Identification of Concussion Subtypes Based on Intrinsic Brain Activity

Key Points

Question

Can brain activity–based subtypes of concussion be objectively identified at the time of injury?

Findings

This cohort study of 771 participants with clinically diagnosed concussion found 5 distinct patterns of intrinsic brain activity at the time of injury, suggesting that these subtypes highlight brain activity alterations compatible with phenotypes of concussion neuropathology.

Meaning

The identification of brain activity–based subtypes of concussion has the potential to advance the understanding of the physiological heterogeneity of concussion, aid in more personalized early diagnosis of concussion and the determination of prognosis, and contribute to optimization of the care path.

This cohort study investigates the presence of intrinsic brain activity–based concussion subtypes, defined as distinct resting state quantitative electroencephalography profiles, at time of injury.

Abstract

Importance

The identification of brain activity–based concussion subtypes at time of injury has the potential to advance the understanding of concussion pathophysiology and to optimize treatment planning and outcomes.

Objective

To investigate the presence of intrinsic brain activity–based concussion subtypes, defined as distinct resting state quantitative electroencephalography (qEEG) profiles, at the time of injury.

Design, Setting, and Participants

In this retrospective, multicenter (9 US universities and high schools and 4 US clinical sites) cohort study, participants aged 13 to 70 years with mild head injuries were included in longitudinal cohort studies from 2017 to 2022. Patients had a clinical diagnosis of concussion and were restrained from activity by site guidelines for more than 5 days, with an initial Glasgow Coma Scale score of 14 to 15. Participants were excluded for known neurological disease or history of traumatic brain injury within the last year. Patients were assessed with 2 minutes of artifact-free EEG acquired from frontal and frontotemporal regions within 120 hours of head injury. Data analysis was performed from July 2021 to June 2023.

Main Outcomes and Measures

Quantitative features characterizing the EEG signal were extracted from a 1- to 2-minute artifact-free EEG data for each participant, within 120 hours of injury. Symptom inventories and days to return to activity were also acquired.

Results

From the 771 participants (mean [SD] age, 20.16 [5.75] years; 432 male [56.03%]), 600 were randomly selected for cluster analysis according to 471 qEEG features. Participants and features were simultaneously grouped into 5 disjoint subtypes by a bootstrapped coclustering algorithm with an overall agreement of 98.87% over 100 restarts. Subtypes were characterized by distinctive profiles of qEEG measure sets, including power, connectivity, and complexity, and were validated in the independent test set. Subtype membership showed a statistically significant association with time to return to activity.

Conclusions and Relevance

In this cohort study, distinct subtypes based on resting state qEEG activity were identified within the concussed population at the time of injury. The existence of such physiological subtypes supports different underlying pathophysiology and could aid in personalized prognosis and optimization of care path.

Introduction

Although there is consensus that concussion is a complex and heterogeneous mild traumatic brain injury (mTBI), current subtype classification relies largely on self-reported batteries of symptoms to group patients according to similar clinical presentations, prevalence, or recovery times. A review¹ including studies from 1990 to 2017 identified mTBI clinical subtypes defined solely on the basis of symptoms reported within 3 to 5 days after the injury, supporting the existence of heterogeneous presentations. The guidelines published from the 2017 Consensus Meeting on Concussion in Sports² stated that neuroimaging techniques have proven that mTBI is associated with changes in brain structure and function and that those changes correlate with postconcussive symptoms and long-term sequelae. It is, therefore, expected to find heterogeneity in the neuronal, axonal, and glial damage and/or microscopic pathology, the extent of which will likely reflect distinctive profiles of underlying pathophysiology of the concussion injury.

The temporal resolution of electroencephalography (EEG) makes it uniquely sensitive to such possible changes in brain function. Features characterizing the EEG signal (ie, quantitative EEG [qEEG] features), can reflect different mechanisms of underlying brain dysfunction. Such features include those showing shifts in the frequency spectra, disruptions in connectivity between regions and networks,^3,4,5,6 and measures of entropy and complexity reflecting neuronal disorganization.⁷ Advances in EEG signal processing and machine learning (ML) methods have been leveraged in the development of EEG-based TBI biomarkers for structural and functional brain injury^8,9 and concussion.^6,10 Although high accuracy has been reported, heterogeneity within the concussion population was not addressed in these previous studies.

The existing literature has demonstrated the use of EEG-derived features to identify subtypes within different neuropsychiatric populations.¹¹ Other works demonstrated how subtype membership could estimate treatment response^12,13 and evolution of cognitive decline.¹⁴ To our knowledge, this study describes the first identification of electrophysiological subtypes of concussion supporting distinct underlying pathophysiology within a population of individuals with concussion.

Methods

Participants and Setting

We performed 2 prospective longitudinal studies (February 2017 to March 2019, and September 2019 to June 2022), at 13 clinical sites across the US (9 universities and high schools and 4 clinical sites; see the full list in eTable 1 in Supplement 1), with approval from primary sites’ institutional review boards. This retrospective analysis included participants with head injuries from these prior studies. Participants were aged 13 to 70 years and clinically deemed concussed at their respective sites. Participants were assessed with a handheld EEG device within 120 hours of injury. Return to activity (RTA) was clinically determined by site standard practice (gradual return protocols). All participants signed written informed consent; for minors, parental written informed consent and adolescent assent were also obtained. Race and ethnicity were self-identified on the basis of single-choice entries and are included in this study only as population descriptors. The longitudinal studies followed Standards for Reporting of Diagnostic Accuracy (STARD) reporting guidelines and were registered on ClinicalTrials.gov. The current retrospective cohort study followed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines for observational studies.

Inclusion and Exclusion Criteria

Participants included in the study consisted of male and female individuals aged 13 to 70 years who met the study definition of concussion and had a Glasgow Coma Scale (GCS) score greater than or equal to 14 at the time of injury and no hospital admission due to either head injury or collateral injuries for more than 24 hours. Excluded participants were those with forehead, scalp, or skull abnormalities or whose clinical condition would not allow electrode placement; those currently taking psychoactive prescription medications daily (with the exception of medications for attention deficit disorder); those with a history of brain surgery or neurological disease (including multiple sclerosis, Alzheimer, or Parkinson disease); pregnant women; those with acute intoxication; those with active fever greater than 100 °F (37.8 °C); and those who were unable to speak or read English. Participants with a concussion were excluded if they lost consciousness for more than 20 minutes, or if there was evidence of abnormality visible on head computed tomography related to the injury.

Study Definition of Concussion and RTA

Participants with concussion were defined as those who had a witnessed head impact and who, by site guidelines, were restrained from normal activity for 5 or more days. RTA determination (number of days to cleared to resume activity date) was made in accordance with a gradual and/or graduated RTA protocol across multiple days, at the end of which a participant was cleared to RTA and/or play. For the nonsport concussion participants (13.2%), the RTA was defined by physician standard of care. For college and high school–based sites, gradual RTA protocols conformed to National Collegiate Athletic Association and policy guidelines.^2,15

Clinical Assessments

Study participants were evaluated with 3 sections of the Sports Concussion Assessment Tool Third Edition¹⁶ or Fifth Edition¹⁷: (1) the GCS; (2) the 22-item Concussion Symptom Inventory (CSI-22) self-rated on a Likert scale (0-6 per item; total score range, 0-132, with a higher score indicating more severe symptoms)¹⁸; and (3) Standard Assessment of Concussion (SAC), a brief neurocognitive screening tool (total score range, 0-30, with a higher score indicating less severe symptoms).¹⁹ History of head injury and concussion was also acquired.

EEG Data Acquisition

Ten minutes of eyes closed resting EEG data was collected. The EEG data were recorded using a disposable headset that included Fp1, Fp2, F7, F8, AFz, A1, and A2 locations of the expanded International 10-20 Electrode Placement System, re-referenced to linked ears, and all electrode impedances were below 10 kΩ throughout the recording. Data were acquired at a sampling rate of 1 kHz. Amplifiers had a band pass filter from 0.3 to 250 Hz (3 dB points) and down-sampled to 100 Hz for feature extraction.

EEG Data Processing and Quantitative EEG Measure Sets

Physiological and nonphysiological contamination (eg, eye movement and electromyography muscle activity) was removed from the 10-minute EEG recording using real-time artifact detection algorithms,²⁰ resulting in 1 to 2 minutes of artifact-free data for each participant. A set of more than 6000 qEEG features was extracted afterward and z-transformed with respect to age-expected normal values.^20,21 These features quantify characteristics of the electrical brain activity of different regions and frequency bands (1.5 to 45 Hz), expressed through measure sets including power (absolute and relative), mean frequency, connectivity (asymmetry, coherence, phase lag, and phase synchrony), complexity (fractal dimension and scale-free activity), and information theory (entropy).

A 2-stage feature reduction was conducted to reduce the number of features and therefore the likelihood of overtraining an ML algorithm. First, only features demonstrated to be replicable over the time period of the study were kept.²⁰ Second, features showing variable ranges within the injured population were included, because such heterogeneity would contribute to the separation of subtypes of concussion. Heterogeneity was computed on the basis of 100 bootstrapped repetitions of a randomized permutation test using P values from the Kolmogorov-Smirnov statstic.²² Features deemed nonsignificant in more than 5% of the repetitions were discarded. As a result of the feature reduction, 471 qEEG features were retained.

Statistical Analysis

Data analysis was performed from July 2021 to June 2023. An initial random division split the cohort into training and testing sets. Spectral coclustering^23,24 was used on the training set to detect distinct partitions (subtypes) of EEG activity. Spectral coclustering performs a 2-way clustering that simultaneously partitions rows and columns of a matrix. The goal is to find subsets of rows that change similarly over a subset of columns, or subsets of rows that have similar values across a subset of columns. The resulting partitions of rows and columns are referred to as biclusters. Spectral coclustering assumes that the data have a checkerboard-like structure, with blocks of high-expression levels and low-expression levels. It finds these distinctive checkerboard patterns by using eigenvectors and singular value decomposition of matrices. The singular vectors for rows and columns are then clustered using K-means. Every row and column is included only in 1 bicluster, with the resulting structure being block-diagonal. To verify the stability of the biclusters, we configured 100 multistart runs of the coclustering algorithm using random seeds to initialize partitions.

After reaching stable partitions in the training set, a supervised classification was performed using the bicluster memberships as class labels. An informed data reduction was performed by a feature subset selection algorithm, recursive feature elimination, which is a stepwise feature elimination method that finds the optimal set of features for a given classification function.²⁵ The algorithm started with the full set of 471 features and iteratively removed the least relevant. Recursive feature elimination was also run on a 100 multistart setup. Each run performed a random 5 times, 5-fold, stratified cross-validation, and 2 functions, support vector machines and logistic regression, were used to guide the search process.

A final classification algorithm was derived over training data using the feature subset with highest classification accuracy. This classifier was derived from a logistic regression using an L2-norm regularization and balanced class learning through adjusting weights inversely proportional to class frequencies in the data. The hold-out participants were classified into subtypes by this algorithm. Any statistical significance reported in the results section was fixed to 2-sided P < .05, unless explicitly stated otherwise. Data analysis was performed with on Python software versions 3.7 to 3.9, with the scikit-learn 1.2.2, pandas 1.5.3, numpy 1.23.5, and scipy 1.11.2 libraries (all from Python Software Foundation).

Results

There were 771 concussed participants (432 male [56.03%]), with a mean age of 20.16 (5.75) years, a mean GCS score of 15.00 (0.06), and median (IQR) time since injury of 50.03 (27.20-70.47) hours. RTA was categorized as rapid recovery (<14 days; 368 participants [47.73%]) or prolonged recovery (≥14 days; 285 participants [36.96%]). There were 118 unexpected failures (15.30%) to follow-up for RTA, approximately one-half of which (55 participants) were due to the COVID-19 pandemic, which reduced the ability for participants to return for follow-up evaluations. Race and ethnicity distributions followed the same percentages as those published by the 2020 US Census Report.²⁶ Full demographic descriptors of the enrolled population are presented in the Table.

Table. Characteristics of the Participants.

Characteristic	Participants, No. (%) (N = 771)
Age, y
Mean (SD)	20.16 (5.75)
Median (range)	19.43 (13.07-70.83)
Sex
Male	432 (56.03)
Female	339 (43.97)
Race and ethnicitya
Black or African American (non-Hispanic)	116 (15.05)
Otherb	32 (4.15)
Unknown	34 (4.41)
White (Hispanic)	81 (10.51)
White (non-Hispanic)	508 (65.89)
Glascow Coma Scale score
Mean (SD)	15.00 (0.06)
Median (range)	15.00 (14.00-15.00)
Total 22-item Concussion Symptom Inventory score at day 0
Mean (SD)	27.83 (22.20)
Median (range)	23.00 (0.00-117.00)
Total Standard Assessment of Concussion score at day 0
Mean (SD)	24.38 (3.88)
Median (range)	25.00 (10.00-30.00)
Return to activity, da
Mean (SD)	15.67 (21.71)
Median (range)	12.00 (2.00-312.00)
≥14 d	285 (36.96)
<14 d	368 (47.73)
Unknown	118 (15.30)

^a All information is not available for all participants.

^b Other refers to Asian, Hispanic Black or African American, Native American or Alaska Native, and Native Hawaiian or Other Pacific Islander.

The training set included 600 participants, and the testing set included 171 participants. Subtypes (data profiles or phenotypes) found by the coclustering runs over the training set were aggregated to analyze stability and identify possible noisy features or participants. We identified 5 stable distinctive data profiles of participants and features across all the runs, with a mean (SD) pairwise consensus score of 98.87% (1.19%). Only 5 participants of the original 600 in training (<1%) and 2 of 471 features (<1%) were systematically assigned into different subtypes over all runs. These participants and features were deemed noisy and removed from further analyses, resulting in 595 participants and 469 features being used to describe the subtypes. The final distribution of features and participants in each subtype are displayed in eTable 2 in Supplement 1, and the defining qEEG measure sets (ie, those that contributed most to the profile) are displayed in Figure 1. As can be seen in Figure 1, distinct patterns of qEEG measure sets characterize the different subtypes, supporting different underlying pathophysiology.

Figure 1. Histogram of the Distribution of Power, Power Ratios, Connectivity, and Complexity Segmented by the Quantitative Electroencephalography (qEEG) Measure Set .

The y-axis shows the percentage distribution within the subtype by measure set (total to 100%).

By labeling each participant with the associated subtype membership from the coclustering analysis, we enabled supervised classification to be performed. The informed data reduction process yielded a list of 257 relevant features of the pool of 469 (54.8% reduction). Figure 2 depicts a t-distributed stochastic neighbor embedding 2-dimensional plot illustrating how samples from each subtype are statistically different from the rest according to the pool of 257 relevant features. It can be clearly seen in Figure 2 how each subtype occupies a different part of the plotting space forming separable groups. Quantitatively, support vector machines and logistic regression were tested as classification functions on a 5-fold cross-validation scheme. Logistic regression showed the best overall performance with a 96.64% classification accuracy when descriminating the 5 subtypes. Sensitivity and precision metrics (positive predictive value) for each individual subtype showed very high performances, with values between 95% and 99% (eFigure 1 in Supplement 1).

Figure 2. t-Distributed Stochastic Neighbor Embedding (t-SNE) 2-Dimensional Plot .

Graph shows the statistical separation between subtypes when using 257 relevant features identified by the informed data reduction procedure.

An in-depth inspection of the key contributing features for subtype 1 and subtype 3 is presented in Figure 3. Divided by qEEG feature domains, Figure 3 shows polar bar charts displaying the mean z score for a common set of the most relevant qEEG features representing each major measure set contributing to the characterization of the subtype. It is important to note that these features do not represent all features used in the coclustering or necessarily those that were most significant in differentiating between subtypes. In Figure 3, the distance from the origin represents the magnitude of the abnormality in both the positive direction (excess relative to age-expected normal values) and the negative direction (deficit relative to age-expected normal values). When comparing polar bar charts for all 5 reported subtypes, distinctive overall patterns of abnormalities characterizing each subtype are clearly observed. For example, subtypes 1 and 4 show opposite values with respect to relative and absolute power, power ratios, mean frequency, and complexity. Thus, although subtype 1 is characterized by extreme excesses for complexity features and mean frequency, and deficits of absolute and relative power, subtype 4 shows the opposite pattern (eFigure 2 in Supplement 1).

Figure 3. Polar Bar Charts of Mean z Values for Subtype 1 and Subtype 3.

Each quadrant displays z values for the 4 most relevant quantitative electroencephalography features segmented by measure set: absolute power, mean frequency, asymmetry, complexity (scale free, fractal dimension, and entropy), coherence, phase stability, phase synchrony, power ratio, and relative power (clockwise).

For subtypes 2 and 3, other domains showed opposite activities—namely phase synchrony, phase stability, and coherence. In both comparisons, the differential activity was significant, with absolute differences over 2 z score units (significance level of a group mean z score can be estimated by considering the square root of the size of the group, and a z score difference of 0.5 is associated with P < .0005). The remaining subtype 5 polar bar chart showed no differential brain activity compared with the other 4 groups, with mean z values close to 0, suggesting an unspecific form of injury (eFigure 2 in Supplement 1).

Validation of the cluster structure was performed using the independent 171 hold-out participants applying the logistic regression classifier derived from the development data set. Prevalence estimations between development and test data demonstrated similar rates. It was noted that subtype 4 was the most common concussion subtype in both development and test populations (136 samples [22.85%] and 51 samples [29.82%], respectively), and that subtype 1 was one of the least common subtypes (90 samples [15.13%] and 28 samples [16.37%], respectively). Occurrence rates for each subtype are presented in eTable 3 in Supplement 1 as a proxy of subtyping prevalence among the concussed population.

Subtype Membership and Clinical Features

A clear association between subtype membership and outcome (days to clinical recovery, or RTA) is shown in eTable 4 in Supplement 1 and Figure 4. In eTable 4 in Supplement 1, RTA days were aggregated by weeks (1, 2, 3, or ≥4) to analyze group patterns and minimize site deviations. From the 766 participants in training and test splits, RTA values for 652 participants were available, with an outlier upper bound of 3 SD over the mean of 85.94 days. To be conservative, we removed outliers beyond 86 days, reducing the number of valid values to 640 with a median (IQR) of 16 (11-23) days. It can be observed that subtype 2 and subtype 4 showed prolonged recovery, with RTA values over 4 weeks in 50 (40.0%) and 55 (35.2%) cases, respectively. On the other end, subtype 3 and subtype 5 presented quick recovery patterns on average, with an aggregate 49 participants (40.8%) with subtype 3 and 53 participants (37.6%) with subtype 5 being cleared for RTA by the end of second week. Finally, most concussed participants in subtype 1 were cleared during the third week after injury. This is the highest rate for week 3 compared with the other 4 subtypes. Overall differences on RTA days across subtypes were statistically significant at a 95% confidence level (Kruskal-Wallis H-test of independent samples). When analyzed by pairs, statistical differences were in accordance with the weekly values displayed in eTable 4 in Supplement 1, with subtype 2 vs 3 and subtype 3 vs 4 reaching statistical significance. All other comparisons reported P > .05 significance level. Figure 4 shows the cumulative distribution for RTA for each subtype. The differences in rates of recovery shown in eTable 4 in Supplement 1 can be observed in this figure. Overall, the fastest rate of recovery is seen in subtype 3 and the slowest rate of recovery in subtype 2. For example, among participants with subtype 2, 75 (60.0%) reached RTA by 21 days but 91 (75.8%) of those with subtype 3 reached RTA by 21 days, with the other subtypes falling in between.

Figure 4. Cumulative Statistical Distribution of Return to Activity Values by Subtype During the First Month After Injury.

The associations of concussion subtype with symptom burden at the time of injury were evaluated with conventional measures of self-report symptom burden using the CSI-22 and the total SAC score. No significant differences were seen between subtype membership and these clinical features. The median (IQR) total CSI-22 score was 23.5 (8.0-37.0) for subtype 1, 23.5 (9.0-45.8) for subtype 2, 23.0 (11.0-38.0) for subtype 3, 24.0 (12.5-40.0) for subtype 4, and 20.0 (9.0-42.0) for subtype 5. The median (IQR) total SAC scores were 26.0 (24.0-27.0) for subtype 1, 24.0 (20.9-27.0) for subtype 2, 26.0 (24.0-28.0) for subtype 3, 25.0 (22.1-27.0) for subtype 4, and 26.0 (23.0-28.0) for subtype 5. This suggests that such clinical characterization does not reflect the reported electrophysiological subtypes.

Discussion

Concussion (mTBI) is a complex and heterogeneous disorder; however, efforts to describe this heterogeneity have mainly focused on clinical symptoms.^1,27 To our knowledge, this cohort study is the first to demonstrate the existence of brain activity–based subtypes within the concussion population. Each subtype is clearly defined by distinct patterns of electrophysiological features, highlighting the expected presence of heterogenous underlying pathophysiology.

The 771 participants in this retrospective analysis were assessed during a 5-year span at 13 clinical sites across the US. The distributions of sex, ethnicity, and rapid or prolonged RTA showed no major deviation from the US expected population²⁶ and state-of-the-art in concussion² (see the Table for details). Although it was noted that the final cohort showed a young mean age, the range included participants in the interval 13 to 70 years.

The 5 subtypes were found to be highly stable, with less than 1% discordance between repeated runs. Quantitatively, using an ML algorithm (logistic regression) for subtype classification on a 5-fold cross-validation scheme, high overall classification accuracy (96.64%) in the discrimination of subtypes was obtained, with sensitivity and precision values per subtype ranging from 95% to 99%.

The existence of different subtypes characterized by distinct patterns of qEEG features supports the hypothesis that these subtypes represent different underlying pathophysiology and that knowledge of subtype membership could provide objective information related to clinical presentation and treatment planning. For example, subtype 3 is characterized by frontal connectivity abnormalities, which have also been observed in noninjured participants with depression.²⁸ It is plausible that participants in this cluster may show higher risk of developing depression than those in other clusters. Furthermore, subtype 4 is characterized by power abnormalities including those within the theta frequency band (3.5-7.5 Hz). Theta activity is prominently generated in the hippocampus,²⁹ and one well-established finding in posttraumatic stress disorder is the presence of hippocampal atrophy.³⁰ It could be hypothesized that compared with participants in other subtypes, participants in this subtype may be more at risk for developing posttraumatic stress disorder over time. Although these secondary conditions were not fully expressed during this study, subtype membership could provide an early indication of an evolving disorder.

Subtype membership was shown to be associated with outcome related to rapid or more prolonged recovery. Subtype 2, those with major disruptions in power features, and subtype 4, those with abnormalities in both power and connectivity features, showed longer recovery periods (≥3 weeks) compared with other subtypes. In the other extreme, subtype 3, which was mostly characterized by disturbances in connectivity measures, coherence, and phase synchrony, showed a rapid recovery trend, with almost 50% of participants cleared to RTA in less than 2 weeks after injury. Since a delay in concussion diagnosis and appropriate concussion management have been shown to lead to much slower recovery from concussion,^31,32 knowledge of the individual’s subtype could positively influence recovery by providing personalized information at the time of injury. Conventional measures of self-reported symptoms (CSI-22 and SAC) were used to estimate the total symptom burden at time of injury and compared with subtype membership. No significant differences in symptom distribution, mechanism of injury, or history of concussion among the subtypes were found. This fact suggests that electrophysiological subtypes of concussion are not simply a reflection of the total symptom burden. Different patterns of underlying disruption of brain activity as the root cause of the differential clinical presentations remains to be explored more extensively.

It is important to note that the focus of this work was to demonstrate the existence of EEG-based physiological subtypes within the concussion population. The use of controls would lead to a solution that maximized homogeneity and not the heterogeneity within the concussed population. In addition, all features were expressed as z scores relative to age-expected normal values; therefore, the qEEG patterns identified take into consideration deviation from normal.

Limitations

This study has limitations that should be acknowledged. The combination of 2 major clinical trials provided a high degree of assurance in the subtype identification. However, this subtyping needs to be replicated in other future analyses using qEEG assessments of concussed participants. The clinical data collected in this study and the follow-up time points limited more extensive study of the associations between subtype membership, clinical presentation, and evolution of clinical sequelae.

Conclusions

In this cohort study of 771 clinically concussed participants, we reported the existence of distinct brain activity–based subtypes at time of injury. Although it is well recognized that concussion is a heterogeneous disorder, attempts to subtype concussion according to subjective reports of symptoms have led to inconclusive results. Our reporting of the existence of physiological subtypes supports different underlying pathophysiology and has the potential to aid in more personalized treatment of concussion, the determination of prognosis, and to optimize care path and improved outcome.

Supplement 1.

eTable 1. Full Name of the Clinical Sites, Number of Participants and Age Descriptors

eTable 2. Distribution of Feature and Subjects for Each Detected Subtype

eFigure 1. Confusion Matrix and Figures of Merit of a Five-Fold Cross Validation Estimation for a Logistic Regression Classifier Evaluated on the Training Data Set (257 Features, 595 Subjects Labeled With the Coclustering Outputs)

eFigure 2. Polar Bar Charts Displaying the Average z-Score for All Five Subtypes Using a Common Set of Highly Relevant qEEG Features

eTable 3. Prevalence Estimation for Each Concussion Subtype Under the Study Population, Both for the Development and Test Sets

eTable 4. Association Between Subtype and Time to Return-To-Activity Event (Counts and Percentages by Subtype Valid Data)

Supplement 2.

Data Sharing Statement