	56	71
LD3	2.7	1.8	42	30.0	30.0	impaired	72	54
LD4	4.0	2.2	45	20.0	26.0	impaired	50	50
LD5	5.7	2.3	35	26.7	30.0	impaired	73	58
LD6	5.7	2.5	46	36.7	36.7	wnl	66	71
LD7	4.0	2.7	34	33.3	36.7	impaired	56	58
LD8	6.9	2.8	37	36.7	20.0	impaired	80	72
LD9	7.0	2.8	45	32.0	36.0	impaired	51	80
LD10	3.6	3.0	37	32.0	20.0	impaired	59	60
LD11	7.8	3.1	54	36.7	26.7	wnl	60	67
LD12	5.8	3.2	42	53.3	36.0	impaired	50	55
LD13	6.6	3.6	38	34.0	28.0	impaired	52	57
LD14	8.9	3.6	37	20.0	23.3	impaired	60	55
LD15	5.1	3.8	40	20.0	20.0	impaired	51	67
LD16	3.1	3.9	36	20.0	20.0	impaired	55	52
LD17	6.6	3.9	42	33.3	28.7	wnl	72	73
LD18	7.2	4.0	41	23.3	44.7	wnl	53	66
LD19	10.0	4.0	42	33.3	38.7	wnl	57	63
LD20	4.1	4.4	36	20.0	20.0	impaired	66	65
LD21	8.4	4.4	37	38.7	30.7	wnl	75	71
LD22	8.3	4.7	38	37.3	48.7	wnl	57	58
LD23	5.1	4.8	38	20.0	20.0	impaired	82	65
LD24	5.5	5.3	45	20.0	30.0	wnl	50	50
LD25	7.3	5.9	37	56.0	26.7	wnl	73	66
LD26	8.1	6.1	37	39.3	28.7	impaired	74	59
LD27	7.8	6.2	60	50.0	36.0	wnl	55	55
LD28	11.3	6.5	42	42.0	32.0	wnl	54	56
LD29	7.8	6.8	37	29.0	25.0	wnl	60	52
LD30	12.9	7.7	48	42.0	25.0	wnl	50	50
LD31	14.0	13.6	48	35.3	26.7	impaired	51	61

^*T-score, mean = 50, SD = 10 (higher scores = better functioning)

^**T-score, mean = 50, SD = 10 (higher scores = higher level of problems); Time interval between assessment of intellect and language functions and MRI was mean = 2.65 and SD = 1.84 years

The present study was approved by the Wayne State University’s Institutional Review Board and written informed consent was obtained from all parents/guardians.

2.3. MRI data acquisition and pre-processing

All MRI scans of LD and TD groups were acquired at the Children’s Hospital of Michigan on a 3T GE-Signa scanner (GE Healthcare, Milwaukee, WI) equipped with an 8-channel head coil and ASSET. DWI was acquired with a multi-slice single-shot diffusion weighted echo-planar imaging sequence at TR = 12,500 ms, TE = 88.7 ms, FOV = 24 cm, 128×128 acquisition matrix, contiguous 3 mm slice thickness, 55 isotropic gradient directions with b = 1000 s/mm², one b = 0 acquisition, and NEX = 1. T1-weighted structural images were also acquired using a 3D fast spoiled gradient-echo sequence at TR/TE/TI of 9.12/3.66/400 ms, with a section thickness of 1.2 mm and a planar resolution of 0.94x0.94 mm².

Before performing the tractography analysis, NIH TORTOISE and FSL topup packages (Andersson, Skare, & Ashburner, 2003) were used to correct motion, noise, physiological artifacts, susceptibility-induced distortion, and eddy current-induced distortion in the DWI data. Whole brain streamline tractography was then reconstructed by probabilistic spherical-deconvolution informed filtering of tractograms (SIFT2) tractography (Smith, Tournier, Calamante, & AlanConnelly, 2015) with second-order integration over fiber orientation distributions (iFOD2) to sample three FOD at every voxel (Tournier, Calamante, & Connelly, 2010). At every voxel of gray/white matter boundary identified by FSL FAST package, 100 dynamically randomized seeding points were applied in the framework of anatomically constrained tractography (Smith, Tournier, Calamante, & Connelly, 2012) in order to reconstruct biologically realistic streamlines from SIFT2 tractography using MRtrix3 package (http://www.mrtrix.org/).

It should be noted that DWI data were obtained at fairly young ages in children with LD (n=23 under 5 years old). These younger ages associated with cognitive language development also coincides with a dynamic development of underlying brain structure. Past research has tracked such structural changes through infancy, childhood, and adolescence using MRI, associating periods of brain growth, axonal myelination, synaptogenesis, synaptic pruning and remodeling with detectable signals on MRI (Reviewed in Lebel & Deoni, 2018). The major white matter tracts are present by the end of normal gestation, but the volume of these tracts continues to grow throughout typical development, increasing most significantly in the first year of life but continuing to grow at a more tapered rate until adulthood. As myelination also continues throughout early life, diffusion MRI metrics such as fractional anisotropy (FA) increase, reflecting the more tightly insulated axonal orientations in developed connections that are beginning to support cognitive functions. This process begins in a stereotyped posterior-to-frontal and caudal-to-cranial fashion, and maturational changes in FA continue even into adulthood, increasing more focally in specific brain regions that may support specific and more advanced cognitive abilities (Erus et al., 2015). Language in particular is a neurocognitive ability supported by a structural network that continues to develop into adulthood, involving white matter tracts like the arcuate fasciculus (AF) that strengthen connections between focal language-critical and often left-lateralized brain regions as linguistic abilities mature (Eichert et al., 2019). In any study of typical or disordered development, it is critical to appreciate the heterogeneity of possible physiological changes underlying changes in MRI signal. It may not be apparent whether structural differences in a population are due to a disruption of existing brain structure, the addition of aberrant structure, or a simply deviation from an age-specific developmental stage, highlighting a need for innovative imaging approaches to developmental disorders that are able probe network properties in a patient-specific, cohort-specific, or age-independent manner.

2.4. Structural Connectome-based Analysis

Whole brain networks were constructed from DWI connectomes based on the automated anatomical labeling template (Tzourio-Mazoyer et al., 2002), parcellated into 116 brain regions including 90 from the cerebrum and 26 from the cerebellum. Two brain regions were considered to be structurally connected when the endpoints of a fiber tract originated from one region and terminated in the other. Pair-wise connections were thresholded such that any link must contain over 5 tracts, to avoid consideration of spurious connections. To extract the language related modular network from the whole brain network, Pearson’s correlation analysis assessed the correlation coefficient, R between total number of streamlines connecting two regions (i.e., connection strength) and expressive or receptive language score. All pair-wise connections having statistically significant values of R identified key nodes in the language modular network. To determine the nodal topological property in the language related sub-network, nodal efficiency (NE) was calculated for each included node (Cao et al., 2013; Widjaja et al., 2015). The NE of a given node i in a brain network is defined as follows:

$$NE(i)=\frac{1}{n-1}{\sum }_{i\ne j}\frac{1}{d(v_i,v_j)}$$ (1)

where d(v_i, v_j) is the shortest weighted path length (Rubinov & Sporns, 2010) which is equal to a minimum value among the total sum of the inverse of individual connection strengths between nodes v_i and v_j for all i ≠ j, n is the number of nodes, and v_i represents the i-th node in the network. This metric is also related to the average path length between a given node i and the remaining nodes in a network (Lawrence, Chung, Morris, Markus, & Barrick, 2014). NE can therefore be thought to reflect the importance of a brain node for information interactions within the network (Widjaja et al., 2015), and alterations in NE reflect a change or possible disruption in the connectivity between that brain node and the remainder of the network (Cao et al., 2013). Supplementary Figure 1 shows two representative examples of language modular networks with relative NE values, one from healthy control and another from LD.

2.5. Canonical correlation analysis

To identify the linear relationship between the neuropsychological language scores of children with LD and measures of altered NE in the modular language network, CCA (Hotelling, 1936) was performed between NE values and receptive or expressive language scores. In short, CCA is a method for exploring the regressions between two multivariate sets of random variables (X and Y) with zero-mean, which finds a pair of linear combinations: W_u and W_v to maximize the correlation coefficient (ρ) between two canonical variables (u = X^TW_u and v = Y^TW_v):

$$\rho =\frac{\text{cov}(u,v)}{\sqrt{\text{var}(u)\text{var}(v)}}=\frac{E[W_u^{T}X{Y}^T{W}_v]}{\sqrt{E[W_u^{T}X{X}^T{W}_u]E[W_v^{T}Y{Y}^T{W}_v]}}$$ (2)

The present study constructed a data matrix X = [x₁∣₂∣,…,∣_N] by concatenating N-NE(i) different column vectors of subjects. These columns contained demeaned values x_i=1,2,…,N showing significant correlations with a demeaned language score column vector of subjects Y, where y_expressive was constructed with expressive language subscores and y_receptive was constructed with receptive language subscores. The resulting W_u and W_v that maximize the canonical coefficient of ρ were finally applied to generate NE-based language score predictions, u_expressive and u_receptive. To optimize the CCA as a function of the modular language network density, we adjusted the network density from 10% to 50% using incremental 10% steps. The network density showing maximal ρ, and minimum mean absolute error (MAE), between the CCA-imaging-derived language score u and psychometrically-derived language score v was then selected to define the modular language network. Also, we evaluated P(AE<10) as a probability value in that absolute error (AE) between the CCA-imaging-derived language score u and psychometrically-derived language score v is less than 10. Many brain network studies consider a wide range of network densities for the following reasons: 1) there is no gold standard for setting a single and definitive network density threshold, 2) the presence of spurious connections can impact the estimation of network characteristics (Bernhardt, Chen, He, Evans, & Bernasconi 2011; Liu, Ke, Liu, Huang, & Hu, 2014; Liu et al., 2012; Tsai, 2018). In the present study, to minimize the number of spurious connections and optimize the CCA as a function of the modular language network density (ratio of the number of connections to the maximum possible number of connections between nodes), we adjusted the network density from 10% to 50% using incremental 10% steps as follows: ROI pairs with fewer than a given number of connections are removed from the modular language network.

2.6. Statistical analysis

ANCOVA controlling for age was utilized to evaluate how significantly CCA-predicted scores: u_expressive and u_receptive differentiated LD from healthy controls. The present study also utilized an affinity propagation clustering algorithm (Frey & Dueck, 2007) to investigate whether distinctive LD subgroups (expressive versus receptive language deficits) could be identified by using CCA-predicted scores and if different profiles of deficit across language domains would be reflected in modular language network efficiency differences. Each subject’s two CCA-predicted scores were selected as two dimensional feature vectors to determine the optimal clusters maximizing an objective function, called “net similarity”, which quantifies the degree of cluster compactness and between-cluster distance in the feature space. In the present study, the central scores of LD clusters (i.e., exemplars) were compared with those of healthy controls in order to assess whether the proposed CCA-based language scores can identify different subtypes which show different degrees (or combinations) of expressive and receptive language deficits in LD populations.

3. RESULTS

Pair-wise connectivity scores between a total of 31 ROIs (Fig 1 and Supplementary Table 1) showed significant correlations with the language scores, y_expressive and y_receptive at p < 0.01. In this 31-node modular language network, the NE values of six brain regions including x₁: left inferior-frontal-opercular, x₂: left insular, x₃: right hippocampus, x₄: left angular gyrus, x₅: left superior-temporal-gyrus, and x₆: right cerebellar-lobule showed significant correlations with either expressive or receptive language scores: respective Pearson’s R², p-values of expressive score, y_expressive = 0.635, .018 (x₁), 0.541, .037 (x₂), 0.659, .014 (x₄), 0.532, .040 (x₅), and 0.877, .001 (x₆); respective Pearson’s R², p-value of receptive score, y_receptive = 0.625, .020 (x₃).

Figure 1.

Identification of automated anatomical labeling nodes associated with the language related subnetwork (i.e., “modular network”) by using Pearson’s correlation analysis between pair-wise connection strength and language score. Red, blue and green spheres of top 3-D visualizations indicate 31 key nodes showing statistically significant Pearson’s correlation coefficient, R, with expressive, receptive, and both scores, respectively, at p < .01, respectively. Black spheres indicate the other 85 nodes having no significant correlations. The central matrix indicates a language related subnetwork whose elements consist of pair-wise connectivity strengths in 31 key nodes. The colorbar indicates the value of pair-wise connectivity strength which is the total number of streamlines connecting a give pair of nodes. From individual nodes of the subnetwork, nodal efficiency values are evaluated from LD children and correlated with expressive and receptive language scores. Five regions including x₁: L.IFGoperc, x₂: L.INS, x₄: L.ANG, x₅: L.STG and x₆: R.CRBL_10 shows significantly correlation with expressive language score (red-colored sphere) while one region, x₃: R.HIP, has significant correlation with receptive language score (blue-colored sphere) as marked in bottom 3-D visualizations. Black spheres indicate the other 29 key nodes having no significant correlation.

The most significant canonical correlation coefficient ρ between the CCA-imaging language score u and language score v was achieved using the sparsest network density (i.e., 10%, Table 2), and was associated with the following W_u and W_v,

$${\mathrm{u}}_{expressive}=0.12\times {\mathrm{x}}_1+1.53\times {\mathrm{x}}_2+3.98\times {\mathrm{x}}_3+3.76\times {\mathrm{x}}_4-0.52\times {\mathrm{x}}_5+0.77\times {\mathrm{x}}_6$$

$${\mathrm{v}}_{expressive}=0.10\times {\mathrm{y}}_{expressive}$$

$${\mathrm{u}}_{receptive}=-2.92\times {\mathrm{x}}_1+1.96\times {\mathrm{x}}_2+6.38\times {\mathrm{x}}_3-0.33\times {\mathrm{x}}_4-1.21\times {\mathrm{x}}_5+0.47\times {\mathrm{x}}_6$$

$${\mathrm{v}}_{receptive}=0.14\times {\mathrm{y}}_{receptive}$$

which provided ρ_expressive = 0.609 (p < .001) with minimum MAE = 6.124 and ρ_receptive = 0.528 (p = .002) with minimum MAE = 4.821.

Table 2.

Canonical correlation coefficient, ρ between the CCA-imaging predicted language score u and measured neuropsychological language score v. Mean absolute error (MAE) and standard deviation of the absolute error (SDAE) of the scaled CCA-imaging predicted score are reported as a function of the modular language network density, where $\overline{v}$ indicates the mean of v.

Network density (%)	CCA-imaging expressive language score, $u_{expressive}∕0.10+{\overline{v}}_{expressive}$			CCA-imaging receptive language score, $u_{receptive}∕0.14+{\overline{v}}_{receptive}$
	ρ (p-value)	MAE	SDAE	ρ (p-value)	MAE	SDAE
10	0.609 (.000)	6.124	5.028	0.528 (.002)	4.821	3.786
20	0.605 (.000)	6.343	4.791	0.518 (.003)	4.996	3.621
30	0.602 (.000)	6.358	4.813	0.519 (.003)	5.003	3.601
40	0.602 (.000)	6.367	4.799	0.516 (.003)	5.017	3.607
50	0.602 (.000)	6.368	4.798	0.516 (.003)	5.016	3.608

Table 3 shows that the probability of the scaled CCA-predicted language subscore to be within actual subscore ± 10 (i.e., one standard deviation of normative controls for standardization): P(absolute error(AE)≤10) reached up to 0.774 and 0.903 for expressive and receptive prediction, respectively, at the whole-group level (n=31). Interestingly, the subsequent subgroup analyses found that the scaled CCA-imaging language score seems to be independent of both the time interval between MRI and assessment of intellect/language functions Δt, and child age at MRI. Similar values of P(AE≤10) were achieved across Δt: 0.800/0.900 expressive/receptive scores at Δt ≤ 1 year vs. 0.762/0.905 expressive/receptive scores at Δt > 1 year. Similar values of P(AE≤10) were also found across scan age: 0.778/0.778 expressive/receptive scores at age of MRI ≤ 3 years vs. 0.773/0.955 expressive/receptive scores at age of MRI > 3 years. These findings provide preliminary evidence that the imaging-based CCA scores may be effective in predicting impaired language functions in younger children (i.e., age ≤ 3 years old) in whom reliable direct assessment of language functions is often not possible.

Table 3.

Prediction probability P of the scaled CCA-imaging language scores to achieve absolute error (AE) less than 5, 10, and 15. Values are listed for five different LD groups/subgroups: whole dataset (n=31), time interval between MRI and neuropsychological assessment (Δt) ≤ 3 year (n=19), Δt > 3 year (n=12), age at MRI ≤ 3 years (n=9), and age at MRI > 3 year (n=22). 10% network density is applied to the modular language networks of all 31 LD children.

Group (n)	CCA-imaging expressive language score, $u_{expressive}∕0.10+{\overline{v}}_{expressive}$			CCA-imaging receptive language score, $u_{receptive}∕0.14+{\overline{v}}_{receptive}$
	P(AE≤5)	P(AE≤10)	P(AE≤15)	P(AE≤5)	P(AE≤10)	P(AE≤15)
Whole (31)	0.484	0.774	0.936	0.613	0.903	1.000
Δt≤3year (19)	0.474	0.790	0.947	0.632	0.895	1.000
Δt>3year (12)	0.500	0.750	0.917	0.583	0.917	1.000
Age of MRI ≤ 3 years old (9)	0.667	0.778	1.000	0.667	0.778	1.000
Age of MRI > 3 years old (22)	0.409	0.773	0.909	0.591	0.955	1.000

The ANCOVA of u_expressive and u_receptive with a “group” factor (LD vs. TD) and age as a covariate showed that the predicted u_expressive and u_receptive were significantly lower in LD patients when compared with healthy controls (F = 15.024, p = .0003 and F = 7.421, p = .009 for u_expressive and u_receptive, respectively, Table 4). In addition, receiver operating characteristic (ROC) curve analysis achieved a statistically significant area under curve (AUC) to discriminate LD children from TD children using our imaging predicted language scores. AUC for expressive/receptive scores were 0.815/0.706 in the whole group (n=31), 0.824/0.733 in subjects with Δt ≤ 1 year (n=10), 0.808/0.692 in subjects with Δt > 1 year (n=21), 0.936/0.814 in subjects with age of MRI ≤ 3 years old (n=9), and 0.764/0.661 in subjects with age of MRI > 3 years old (n=22) (Table 4). The highest sensitivity/specificity values were achieved at age of MRI ≤ 3, 0.876/0.875 (expressive score) and 0.765/0.765 (receptive score), indicating that the most accurate predictions can be achieved when MRI scanning is completed at an earlier age.

Table 4.

ANCOVA F-statistic corrected for age, area under curve (AUC), sensitivity, and specificity to discriminate LD children from TD children using receiver operating characteristic (ROC) curve analysis. 10% network density is applied for modular language networks of 31 LD children.

Group (n)	CCA-imaging receptive language score, $u_{expressive}∕0.10+{\overline{v}}_{expressive}$			CCA-imaging receptive language score, $u_{receptive}∕0.14+{\overline{v}}_{receptive}$
	F-statistic (p value)	AUC (p-value)	Sensitivity /Specificity	F-statistic (p value)	AUC (p-value)	Sensitivity /Specificity
Whole (31)	15.02 (.000)	0.815 (.000)	0.766/0.766	7.42 (.009)	0.706 (.004)	0.680/0.681
Δt≤3 year (19)	14.54 (.000)	0.822 (.000)	0.762/0.762	6.01 (.020)	0.710 (.008)	0.683/0.683
Δt>3 year (12)	8.39 (.008)	0.792 (.000)	0.770/0.770	4.75 (.039)	0.701 (.025)	0.676/0.676
age of MRI ≤ 3years old (9)	23.26 (.000)	0.936 (.000)	0.876/0.875	6.448 (.018)	0.814 (.000)	0.765/0.765
age of MRI >3 years old (22)	8.840 (.005)	0.764 (.000)	0.720/0.720	5.406 (.026)	0.661 (.034)	0.647/0.647

Figure 2 shows cluster plots of the two pair-wise NE-based scores u_expressive and u_receptive maximizing the net similarity in LD group (net similarity = −28.61 from 5 clusters) and control group (net similarity= −15.60 from 5 clusters). A dot is plotted for each subject, the color indicates the subject’s group, and lines connect subjects within a cluster. Clusters are defined by an exemplar subject (i.e., the most representative subject having central scores), who is marked by a bigger circle. Five imaging-based clusters of patients were found within LD group, which roughly reflect different degrees of language disorder severity: 1) LD₁, having extremely low score in u_expressive and a low score u_receptive, 2) LD₂, having low scores in both u_expressive and u_receptive, 3) LD₃, having a low score in u_expressive and a mid-range score in u_receptive, 4) LD₄, having a low score in u_expressive and a high score in u_receptive, and 5) LD₅, having high scores in both u_expressive and u_receptive. LD₅ had a u_expressive and u_receptive within normal range, which could be misclassified as healthy controls (TD) in the subsequent ROC curve analysis (Table 5). Note that the proposed predictors u_expressive and u_receptive showed the lowest Fisher’s exact test probability to differentiate these LD₅ subjects from healthy controls (p = .611) while they provided excellent discrimination for each of other four subtypes, LD_1-4, yielding a high sensitivity/specificity of 0.99, 0.99, and 0.88, 0.77, respectively.

Figure 2.

Affinity propagation clustering plots of the two pairwise NE-based scores [u_expressive, u_receptive] maximizing the net similarity within the LD group (net similarity = −28.61 from 5 clusters, LD_1-5) and healthy control group (net similarity= −15.60 from 5 clusters). A dot is plotted for each subject, the color indicates the subject’s group, and lines connect subjects within a cluster. Clusters are defined by an exemplar subject (i.e., the most representative subject, having central scores), who is marked by a bigger circle.

Table 5.

Accuracy of the scaled CCA-imaging language score, $[u_{expressive}∕0.10+{\overline{v}}_{expressive},u_{receptive}∕0.14+{\overline{v}}_{receptive}]$, corrected for age, in differentiating LD subtypes LD_1-5 from TD children.

Subgroup (n)	LD1 (n=6)	LD2 (n=10)	LD3 (n=7)	LD4 (n=3)	LD5 (n=5)	LD1-5 (n=31)
AUC	0.999	0.999	0.943	0.836	0.682	0.826
Asymptotic p-value	0.000	0.000	0.000	0.014	0.110	0.000
Sensitivity	0.996	0.996	0.876	0.774	0.643	0.776
Specificity	0.996	0.996	0.876	0.774	0.643	0.776
Fisher’s exact test p-value	.000	.000	.002	.050	.611	.001

4. DISCUSSION

Two major findings emerge from the present study. First, conventional DWI connectome combined with language measures correlates using CCA could identify a modular connectivity network including frontal-temporal language areas, hippocampus, and cerebellum, and which showed significantly altered efficiency in both long- and short-range axonal connections. Nodal efficiency values of the identified regions showed high correlations with measures of language functions, up to ρ_expressive and ρ_receptive of 0.609 and 0.528. These values provided a high prediction probability of the CCA-imaging language metric in predicting the actual language scores within (±) one standard deviation, with values of 0.774 and 0.903 for expressive and receptive function, respectively. Secondly, these CCA correlations successfully differentiated LD children from healthy controls, p = .0003 and .009 for expressive and receptive scores, respectively – these findings suggest that the language outcomes for children may differ according to the extent of white matter connections between multiple cortico-subcortical areas that are involved in distinct language functions.

The six regions reported as important predictors of language scores represent well-established language-critical areas in right-handed individuals: in particular, structural and functional abnormalities in left inferior frontal, superior temporal, and angular cortical areas have repeatedly been linked with language impairment. When considering these regions in the current study, it is also worth noting that NE values in left frontal operculum and insular cortices are inversely related to language scores, while NE values in left angular and superior temporal regions are directly related. In other words, patients with more severe language deficits showed inferior frontal regions with more network connections or shorter path lengths, and temporal regions with fewer network connections or longer path lengths. The involvement of right hippocampus and cerebellum is also consistent with some previous research emphasizing the role of these regions in nonclassical language networks (Hodge et al., 2010; Sass et al., 1992). The direct relationship between the right hippocampus NE and language scores suggests that a more directly connected hippocampus within this language network is associated with higher language scores, and that this feature also helps distinguish LD patients from controls. The relationship between language scores and right cerebellar nodule in this patient population is more nuanced, suggesting that a high NE in this region is associated with higher receptive and expressive language scores.

It is difficult to attribute these divergent effects on known language network regions to any specific cause, especially given that the subjects of this study are of a relatively wide age range, and in varying stages of pathological and compensatory changes in brain structure. Importantly, though, a role of the cerebellum in non-motor functions, including language functions has been reported previously (Mariën et al., 2014). Nevertheless, it speaks to the importance of this modular network, and to the importance of applying network analyses to these conditions more generally.

Brain development is most dynamic in the first few years of life with cortical expansion and myelination of white matter tracts (Gogtay et al., 2004). This includes the development of both cortex and white matter tracts/networks associated with language expression and reception (Gopal et al., 2012; Jeong et al., 2016; Sundaram, Sivaswamy, Makki, Behen, & Chugani, 2008). Our working hypothesis was that regional cortical and white matter abnormalities in children with LD are selectively disrupted potentially leading to specific patterns of language impairment identified in this population, thus allowing us to elucidate specific subtypes of LD for targeted interventions. Such an approach might be useful for conceptualizing new disease mechanisms in LD subtypes, as specific connectivity disorders may result in specific speech and language deficits. Indeed, the present study provided preliminary evidence that children with LD might be successfully subtyped via integrating an advanced brain connectome analysis with direct assessment of language functions; in fact, such subtyping might be accomplished with MR imaging even very early age and very early in the development of language functions (see below). Specifically, our analysis revealed 5 imaging-defined severity-reflecting subtypes in our dataset: “profound expressive and receptive impairment (LD₁)”, “moderate expressive and receptive impairment (LD₂)”, “mild expressive and receptive impairment (LD₃)”, “expressive impairment, with typical receptive language function (LD₄)” and “receptive impairment with typical expressive function (LD₅).” In this way, imaging data can extract “psychometry-driven knowledge” of heterogeneous LD subtypes using DWI connectome features measured in cortical, subcortical, and white matter structures. This integration was performed using CCA that statistically combines reginal efficiencies of axonal connectivity in a language modular network, taking advantage of their covariance together, thus unveiling “subthreshold abnormalities” that have not been readily identified by single topological parameters. This combination of connectome data and assessment of intellect and language measures may allow us to define specific LD phenotypes. Conventional MRI has been of limited value in this task, except to rule out lesional/structural etiologies (Burden et al., 1996).

It is especially critical to develop such non-psychometric methods of assessing children with LD, particularly in very young (≤3 years) children, given that direct psychometric assessment is often unreliable at younger ages and when intervention is most critical. In the present study, DWI imaging metrics were able to predict later language impairment even in children scanned at a very young age (10 of the children in the present study were scanned at 3 years of age or younger). Thus, the findings of the present study may directly translate scientific discoveries into public health and education systems by crafting individualized diagnostic tools that allow children with LD to receive specialized treatments or care plans designed to focus on their individual needs and with onset of intervention at a critical and early time. The proposed sub-classification system might be able to identify children with language impairment, but who appear to have more normal white matter, and perhaps a milder pattern of language impairment (such as the LD₃ subtype identified in the current analysis), and which may suggest a higher likelihood of successful and aggressive therapeutic intervention (Hustad et al., 2018). It should be noted that our aim was to predict children with LD that would have enduring language disorders from those evidencing late language acquisition. In fact, the finding that the data might support some distinction between those children evidencing more receptive versus expressive delay was not entirely expected. Given the different brain structures that support expressive and receptive language functions, it is not entirely surprising that some sort of structurally-salient difference exists between these different phenotypes. However, given the heterogeneous and often experience-dependent plasticity involved in language disorder development, the relatively small sample size of this study cannot explore.

There are several limitations in the present study. First, as noted in the Subjects section, all data on children in the LD group were obtained for clinical management of language disorder. Thus, the findings of this study were naturally based on a retrospective and observational design. Such design inherently limits the sample size of participants who underwent the detailed intellectual and language assessments, and this makes it difficult to interpret the interaction of imaging metrics with biological and clinical variables such as age, gender, and therapeutic intervention. More studies and replication in larger numbers of children with a wider variety of language problems and comparisons and with age-matched controls are needed to further establish detailed and specific imaging-phenotype trait relationships. Second, despite the exceptional quality of DWI data using high angular resolution diffusion imaging (HARDI), it still remains controversial whether current DWI tractography techniques can accurately derive long-range anatomical connections (Reveley et al., 2015; Thomas et al., 2014). Considering this inherent uncertainty, the proposed approach is limited in its ability to further investigate ideas like language reorganization and crowding effects (Kim et al., 2018) in developing brain networks, which involve long-range connections between cortical and subcortical areas.

Importantly, a number of children in the study evidence comorbid impairments, including in motor functions, and behavioral concerns, including overactivity, and anxiety. While such comorbidities are commonly identified in children with language disorders (Bishop, Snowling, Thompson, & Greenhalgh, 2017), they do add error to interpretation of DWI network findings, as some of the findings could be attributable to the comorbid conditions. Again, further study with more homogeneous samples of children with language disorder are needed to bolster the findings of the present study. Additionally, as noted in the methods, we excluded children with language disorder who also have impaired global intellect, autism spectrum disorders, and medical conditions. As noted by the Bishop et al. (2017), this may reduce the generalizability of the present results; such a choice was intentional in order to try and reduce error in the pursuit of imaging substrates that might predict language outcomes, but this experimental control does admittedly distinguish our study cohort from the broader population of children with language impairment/disorder. Further research will be necessary to identify whether such findings are specific to this phenotype (language disorder in context of typical nonverbal intellect) or can be generalized to the population of children with language disorder. Finally, we only included the more broad categories of language deficit – expressive and receptive functions, and which do not allow for much more specific or fine grained description of language deficits, such as word structure, grammar, speech and/or pragmatic aspects of language, and which might also provide for increased/greater specificity in findings -limiting the potential findings of the study as well; further research and with more fine grained metrics might provide additional important information.

5. CONCLUSION

This study translates advanced DWI techniques with a clinical application in a pediatric population, where currently available neuroimaging approaches are suboptimal to objectively differentiate young patients with language disorder from healthy children. Continued and systematic investigation of the proposed methods may improve early identification of language problems, even in very young children, and also may allow for identifying distinct patterns of language impairment, allowing earlier and more targeted intervention for improving language outcomes in children with language disorder.