Metabolomics

, 12:185

Metabolomics profiling of concussion in adolescent male hockey players: a novel diagnostic method

  • Mark Daley
  • Greg Dekaban
  • Robert Bartha
  • Arthur Brown
  • Tanya Charyk Stewart
  • Timothy Doherty
  • Lisa Fischer
  • Jeff Holmes
  • Ravi S. Menon
  • C. Anthony Rupar
  • J. Kevin Shoemaker
  • Douglas D. Fraser
Original Article

DOI: 10.1007/s11306-016-1131-5

Cite this article as:
Daley, M., Dekaban, G., Bartha, R. et al. Metabolomics (2016) 12: 185. doi:10.1007/s11306-016-1131-5

Abstract

Introduction

Concussions are a major health concern as they cause significant acute symptoms and in some athletes, long-term neurologic dysfunction. Diagnosis of concussion can be difficult, as are the decisions to stop play.

Objective

To determine if concussions in adolescent male hockey players could be diagnosed using plasma metabolomics profiling.

Methods

Plasma was obtained from 12 concussed and 17 non-concussed athletes, and assayed for 174 metabolites with proton nuclear magnetic resonance and direct injection liquid chromatography tandem mass spectrometry. Data were analysed with multivariate statistical analysis and machine learning.

Results

The estimated time from concussion occurrence to blood draw at the first clinic visit was 2.3 ± 0.7 days. Using principal component analysis, the leading 10 components, each containing 9 metabolites, were shown to account for 82 % of the variance between cohorts, and relied heavily on changes in glycerophospholipids. Cross-validation of the classifier using a leave-one out approach demonstrated a 92 % accuracy rate in diagnosing a concussion (P < 0.0001). The number of metabolites required to achieve the 92 % diagnostic accuracy was minimized from 174 to as few as 17 metabolites. Receiver operating characteristic analyses generated an area under the curve of 0.91, indicating excellent concussion diagnostic potential.

Conclusion

Metabolomics profiling, together with multivariate statistical analysis and machine learning, identified concussed athletes with >90 % certainty. Metabolomics profiling represents a novel diagnostic method for concussion, and may be amenable to point-of-care testing.

Keywords

Concussion Diagnosis Biomarker Metabolomics Adolescents Ice hockey 

1 Introduction

Concussions are a major public health concern and they disproportionately affect youth, with more than half occurring in children and adolescents (http://www.cdc.gov/TraumaticBrainInjury). Concussions often result in significant acute symptoms and long-term consequences for daily functioning, and ongoing neurologic and cognitive development. Diagnosis of concussion can be difficult, as are the decisions to stop play (Harmon et al. 2013). Thus, there is great interest in the discovery of blood biomarkers to aid in concussion diagnoses (Di Battista et al. 2013), but a single biomarker with sufficient sensitivity and specificity for widespread use has not been identified (Jeter et al. 2013).

Metabolomics is a field of study that measures a person’s small metabolite profile (<1500 Daltons), including amino acids, acylcarnitines, glycerophospholipids, sphingolipids and sugars (Bujak et al. 2014). Two complementary analytical methods for metabolomics are proton nuclear magnetic resonance (1H NMR; μM range) spectroscopy and mass spectrometry (MS; pM range). The attraction of metabolomics lies with the concept that metabolites fall downstream of genetic, transcriptomic, proteomic and environmental variation, thus providing an integrated and dynamic measure of a medical condition (Bujak et al. 2014).

Our aims were: (1) to recruit adolescent male ice hockey players who had suffered a recent concussion and age-, sex- and activity-matched controls; (2) to measure a large number of their plasma metabolites with 1H NMR and direct injection-liquid chromatography/MS/MS (DI-LC/MS/MS); and (3) to determine whether metabolomics profiling could accurately identify those athletes that suffered a concussion.

2 Materials and methods

2.1 Subject recruitment

Male adolescent ice hockey athletes (Bantam Division; aged 12–14 years) were recruited to participate in this study. All subjects were assessed by Primary Care Sport Medicine Physicians at the Fowler Kennedy Sport Medicine Clinic at Western University. Concussion was diagnosed when there was an observed mechanism of injury followed by onset of typical concussive symptoms, and the absence of structural injury (i.e. no focal neurological abnormalities on examination). Control athletes consisted of non-injured hockey players that were age- and sex-matched, and that had not suffered a concussion in the past 6 months. Any subject with a reported neurological disease was excluded.

Concussed and control athletes, including their parents/guardians, completed a Sport Concussion Assessment Tool–3rd edition [SCAT3; 13–14 years of age] (Guskiewicz et al. 2013) or a Child-SCAT3[(a modified tool recommended for children 12 years of age or younger that takes into account developmental differences in performance] (Glaviano et al. 2015). All athletes underwent a complete history, physical and neurologic examination by a Primary Care Sport Medicine Physician with expertise in concussion management, and all injured athletes were provided with standardized concussion care.

All athletes on the first clinic visit had 20 ml of blood drawn into EDTA Vacutainer® tubes. The blood was centrifuged, the plasma aliquoted into cryovials at a volume of 500 μl and stored at −80 °C. Freeze/thaw cycles were avoided. Plasma was collected by strict standard operating procedures, with equal processing times between cohorts (Brisson et al. 2012; Gillio-Meina et al. 2013). No restrictions were placed on any subjects (i.e. fasting), and thus, they should be considered in their natural state.

2.2 DI-LC/MS/MS

A targeted quantitative metabolomics approach was applied to analyze the plasma samples using a combination of direct injection mass spectrometry (AbsoluteIDQ™ Kit) with a reverse-phase LC/MS/MS Kit (BIOCRATES Life Sciences AG, Austria). The method combines the derivatization and extraction of analytes, and selective mass-spectrometric detection using multiple reaction monitoring pairs (standards are integrated in the Kit plate filter for metabolite quantification). Briefly, plasma samples were thawed on ice and then vortexed and centrifuged at 13,000×g. Each plasma sample (10 µl) was loaded onto the center of the filter on the upper 96-well kit plate and dried in a stream of nitrogen. Subsequently, 20 µl of a 5 % solution of phenyl-isothiocyanate was added for derivatization. After incubation, the filter spots were dried again using an evaporator. Extraction of the metabolites was then achieved by adding 300 µl methanol containing 5 mM ammonium acetate. The extracts were obtained by centrifugation into the lower 96-deep well plate, followed by a dilution step with kit MS running solvent. Mass spectrometric analysis was performed on an API4000 Qtrap® tandem mass spectrometry instrument (Applied Biosystems/MDS Analytical Technologies, Foster City, CA) equipped with a solvent delivery system. The samples were delivered to the mass spectrometer by LC followed by a DI. The Biocrates MetIQ software was used to control the entire assay workflow, from sample registration to automated calculation of metabolite concentrations. A targeted profiling scheme was used to quantitatively screen for known small molecule metabolites using multiple reaction monitoring, neutral loss and precursor ion scans.

2.3 1H NMR

Plasma samples were deproteinized by ultra-filtration as previously described (Psychogios et al. 2011). Prior to filtration, 3 kDa cut-off centrifugal filter units (Amicon Microcon YM-3) were rinsed five times each with 0.5 ml of H2O and centrifuged (10,000 rpm for 10 min) to remove residual glycerol bound to the filter membranes. Aliquots of each plasma sample were then transferred into the centrifuge filter devices and centrifuged (10,000 rpm for 20 min) to remove macromolecules (primarily protein and lipoproteins) from the sample. The filtrates were collected and the volumes were recorded. The volume of the sample was adjusted with addition of 50 mM NaH2PO4 buffer (pH 7.0) until the total volume of the sample was 600 µl. Any sample that had to have buffer added to bring the solution volume to 600 μl, was annotated with the dilution factor and metabolite concentrations were corrected in the subsequent analysis. Subsequently, 70 µl of D2O and 30 µl of a standard buffer solution (11.7 mM DSS (disodium-2, 2-dimethyl-2-silcepentane-5-sulphonate], 730 mM imidazole, and 0.47 % NaN3 in H2O) was added to the sample.

The plasma sample (700 µl) was then transferred to a standard NMR tube for subsequent spectral analysis. All 1H NMR spectra were collected on a 500 MHz Inova (Varian Inc. Palo Alto, CA) spectrometer equipped with a 5 mm HCN Z-gradient pulsed-field gradient room-temperature probe. Proton NMR spectra were acquired at 25 °C using the first transient of the NOESY-pre-saturation pulse sequence, chosen for its high degree of quantitative accuracy (Saude et al. 2006). All free induction decays were zero- filled to 64 K data points and subjected to line broadening of 0.5 Hz. The singlet produced by the DSS methyl groups was used as an internal standard for chemical shift referencing (set to 0 ppm) and for quantification all 1H NMR spectra were processed and analyzed using the Chenomx NMR Suite Professional Software package version 7.1 (Chenomx Inc, Edmonton, AB). The Chenomx NMR Suite software allows for qualitative and quantitative analysis of an NMR spectrum by manually fitting spectral signatures from an internal database to the spectrum. Specifically, the spectral fitting for metabolites was performed using the standard Chenomx 500 MHz metabolite library. Typically 90 % of visible peaks were assigned to a compound and more than 90 % of the spectral area could be routinely fit using the Chenomx spectral analysis software. Most of the visible peaks are annotated with a compound name. It has been previously shown that this fitting procedure provides absolute concentration accuracy of 90 % or better. Each spectrum was processed and analyzed by at least two NMR spectroscopists to minimize compound misidentification and mis-quantification. We used sample spiking to confirm the identities of assigned compounds. Sample spiking involves the addition of 20–200 µM of the suspected compound and examination of the resulting spectra to determine whether the relative NMR signal intensity changed as expected.

2.4 Data analyses

Demographic and concussion tool data were reported as mean ± standard deviation (SD), with a P value <0.05 taken as our standard of statistical significance. Raw NMR and MS data for each subject were ingested and normalized within each metabolite marker, across subjects. More specifically, the data for each metabolic marker were scaled to have unit norm. Initial exploratory analysis involved performing Principal Component Analysis (PCA) directly on the subjects by a metabolites matrix. Motivated by the observation that the inherent dimensionality of the data was significantly lower than the number of metabolite markers, we performed nonlinear dimensionality reduction on the full data matrix using the t-distributed stochastic nearest neighbour (t-SNE) embedding algorithm (van der Maaten and Hinton 2008). Unlike PCA which enforces a brittle, orthogonal, linear refactorization of the data, t-SNE assumes that the ‘optimal’ representation of the data lies on a manifold with complex geometry, but low dimension, embedded in the full dimensional space of the raw data. We used t-SNE to reduce the full metabolic dataset down to only two dimensions. We then trained separate support vector machines (SVM), with linear kernels, on the dimensionality reduced and full datasets to classify subjects as concussed or non-concussed. We cross-validated our classifier using a leave-one-out approach and assessed statistical significance against a null distribution generated by resampling. To investigate the robustness of the trained classifier, the areas under the receiver operating characteristic (ROC) curves were determined (Hajian-Tilaki 2013); one curve was generated for each fold of a fourfold cross-validation approach where the classifier for each fold is trained on a subset of the data and then tested on a disjoint, withheld subset that was not used for training.

3 Results

Plasma was assayed from male adolescent ice hockey players: 12 concussed (13.4 ± 2.3 years of age) and 17 non-concussed (12.9 ± 1.0 years of age; P = 0.213). The estimated time from concussion occurrence to blood draw at the first clinic visit was 2.3 ± 0.7 days.

Self-reported symptom evaluation as per SCAT3 (n = 11) revealed a total symptom score and a total symptom severity of 11.6 ± 4.8 and 29.3 ± 22.8, respectively (Table 1). One concussed athlete was evaluated with the Child SCAT; total symptom score of 6 and a total symptom severity of 12 (the parent score indicated a total symptom score of 3 and a total symptom severity of 8). All non-concussed athletes were evaluated with a SCAT3 (n = 17), which revealed a total symptom score and a total symptom severity of 0.5 ± 1.5 and 0.6 ± 1.8, respectively.
Table 1

Symptom evaluation in adolescent concussion patients via SCAT3 (Guskiewicz et al. 2013)

Self-reported symptom

# of concussion patients

with symptoms (n = 11)

% of concussion patients

with symptoms

Headache

10

91

Dizziness

9

82

Pressure in head

9

82

Sensitivity to light

9

82

Don’t feel right

9

82

Difficulty concentrating

8

73

Fatigue or low energy

8

73

Sensitivity to noise

8

73

Feeling slowed down

8

73

Drowsiness

7

64

Balance problems

7

64

Trouble falling asleep

7

64

Difficulty remembering

6

55

Neck Pain

5

45

Blurred vision

4

36

Feeling like in a fog

4

36

Confusion

4

36

Irritability

3

27

Nausea or vomiting

2

18

More emotional

1

9

Sadness

1

9

Nervous or Anxious

1

9

1 patient is not shown as they had symptom evaluation via a Child SCAT3 (reported symptoms: difficulty paying attention; I daydream too much; headache; dizzie; tired a lot; and tired easily) (Glaviano et al. 2015)

Plasma was assayed for 143 metabolites by DI/LC–MS/MS (Supplementary Table 1) and 31 metabolites by 1H NMR (Supplementary Table 2). Using PCA, the leading 10 components (each containing 9 metabolites) were demonstrated to account for 82 % of the variance between concussed and non-concussed athletes (Table 2). The most striking observation was the high variance in plasma glycerophospholipids between concussed and non-concussed athletes.
Table 2

PCA identified the top 10 weighted components, each containing 9 metabolites

Component 1 (Explained variance: 28.21 %)

Component 2 (Explained variance: 14.83 %)

Component 3 (Explained variance: 9.7 %)

Component 4 (Explained variance: 6.1 %)

Component 5 (Explained variance: 5.76 %)

PC aa C36:0

C5-OH

PC aa C36:0

C14:2

Isoleucine

PC aa C36:5

lysoPC a C18:2

PC aa C36:5

C2

Leucine

PC aa C36:6

PC aa C36:0

PC aa C38:0

SM C22:3

α-Aminoadipic

PC aa C38:0

PC aa C36:5

PC aa C38:1

Arginine

Putrescine

PC aa C38:6

PC aa C36:6

PC aa C38:6

Citrulline

3-OH-butyrate

PC aa C40:6

PC aa C38:1

PC aa C40:6

Putrescine

Creatine

PC ae C38:0

PC ae C38:1

PC aa C42:0

Acetone

Isoleucine

PC ae C38:1

PC ae C38:2

PC ae C38:6

Carnitine

Leucine

PC ae C40:6

Citrulline

PC ae C40:6

Glycerol

Proline

Component 6 (Explained variance: 5.61 %)

Component 7 (Explained variance: 3.84 %)

Component 8 (Explained variance: 2.97 %)

Component 9 (Explained variance: 2.63 %)

Component 10 (Explained variance: 2.46 %)

C14:2

lysoPC a C26:0

C3

PC aa C40:3

PC ae C36:2

C18:1

lysoPC a C26:1

C5

PC ae C38:1

PC ae C36:5

C18:2

lysoPC a C28:1

PC aa C30:2

PC ae C42:4

SM (OH) C14:1

C3

PC ae C30:1

PC aa C32:2

ADMA

SM C26:0

C5

PC ae C38:1

Proline

Putrescine

trans-OH-Proline

trans-OH-Proline

Putrescine

trans-OH-Proline

3-OH-isovalerate

3-OH-isovalerate

Putrescine

Acetone

Acetone

Acetone

Methanol

3-OH-butyrate

Carnitine

Carnitine

Carnitine

Propylene glycol

3-OH-isovalerate

Succinate

Proline

Propylene glycol

Succinate

Explained variance, difference between concussed and non-concussed athletes

ADMA asymmetric dimethylarginine

The full metabolic dataset was then reduced down to two dimensions using t-SNE, as the inherent dimensionality of the data was significantly lower than the number of metabolites (Fig. 1). Following this dimensionality reduction step, a SVM was trained, with a linear kernel, to classify subjects as concussed or non-concussed. Cross validation of our classifier using a leave-one-out approach demonstrated a 92 % accuracy rate in diagnosing a concussion in adolescent ice hockey players.
Fig. 1

Individual subjects are plotted using the t-SNE dimensionality reduction step to illustrate the clustering of cohorts [12 concussed athletes—closed circles, 17 non-concussed (“control”) athletes—open circle]. The axes are dimension-less

Taking classification accuracy as our test statistic, we investigated the significance of our observed accuracy via permutation testing. We generated a null distribution by randomly shuffling class labels; training and testing a new classifier for each shuffled label set and recording the classification rate. Comparing our observed 92 % accuracy rate to a 10,000 sample null distribution in which none of the null classifiers reached a 92 % accuracy rate, we calculated a p < 0.0001.

The number of metabolites required to achieve reasonable classification accuracy were minimized using two independent statistical techniques. First, a Chi square test was used to select informative metabolites in a univariate manner, and 92 % classification accuracy was maintained with only 17 metabolites (Table 3, Column 1). Second, recursive feature elimination was then used to verify accuracy, and yielded a similar classification accuracy of 90 % with 31 metabolites (Table 3, Column 2).
Table 3

Classification accuracies of 90–92 % achieved with fewer metabolites

92 % Accuracy determined with a χ2 test

90 % Accuracy determined with recursive feature elimination

1. C5

1. C5

2. PC aa C32:1

2. PC aa C30:2

3. PC aa C32:2

3. PC aa C32:0

4. PC aa C36:5

4. PC aa C32:1

5. PC aa C36:6

5. PC aa C32:2

6. PC ae C34:0

6. PC aa C32:3

7. PC ae C34:3

7. PC aa C34:4

8. PC ae C36:0

8. PC aa C36:6

9. PC ae C36:1

9. PC aa C42:6

10. PC ae C36:2

10. PC ae C30:0

11. PC ae C38:1

11. PC ae C30:1

12. PC ae C38:2

12. PC ae C32:1

13. PC ae C38:3

13. PC ae C34:0

14. Putrescine

14. PC ae C34:2

15. Formate

15. PC ae C34:3

16. Methanol

16. PC ae C36:0

17. Succinate

17. PC ae C36:2

 

18. PC ae C38:1

 

19. PC ae C38:3

 

20. SM C22:3

 

21. SM C24:0

 

22. SM C24:1

 

23. α-Aminoadipic acid

 

24. trans-OH-Proline

 

25. Putrescine

 

26. Betaine

 

27. Formate

 

28. Glucose

 

29. Glycerol

 

30. Methanol

 

31. Serine

As a final step, concussed and non-concussed athletes were clustered by direct comparison of their metabolomic profiles. We computed the Pearson product-moment coefficient for each pair of (normalized) subject metabolic profiles to yield a correlation matrix. Clusters were optimally identified in this correlation matrix with agglomerative complete-linkage hierarchical clustering (Fig. 2), and graphically illustrated that the metabolomic profiles were dissimilar between concussed and non-concussed subjects.
Fig. 2

Agglomerative complete-linkage hierarchical clustering, yielding 3 second-level clusters: green cluster—all non-concussed (“controls”) athletes but 2; cyan cluster—all concussed athletes; and red cluster—two concussed athletes and one non-concussed (“control”) athlete). Dissimilarities are evident between concussed and non-concussed athletes

Lastly, ROC curves were generated over a fourfold cross-validation with a SVM (Fig. 3), demonstrating the performance of a binary classifier as its discrimination threshold is varied, and yielding a mean area under the curve of 0.91. When evaluating the diagnostic ability of a test to discriminate the true state of the subjects (concussed vs. non-concussed athletes), areas under the curve >0.90 are generally considered to provide excellent accuracy classification (with 1.0 being a “perfect” test and 0.5 attributed to “luck”) (Zhu et al. 2010).
Fig. 3

Receiver operating characteristic (ROC) curves for concussion classification based on metabolomics profiling (area under the curve = 0.91). The ROC curve plots the true positive rate against the false positive rate of the classifier; the top left corner of the plot is “ideal” (area under the curve = 1.00; ROC fold 0 to 2 are overlapping with only red visible) and the diagonal line would be occupied by a classifier that simply guessed randomly at labels, or “luck” (area under the curve = 0.50). Excellent classification is indicated by an area under the curve >0.90

4 Discussion

We report that plasma metabolomics profiling, together with multivariate statistical analysis and machine learning, is a novel concussion diagnostic method with a high level of accuracy (>90 %). One of the most striking patterns observed was the reliance of the model on changes in plasma glycerophospholipids, accounting for approximately 50 % of the variance between concussed and non-concussed athletes.

We specifically investigated concussion in adolescent male ice hockey players. In our region, adolescent males are at the highest risk for concussion, and most frequently concussed at ice hockey arenas (Stewart et al. 2014). Concussions in these adolescent patients are of particular concern as their brains are still developing (Halstead and Walter 2010; Toledo et al. 2012). Younger patients are also more susceptible to injury due to thinner skulls, weaker neck muscles, less myelination, greater brain water content, higher metabolic requirements and a larger subarachnoid space (Karlin 2011; Morrison et al. 2013). The time to recovery is prolonged in younger patients relative to adults (Lovell et al. 2004; Pellman et al. 2006), and brain injury may have life-long consequences for adolescents via interrupted intellectual and social development (Toledo et al. 2012). Accurate concussion diagnosis is particularly important for adolescents, as rapid deployment of treatment and rehabilitation services could be life-changing.

Concussions were identified by a mechanism of injury associated with typical concussion symptoms, and in the absence of focal/structural injury. Athletes were assessed with either the SCAT3, or the Child-SCAT3 for one 12 year old subject, as these are the recommended concussion assessment tools for these age groups (Glaviano et al. 2015; Guskiewicz et al. 2013). Based on the average number of self-reported symptoms and the symptom severity scores, the athletes suffered a mild-moderate injury. For comparison, we also report normative SCAT3 values for the non-concussed athletes.

Concussion diagnostics remains problematic, with clinical judgment as the gold standard (McCrory et al. 2013). Thus, there has been an active search for diagnostic blood biomarkers (e.g. GFAP, Tau, NFL), but no single biomarker or biomarker panel has been identified for widespread use, likely due to inadequate sensitivity and/or specificity (Jeter et al. 2013). A single biomarker or a small number of biomarkers may not accurately reflect patient and injury heterogeneity (Papa et al. 2015). Metabolomics profiling with as few as 17 metabolites were required for classification accuracy and may be useful for developing a future concussion diagnostic test.

Conventional statistics are model-driven; they are based on the assumption that there are a relatively small number of important variables and that careful variable selection is the key to good model performance. This approach provides important clinical information on populations, but is significantly limited for understanding injury/disease in individuals. A supplement to conventional statistics is machine learning, which allows the data to create the model by detecting underlying patterns (Shouval et al. 2014). Metabolomics is ideally suited for machine learning techniques, as the final performance of the model relies on how much information each dataset contains.

Metabolomics profiling requires analyses of all detected metabolites simultaneously, with PCA analysis techniques used most commonly (Bujak et al. 2014). Unlike PCA which enforces a brittle, orthogonal, linear refactorization of the data, t-SNE assumes that the ‘optimal’ representation of the data lies on a manifold with complex geometry, but low dimension, embedded in the full dimensional space of the raw data (van der Maaten and Hinton 2008).

Using the aforementioned analytics, we determined that the variance in metabolites between concussed and non-concussed athletes was most pronounced for the glycerophospholipids. Glycerophospholipids are dynamic molecules, which turn over at different rates depending on their structure, composition and localization in cellular membranes (Farooqui et al. 2000).

The human brain is nearly 60 percent lipid by dry weight; with the highest concentrations of lipid in myelin (78–81 %), followed by white matter (49–66 %) and grey matter (36–40 %) (O’Brien and Sampson 1965). The primary lipid classes in brain, as a percentage of dry weight, include glycerophospholipids (20–32 %), sphingolipids (4–29 %), cholesterol (7–22 %) and sphingomyelin (1–5 %). The concentration of each lipid class is variable with age and between white matter, grey matter and myelin. The lipid composition in white (glia and neuronal axons) and gray matter (neuronal soma) is driven primarily by their structural roles. The total glycerophospholipid content, as a percentage of dry weight, in white (20–26 %; glia and neuronal axons) matter is only slightly greater than in gray matter (20–23 %; neuronal soma).

Once the number of metabolites was reduced, but still maintained high classification accuracy, the most informative glycerophospholipids were the choline plasmalogens (e.g., PCaeC34:0, PCaeC34:3, PCaeC36:0, PCaeC36:1, PCaeC36:2, PCaeC38:1, PCaeC38:2 and PCaeC38:3). Plasmalogens are present in significant amounts in myelin, with >70 % of myelin glycerophospholipids being plasmalogens (Braverman and Moser 2012). Plasmalogens contribute to membrane structure, act as membrane antioxidants and are a source of second messenger molecules.

The acylcarnitine C5 also had a prominent role in classification accuracy. C5 is produced primarily during the catabolism of the branched chain amino acid leucine. Leucine is rapidly transported into the brain, and it is important for the synthesis of the excitatory neurotransmitter glutamate, as well as energy metabolism, fatty acid transport and mitochondrial fatty acid oxidation, ketosis, oxidative stress and mitochondrial membrane damage (Yudkoff 1997). The differences in C5 could reflect different activity levels, but previous literature has suggested that medium chain acylcarnitines (C8–C12) were the dominating plasma biomarkers of moderate intensity exercise (Lehmann et al. 2010).

Other metabolites of importance for accurate classification include putrescine, methanol, formate and succinate, which are implicated in diverse biochemical reactions including protein breakdown and energy metabolism. The metabolites identified may reflect secondary consequences to the primary concussive injury or themselves may have common secondary metabolic impacts.

Our study has several limitations. First, our study evaluated only a small number of adolescent athletes. However, to identify a novel diagnostic method with such a small number illustrates the power of metabolomics profiling. Second, we are unclear of the origin of the metabolite changes measured. Further studies should investigate plasma together with cerebrospinal fluid (consenting adults), or animal TBI models where brain tissue can also be harvested and examined. Third, it is unclear which metabolite changes identify injured cells, pathways and/or metabolic processes. Fourth, our concussed athletes were otherwise healthy adolescent males and the study results may not be generalizable to other populations, including females, older athletes or athletes participating in non-contact sports. Additional control groups should be added in follow up studies (i.e. orthopedic trauma without concussion). Future metabolomics profiling should account for key factors such as age, sex, diet, body mass index, type of sport, comorbidities and any disease therapies or interventions. Also, variations in brain maturation could be partially controlled by normalizing metabolite values to blood 24-OH cholesterol, a potential marker of brain development (Lutjohann et al. 1996; Vitali et al. 2014). Finally, the self-reporting of symptoms is complicated by the subjective nature of the assessment, and athletes typically underreport the symptoms (Lovell and Solomon 2013; Meier et al. 2015). Sub-clinical brain injuries would not have been accurately represented.

5 Conclusion

Using plasma metabolomics profiling, together with multivariate statistical analysis and machine learning, we identified concussed individuals with >90 % certainty. Of the two analytic techniques used, 1H NMR and DI-LC/MS/MS, the metabolites measured with MS offer greater predictive ability (i.e. glycerophospholipids and C5). Metabolomics profiling represents a novel diagnostic method for concussion, and may be amenable to point-of-care MS testing in the near future.

Acknowledgments

We thank Ms. Christy Barreira and Ms. Sandra Shaw for excellent technical support, and Ms. Kathryn Manning and Mr. Kevin Blackney for assistance with data. We graciously acknowledge analytic support from The Metabolomics Innovation Centre at the University of Alberta, Edmonton, AB (Ms. Rupasri Mandal, Ms. Jennifer D. Reid and Dr. David Wishart). This study was supported by the Children’s Health Foundation (http://childhealth.ca/) grant to DDF.

Compliance with ethical standards

Conflict of interest

The authors have filed a patent application for metabolomics profiling of central nervous system injury (US Trade and Patent Office No. 62/135886).

Ethical approval

This study was approved by the Human Research Ethics Board at Western University (#103365).

Informed consent

Written informed consent was obtained from the legal guardians and assent was obtained from adolescent subjects.

Supplementary material

11306_2016_1131_MOESM1_ESM.pdf (28 kb)
Supplementary material 1 (PDF 28 kb)
11306_2016_1131_MOESM2_ESM.pdf (7 kb)
Supplementary material 2 (PDF 6 kb)

Funding information

Funder NameGrant NumberFunding Note
Children's Health Foundation

    Copyright information

    © Springer Science+Business Media New York 2016

    Authors and Affiliations

    • Mark Daley
      • 1
    • Greg Dekaban
      • 2
    • Robert Bartha
      • 3
    • Arthur Brown
      • 4
    • Tanya Charyk Stewart
      • 5
    • Timothy Doherty
      • 6
    • Lisa Fischer
      • 7
    • Jeff Holmes
      • 8
    • Ravi S. Menon
      • 3
    • C. Anthony Rupar
      • 9
    • J. Kevin Shoemaker
      • 10
    • Douglas D. Fraser
      • 4
      • 11
      • 12
    1. 1.Department of Computer ScienceWestern UniversityLondonCanada
    2. 2.Molecular MedicineRobarts Research Institute, Western UniversityLondonCanada
    3. 3.Department of Medical BiophysicsWestern UniversityLondonCanada
    4. 4.Department of Physiology and PharmacologyWestern UniversityLondonCanada
    5. 5.Department of SurgeryWestern UniversityLondonCanada
    6. 6.Department of Physical Medicine and RehabilitationWestern UniversityLondonCanada
    7. 7.Department of Family MedicineWestern UniversityLondonCanada
    8. 8.Department of Occupational TherapyWestern UniversityLondonCanada
    9. 9.Departments of Biochemistry and Pathology and Laboratory MedicineWestern UniversityLondonCanada
    10. 10.School of KinesiologyWestern UniversityLondonCanada
    11. 11.Departments of Pediatrics and Clinical Neurological SciencesWestern UniversityLondonCanada
    12. 12.Paediatric Critical Care Medicine, Children’s Hospital, London Health Sciences CentreWestern UniversityLondonCanada

    Personalised recommendations