Examining post-concussion white matter change in a pediatric sample

Highlights

• •Diffusion-Weight Imaging (DWI) has been used to explore pediatric concussion outcomes but findings have been mixed.
• •We combine multiple DWI metrics using Linked Component Analysis to explore outcomes more consistently.
• •Our findings suggest that existing diffusion techniques may not be sufficient to detect changes in white matter microstructure in pediatric concussion.

Abstract

Diffusion-Weight Imaging (DWI) is increasingly used to explore a range of outcomes in pediatric concussion, particularly the neurobiological underpinnings of symptom recovery. However, the DWI findings within the broader pediatric concussion literature are mixed, which can largely be explained by methodological heterogeneity. To address some of these limitations, the aim of the present study was to utilize internationally- recognized criteria for concussion and a consistent imaging timepoint to conduct a comprehensive, multi-parametric survey of white matter microstructure after concussion. Forty-three children presenting with concussion to the emergency department of a tertiary level pediatric hospital underwent neuroimaging and were classified as either normally recovering (n = 27), or delayed recovering (n = 14) based on their post-concussion symptoms at 2 weeks post-injury. We combined multiple DWI metrics across four modeling approaches using Linked Independent Component Analysis (LICA) to extract several independent patterns of covariation in tissue microstructure present in the study cohort. Our analysis did not identify significant differences between the symptomatic and asymptomatic groups and no component significantly predicted delayed recovery. If white matter microstructure changes are implicated in delayed recovery from concussion, these findings, alongside previous work, suggest that current diffusion techniques are insufficient to detect those changes at this time.

1. Introduction

There is a growing academic and clinical interest in understanding the neurobiological and physiological mechanisms that underpin delayed recovery post-concussion. Indeed, key questions remain regarding the acute and delayed neurobiological effects of concussion on the brain. Utilizing the Concussion in Sport Group’s (CISG) Berlin Consensus Statement definition (McCrory et al., 2017), concussion symptoms largely reflect a functional disturbance rather than a structural brain injury and therefore no abnormalities are identified using standard structural neuroimaging (i.e., Computed Tomography, Magnetic Resonance Imaging). More sensitive neuroimaging techniques and methods are required to investigate potential subtle changes post-injury.

Relatively few studies have examined the neurobiological outcomes of pediatric concussion, but Diffusion-Weighted Imaging (DWI) has been the most widely-used technique. DWI is a Magnetic Resonance Imaging (MRI) technique where image contrast depends upon the extent and direction of diffusion of water molecules within tissue. As myelinated white matter fibres restrict the free diffusion of water, this technique allows for the estimation of the organisation and structure of white matter tracts. Commonly used models of diffusion include the diffusion tensor model (DTI) (Basser, 1995) to derive estimates of directionality (fractional anisotropy; FA) or magnitude of diffusion (mean diffusivity; MD), both along (axial diffusivity; AD) and perpendicular to (radial diffusivity; RD) white matter fiber tracts (Dennis et al., 2015, Huisman et al., 2004). Despite its popularity, DTI modeling has a number of noted limitations including an inability to accommodate complex fiber configurations (Jeurissen et al., 2013, Jones et al., 2013). More complex methods, requiring specialized DWI acquisitions, attempt to model microscopic biophysical properties of white matter such as the average degree of diffusion within neurites, neurite orientation and neurite density, and include NODDI (Neurite Orientation Dispersion and Density Imaging) (Zhang et al., 2012) and SMT (Spherical Mean Technique) (Kaden et al., 2016). Further, local fiber densities may be estimated through deconvolution of the diffusion signal to recover multiple fiber populations (Raffelt et al., 2012). Together, these methods provide sensitive measures of white matter microstructure in both healthy and injured brains (Beck et al., 2021, Bourke et al., 2021, Ehrler et al., 2021).

Imaging, particularly DWI, may have the potential to play a significant role in concussion management. This includes the diagnosis of concussion (i.e., differentiating between concussed and non-concussed), monitoring recovery from concussion (i.e., improvement over time), predicting slow or poor recovery, and identifying long-term structural injury (Rausa et al., 2020). However, currently, DWI findings in pediatric concussion are mixed, with some identifying significant increases in DWI metrics (i.e., FA) relative to controls (Babcock et al., 2015, Bartnik-Olson et al., 2014, Manning et al., 2017, Mayer et al., 2015, Van Beek et al., 2015, Wilde et al., 2008, Yallampalli et al., 2013), some finding a decrease in DWI metrics (King et al., 2019, Lancaster et al., 2016, MacDonald and Duerson, 2015) and others reporting no changes (Maugans et al., 2012, Shapiro et al., 2020). Previously, our research group utilized diffusion imaging and voxelwise tract-based spatial statistics (TBSS) to assess diffusion neuroimaging correlates of delayed recovery post-concussion in children (Shapiro et al., 2020). Forty-three children underwent magnetic resonance imaging (MRI) at 2 weeks post-injury and were classified (based on parent and child reported symptoms) as either normally recovering (n = 26) or delayed recovering (n = 17). Diffusion imaging comparison using voxelwise tract-based spatial statistics (TBSS) analysis found no difference between the groups in fractional anisotropy, axial diffusion, radial diffusion, or mean diffusivity metrics. Post-hoc tract-based Bayesian analysis found evidence for the null hypothesis across 11 white matter tracts. We concluded that there was no evidence that delayed recovery from post-concussive symptoms in children is caused by white matter microstructural damage.

With advances in diffusion imaging and modelling, alternative measures of tissue microstructure may prove more sensitive to subtle changes in white matter post-concussion. As spatial correlations often exist between different imaging modalities, when viewed together, multiple metrics can reveal different aspects of a common neurobiological or pathological process. A number of ‘data fusion’ methods have been introduced that seek to combine metrics from different sources to identify patterns of shared variance across different imaging modalities or measures (Gong et al., 2021, Groves et al., 2011, Groves et al., 2012, Qi et al., 2022). These unsupervised, data-driven methods provide an alternative to performing multiple parallel unimodal analyses, are robust to different noise characteristics across image types and provide a compact representation of the full multimodal dataset for further analysis. This makes them an ideal methodical choice to identify the presence of structured patterns of variation across multiple imaging metric simultaneously (Groves et al., 2011, 2012).

Within the broader literature, the heterogeneity of findings can also, to a large degree, be explained by methodological limitations of the pediatric concussion literature. These include injury definitions (concussion versus mTBI), variations in outcome measures and approaches to imaging analysis, discrepant timing of imaging relative to timing of injury, small sample sizes, and differences in study design (e.g., recruitment from the emergency department versus dedicated concussion clinics) (Rausa et al., 2020). These caveats can have significant implications on the findings that emerge from pediatric concussion studies.

To address some of these limitations, the present study utilized internationally- recognized criteria for concussion and a consistent imaging timepoint to conduct a comprehensive, multi-parametric survey of white matter microstructure after concussion. Using an unsupervised data fusion approach, Linked ICA, we combined multiple DWI metrics across four modeling approaches to test the hypothesis that there would be no differences between concussed children with persistent symptoms and concussed children with symptom resolution at two weeks post-concussion.

2. Materials and methods

2.1. Design

This study was part of a larger single site, prospective, longitudinal study (Take CARe Biomarkers – Concussion Assessment and Recovery research) conducted at a state-wide tertiary pediatric hospital in Melbourne, Australia (see Takagi et al., (2019) for full study details). In brief, the study recruited and assessed children presenting with concussion to the emergency department (ED) and then followed up with the children 1–4 days, two weeks, one month, and three months post-injury. Multiple domains of functioning were assessed at each time point (concussive symptoms, neurocognition, behavior, quality of life, fatigue, etc.) Neuroimaging of participants occurred at the 2-week time point.

2.2. Participants

Participants were a subset (n = 45) of the Take CARe Biomarkers study who were part of the concussion sample and underwent standardized neuroimaging at two weeks post-injury (Sady et al., 2014). Concussion was defined according to the CISG’s Berlin Consensus Statement (McCrory et al., 2017). Participants aged between 5 and 18 years who presented to the ED within 48 h of sustaining a concussion were eligible for inclusion. Exclusion criteria included Glasgow Coma Scale score (GCS) < 13; structural/hemorrhagic intracranial injury; clinical evidence of a cerebrospinal fluid (CSF) leak; requiring intubation, general anaesthesia, or neurosurgical intervention; developmental or intellectual disability; a history of moderate or severe Traumatic Brain Injury (TBI) or neurological conditions; history of psychosis or bipolar disorder; medication use (anti-depressants, stimulants); intoxication upon presentation to the ED; fever; injury resulting from abuse or assault; multiple injuries; no clear history of trauma as the primary event; having an insufficient understanding of English; and being currently enrolled in the study. One participant was excluded from the sample due to poor-quality image acquisition caused by orthodontic braces. Another was excluded due to an MRI acquisition error. The final sample of 43 participants was included in the analysis.

Participants above the age of 12 years and their parents/guardians gave written informed consent to be a part of the study at the time of recruitment. For children at or below the age of 12 years, written informed consent was obtained from the parent/guardian and verbal assent obtained from the child. Ethical approval for this study was given by The Royal Children’s Hospital (RCH) Human Research Ethics Committee.

2.3. Measures of Post-Concussion symptoms

Patient symptoms were assessed using the Post-Concussion Symptom Inventory (PCSI), a measure of post-concussive symptoms that has strong psychometric properties (internal consistency α = 0.8 – 0.9; test–retest reliability intraclass coefficient = 0.79 – 0.89; convergent validity r = 0.8) (Sady et al., 2014). It includes both a parent and child self-report Likert scale endorsing levels of post-concussive symptomology, with different versions of the child self-report for younger (5–12 years) and older (12–18 years) children.

Recovery was assessed using symptoms compared to pre-injury level, defined as delayed (≥2 symptoms) and normal (<2 symptoms) at 2 weeks post-injury (Barlow et al., 2015, Hearps et al., 2017, Zemek et al., 2016). The CISG defines delayed recovery in children as persisting symptoms at or beyond 4 weeks post-injury. In our sample, the vast majority of participants who were symptomatic at two weeks remained symptomatic at 4 weeks (Anderson et al., 2020). Furthermore, based on the clinical experience of the authors at The Royal Children’s Hospital, parents of children with symptoms persisting at two weeks tend to seek help at two weeks post-injury. We selected two weeks as our MRI timepoint to capture data that may translate to early intervention and prediction.

2.4. MRI acquisition

All imaging was completed at The Royal Children’s Hospital in Melbourne, Australia on a 3 T Siemens Magnentom TimTrio running B17 software with a 32-channel head coil. As part of a larger image acquisition protocol that included functional and structural sequences (Andersson & Sotiropoulos, 2016), we acquired high-angular resolution DWI using an echo-planar sequence with the following parameters: b = 1000/2800 s²/mm; directions = 25/60; TE = 110 ms; TR = 3200 ms; voxel size = 2.4 × 2.4 × 2.4 mm³, alongside five diffusion-weighted images with no diffusion weighting (b = 0 s²/mm) to increase signal-to-noise ratio. Diffusion data were acquired with a multiband factor of 3 for accelerated acquisition. Reverse phase-encoded images were collected to enable susceptibility-induced distortion correction. In addition, we acquired high-resolution T1 weighted MPRAGE structural volumes with: TR = 2530 ms, TE = 1.77/3.51/5.32/7.2 ms, slice thickness = 0.9 mm, voxel size = 0.9 × 0.9 × 0.9 mm³.

2.5. Procedure

Patients presenting to the RCH ED with concussion were screened by trained research assistants daily (weekdays 9 a.m.-10p.m., weekends 12p.m.-10p.m.) by reviewing the digital record of their ED visit. The treating physician of eligible patients was then contacted to confirm their eligibility. Eligible patients’ families were approached to participate in the study. Once consented, pre-injury and demographic information were obtained from parents in the ED and both the parents and child completed the PCSI. Injury-related information was obtained from the study clinical report form. All participants attended a dedicated concussion clinic at the RCH at 1–4 days post-recruitment and again at 2 weeks post-injury where they completed the PCSI, brief neuropsychology testing, and underwent clinical assessment with a clinician with expertise in concussion management. At this time, participants were classified into two groups based on their parent-rated PCSI scores according to the criteria described by Hearps et al.: i) recovered group (n = 26); and ii) delayed recovery group (n = 17) (Hearps et al., 2017). Participants also underwent MRI imaging at the 2-week appointment (range = 8–24 days post ED injury).

2.6. Image processing

Image pre-processing was completed by trained research staff blinded to group membership. DWI data were first pre-processed within shell by denoising and correcting for Gibb’s ringing artefacts, followed by correction for geometric distortions, B1 field inhomogeneity and subject motion outliers using FSL’s eddy and ANTS N4 as implemented in MRtrix3 (using dwipreproc; dwibiascorrect) (Andersson and Sotiropoulos, 2016, Kellner et al., 2016, Tournier et al., 2019, Tustison et al., 2010, Veraart et al., 2016). We then used four independent diffusion models (DTI, NODDI, SMT, CSD) to extract nine parametric maps of white matter microstructure for each subject.

2.7. Diffusion tensor imaging

DTI is a relatively simple model of anisotropic diffusion that fits a 3D tensor to the diffusion signal at each voxel (Basser, 1995). Using data acquired at b = 1000 s²/mm (n = 25 directions), we fit a DTI model to the diffusion data and extracted parametric maps of fractional anisotropy (FA) and mean diffusivity (MD) using MRtrix3 (dwi2tensor; tensor2metric).

2.8. NODDI

The NODDI model is a multi-compartment model that aims to resolve the contribution of neurite density and orientation, as well as free water diffusion, to the diffusion signal (Zhang et al., 2012). Multi-shell data were aligned and merged into a single volume (n = 85 directions) before fitting the NODDI model at each voxel, implemented with the NODDI MATLAB toolbox (v0.9) using default parameters. From this, we extracted parametric maps of free water volume fraction (f_ISO), intracellular volume fraction (f_ICVF), and neurite orientation dispersion (ODI).

2.9. SMT

The Spherical Mean Technique aims to remove the orientation dependence of the diffusion signal to provide maps of tissue microstructure that are independent of the underlying fiber orientations (Kaden et al., 2016). Using the SMT package (https://github.com/ekaden/smt), we fit a multicompartment SMT model with Rician noise correction (fitmcmicro) to the multi-shell DWI data, extracting maps of intra-neurite volume fraction (intra), intrinsic diffusivity (diff), and extra-neurite mean diffusivity (extramd).

2.10. CSD

Constrained Spherical Deconvolution is a non-parametric approach to estimate white matter fiber orientations from the diffusion signal, from which measures of apparent fiber density can be derived (Dell’Acqua and Tournier, 2019, Raffelt et al., 2012). Using tissue segmentation maps derived from each subject’s T1 image (5ttgen), we used the high b-value DWI data (n = 60 directions) to estimate single fiber response functions for GM, WM, and CSF tissue types (dwi2response) before calculating fiber orientation distributions for all voxels (dwi2fod) using multi-tissue CSD as implemented in MRtrix3 (Jeurissen et al., 2014). Finally, maps of Apparent Fiber Density (AFD) were calculated as the mean amplitude of fiber orientation distributions within a given voxel (fod2fixel; fixel2voxel).

2.11. Template creation and registration

Together, this process resulted in nine parametric maps of white matter microstructure per subject (Fig. 1). We aligned each subject’s set of parametric maps to a study-specific template using DTI-TK, a nonlinear registration optimized for diffusion imaging (Zhang et al., 2007a, Zhang et al., 2007b). We created a template by first averaging DWI volumes of a randomly selected subset of subjects (n = 10), before aligning all subject’s DWI data to the template using rigid, affine and diffeomorphic registration using default parameters. Once aligned, the DWI volumes were averaged to update the template and the diffeomorphic registration process repeated for a total of 6 iterations (dti_diffeomorphic_population). After calculating transformations to the common template space, all subject’s parametric maps were transformed to template space before smoothing with a 5 mm FWHM Gaussian kernel. Finally, we masked each map using a probabilistic white matter mask thresholded at 50% to ensure that only white matter voxels were included in the analysis.

Fig. 1.

Parametric white matter maps for a single subject. Nine maps from four separate diffusion models (DTI, CSD, NODDI, and SMT) are shown after transformation into a common template space. Intensity is scaled for comparison across maps.

2.12. Linked independent component analysis

Although each map provides a different measure of tissue microstructure, there is clear overlap evident in the relative distribution of metrics across tissue types and regions (e.g, FA and AFD; Fig. 1). This similarity suggests that each map represents a different view of a set of latent tissue properties that are reflected similarly across some metrics and models. To account for spatial covariance across maps, we aimed to identify latent features common to the white matter of all subjects and represented by common patterns of variation present across multiple parametric maps. To achieve this, we applied Linked Independent Component Analysis (LICA), an unsupervised data fusion model designed to identify imaging patterns from a combination of different imaging maps and modalities (Groves et al., 2011, 2012).

We performed LICA (Fig. 2) using FSL’s LICA (FLICA) toolbox, implemented in MATLAB R2015b using code available at: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FLICA. We first concatenated individual parametric maps into nine modality groups, before performing variance normalisation separately on each group of aligned maps (Beckmann & Smith, 2004). We then decomposed the full multi-parametric dataset (43 subjects × 9 maps) into a set of imaging components using LICA (iterations = 10,000). Following previous examples, and given our sample size, we specified the model to estimate 10 spatially-independent components (Ball et al., 2019, Francx et al., 2016, Groves et al., 2012, Shapiro et al., 2021).

Fig. 2.

Linked ICA analysis pipeline. Linked ICA models multimodal imaging data as a set of spatial maps linked by a shared set of subject weights. Individual images are concatenated into modality groups (left, one per diffusion metric). The imaging data are then decomposed into a set of multimodal components. Each component comprises a set of spatially independent maps (middle), one per imaging metric, linked together by a single shared subject weight (right). The contribution of each metric to each component is determined by a modality weight (middle, bottom). The component’s subject-specific weight describes the extent to which the linked patterns of variation described by each component map are expressed in each subject’s multimodal data. To test associations with concussive symptoms, subject-specific weights for each component were entered into a logistic regression model alongside age and sex to predict clinical diagnosis and symptom severity (PCSI 2 + ) (inset).

This process resulted in a set of 10 multimodal components reflecting patterns of shared variance present across all parametric maps. Each component comprises a set of spatially independent maps, one per microstructural metric, linked together by a single shared subject weight (Fig. 2). The component weight describes the extent to which the patterns of variation in each map are represented in each subject's diffusion data and can be used as a simplified representation of the full multi-parametric dataset for further analysis.

2.13. Statistical analysis

To determine the derived components’ putative relationship with symptomology at 2 weeks post-injury, exploratory logistic regression models predicted clinician-diagnosed delayed recovery, as well as an established PCSI cut-off score (≥2 worsened symptoms above pre-injury ratings). The association between post-concussion symptomology and subject-specific component weights for each LICA component were modeled separately, adjusting for child age and sex. Odds ratios (OR) and 95% confidence interval (95% CI) were calculated and presented.

2.14. Ethical approval

This study has been approved by the Human Research Ethics Committee of The Royal Children’s Hospital Melbourne.

3. Results

3.1. Cohort demographics

See Table 1.

Table 1.

Normal and Delayed Recovery Group Demographics and Method of Injury.

	Normal Recovery	Delayed Recovery	Total Sample	t (df)/ χ2	p	Cohen’s d/ φ
N	26	17	43
Age at recruitment, M, (SD)	12.93 (2.61)	13.28 (2.31)	13.07 (2.47)	-0.46 (41)	.65	0.14
Sex, n (%) male	21 (80.8)	10 (58.8)	31 (72.1)	2.46	.12	0.24
Handedness, n (%) right	24 (92.3)	17 (100)	41 (95.3)	1.37	.24	0.18
Number of previous concussions, n (%)
0	16 (61.5)	12 (70.6)	28 (65.1)	0.61 (41)	.55	0.19
1	2 (7.7)	3 (17.6)	5 (11.6)
2	5 (19.2)	0 (0)	5 (11.6)
3+	3 (11.5)	2 (11.8)	5 (11.6)
Number of days between recruitment and imaging*, M (SD)	14.62 (3.6)	14.59 (3.9)	14.60 (3.7)	0.02 (41)	.98	0.008
CT scan in ED, n (%)	3 (11.5)	2 (11.8)	5 (11.6)	0.001	.98	0.005
Method of Injury, n (%)
Fall	19 (73.1)	8 (47.1)	27 (62.8)
Fall from bike/motorbike	4 (15.4)	0 (0)	4 (9.3)
High speed projectile	2 (7.7)	1 (5.9)	3 (7.0)
High impact object	2 (7.7)	1 (5.9)	3 (7.0)
Collision with child or object	11 (42.3)	7 (41.1)	18 (41.9)

Age, number of days between recruitment and imaging, and number of previous concussions analysed with independent samples t tests. Sex, Handedness, CT scan analysed with χ² tests. Method of injury not analysed due to participants belonging to multiple groups and missing data. *Recruitment occurred within 48 h of injury. SD, standard deviation; CT, computed tomography; ED, emergency department.

3.2. Linked ICA of white matter microstructure

Image components resulting from LICA are shown in Fig. 3. Each component represented combination of spatial patterns shared across microstructural maps. Component 1 revealed widespread increases in intra-neurite volume fractions (indexed by fICVF and intra) and FA, with concomitant decreases in extra-neurite diffusivity and mean diffusivity (component 1; Fig. 3). Expression of this component was significantly associated with subject age (R2 = 0.20, p < 0.01) and captured known developmental changes in the white matter with fICVF, intra and FA increasing with age. Other age-associated changes were observed in components 6 (increased diffusivity near the grey/white matter border; R² = 0.15, p < 0.01) and 7 (oppositional effects on fibre density and microstructural diffusivity in the internal capsule; R² = 0.19, p < 0.01). Other components captured widespread increases across multiple measures of tissue diffusivity (component 3) and apparent fiber density (component 4). Regional patterns of variation in white matter microstructure were observed around the ventricles (component 8), in the internal capsule (component 7) and in posterior periventricular white matter (component 9).

Fig. 3.

Linked ICA decomposition of white matter microstructure. Component maps are shown for each of 10 linked ICA components derived from the whole group’s DWI data. Each component comprises nine coefficient maps (one per DWI metric) along with a subject-specific weight that describes the contribution of that component to the full diffusion dataset of each individual. For each map, warm colors indicate regions where the specified parameter increases in line with subject weight. Cold colors indicate the opposite. Blank maps indicate that the parameter does not contribute significantly to the component – i.e.: doesn’t vary across the same regions in the same way as the other parameters. Scatterplots show the relationship between component weight and age. Circles denote asymptomatic individuals, Xs represent those with clinically-diagnosed concussion symptoms. Lines show linear regression (shaded area: 95% C.I.; R², ** indicates p < 0.01).

3.2. Relationship between post-concussion recovery and white matter imaging

To test the association between white matter tissue microstructure, represented by subject-specific component weights of the 10 LICA components, we fit a series of logistic regression models exploring the predictive value of each component in relation to delayed recovery, as measured by clinician diagnosis as well as ≥ 2 + worsened symptoms on the PCSI. Table 2 shows the OR for each component, adjusting for child age and sex. We did not find any significant associations between white matter components and post-concussion symptomology (Table 2).

Table 2.

Component Prediction of Delayed Recovery (Clinician-Diagnosed and PCSI Scored), Adjusting for Child Age and Sex.

	Clinician Diagnosis				PCSI 2+
Component	OR	(95%CI)	p		OR	(95%CI)	p
0	0.57	(0.23-1.42)	0.23		0.89	(0.43-1.85)	0.76
1	1.11	(0.51-2.39)	0.8		0.79	(0.40-1.55)	0.49
2	1.87	(0.78-4.49)	0.16		1.93	(0.91-4.09)	0.09
3	0.87	(0.41-1.87)	0.73		0.81	(0.43-1.53)	0.52
4	1.72	(0.82-3.61)	0.15		1	(0.52-1.92)	>0.99
5	1.15	(0.54-2.43)	0.72		1.75	(0.74-4.16)	0.2
6	2.29	(0.87-6.02)	0.09		1.51	(0.73-3.12)	0.27
7	0.37	(0.09-1.50)	0.17		1.28	(0.62-2.64)	0.51
8	0.68	(0.27-1.71)	0.41		0.9	(0.44-1.84)	0.77
9	0.68	(0.32-1.46)	0.32		0.83	(0.42-1.62)	0.58

OR,odds ratio; CI,confidence interval; PCSI,Post-Concussive Symptoms Inventory.

Analyses indicate no statistically significant component OR when predicting delayed recovery.

4. Discussion

The aim of the present study was to investigate a possible association between white matter microstructure and delayed recovery, as represented by persisting post-concussion symptoms at 2 weeks post-injury in children. Based on our previous work, we hypothesized that there would be no association between white matter microstructure and persistent symptoms.

The present study utilized a series of logistic regression models to explore the predictive value of patterns of variation in white matter microstructure derived from LICA in relation to delayed recovery, as measured by clinician diagnosis as well as 2 + worsened symptoms as described in Hearps and colleagues (2017). Our analysis did not identify between-group differences (i.e., component differences between symptomatic versus asymptomatic) and no component significantly predicted delayed recovery.

These results are consistent with our previous work which did not show a difference between normally-recovering children and children with delayed recovery from concussion using a diffusion model (Shapiro et al., 2020, 2021). In addition, in their longitudinal study, King and colleagues did not find a difference between symptomatic and non-symptomatic children in terms of their region of interest diffusion tractography analyses of their corpus callosum, uncinate fasciculi, and corticospinal tracts (King et al., 2019). They also could not identify a within- or between-group difference in diffusion metrics between acquisitions at a one month and two month timepoint post-injury, suggesting that the limitations to the use of diffusion imaging to predict delayed recovery also likely extend longitudinally as an inability to predict recovery from persistent post-concussive symptoms.

In contrast, a study by Manning and colleagues (2017) identified post-concussion diffusion abnormalities within multiple white matter tracts in a small cohort of youth hockey players. Notably, tract-specific spatial statistics revealed a large region along the superior longitudinal fasciculus with the largest decreases in diffusivity measures, which significantly correlated with clinical deficits. The authors concluded that adolescents may be particularly vulnerable to post-concussive injury while axons continue to myelinate and mature. However, an important point of differentiation between Manning and colleagues (2017) and the present study is our definition of concussion. We conceptualise concussion as a subset of mTBI; most notably, that an abnormality identified on standard clinical imaging precludes a concussion diagnosis. In Manning and colleagues (2017), their sample includes adolescents with, “a mild traumatic brain injury or concussion”. It is therefore likely that Manning’s sample includes adolescents with more severe injuries, which could explain the differences in findings.

We have extended upon our previous study by incorporating multiple parametric measures of tissue microstructure derived from multiple modeling techniques. To account for the covariance between multiple measures, we employed a data fusion method to extract several independent patterns of covariation in tissue microstructure present in the study cohort. Linked independent component analysis is an unsupervised, data-driven technique used for analysis and dimension reduction in multimodal data (Groves et al., 2011). Its objective is to uncover structured patterns of variation across multiple imaging modalities. Spatial correlations between different imaging modalities are common when each imaging metric or modality offers a different view of a shared underlying process. While performing separate analyses on individual image metrics may enhance sensitivity to potential differences, it can also elevate the risk of false positives when performed over multiple metrics. Furthermore, due to variations in noise characteristics, the spatial extent of any differences in modalities can differ even if they are driven by the same factor, hindering accurate interpretation. As such, a single generative pathological process can result in distinct patterns and interpretations across modalities when considered only in isolation (Groves et al., 2011, 2012). In our study, each component received contributions from at least four metrics, highlighting the high degree of covariance in the multimodal data. While the decision to perform a joint or separate analysis depends on the specific research questions being addressed, in this study, we hypothesised that any difference in white matter tissue microstructure would be captured jointly across a number of diffusion metrics and therefore, Linked ICA provides a suitable modelling choice to test this hypothesis.

Using logistic regression models, we found that no components were differentially expressed in case and control groups. Together, our findings lead to the conclusion that, given current technological limitations, common diffusion imaging techniques are unable to predict whether children will have a delayed recovery from concussion at the 2-week timepoint. At a clinical level, these findings suggest that diffusion imaging may not be a useful or cost-effective tool for predicting, diagnosing, or tracking the recovery of delayed recovery post-concussion, especially considering the high costs associated with MRI. Further, while preliminary evidence suggests that diffusion imaging may be able to differentiate between concussed children and non-concussed children (Bartnik-Olson et al., 2014, Mayer et al., 2012), this can be achieved far more economically, and with less distress to the child, by completing a thorough clinical assessment. Further research is required to determine if alternative diffusion modelling approaches (Assaf and Basser, 2005, Dhollander et al., 2021, Li et al., 2022) prove sensitive to mTBI in children.

Despite the strengths of our prospective, longitudinal design and implementation of rigorous definitions of concussion and consistent timing of imaging, there are several limitations of the present study that are important to consider. First, the relatively small sample size limits the generalizability of our findings, as does the sample population (i.e., children presenting to the emergency department for their concussion). The prospective nature of the recruitment of this study allows for the investigation to center on the cases of pediatric concussion that typically present to the ED. However, it is well-understood that only a fraction of children who sustain a concussion receive medical care (Crowe et al., 2009). It is therefore likely that the sample analyzed may be biased towards more severe presentations and symptom clusters that are more overt. It is therefore likely that the sample analysed may be biased towards more severe presentations and symptom clusters that are more overt. It is perhaps in the cases of less overt persistent post-concussive symptoms that neuroimaging may provide insight into symptom etiology, however more research in this space is still required.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.