Multi-modal proton and sodium MRI for outcome prediction in mild traumatic brain injury

Abstract

Objectives:

In mild traumatic brain injury, imaging biomarkers are needed to support clinical management. In four antecedent publications, we used two new (sodium and fingerprinting) and two established (spectroscopy and diffusion) MR techniques in a longitudinally followed patient cohort. Here we report final results and combine all data to determine which marker(s) from the four modalities offer the greatest utility for detecting injury and predicting outcomes. We also leverage the independent specificities offered by each modality to explore injury mechanisms.

Materials and Methods:

The longitudinal spectroscopy data was analysed to complete a full dataset of proton (spectroscopy, fingerprinting, diffusion) and sodium MRI, acquired alongside symptomatic, cognitive, and functional assessments in 27 patients at 1-, 3-, and 12-months following injury. Twenty-three matched controls were scanned once. Testing for associations between nine MR markers and three outcome measures was standardized across the entire dataset, and performed using Spearman correlations and logistic regression.

Results:

Previously elevated white matter choline and creatine from spectroscopy (markers of membrane turnover and cellular energetics, respectively) normalized to control levels by 3 months, at rates which correlated with the rate of symptom recovery. Sodium and spectroscopy showed findings coinciding in pattern and time point, but there were no associations between them, suggesting independent origin. Choline and creatine met the greatest number of biomarker properties, followed by water T₁ from fingerprinting (marker of the cellular microenvironment).

Conclusions:

We identified independent, dynamic, metabolic and ionic changes, with choline and creatine from spectroscopy fulfilling the most criteria for a clinical biomarker.

Introduction

A mild traumatic brain injury (mTBI) is caused by a mechanical insult to the head and may result in a heterogeneous combination of clinical features that span multiple domains [1]. While most patients are expected to recover, a growing number of studies have demonstrated that a subset, reaching as high as 30-40% [2, 3], will continue to suffer from persistent post-concussive symptoms and potentially chronic disability [4-7]. The possible long-term health consequences and economic burden [8, 9] associated with incomplete recovery have therefore motivated the development of prognostic biomarkers [10, 11] that can identify patients at risk for poor clinical outcomes and guide rehabilitation strategies.

Diffuse axonal injury is the main pathophysiological phenomenon underlying mTBI [12, 13], but is difficult to detect using CT and MRI [14]. Therefore, various quantitative MR (qMR) metrics have been proposed as surrogate radiological markers for injury in otherwise normal-appearing tissues. Fractional anisotropy (FA) and apparent diffusion coefficient (ADC) from diffusion tensor imaging (DTI) are most utilized, given their ability to probe the microstructural organization and integrity of white matter (WM). MR spectroscopy (MRS) has been leveraged for its sensitivity to changes in cellular biochemistry, most often through quantification of N-acetyl-aspartate (NAA), choline (Cho), creatine (Cr) and myo-inositol (mI), the respective markers for neuronal health, membrane turnover, energy, and astroglial metabolism. Unfortunately, despite the sensitivity of DTI and MRS, they are uncommonly integrated in clinical practice, due to potentially inconsistent results, small effect sizes, and incomplete consensus for their use. Research efforts are therefore currently concentrated on two fronts: to address these pitfalls for the established modalities, and to investigate the potential of new qMR modalities for TBI assessment.

These two general directions guided a prospective, longitudinal, multi-modal proton and sodium (²³Na) MR study of a single mTBI cohort, recently completed at our institution. The project consisted of a replicability study addressing the generalizability of previous MRS findings [15]; and exploratory studies with ²³Na MRI [16, 17] and magnetic resonance fingerprinting (MRF) [18], two advanced techniques not previously applied to TBI, and thus benchmarked against DTI. The broad goal of the project was to quantify each modality’s utility to discriminate patients from controls, and the extent of its cross-sectional and serial correlations with outcome measures. A secondary objective was to investigate injury mechanisms leveraging the rich multi-modal data, containing markers of microstructural integrity (ADC, FA), metabolism (NAA, Cr, Cho, mI) ionic homeostasis (apparent total sodium concentration, aTSC), and tissue microenvironment (longitudinal and transverse water relaxation times, T₁ and T₂). Cross-sectional results have been published for all modalities (²³Na MRI [16], MRF [18], DTI [18, 16], MRS [15]) and longitudinal data have also been published for all modalities (²³Na MRI [17], MRF [18], DTI [18]), except for MRS. This paper, therefore, has three objectives. First, to present the longitudinal MRS data, testing several hypotheses stemming from the cross-sectional (replicability) study. Second, to explore associations between relevant markers from different modalities to elucidate injury mechanisms. Third, to provide a compendium for the cross-sectional and predictive associations between all markers and outcome measures gathered over the course of the project, to enable comparisons for guiding future work.

Materials and Methods

Study design

Recruitment and longitudinal design

This prospective, longitudinal study was approved by the Institutional Review Board at the New York University Grossman School of Medicine, and reported according to the STROBE checklist for cohort studies. All participants gave written informed consent.

Patients between the ages of 18 to 65 were eligible if they sustained an injury that met the diagnostic criteria for mTBI according to the definition provided by the American Congress of Rehabilitation Medicine [19], less than two months prior to enrollment, as stated previously [16-18, 15]. Controls were eligible according to the same criteria as for patients, but excluding any TBI history. Additional criteria are listed in the Supplementary Methods.

Patients underwent MR imaging and outcome testing at approximately 1 month (Visit 1), 3 months (Visit 2), and 12 months (Visit 3) following injury. Controls only underwent MR imaging at a single visit.

Patient outcome measures

Patients were assigned a headache diagnosis (migraine, probable migraine, or tension-type) following a structured headache questionnaire known as the International Classification of Headache Disorders, 3^rd Edition (ICHD-3) [20], which assessed headache history and characteristics (i.e., frequency, location, intensity, quality, and accompanying symptoms).

Patients were evaluated for symptom burden using the Rivermead Post-Concussion Symptoms Questionnaire (RPQ) [21], which yields a total score that measures the presence and severity of 16 post-concussion symptoms, with higher scores reflecting greater symptom severity. Given the heterogeneity of symptom profiles following TBI, alternative scoring methods have been developed to improve sensitivity. These include: (1) the RPQ-3, which captures the three early physical symptoms commonly associated with concussion (e.g., headache, dizziness, nausea and/or vomiting) [22]; (2) the RPQ-13, which captures the remaining 13 somatic, emotional, and cognitive symptoms that may develop later [22]; and the three-factor model that further classifies symptoms into (3) RPQ-somatic, (4) RPQ-emotional, and (5) RPQ-cognitive domains [23].

Patients were evaluated for cognitive functioning using the Brief Test of Adult Cognition by Telephone (BTACT) [24], which produces a composite z-score that reflects overall cognition, with higher scores indicating better cognitive performance relative to normative values.

Patients were evaluated for global functioning using the Glasgow Outcome Scale – Extended (GOSE) [25], which categorizes patients into one of eight outcome levels that reflect varying degrees of disability and independence in daily life (e.g., death to upper good recovery), with higher scores indicating better functional recovery.

Additional details related to each of these four instruments are provided in the Supplementary Methods.

Imaging protocol

Scans were performed at 3 Tesla (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany) using a 20-channel transmit-receive head coil to acquire structural MRI, MRS, MRF, and DTI data, and using a custom-engineered dual-resonance ¹H/²³Na transmit-receive birdcage head coil to acquire ²³Na MRI data. Sequence parameters are listed in Tables S1-2.

Imaging post-processing and qMR markers

Nine markers were derived from the multi-modal imaging protocol (Table 1), following the data processing and analysis pipeline described in the Supplementary Methods. Briefly, aTSC was processed using a linear regression analysis approach over all voxels within the entire brain to yield global gray matter (GM) and WM concentrations [16, 17]. MRS metabolites, referenced to internal water, were processed using a similar linear regression analysis approach, but over voxels within atlas-defined frontal, parietal, temporal, and occipital lobes to yield lobar WM values [15]. T₁, T₂, ADC, and FA were processed using a voxel averaging approach [16, 18] to obtain values from seven WM regions (global, frontal, and posterior WM, the corona radiata, and the body, genu, and splenium of the corpus callosum) [26, 27].

Table 1.

Quantitative magnetic resonance imaging techniques and their markers

Marker	Units	Description	Biological feature
Proton (1H) protocol
Magnetic resonance spectroscopy (MRS)
N-acetyl-aspartate (NAA)	i.u.	A derivative of aspartic acid, with roles in myelin production (as an acetate contributor), mitochondrial function, and osmoregulation. Predominantly found within neurons.	Neuronal health and integrity
Choline (Cho)	i.u.	An essential nutrient and precursor to membrane phospholipids and acetylcholine. Present in higher concentrations within glial cells.	Membrane turnover
Creatine (Cr)	i.u.	A chemical that, when phosphorylated, serves as a substrate for adenosine triphosphate (ATP) metabolism. Present in higher concentrations within glial cells.	Cellular energy
Myo-inositol (mI)	i.u.	A sugar molecule with roles in cell signaling and osmoregulation (particularly in astrocytes).	Astroglial marker
Magnetic resonance fingerprinting (MRF)
Longitudinal relaxation time (T1)	ms	Time constant that describes the rate at which the longitudinal component of the magnetization vector returns to thermal equilibrium, termed spin-lattice relaxation to reflect the transfer of energy from the nuclear spins to their environment.	Biophysical tissue properties (e.g., fluid and macromolecule content, tissue density)
Transverse relaxation time (T2)	ms	Time constant that describes the rate at which the transverse component of the magnetization vector decays to zero, termed spin-spin relaxation to reflect the loss of coherence caused by interactions between spins.
Diffusion tensor imaging (DTI)
Apparent diffusion coefficient (ADC)	mm2/s	The average diffusion of water within tissue.	Tissue microstructure (e.g., white matter integrity, cellularity)
Fractional anisotropy (FA)	-	The degree of anisotropic diffusion of water molecules within tissue, represented by a value between zero and one. FA = 0 represents isotropic diffusion (free movement of water molecules in all directions) whereas FA = 1 (completely restricted movement of water molecules in one direction) represents anisotropic diffusion.
Sodium (23Na) protocol
23Na MRI
Apparent total sodium concentration (aTSC)	mM	The average of the intra- and extra-cellular sodium concentrations, weighted by their respective tissue volume fractions. Healthy cells typically maintain a sodium concentration gradient of 15 mM intracellularly, and 140 mM extracellularly, which is critical for various cellular functions, such as cell signaling.	Ionic homeostasis

Longitudinal MRS

The first objective of this work was to characterize longitudinal metabolite changes following mTBI, their relationships with outcomes at two time points along the recovery trajectory, and their potential to predict future impairment. Given that we found elevated WM Cho and Cr in patients at Visit 1 compared to their age- and sex-matched controls, and no group differences in levels of any other metabolite (NAA, mI) [15], we expected our longitudinal findings to solely involve WM Cho and Cr. Specifically, we hypothesized that elevated WM Cho and Cr would decline from Visit 1 to Visit 2 (H1), and that their rates of decline would correlate with improvements in symptomatic and cognitive outcomes (H2), consistent with the pattern of recovery that has been shown in the literature and previously in this cohort [17]. We expected this decline to reflect a return to metabolic homeostasis, which would be evidenced by the absence of differences in metabolite levels between patients at Visit 2 and their matched controls (H3), as well as the presence of cross-sectional correlations between metabolite levels and outcomes at Visit 2 (H4). Moreover, the elevated WM Cho and Cr at Visit 1 were driven by patients who endorsed post-concussive symptoms and experienced an incomplete return to pre-injury levels of global functioning, suggesting that metabolic alterations through the subacute phase of injury may be associated with more severe long-term impairments and prolonged recovery trajectories. Therefore, we hypothesized that elevated WM Cho and Cr at Visit 1 would correlate with worse symptomatic and cognitive outcomes at Visit 2 and at Visit 3 (H5) and predict worse global functioning at Visit 2 (H6).

Multi-modal associations to uncover injury mechanisms

The second objective of this work was to gain insight into injury mechanisms by leveraging the wide range of specificities of the nine markers (Table 1). This was done by assessing the relationships between markers which, for a given region and visit, showed differences in patients compared to controls. Such correspondence was only found once: in global WM, for lower aTSC and higher Cho and Cr, in patients at Visit 1 compared to controls. It suggested that, compared to the microstructural (DTI) phenotype of injury, physiological sequelae (metabolic and ionic) may have a more diffuse distribution. To test whether they also share a common pathophysiological origin, we hypothesized indirect correlations between global WM aTSC and WM Cho and Cr, exhibited in patients at Visit 1, but not in controls (H7).

Cross-modality comparisons of utility to correlate with or predict patient outcomes

Our four published studies on the application of MRS, ²³Na MRI, MRF, and DTI in this mTBI cohort revealed correlations with outcomes among Cho, Cr, aTSC, T₁, T₂, ADC, and FA. To determine which marker(s) demonstrate the greatest utility to correlate with or predict meaningful clinical endpoints comprising symptomatic, cognitive, and functional outcomes, a comprehensive comparison is needed. While a qualitative review of all prior findings is feasible, previous studies differed in terms of the number of subjects and the types of statistical methods. Additionally, predictive correlations between MRF and DTI markers and symptomatic and cognitive outcomes at Visit 3 were not previously performed, nor was predictive modelling of ²³Na MRI markers and functional outcomes at Visit 2. We can now address these inconsistencies by using data from the final full cohort and by re-examining all multi-modal qMR data with the same statistical methods outlined in H1 to H6. Therefore, the third objective of this work was to standardize the statistics of our antecedent work (alongside the final findings reported here), in order to draw the most meaningful conclusions in regards to the strongest cross-sectional and longitudinal associations with outcome after mTBI.

Statistical analyses

The current study’s sample size was determined using the power analyses described previously [16, 18]. Additional details are provided in the Supplementary Methods.

To test H1, the Wilcoxon signed-rank test (WSRT) was used to examine whether the within-subject change in each marker, from Visit 1 to Visit 2, was different from zero, and a linear mixed-effects model was applied to account for time from injury. Only patients providing data at both time points were included in the computations. Sensitivity analysis assessed baseline differences between patients who returned at Visit 2 and patients who were lost to follow-up.

To test H3, the Mann-Whitney U (MWU) test was used to examine whether patients were different from their age- and sex-matched controls in terms of each marker at Visit 2.

To test H2, H4, and H5, direct Spearman correlations were used to examine the relationships between markers and continuous symptomatic (RPQ) and cognitive (BTACT) outcome scores at different visits. Specifically, these included correlations between (i) markers’ rates of change and outcomes’ rates of change from Visit 1 to Visit 2 (H2); (ii) Visit 2 markers and Visit 2 outcomes (H4); and (iii) Visit 1 markers and Visit 2 and Visit 3 outcomes (H5). Additionally, to examine whether specific symptoms were driving any observed effects with RPQ, the following post hoc analyses were performed: direct Spearman correlations between (i) rates of change in marker levels and rates of change in RPQ subtest scores (e.g., RPQ-3, RPQ-13, RPQ-somatic, RPQ-emotional, and RPQ-cognitive) from Visit 1 to Visit 2; (ii) marker levels and RPQ subtest scores at Visit 2; and (iii) marker levels at Visit 1 and RPQ subtest scores at Visit 2 and Visit 3. To test H6, a univariate logistic regression analysis was conducted to model the relationship between Visit 1 markers and GOSE-defined Visit 2 functional recovery. This was followed by a receiver operating characteristic curve (ROC) analysis to evaluate the overall performance of the classifier in discriminating recovered (GOSE = 8) from non-recovered (GOSE ⩽ 7) patients at Visit 2. Clinically useful performance was defined by an area under the curve (AUC) value > 0.80.

Of note, both H5 and H6 test predictive ability, i.e., whether markers measured from a preceding visit are associated with outcome at a future visit, but different statistical methods were employed, owing to the different nature of the outcome scales. “Predictive correlations” are used to refer to analyses concerning the continuous outcomes (RPQ, BTACT), and “predictive modelling” is used to refer to analyses concerning the binary outcome (GOSE).

To test H7, lobar WM metabolites, derived from linear regression, were first averaged to yield “global” WM metabolites. Direct Spearman correlations were then estimated separately in patients at Visit 1 and in controls to examine the relationships between global WM aTSC and WM metabolites in these two populations.

A detailed explanation of our standardization approach, specifying the tests that were repeated or newly conducted for any given qMR dataset, is provided in Table S3. Results from within- and between-group analyses were compared using percent difference, a measure of relative magnitude change, and Cohen’s d effect sizes. Specifically, the standardized summary statistics, Cohen’s d_z [28] and Cohen’s d [29], were used to interpret the magnitude of the observed differences within and between groups, respectively. Results from correlation analyses were compared using the Spearman correlation coefficient (r), a measure of the strength of the monotonic relationship. Results from predictive modelling (logistic regression) were compared using the AUC to assess model performance, alongside parameter estimates, standard errors, odds ratios, and their 95% confidence intervals.

All statistical analyses were performed at the two-sided 5% significance level, using SAS software (version 9.4, SAS Institute). P values are reported without any adjustment for multiple statistical testing.

Results

Participants

Of the 31 patients enrolled, 27 were examined at Visit 1 (mean 22 days post-injury); of these, 19 returned for Visit 2 (mean 109 days from injury) and 15 for Visit 3 (mean 418 days from injury) (Fig. S1, Table 2). At Visit 2, 17 completed imaging and outcome testing and 2 completed outcome testing only. At Visit 3, 9 completed imaging and outcome testing, and 6 completed outcome testing only. All data were included in the analyses of H1 to H6 except for Visit 3 imaging, which had a small sample size (n = 9), and therefore precluded full hypothesis testing. Of the 24 controls enrolled, 23 completed imaging (Fig. S1, Table 2).

Table 2.

Participant demographics, patient injury characteristics, and patient outcomes

Parameters	Patients			Controls (n = 23)
	Visit 1 (n = 27)	Visit 2 (n = 19)	Visit 3 (n = 15)
Age, years	36 (12) [18, 60]	36 (12) [18, 60]	37 (13) [19, 61]	35 (13) [23, 61]
Sex, n
Female	20 (74%)	13 (68%)	11 (73%)	15 (65%)
Education, years	16 (2) [11, 22]	15 (2) [11, 18]	15 (2) [11, 18]	16 (3) [10, 26]
Race, n
Asian	1 (4%)	-	-	3 (13%)
Black or African American	5 (18%)	5 (26%)	4 (27%)	2 (9%)
White	16 (60%)	9 (47%)	7 (47%)	14 (61%)
Other (e.g., Mixed Race, Not Reported)	5 (18%)	5 (26%)	4 (27%)	4 (17%)
Time from injury, days	22 (10) [5, 53]	109 (18) [77, 142]	418 (50) [348, 499]	-
Cause of injury, n
Assault	1 (4%)	1 (5%)	1 (7%)	-
Bike fall	3 (11%)	1 (5%)	-	-
Fall	9 (33%)	8 (42%)	8 (53%)	-
Motor vehicle accident	2 (7%)	-	-	-
Pedestrian-related collision (e.g., Ped/Auto, Ped/Bike)	3 (11%)	3 (16%)	-	-
Sport collision	4 (15%)	3 (16%)	4 (27%)	-
Other (e.g., injury involving an inanimate object)	5 (19%)	3 (16%)	2 (13%)	-
Loss of consciousness, n
None	14 (52%)	-	-	-
<1 min.	5 (19%)	-	-	-
1-29 min.	4 (15%)	-	-	-
Unknown	4 (15%)	-	-	-
Post-traumatic amnesia, n
None	15 (56%)	-	-	-
<1 min.	7 (26%)	-	-	-
1-29 min.	3 (11%)	-	-	-
1-24 hrs.	2 (7%)	-	-	-
Alteration of consciousness, n
None	1 (4%)	-	-	-
<1 min.	5 (19%)	-	-	-
1-29 min.	15 (56%)	-	-	-
30-59 min.	2 (7%)	-	-	-
1-24 hrs.	3 (11%)	-	-	-
1-7 days	1 (4%)	-	-	-
MRI findings, n
Diffuse axonal injury	2 (7%)	-	-	-
Hemorrhage	1 (4%)	-	-	-
Headache phenotype, n
Migraine-like	12 (44%)	1 (5%)	2 (13%)	-
Probable migraine-like	8 (30%)	6 (31%)	3 (20%)	-
Tension-type headache	1 (5%)	2 (11%)	0 (0%)	-
No headache	6 (22%)	10 (53%)	10 (67%)	-
Functional Outcome (GOSE), n
5 (Lower moderate disability)	1 (4%)	-	-	-
6 (Upper moderate disability)	14 (52%)	2 (10%)	1 (7%)	-
7 (Lower good recovery)	4 (15%)	6 (32%)	8 (53%)	-
8 (Upper good recovery)	8 (30%)	11 (58%)	6 (40%)	-
Symptomatic Outcome (RPQ), score
RPQ (total score)	20 (11) [0, 43]	11 (13) [0, 43]	11 (13) [0, 39]	-
RPQ-3	4 (3) [0, 10]	2 (2) [0, 8]	2 (3) [0, 8]	-
RPQ-13	17 (10) [0, 33]	10 (11) [0, 35]	9 (11) [0, 33]	-
Three-factor model, RPQ-Somatic	11 (6) [0, 26]	5 (7) [0, 24]	6 (8) [0, 22]	-
Three-factor model, RPQ-Emotional	5 (4) [0, 15]	3 (4) [0, 14]	2 (4) [0, 14]	-
Three-factor model, RPQ-Cognitive	5 (4) [0, 12]	3 (3) [0, 9]	2 (3) [0, 9]	-
Cognitive Outcome (BTACT), z-score
BTACT (composite z-score)	−0.50 (1.11) [−2.26, 1.48]a	−0.15 (1.32) [−2.14, 2.55]	0.63 (1.08) [−0.62, 2.37]b	-

^aAt Visit 1, BTACT composite z-scores could only be computed for 25 patients, due to external interruption during administration (n = 2).

^bAt Visit 3, BTACT composite z-scores could only be computed for 13 patients, due to loss to follow-up (n = 2).

Values are given as either mean (standard deviation) [range] or count (percentage).

At Visit 1, the majority of patients (74%) were characterized as having migraine/probable migraine, and nearly all (96%) reported experiencing at least one symptom on the RPQ that either had not been present prior to their injury or had worsened following their injury (Table 2). The most reported post-concussive symptoms at this visit were fatigue (85%), headache (78%), difficulty concentrating (78%), slowed thinking (63%), and forgetfulness (59%).

Notably, the proportion of migraineurs and patients endorsing symptoms declined at a similar rate over time, with approximately 40% at Visit 2, and approximately 30% at Visit 3 (Table 2).

Longitudinal changes in metabolite levels (H1)

Of the 17 patients who received imaging at Visit 2, one was excluded due to outlying values (following a priori exclusion criteria in Table S2), and another had an incomplete MRS dataset at Visit 1 (Fig. S1). This resulted in a final sample of 15 patients with complete, adequate data at both Visits 1 and 2, who exhibited reductions in WM Cho (frontal: d_z = −0.86, p = 0.02; parietal: d_z = −0.81, p = 0.01) and Cr (temporal: d_z = −0.68, p = 0.02) over time (Fig. 1, Table 3A), even after adjusting for the potential confounding effect of days between visits (Table S4).

Fig. 1.

Longitudinal changes in white matter metabolites from Visit 1 to Visit 2. Boxplots showing the distributions of (A) lobar WM Cho and (B) lobar WM Cr in patients who provided data at Visit 1 (n = 15, red) and Visit 2 (n = 15, green). Within each box, the boundary lines illustrate the interquartile range, the horizontal line illustrates the median value, and the closed circles represent data from individual subjects. Each connecting line between two circles (one at Visit 1 and the other at Visit 2) highlights values obtained from the same patient across visits. For each statistically significant finding (*p < 0.05), the associated Cohen’s d_z effect size and 95% confidence interval (CI) are shown. Note that patients’ WM Cho and Cr were unidirectionally lower at Visit 2 than at Visit 1, with large reductions (∣d_z∣ > 0.8) in frontal and parietal WM Cho, and moderate reductions (0.5 ≤ ∣d_z∣ < 0.8) in temporal WM Cr.

Table 3.

Multi-modal findings in relation to hypotheses H1 to H6

A. Longitudinal change from Visit 1 to Visit 2
Marker	Region	Sample size (n)		Difference (%)	Effect size (dz)a	p
Cho	Frontal Lobe	15		4.84	−0.86	0.02*
	Parietal Lobe	15		4.35	−0.81	0.01*
	Temporal Lobe	15		6.21	−0.44	0.12
	Occipital Lobe	15		4.81	−0.59	0.08
Cr	Frontal Lobe	15		3.35	−0.55	0.08
	Parietal Lobe	15		2.52	−0.41	0.10
	Temporal Lobe	15		6.21	−0.68	0.02*
	Occipital Lobe	15		2.36	−0.40	0.15
aTSC	Global GM	17		9.98	1.20	<0.01*
	Global WM	17		3.77	0.50	0.11
B. Rate of change relationships from Visit 1 to Visit 2
Marker	Region	Outcome		Sample size (n)	r	p
Cho	Frontal Lobe	RPQ		15	0.52	0.05
	Parietal Lobe			15	0.68	<0.01*
	Temporal Lobe			15	0.50	0.06
	Occipital Lobe			15	0.75	<0.01*
	Frontal Lobe	BTACT		13	−0.33	0.27
	Parietal Lobe			13	−0.49	0.09
	Temporal Lobe			13	−0.48	0.10
	Occipital Lobe			13	−0.26	0.38
Cr	Frontal Lobe	RPQ		15	0.55	0.03*
	Parietal Lobe			15	0.52	0.05
	Temporal Lobe			15	0.65	<0.01*
	Occipital Lobe			15	0.43	0.11
	Frontal Lobe	BTACT		13	−0.46	0.11
	Parietal Lobe			13	−0.36	0.23
	Temporal Lobe			13	−0.45	0.12
	Occipital Lobe			13	−0.30	0.32
aTSC	Global GM	RPQ		17	0.45	0.07
	Global WM			17	0.36	0.12
	Global GM	BTACT		15	−0.60	0.02*
	Global WM			15	0.10	0.74
T2	Body CC	RPQ		13	−0.61	0.03*
FA	Corona radiata	RPQ		17	0.79	<0.01*
		BTACT		15	−0.65	0.01*
C. Group differences at Visit 2
Marker	Region	Sample size (n)		Difference (%)	Effect size (d)b	p
Cho	Frontal Lobe	16 patients / 23 controls		1.77	−0.13	0.66
	Parietal Lobe			0.85	−0.06	0.97
	Temporal Lobe			1.79	−0.11	0.92
	Occipital Lobe			0.83	−0.05	0.88
Cr	Frontal Lobe			1.50	0.20	0.66
	Parietal Lobe			2.62	0.29	0.50
	Temporal Lobe			1.85	−0.15	0.56
	Occipital Lobe			0.96	0.12	0.80
aTSC	Global GM	17 patients / 23 controls		1.36	−0.24	0.51
	Global WM			0.67	−0.08	0.77
D. Cross-sectional relationships at Visit 2
Marker	Region	Outcome		Sample size (n)	r	p
Cho	Frontal Lobe	RPQ		16	0.41	0.11
	Parietal Lobe			16	0.27	0.31
	Temporal Lobe			16	0.53	0.03*
	Occipital Lobe			16	0.42	0.11
	Frontal Lobe	BTACT		16	0.14	0.59
	Parietal Lobe			16	0.11	0.68
	Temporal Lobe			16	0.17	0.52
	Occipital Lobe			16	0.04	0.90
Cr	Frontal Lobe	RPQ		16	0.45	0.08
	Parietal Lobe			16	0.02	0.94
	Temporal Lobe			16	0.56	0.03*
	Occipital Lobe			16	0.09	0.75
	Frontal Lobe	BTACT		16	−0.10	0.72
	Parietal Lobe			16	−0.06	0.84
	Temporal Lobe			16	0.17	0.52
	Occipital Lobe			16	0.50	0.05
aTSC	Global GM	RPQ		17	−0.15	0.56
	Global WM			17	0.42	0.10
	Global GM	BTACT		17	−0.01	0.96
	Global WM			17	0.11	0.68
T1	Corona radiata	RPQ		16	0.70	<0.01*
	Body CC			16	0.51	0.04*
	Genu CC			16	0.66	<0.01*
	Splenium CC			16	0.58	0.02*
	Frontal WM			16	0.56	0.03*
T2	Genu CC	RPQ		16	0.58	0.02*
ADC	Corona radiata	RPQ		17	0.64	<0.01*
FA	Frontal WM	RPQ		17	−0.56	0.02*
E. Predictive relationships with Visit 2 symptomatic and cognitive outcomes
Marker	Region	Outcome		Sample size (n)	r	p
Cho	Frontal Lobe	RPQ		18	0.18	0.48
	Parietal Lobe			18	0.23	0.37
	Temporal Lobe			18	0.29	0.24
	Occipital Lobe			18	0.22	0.38
	Frontal Lobe	BTACT		18	0.08	0.75
	Parietal Lobe			18	0.11	0.65
	Temporal Lobe			18	0.30	0.23
	Occipital Lobe			18	−0.02	0.93
Cr	Frontal Lobe	RPQ		18	−0.16	0.51
	Parietal Lobe			18	0.20	0.43
	Temporal Lobe			18	0.24	0.35
	Occipital Lobe			18	−0.04	0.87
	Frontal Lobe	BTACT		18	0.32	0.20
	Parietal Lobe			18	0.30	0.23
	Temporal Lobe			18	0.25	0.32
	Occipital Lobe			18	0.32	0.19
aTSC	Global GM	RPQ		19	−0.13	0.60
	Global WM			19	0.02	0.94
	Global GM	BTACT		19	0.39	0.10
	Global WM			19	0.15	0.54
T1	Corona radiata	BTACT		15	0.58	0.02*
	Body CC			15	0.57	0.03*
	Genu CC			15	0.82	<0.01*
	Splenium CC			15	0.57	0.03*
	Global WM			15	0.58	0.03**
ADC	Corona radiata	RPQ		19	0.51	0.02*
	Splenium CC	BTACT		19	0.48	0.04*
FA	Frontal WM	RPQ		19	−0.53	0.02*
F. Predictive relationships with Visit 3 symptomatic and cognitive outcomes
Marker	Region	Outcome		Sample size (n)	r	p
Cho	Frontal Lobe	RPQ		14	0.31	0.28
	Parietal Lobe			14	0.36	0.21
	Temporal Lobe			14	0.24	0.41
	Occipital Lobe			14	0.23	0.43
	Frontal Lobe	BTACT		12	0.59	0.04*
	Parietal Lobe			12	0.56	0.06
	Temporal Lobe			12	0.76	<0.01*
	Occipital Lobe			12	0.61	0.04*
Cr	Frontal Lobe	RPQ		14	−0.17	0.55
	Parietal Lobe			14	0.11	0.71
	Temporal Lobe			14	0.20	0.50
	Occipital Lobe			14	0.13	0.66
	Frontal Lobe	BTACT		12	0.45	0.14
	Parietal Lobe			12	0.46	0.13
	Temporal Lobe			12	0.57	0.05
	Occipital Lobe			12	0.85	<0.01*
aTSC	Global GM	RPQ		15	−0.30	0.29
	Global WM			15	0.01	0.98
	Global GM	BTACT		13	0.19	0.52
	Global WM			13	0.61	0.03*
T1	Genu CC	BTACT		10	0.71	0.03*
ADC	Genu CC	BTACT		13	0.63	0.02*
	Splenium CC			13	0.64	0.02*
G. Predictive relationships with Visit 2 functional outcome
Marker	Region	Estimate	SE	Odds Ratio (95% Confidence Interval)	p	AUC
Cho	Frontal Lobe	−1.14	0.74	0.32 (0.08, 1.34)	0.12	0.65
	Parietal Lobe	−1.51	0.85	0.22 (0.04, 1.17)	0.08	0.72
	Temporal Lobe	−0.84	0.69	0.43 (0.11, 1.68)	0.22	0.67
	Occipital Lobe	−0.07	0.61	0.94 (0.28, 3.10)	0.91	0.54
Cr	Frontal Lobe	−0.20	0.34	0.82 (0.43, 1.58)	0.55	0.59
	Parietal Lobe	−0.16	0.31	0.85 (0.46, 1.58)	0.61	0.58
	Temporal Lobe	0.12	0.25	1.13 (0.69, 1.86)	0.63	0.53
	Occipital Lobe	−0.04	0.27	0.96 (0.57, 1.64)	0.89	0.55
aTSC	Global GM	−0.13	0.22	0.88 (0.58, 1.34)	0.56	0.56
	Global WM	0.11	0.15	1.12 (0.83, 1.51)	0.46	0.54

^ad_z < 0 denotes decrease (Visit 1 to Visit 2)

^bd < 0, denotes lower values in patients vs. controls

CC = corpus callosum; SE = standard error; AUC = area under the curve

Note. Markers were predominately measured within WM with the exception of aTSC. For brevity, only the statistically significant (*p < 0.05) findings involving ²³Na MRI, MRF, and DTI markers are presented (see Table S3 for overlap with prior publications and Table S6 for NAA and mI). (A) The Wilcoxon signed-rank test (WSRT) was used to evaluate whether markers from Visit 1 were significantly different from those at Visit 2, following H1. (B) Spearman correlations were used to evaluate whether rates of change in markers were associated with rates of change in symptomatic and cognitive outcomes, from Visit 1 to Visit 2, following H2. (C) The Mann-Whitney U (MWU) test was used to evaluate whether markers from Visit 2 were significantly different in patients compared to controls, following H3. (D) Spearman correlations were used to evaluate whether markers at Visit 2 cross-sectionally correlated with symptomatic and cognitive outcomes at Visit 2, following H4. (E) Spearman correlations were used to evaluate whether markers at Visit 1 serially correlated with symptomatic and cognitive outcomes at Visit 2 or (F) Visit 3, following H5. (G) Logistic regression was used to evaluate whether markers at Visit 1 were significantly associated with functional outcome at Visit 2, following H6. Additionally, receiver operating characteristic curve analysis assessed the overall discriminative ability of the logistic regression models for predicting functional outcome at Visit 2.

Rate of change relationships with symptomatic and cognitive outcomes (H2)

The rates of change in WM Cho (parietal: r = 0.68, p < 0.01; occipital: r = 0.75, p < 0.01) and Cr (frontal: r = 0.55, p = 0.03; temporal: r = 0.65, p < 0.01) directly correlated with the rate of change in RPQ (Fig. 2, Table 3B), as well as with the rates of change in all RPQ subtest scores (Table S5A). No correlations were observed between the rates of change in WM Cho or Cr and the rate of change in BTACT.

Fig. 2.

Rate of change relationships between white matter metabolites and symptomatic outcome from Visit 1 to Visit 2. Associations between (A) the rate of change in lobar WM Cho and the rate of change in RPQ total scores, and (B) the rate of change in lobar WM Cr and the rate of change in RPQ total scores. Note that a greater rate of reduction in lobar WM Cho and Cr levels correlated with a greater rate of reduction in RPQ total scores (indicating a reduction in the number of endorsed symptoms or the degree of symptom severity), particularly in the occipital lobe (r > 0.70). Only statistically significant associations (p < 0.05) are shown alongside their Spearman correlation coefficients (r) and 95% confidence intervals (CI).

Group differences in metabolite levels (H3)

The 16 patients at Visit 2 demonstrated no differences in WM Cho or Cr in comparison to their 23 age- and sex-matched controls (Fig. 3, Table 3C).

Fig. 3.

Group differences in white matter metabolites at Visits 1 and 2. Boxplots showing the distributions of (A) lobar WM Cho and (B) lobar WM Cr in controls (n = 23, blue), patients at Visit 1 (n = 26, red), and patients at Visit 2 (n = 16, green). Within each box, the boundary lines illustrate the interquartile range, the horizontal line illustrates the median value, and the closed circles represent data from individual subjects. For each statistically significant finding (*p < 0.05), the associated Cohen’s d effect size and 95% confidence interval (CI) are shown. Using the complete dataset provided for the current study, which includes a slightly larger sample of control subjects compared to that of Chen et al. [15], we replicated group differences between controls and patients at Visit 1, and found no group differences between controls and patients at Visit 2.

Cross-sectional relationships with symptomatic and cognitive outcomes (H4)

WM Cho (temporal: r = 0.53, p = 0.03) and Cr (temporal: r = 0.56, p = 0.03) directly correlated with RPQ at Visit 2 (Fig. 4A-B, Table 3D), as well as with RPQ subtest scores at Visit 2 except for RPQ-3 and RPQ-Somatic, respectively (Table S5B). No correlations were observed between WM Cho or Cr and BTACT at Visit 2.

Fig. 4.

Cross-sectional and predictive relationships between white matter metabolites and outcomes. Cross-sectional associations between (A) Visit 2 WM Cho and (B) Visit 2 WM Cr, and Visit 2 RPQ total scores. Note that elevated lobar WM Cho and Cr levels correlated with higher RPQ total scores (indicating a higher number of endorsed symptoms or worse symptom severity). Predictive associations between (C) Visit 1 WM Cho and (D) Visit 1 WM Cr, and Visit 3 BTACT composite z-scores, which are strongest (r > 0.70) in the temporal and occipital lobes, respectively. Note that elevated lobar WM Cho and Cr levels correlated with higher BTACT composite z-scores (counterintuitively indicating above average cognitive performance). Only statistically significant associations (p < 0.05) are shown alongside their Spearman correlation coefficients (r) and 95% confidence intervals (CI).

Predictive relationships with symptomatic and cognitive outcomes (H5)

Visit 1 WM Cho (frontal: r = 0.59, p = 0.04; temporal: r = 0.76, p < 0.01; occipital: r = 0.61, p = 0.04) and Cr (occipital: r = 0.85, p < 0.01) directly correlated with Visit 3 BTACT (Fig. 4C-D, Table 3F). No correlations were observed between Visit 1 WM Cho or Cr and Visit 2 BTACT (Table 3E), Visit 2 RPQ (Table 3E), or Visit 3 RPQ (Table 3F).

Predictive relationships with global functional outcome (H6)

Of the 19 patients who underwent outcome testing at Visit 2, 11 were defined as functionally recovered and 8 were defined as functionally non-recovered. After imputation, following the rule established in the Supplementary Methods, a total of 10 patients were defined as functionally non-recovered. No Visit 1 metabolite predicted Visit 2 functional recovery (Tables 3G, S5G).

Relationships between MRS and ²³Na MRI markers (H7)

No correlations were observed between WM Cho or Cr and global WM aTSC at Visit 1, in either patients or controls.

Compendium of cross-sectional and predictive associations of qMR markers with patient outcomes

Summaries of MRS, ²³Na MRI, MRF, and DTI findings obtained from standardized Spearman correlation and logistic regression analyses are provided in Tables 3, S5, and S6. Any differences compared to the previously published findings, such as changes in statistical significance following standardization, are described in Table S3. Below we highlight the statistically significant associations with RPQ, BTACT, and GOSE that can be considered “strong” in accordance with Schober et al. [30] (Spearman ∣r∣ ⩾ 0.70).

In terms of RPQ, the strongest associations involved rate of change correlations for FA in the corona radiata (r = 0.79; Table 3B), WM Cho (occipital: r = 0.75; Table 3B), and WM mI (parietal: r = 0.73; Table S6B), followed by a cross-sectional correlation at Visit 2 for T₁ in the corona radiata (r = 0.70; Table 3D). There were no strong cross-sectional correlations at Visit 1 (Table S7), nor any strong predictive correlations with Visit 2 RPQ (Tables S3E, S5E) or Visit 3 RPQ (Tables 3F, S5F).

In terms of BTACT, the strongest associations involved predictive correlations with Visit 2 BTACT (Table 3E), for T₁ in the genu of the corpus callosum (r = 0.82); and with Visit 3 BTACT (Table 3F), for WM Cho (temporal: r = 0.76), WM Cr (occipital: r = 0.85), and T₁ in the genu of the corpus callosum (r = 0.71). There was also a strong rate of change correlation for WM mI (parietal: r = −0.82; Table S6B). There were no strong cross-sectional correlations at Visit 1 (Table S7) nor at Visit 2 (Tables 3D, S6D).

In terms of GOSE, there were no statistically significant predictors of Visit 2 functional recovery (Tables 3G, S5G).

Discussion

Our prospective longitudinal case-controlled study included new (²³Na MRI, MRF) and established (MRS, DTI) modalities to assess their potential clinical utility, and to leverage their varied strengths as biomarkers to examine mechanisms of injury and predictors of unfavorable outcomes. The current work focused on aims that arose from key findings from our four initial studies [16-18, 15], starting with the longitudinal evaluation of MRS markers, which enabled the investigation of possible injury mechanisms, along with a standardized comparison of each marker’s cross-sectional and predictive relationships with outcome.

Longitudinal MRS

WM Cho and Cr normalized to control levels from Visit 1 to Visit 2 (in support of H1 and H3), at rates that correlated with the rate of symptomatic improvement in patients (in partial support of H2). Higher WM Cho and Cr cross-sectionally correlated with worse symptomatology at Visit 2 (in partial support of H4). Yet, elevated Visit 1 WM Cho and Cr were not predictive of worse Visit 2 or Visit 3 symptomatic, cognitive, or functional outcomes (in rejection of H5 and H6).

A mechanistic explanation for these results is speculative, but the normalization of metabolism over time shows that the higher Cho and Cr at Visit 1 did not reflect irreversible injury. One possibility is that cell membrane changes were responsible for the release of Cho from its bound state [31-33] and Cr was recruited to generate ATP for the energy-demanding processes of membrane stabilization and repair [34]. As cell membrane dynamics returned to normal, so did metabolite levels, consistent with observed metabolite trajectories in past serial MRS studies in mild-to-moderate TBI [35-42] and histological evidence of recovery [43-45].

Notably, normalization of metabolite levels over time has been reported in past serial MRS studies in mild-to-moderate TBI [35-42], consistent with histological evidence of recovery [43-45], but with some exceptions [46, 39]. The most relevant literature to which we can compare our results are two studies which utilized a similar multi-voxel MRS acquisition and global linear regression analysis approach to examine diffuse injury. Yeo et al. [36] also observed elevated WM Cr (alongside elevated WM Glx) in adults within 1 month of mTBI compared to controls, which decreased over 3 – 5 months, and correlated with reduced symptoms at follow-up. Berger et al. [42] observed decreased WM NAA/Cr and NAA/Cho ratios in children within 3 weeks of mild-to-moderate TBI compared to controls, both of which returned to control levels at approximately 12 months. Furthermore, increases in both ratios were associated with better future cognitive outcomes. These observations, alongside our cross-sectional and serial results suggest that TBI results in global injury, which abates over time. Previously, we concluded that using an MRS approach with high sensitivity may be more important than the choice of region [15], and this assertion is supported by the serial results here and cited literature.

Injury mechanisms

Although the MRS findings above captured a pattern of diffuse WM injury consistent with our ²³Na MRI findings [16, 17], using the same high sensitivity global linear regression approach [47] in the same patient cohort, the absence of multi-modal associations (in rejection of H7) suggests that there is no shared pathophysiologic etiology. In other words, ionic disequilibrium (lower WM aTSC) may arise from derangements not directly related to processes involving increased membrane turnover (higher WM Cho) and energy production (higher WM Cr).

To our knowledge, this is the first study to have examined the relationship between MRS and ²³Na MRI markers in healthy subjects or in mTBI.

Considerations for using the compendium

We developed a compendium that takes into consideration all relevant qMR markers and outcome measures, as follows: First, only global markers for MRS and ²³Na MRI are included. Regional results for these two modalities were mostly not statistically significant, but the differences in the means between patients and controls were invariantly unidirectional, suggesting global injury with insufficient statistical power to detect regional changes. The precision was boosted by a global approach [47] applied to both MRS and ²³Na MRI, resulting in statistically significant results with large effect sizes; Second, GM markers were not included for MRS, MRF, and DTI, because all either lacked statistically significant findings or showed only single findings which were not deemed relevant; Third, subtest scores were not included for RPQ and BTACT, since there were no domain-specific findings for any qMR marker. The purpose of the above selection criteria was to maximize the relevance of the compendium and to decrease the probability of type I errors in the updated statistical analyses. We summarize main findings below (which are further discussed in the Supplementary Discussion), and suggest a framework for comparing biomarker performance.

Assessing biomarker utility based on the compendium

Conclusions on biomarker utility need to be considered after multiple levels of evaluation. Identifying the strongest correlations is a useful summary (see Results), but is insufficient. One definition of a biomarker requires that it deviates from normal values, correlates with clinical presentation, changes in accordance with changes in clinical presentation, and predicts future outcomes. It is possible for any qMR marker to meet individual criteria within this definition owing simply to type I error. Therefore, it is also useful to look at whether a biomarker performed well in most domains, even if it did not show strong correlations in all of them. In the context of this study, consistent biomarker behaviour is ascribed to a biomarker that correlates with outcome the same way (direction) at both Visits 1 and 2 (“Domain 1”); that changes according to changes to the outcome measure (e.g., biomarker rate of change is directionally consistent with the outcome measure’s rate of change) (“Domain 2”); and predicts similar outcome(s) at both future Visits 2 and 3 (“Domain 3”).

The MRF marker T₁ showed consistently positive correlations with RPQ at Visit 1 and RPQ at Visit 2 (addressing Domain 1). However, differentiation between cohorts is considered a core feature of diagnostic biomarkers [48, 49], and this property was not seen for T₁, i.e., T₁ was not abnormal at Visit 1. In contrast, MRS and ²³Na MRI markers were abnormal at Visit 1, showing altered levels in patients compared to controls. Moreover, for WM Cho and Cr, changes at Visit 1 were found only in symptomatic individuals [15], and showed consistently positive correlations with RPQ at Visit 2 and with the rate of change in RPQ from Visit 1 to Visit 2 (addressing Domains 1 and 2). For GM aTSC, changes at Visit 1 showed a consistent, positive correlation with BTACT at Visit 1, but only correlated with BTACT subtest scores at Visit 2 [17] (partially addressing Domain 1), and showed an inconsistent (negative) correlation with the rate of change in BTACT from Visit 1 to Visit 2. The MRF marker T₂ and the DTI marker FA showed inconsistent correlations across domains. No markers yielded findings that addressed Domain 3, i.e., none consistently predicted future symptomatic (RPQ), cognitive (BTACT), or functional (GOSE) outcomes. Altogether, following this framework, WM Cho and Cr fulfilled the most criteria, followed by regional WM T₁.

Limitations

The small sample size is an obvious limitation, but it is one that is unavoidable, as the trade-off is the large number of advanced MRI modalities applied to a single cohort, two of them for the first time in TBI. Almost all modalities required local technical expertise, while ²³Na MRI required multi-nuclear scanner capabilities and custom-built hardware, as well as an additional scanning session. Hence, the data here are a unique combination, not available in data repositories.

The attrition that occurred from Visit 1 to Visit 2, and again from Visit 2 to Visit 3, was in part a consequence of the COVID-19 pandemic. Consequently, patients’ Visit 3 imaging data were excluded from the analyses, and control subjects were only scanned once (but used for group-level comparisons with patients at both visits). In addition, patients returned for their follow-up visits at widely different intervals, ranging from 77 to 142 days from injury for Visit 2, and ranging from 348 to 499 days from injury for Visit 3 (Table 2). However, this variation was addressed using sensitivity analysis, which observed no differences in baseline values between patients with and without follow-up data (Table S8), and possible time-dependent associations (in H1) were addressed using a linear mixed-effects regression model.

While comorbid conditions such as migraine or mood disorders could have influenced our cross-sectional findings in H2 and H4, we addressed this concern through targeted post hoc analyses using RPQ subtest scores (Table S5), which demonstrated that elevated WM Cho and Cr remained significantly associated with symptomatology across different (somatic, emotional, and cognitive) domains. These findings support the interpretation that the observed metabolic changes are, indeed, related to injury. Nonetheless, future studies with larger and more phenotypically diverse cohorts may be better suited for examining symptom-specific contributions to metabolite alterations.

Only univariate models were performed to evaluate the predictive associations between qMR markers at Visit 1 and functional recovery at Visit 2, even though clinical characteristics (e.g., injury severity) and demographic parameters (e.g., age, sex, education) may, themselves, serve as predictors of future post-concussive symptoms [2] and incomplete functional recovery [50-52] following mTBI. We did not build more complex, multivariate models to account for these fixed factors due to our limited number of subjects.

The biomarker assessment criteria’s focus on consistent findings may not be appropriate for all imaging markers, as some have multiple specificities. For example, changes in different directions may be due to different pathophysiology. However, for utility in the clinical setting, consistency in directionality of changes and in their relationship with clinical outcome is the preferred biomarker phenotype.

Conclusion

Overall, this study delivers a comprehensive cross-sectional and longitudinal assessment of proton and sodium qMR markers for the evaluation and prognosis of symptomatic, cognitive, and functional outcomes after mTBI. At Visit 1, patients had global metabolic (MRS) and ionic (²³Na MRI) abnormalities, but these findings were unrelated, suggesting a different pathophysiological origin. By Visit 2, both types of abnormalities had resolved, and outcome scores had improved. Among all qMR markers, white matter choline and creatine from MRS met the most criteria for a clinical biomarker. This determination was made based on their associations with symptomatic outcome. White matter T₁ from MRF also fulfilled some biomarker criteria, again in the symptomatic domain. The widely used markers from DTI did not fulfil a comparable number of biomarker criteria. No markers consistently predicted future outcomes. A compendium containing associations of all markers with outcome is included to guide future studies.

Supplementary Material

Supplementary information: This manuscript has accompanying supplementary material.

Data availability:

The data that support the findings of this study are available from the corresponding author upon reasonable request.