Key words mild traumatic brain injury - EEG - phase locking value
Introduction
Traumatic brain injury (TBI) is defined as “an alteration in brain function
or other evidence of brain pathology, caused by an external force [1 ].” It may be caused by accidents and
violence, and accounts for one-third of all injury-related deaths in the United
States. About 3% of the population experience chronic disability related to
TBI. It is associated with a broad spectrum of symptoms that unfold over hours to
months and, in some cases, indefinitely. Mild TBI (mTBI) patients constitute a large
majority of all recorded cases of TBI [2 ]
[3 ]
[4 ]. Although most mTBI
patients recover, a substantial minority (7–33%) develop persistent
disabilities (post-concussive syndrome or PCS) in the form of somatic (headaches,
dizziness), cognitive (attention and memory impairment), and emotional
(irritability, depression) problems [5 ]. The diagnosis
of mTBI is difficult and exacerbated by the fact that individual patients experience
different subsets of clinical symptoms. Cumulative effects of multiple concussions
demonstrate that latent dysfunction lingers even in patients whose symptoms have
resolved. Reports suggest that 11–23% of deployed service members
have mTBI [6 ].
Brain imaging technologies have been used to assess mTBI, and because computed
tomography (CT) is practical and ubiquitous in emergency medicine, CT is commonly
the first neuroimaging procedure to be used for the mTBI patient. However, CT
findings or other structural imaging have been poor predictors of long-term
sequelae, perhaps because mTBI is not associated with bleeding or other gross
trauma, unlike more severe brain injuries [7 ]
[8 ]. This is partly due to the fact that mTBI injuries
are primarily microscopic and diffuse and most mTBI patients have normal CT and
structural magnetic resonance imaging (MRI) scans [9 ]
[10 ]. Diffusion tensor imaging (DTI),
another structural assessment, could not discriminate between normal and mTBI
patients even when the patients had post-concussive symptoms such as verbal memory
deficits [11 ]. Although more recent DTI efforts have
had some success in distinguishing mTBI and control patients, mTBI appears to impact
brain function to a greater extent than it impacts brain structure [12 ].
Concussion is common and physicians face serious challenges in the management and
prognostication of mTBI. In the US, emergency departments (EDs) receive 1.2 million
patients with TBI, accounting for 1.2% of all ED visits, with 38% of
these patients being discharged home without specific recommendations for follow-up
[2 ]
[13 ]
[14 ]. In a longitudinal study of mTBI patients
presenting to the ED, 82% of patients reported at least one subsequent
post-concussive symptom and>40% had significantly reduced
Satisfaction with Life scores post-injury, indicating mTBI produces significant
disability [4 ].
Sports-related concussion is one of the major causes of TBI and increasing media
coverage of professional athletes whose careers were ended by brain injury and
recent relevant changes to the National Football League policy reflect only one
aspect of the overall societal impact of mTBI [15 ]
[16 ]
[17 ].
High school football players had alterations in default mode network (DMN) activation
measured by fMRI functional covariance as compared to non-contact sports athletes
[18 ]. Significant reductions in DMN activation of
football players were shown at 7 days after injury [19 ]. Significant cerebral glucose metabolism changes were found in former
National Football League players who had suffered repetitive sub-concussive injury
as well as concussions during their careers [20 ]. Many
athletes with no single mTBI event may accumulate thousands of sub-concussive
incidents eventually resulting in neurological impairment [21 ]. In athletes with sports-related injuries, EEG abnormalities
extracted by analysis persisted after the post-concussive symptoms had resolved,
suggesting that EEG can be useful to detect and manage mTBI [22 ]
[23 ].
Could EEG be used to manage mTBI in athletes? For EEG to be practical in managing
sports concussion, it should be easy to record high-quality, clinical-grade EEG at
sports facilities, and quantitative measures from the EEG should be sufficient to
accurately distinguish between injury and non-injury EEGs. Resting-state EEG
requires that the subject sit or lie still, making it a promising methodology for
determining the presence and severity of mTBI. Resting-state recordings are easy for
the subjects as well as for the EEG operator to collect, because no motor,
cognitive, or other neurological tests are administered. In the present study, EEGs
were recorded at a university sports facility by athletic trainers so we could
assess whether it is practical and convenient to collect high-quality EEGs from
athletes. We used commercially available systems, microEEG, a portable, wireless
battery-operated amplifier device as well as StatNet, a rapid-to-apply, single-use
disposable EEG electrode cap. A resting-state EEG was collected from college
athletes at baseline (BL) before the sports season, after a concussive injury (IN),
and post-season (PS). Analysis of the recordings by extracting quantitative measures
of local and inter-area neural activity synchronization and binary
injury/non-injury classification demonstrates that the EEGs could be
discriminated with 81% accuracy. By examining the individual quantitative
features extracted from the EEG, we find that inter-hemispheric phase
synchronization is significantly lower in the Injury group.
Materials and Methods
Participants, data collection, and study design
A total of 15 female and 15 male subjects participated; each was a member of a
collegiate sports team at Massachusetts Institute of Technology (MIT) and
consented to participate according to IRB and NIH guidelines. In general, the
study was conducted in conformity with the declaration of Helsinki and Good
Clinical Practice [24 ]. Resting-state EEG was
recorded in the sports facility by athletic trainers who had previously learned
to use the equipment and administer the EEG during a 5-hour training session.
The training duration for the athletic trainers was up to five hours, which
included the vetting process of their EEG readouts prior to authorizing them to
record EEG signals from the athletes.
A minimum of 15-min resting-state EEG with eyes closed was recorded while the
subject sat in a chair. EEGs were recorded using StatNet, a single-use
disposable system of 16 electrodes that takes less than 5 minutes to
apply to the subject [25 ]. The StatNet was
connected to microEEG, a portable, battery-operated, wireless, digital EEG
recording system [26 ]
[27 ]. The data were stored on a hard disk for offline clinical review
by a board-certified neurologist, as well as separate quantitative analysis
using digital signal processing techniques that are detailed next.
Data analysis
The digital signal processing and calculations described in this paper used
Matlab v.9.3.0.713579 (The MathWorks, Inc., Natick, Massachusetts, United
States).
Pre-processing
The raw EEG was first pre-processed to minimize or eliminate signal artefacts
originating from outside the brain. Initially the recordings were visually
inspected in the frequency and time domains. The power spectra of all the
recordings contained an alpha rhythm peak (8–12 Hz), especially
in the parietal and occipital areas, and the amounts of relative power in any
frequency bands contained no gross inconsistency relative to expectations [27 ]. The 0.16–70 Hz
bandpass-filtered voltage time-series were inspected for the presence of
artefacts such as eye blinks, eye movements, electromyogram or muscle activity,
motion effects, and electrocardiogram. EEG time segments contaminated with such
patterns constituted < 10% of any single recording.
Consequently, no recordings were excluded from the study.
The effects of artefacts on subsequent quantitative analyses were minimized and
signal-to-noise was maximized by an artefact-rejection data pre-processing
pipeline that included an independent component analysis (ICA)-based artefact
rejection scheme [28 ]
[29 ]. Bandpass filtering (0.16 Hz to 40 Hz) reduced
slow drifts and high frequency artefacts with a zero phase-shift Hamming
windowed-sinc FIR filter (EEGLAB function ‘pop_eegfiltnew’).
Then artefacts were identified and removed objectively using the ADJUST method
[30 ]. The filtered continuous EEG data first
undergoes ICA decomposition with an Extended-Infomax algorithm [31 ]. ADJUST then detects and removes the
artefact-independent components associated with eye blinks, horizontal and
vertical eye movements, and generic discontinuities. It then reconstructs a
cleaned version of the original data. [Fig. 1 ]
shows a fragment of multichannel EEG data with an eye blink artefact before and
after removing the artefact by the ADJUST method.
Fig. 1 The effect of artefact-detection preprocessing. A fragment
of multichannel EEG data contaminated by blink artefact (a )
before and (b ) after the data are cleaned by the ADJUST method.
The highlighted segments of the data show the effect of blinking on EEG
channels.
Feature extraction
Both univariate (based on data from a single sensor) and multivariate (multiple
sensors) features were used to develop methods for objectively discriminating
between groups of EEGs. Pre-processed EEG was first divided into a set of
adjacent, non-overlapping feature time windows of fixed size ΔT
For each window, we computed the univariate frequency band-power (FBP),
multivariate phase locking value (PLV), and univariate modulation index (MI) to
estimate phase-amplitude coupling (PAC). These variables provide metrics to
decode mental states as well as distinguish between normal and other brain
activity patterns [32 ]
[33 ]
[34 ].
FBP : The frequency band power is calculated by integrating the power
spectral density of EEG (within a feature time window) under a particular
frequency range, then dividing it by the total power across the spectra. The
magnitude of FBP estimates the local spatio-temporal synchronization of
extracellular inhibitory and excitatory currents [35 ]
[36 ]. Our frequency ranges were
chosen in 4 Hz increments from 0 to 32 Hz, namely delta
(0–4 Hz), theta (4–8 Hz), alpha
(8–12 Hz), beta1 (12–16 Hz), beta2
(16–20 Hz), beta3 (20–24 Hz), beta4
(24–28 Hz), beta5 (24–28), and low gamma
(28–32 Hz). The upper bound was chosen to be low gamma since
cerebral activity contributes negligibly to scalp EEG beyond this range [37 ].
PLV: The phase locking value (PLV) measures the phase synchronization
between the narrow-band filtered EEG recorded from a pair of distinct
electrodes. Accordingly, PLV quantifies long-range, frequency-specific,
amplitude-independent phase synchronization of EEG oscillations between brain
areas to estimate inter-area neural integration [38 ]. We computed PLV after filtering the EEG in 2-Hz wide bands that
were centered at 4, 10, 20, and 40 Hz, which allows for precise
estimates of phase [32 ]. The signal pairs were
selected to be intra-hemispheric (F7-T5, F3-P3, F4-P4, F8-T6, FP1-O1, FP2-O2),
inter-hemispheric symmetric (FP1-FP2, F7-F8, F3-F4, T3-T4, T5-T6, O1-O2), and
inter-hemispheric asymmetric (FP1-O2, FP2-O1, F7-T6, F8-T5). PLV values of 1
indicate that the phase difference between the two signals is constant, whereas
PLV values of zero indicate that the phase difference is uniformly distributed
between 0 and 2π [33 ]
[39 ]. To calculate the PLV, the pair of EEG signals
were bandpass-filtered at a given frequency within a 2 Hz wide band,
their phases extracted via the Hilbert transform, the phase difference,
Δθ , between the two signals was computed at every
data point, and the PLV for a particular window was found as the absolute value
of the mean of exp{i
θ } over the window [40 ]. (Note i=√-1
PAC: The phase-amplitude co-modulation estimates phase-amplitude coupling
between the phase of a low frequency oscillation in the EEG and the amplitude of
a high frequency oscillation, providing an estimate of local, multi-frequency
organization of neuronal activity [41 ]
[42 ]
[43 ].
Following previous work, we chose the low frequency ranges theta and alpha, and
high frequency ranges 15–25 Hz and 30–40 Hz. PAC
was computed using the Kullback-Leibler divergence between the uniform
distribution and the phase-amplitude distribution obtained from
bandpass-filtered signals [44 ]
[45 ].
Classification and validation
We used support vector machine (SVM) binary classification to make inferences
about whether particular recordings belonged to the subgroups designated as
Baseline (BL), Injury (IN), or Post-Season (PS). Since the smallest data set
(IN) contained 12 recordings, we randomly selected 12 recordings from each of
the BL and PS sets to have an equal number of recordings from each class. This
ensured that the chance value for correct classification would be 0.5. After
feature extraction the data were in the form of a matrix whose columns are
different features and rows are time windows (or an observation). Every
observation in our experiments has an associated label, BL, IN, or PS, thus the
problem was suitable for analysis by supervised machine learning (ML).
Supervised ML approaches in general, and SVM in particular, take advantage of
distributed and subtle patterns of activation that are otherwise possibly
undetectable, to allow inference at the level of the individual subject and
specific time window, rather than at the group level. SVM classifies data points
by maximizing the margin between classes in a high-dimensional feature space
(for a review and references [46 ]. We used SVM for
pairwise binary classification.
For testing we used leave-one-out cross-validation (LOOCV). This was implemented
by leaving out one observation for training the SVM, which was then tested on
the left-out observation. This process was repeated for each observation. The
Matlab functions svmtrain, svmclassify, and cvpartition were used to implement
the above calculations. To evaluate performance, accuracy was determined as the
fraction of observations that could be correctly identified in the test
data.
Statistical analysis
We used statistical tests to determine the significance of our results.
Specifically, we used a binomial test which calculates the p-value based on the
total number of observations in each class and the chance level of accuracy
[33 ]
[47 ]. To
test the significance of phase locking values (PLV), we created a null
distribution assuming random phase differences. The p-value of PLV was
calculated for each subject and time window from the null distribution. We then
considered a global null hypothesis that the PLV was no better than chance for
any subject and time window. We used the Bonferroni procedure to control for
multiple comparisons and determined significance if p < alpha/k,
where alpha was set to 0.05 and k was the total number of tests.
We chose the relatively conservative, non-parametric Kolmogorov-Smirnov (KS) test
to evaluate this null hypothesis.
Results
Forty-seven EEGs were recorded from 30 distinct subjects. All EEGs were evaluated by
a board-certified neurologist EEG reader and all were assessed to be technically
valid and interpreted as normal, including 14 post-season EEGs, and 12 injury EEGs
from 9 subjects. Three subjects contributed two injury EEGs each, and because the
intervals between the two EEGs were 6, 26, and 56 days, they were treated as
independent samples. None of the injury subjects had provided baseline EEGs and none
provided post-season EEGs. The median duration of the EEG recordings was
21.3 minutes, with a minimum duration of 19.6 minutes and with two
recordings lasting an hour or longer. It is therefore demonstrated that with the
StatNet and microEEG systems, it is feasible for athletic trainers to record
clinically valid resting-state EEGs from student athletes in the sports
facility.
We then used machine learning to evaluate whether there were differences between the
three classes of EEG. [Fig. 2 ] shows the values of
accuracy of binary discrimination between pairs of EEG classes for feature windows
of fixed size ΔT =120 s (shorter
ΔT =60 s and longer
ΔT =240 s feature windows did not lead to different
conclusions). [Fig. 2 ] shows the classification
results after partitioning the data in different ways. When as much of the available
data are used with the limit in each class set by the class with the lowest number
of EEGs (IN=12), we find that BL and IN could be discriminated with
81% accuracy using the FBP, PLV, and PAC EEG features together, even though
discrimination based on any one feature was no better than chance. This suggests
there may be a potential for additive effects, at least with small datasets such as
this ([Fig. 2a ]). Selecting only the highest quality
EEGs (N=8 for each class) did not change the result, but now FBP and PLV
measures alone were able to discriminate BL and IN better than chance, in addition
to the three measures together ([Fig. 2b ]). Limiting
one injury EEG to a subject reduced the accuracy of SVM discrimination to chance
([Fig. 2c ]; N=10), unless the data set
was limited to the highest quality EEGs, in which case eliminating the redundant
EEGs had no effect ([Fig. 2d ]; N=8). Given
that BL and PS EEGs could not be discriminated using the three EEG measures together
(discrimination using one measure but not all three are likely the result of
overfitting and limited data) and BL and IN EEG can be discriminated, one might
expect that IN and PS EEGs to also be discriminable. However as shown in [Fig. 2 ], this was not reliably the case. Although the
pattern of results presented [Fig. 2 ] demonstrates
the possibility of successfully discriminating BL and IN EEGs, the inability to
discriminate PS and IN EEGs, as well as BL and PS, indicates there are non-linear
complexities in using the SVM approach, likely due to the small data set, and
sensitivity to artefacts in the EEGs.
Fig. 2 Accuracy of automated binary classification of Baseline (BL)
vs. Injury (IN), Baseline vs. Post-Season (PS), and PS vs. IN, with feature
window size ΔT =120 s. The horizontal shaded
areas in the background indicate 95% of the null density of
accuracy. (*p<0.05.) (a) All available EEGs in class
IN were used regardless of quality so that there were 12 recordings in every
class. (b) EEGs were selected for quality so that there were 8
recordings in each class. Multiple recordings from subjects were allowed in
(A) and (B); there were 2 such recordings, both in class IN. Decreasing
quality was measured in terms of the number of segments that had been
removed due to artefacts (maximum of 60 cuts being used as the quality
threshold). (c) All available EEGs in class IN were used regardless
of quality, but multiple recordings per subject were disallowed, so that
there were 10 recordings in every class. (d) EEGs were selected for
quality and multiple recordings per subject were disallowed, so that there
were 8 recordings in each class.
Having examined the classification accuracy of various feature types, we investigated
the spatiotemporal patterns in the values of the various types of features in the
different EEG classes BL, IN, and PS.
[Fig. 2 ] suggested that FBP may differ significantly
between the baseline and injured subjects. [Fig. 3 ]
compares the topographic distribution of this feature for the delta to alpha
frequency ranges. The figure indicates clear differences in the distribution of
delta, theta, and alpha power between the BL and IN groups. Meanwhile the BL and PS
groups appear to differ, to a smaller extent, only in alpha power distribution.
Overall, the IN EEGs appear to express reduced frontal delta and increased frontal
theta, providing a basis for SVM discriminations.
Fig. 3 Topographic distribution of frequency band power
ΔT=60 s. Electrode locations are indicated by small
grey dots. The values shown (color bar at right) represent averages over the
entire set of recordings in the class BL (left column), IN (middle), or PS
(right).
To further investigate this hypothesis, we considered the channel-averaged FBP in the
different frequency ranges in [Fig. 4a ]. Overall
delta, theta, and alpha power do not appear different in the three groups of EEG,
not only is the inter-subject variability large, but as can be seen in [Fig. 4 ], the abnormal power is not homogeneously
distributed across the scalp and thus not measurable as a spatial average. We
nonetheless examined whether combinations of power in various frequency bands could
provide a succinct discriminator of non-injured and injured EEGs. To visualize this
possibility, we show in [Fig. 4b ] the recordings as a
scatter plot in the plane of Theta/Gamma (x-axis) and Alpha/Delta
power (y-axis). The figure indicates that injured subjects may have greater values
of these two composite variables, relative to the BL and PS recordings, which appear
to be mostly lumped together close to the origin. However, although the regions
occupied by the IN versus other classes of recordings are somewhat distinct, they
are not disjointed and there is substantial overlap.
Fig. 4 Indicators derived from time averaged band power in standard
ranges of frequency. a : Error bars indicate the standard deviation of
intersubject variability within the same class (BL, IN, or PS). b :
Time averaged FBP shown in the plane of Theta/Gamma (x-axis) and Alpha/Delta
power (y-axis). Each point represents a distinct recording in the class BL
(circle), IN (triangle), and PS (square).
[Fig. 2 ] highlighted PLV as an important discriminator
of IN, suggesting that frequency-specific inter-area synchrony may be altered in
injured subjects. We now investigate the specific frequencies and pairs of sites
that may differentially contribute to this result. [Fig.
2 ] shows PLV calculated for 4 Hz (top row), 10 Hz
(2nd row), 20 Hz (3rd row), and 30 Hz
(bottom row). The error bars show the standard deviation of intersubject
variability. Values of PLV differing significantly between any pair of groups are
indicated by an asterisk above the pair of bars (*p<0.05.). The
figure clearly flags mostly frontal inter-hemispheric connectivity as an important
discriminator of the IN state, since most of the significantly differing PLV values
occur in pairs that symmetrically connect the frontal hemispheres (FP1-FP2, F7-F8,
and T3-T4).
Additional details are provided in [Table 1 ], which
displays the PLV at multiple frequencies and locations. The value of the PLV is
provided for the Baseline, Injury, and Post-Season states. Each one of these PLVs is
statistically significant in accordance with the Bonferroni corrected procedure
described in the Methods section. To the right of the three PLV columns, an
additional three columns show whether the differences in PLV between IN and BL, PS
and IN, and PS and BL were statistically significant. The table includes only
frequencies and locations for which at least one of these differences was
significant. Significance was calculated by a non-parametric method explained in the
Methods section. It is indicated by an asterisk in the table. The differences are
also color-coded, so that red indicates a positive and blue a negative difference.
For example, if the PLV for injured recordings was lower than baseline, then the
table entry under IN-BL is blue.
Table 1 Phase locking value (PLV) for the Baseline (BL),
Injury (IN), and Post-Season (PS) averaged over subjects.
Frequency (Hz)
Electrode Pairs
PLV
PLV Difference
Baseline (BL)
Injury (IN)
Post-Season (PS)
IN-BL
PS-IN
PS-BL
4
FP1-FP2
0.758
0.66
0.614
*
*
F7-F8
0.854
0.754
0.863
*
*
T3-T4
0.578
0.485
0.586
*
10
FP1-FP2
0.821
0.795
0.728
*
*
F7-F8
0.81
0.715
0.823
*
*
15
FP1-O1
0.275
0.24
0.279
*
FP1-FP2
0.757
0.628
0.595
*
F7-F8
0.789
0.685
0.811
*
FP1-O2
0.223
0.195
0.251
*
20
FP1-FP2
0.732
0.619
0.567
*
F7-F8
0.791
0.724
0.815
*
25
FP1-FP2
0.712
0.562
0.513
*
F7-F8
0.79
0.732
0.816
*
30
F8-T6
0.325
0.375
0.365
*
FP1-FP2
0.703
0.498
0.45
*
F7-F8
0.787
0.716
0.811
*
F8-T5
0.271
0.36
0.328
*
35
F8-T6
0.289
0.329
0.325
*
*
FP1-FP2
0.69
0.452
0.403
*
F7-F8
0.779
0.701
0.802
*
*
F8-T5
0.235
0.322
0.291
*
40
FP1-FP2
0.687
0.448
0.407
*
F7-F8
0.777
0.706
0.804
*
F8-T5
0.232
0.331
0.294
*
The table shows that PLV of injured subjects was generally lower than baseline in the
lower frequencies (up to low gamma range) in some frontal inter-hemispheric
electrode pairs (blue squares in column titled IN-BL). They were also lower than the
PLV measured post-season (red squares in column titled PS-IN). In addition, the
table indicates that in these frequencies and pairs of sites, post-season PLV was
lower than baseline (blue squares in column titled PS-BL).
[Table 1 ] Phase locking value (PLV) for the Baseline
(BL), Injury (IN), and Post-Season (PS) averaged over subjects, and the statistical
significance (*p>0.05) and positive/negative
(red/blue) in difference in the PLV of two states. The most available
quality-selected data (N= 8, as in [Fig.
2b ]) were used for these comparisons.
To visualize the changes in functional connectivity implied by the values in [Table 1 ] and [Fig. 5 ],
we plot each statistically significant PLV difference as a line that connects the
corresponding pair of electrodes in [Fig. 6 ]. The
figure shows that PLV changes are associated mostly with the inter-hemispheric pair
F7-F8 in the delta, alpha, and beta frequencies, while they are also associated with
intra-hemispheric connections centering on F8 in the low-gamma frequencies.
Fig. 5 PLV at multiple frequencies and pairs of electrodes. The error
bars indicate the standard deviation of intersubject variability.
(*p<0.05)
Fig. 6 The difference in PLV between subject states visualized across
multiple pairs of topographic sites and frequencies. The thickness of the
line connecting a pair of electrodes is linearly proportional to the
absolute value of the PLV change. As in [Table
1 ], positive (negative) differences are shown in red (blue). Only
PLV changes that are significantly above chance are shown.
Discussion and conclusions
Discussion and conclusions
This paper addressed the diagnosis of mTBI, a public health concern affecting a very
large population including children, athletes, and military personnel. The study was
designed as a proof-of-principle feasibility trial, importantly without an
independent control group. Nonetheless, this design yielded results demonstrating
that it is feasible for athletic trainers at a sports facility to record clinically
valid, high-quality resting-state EEGs from student athletes. The findings also show
that spatially-localized metrics of EEG synchrony can discriminate mTBI-associated
EEGs from control EEGs. These findings provide proof-of-concept evidence that
resting-state EEG is a practical, non-invasive measurement technique that can be
implemented by minimally trained personnel to help diagnose mTBI. Performing on-site
head trauma screening or assessment on the sidelines is preferred for timely
evaluation and decision-making. This approach is particularly important for mild
injuries where symptoms may not be immediately apparent and there is a risk of
second impact syndrome.
There were differences among the Baseline, Injury, and Post-Season EEGs in the
frequency band power features, which estimates the degree of synchronized synaptic
currents local to the electrode site, effects of volume conduction notwithstanding.
Reduced delta and increased theta at frontal sites characterized injury EEGs, with
only a weak tendency to increased alpha at posterior sites ([Fig. 3 ]). With respect to the Baseline EEGs, the
Injury EEGs had higher theta and lower delta power. Interestingly, relative to
Baseline EEGs, we also found that the Post-Season EEGs had frequency-band- and
location-specific changes that were in the opposite direction as the
injury-associated EEGs. Such alterations were sufficient for 25% above
chance accuracy of classifying Baseline and Injury EEGs ([Fig. 2 ]). While the observed power differences were relatively specific,
they did not differentiate between the EEGs of the Baseline and Post-Season groups.
This is not unexpected, as both groups were comprised of athletes without head
injuries and were anticipated to have comparable features. Furthermore, these power
differences were also insufficient to distinguish between the EEGs of the Injury and
Post-Season groups.
In contrast, we found no systematic EEG class differences in estimates of PAC
computed locally, i. e., at single electrode sites. It is worth noting that
inter-area PAC based on resting-state magnetoencephalography of mTBI patients was
found to differ between mTBI patients and controls, suggesting that mTBI might
affect measures of inter-areal neural synchronization, which depend on the
coordination of long-range conduction velocities and delays that are mediated in
part by the integrity of myelination [48 ].
Consistent with inter-areal disturbances in mTBI, prominent differences among the
three EEG classes were found in the phase locking index that estimates the degree of
frequency-specific phase synchronization between distinct brain areas. Although we
calculated PLV for multiple pairs of sites, the statistically significant
injury-associated alterations were found mostly in inter-hemispheric symmetric sites
at frontal electrodes and mid-line ([Fig. 6 ]). These
are populations located on either side of and equally distant from the midline
sagittal plane, such as the left and right orbitofrontal sites FP1 and FP2. The
inter-group changes were particularly strong at 4 and 10 Hz, which are EEG
oscillation frequencies known to have non-local biophysical origins, unlike gamma,
which is generated locally [49 ]. As in the case of
FBP, these differences are sufficient for a 25% greater than chance accuracy
of classifying Baseline and Injury EEGs ([Fig.
2 ]).
Thus, the present findings are consistent with reduced inter-hemispheric synchrony
that can be attributed to changes in white matter microstructure, including weakened
alpha frequency inter-hemispheric synchrony that can be detected by resting state
magnetoencephalography of mTBI patients [34 ]
[50 ]
[51 ]
[52 ].
mTBI patients often suffer from problems in attention, memory, and executive
function, all high-level functions enabled by multiple brain regions integrated by
long-range connections [53 ]. Electrographic findings
from mTBI patients include slowing and reduced alpha activity and such findings
within the first 24 h have been associated with worse recovery, however
traditional EEG interpretation needs to be enhanced by quantitative analysis to
reveal changes to functional connectivity, as suggested by the present findings
[54 ].
Thus, our findings, though preliminary, are nonetheless consistent with the reduced
inter-hemispheric functional connectivity that was reported in a fMRI study of
athletes with mTBI at 10 days after injury [50 ].
Resting-state fMRI of mTBI patients also showed reduced connectivity in the default
mode network pattern in the days and weeks following injury [49 ]
[55 ]
[56 ].
Despite its advantages in superior spatial resolution, the expense and lack of
portability make fMRI largely unsuitable for many applications, whereas as
demonstrated here, resting-state EEG with portable and easy-to-deploy equipment may
be convenient, feasible, and similarly effective to screen for mTBI and for
evaluating return-to play decisions.
Our results may form the basis for a rapid and practical method to diagnose mTBI in
the clinic as well as in the field but additional, controlled studies with
independent injury and control groups will be important for validating this
possibility. Better diagnosis of mTBI would not only help in the management of
athletes but also help increase the efficiency of clinical trials through methods
suitable for identifying a cohort of mTBI patients at higher risk of developing
long-term problems [57 ]. The study represents a step
towards increasing access to brain injury care and reducing inequity in various
settings, including the military, and athletic facilities. This approach has the
potential to contribute to the development of more effective and accessible tools
for the detection of mild traumatic brain injuries, with broad implications for the
improvement of functional outcomes and quality of life of those affected.
The ability to identify patients at high risk of long-term problems would help to
appropriately channel scarce clinical resources such as specialty follow-up and
rehabilitation. Clinical trials are likely to be enriched if they recruit patients
with a high probability of brain dysfunctions post-injury.
