A Routine Electroencephalography Monitoring System for Automated Sports-Related Concussion Detection

Abstract

Cases of concussions in the United States keep increasing and are now up to 2 million to 3 million incidents per year. Although concussions are recoverable and usually not life-threatening, the degree and rate of recovery may vary depending on age, severity of the injury, and past concussion history. A subsequent concussion before full recovery may lead to more-severe brain damage and poorer outcomes. Electroencephalography (EEG) recordings can identify brain dysfunctionality and abnormalities, such as after a concussion. Routine EEG monitoring can be a convenient method for reducing unreported injuries and preventing long-term damage, especially among groups with a greater risk of experiencing a concussion, such as athletes participating in contact sports. Because of the relative availability of EEG compared to other brain-imaging techniques (e.g., functional magnetic resonance imaging), the use of EEG monitoring is growing for various neurological disorders. In this longitudinal study, EEG was analyzed from 4 football athletes before their athletic season and also within 7 days of concussion. Compared to a control group of 4 additional athletes, a concussion was detected with up to 99.5% accuracy using EEG recordings in the Theta-Alpha band. Classifiers that use data from only a subset of the EEG electrodes providing reliable detection are also proposed. The most effective classifiers used EEG recordings from the Central scalp region in the Beta band and over the Temporal scalp region using the Theta-Alpha band. This proof-of-concept study and preliminary findings suggest that EEG monitoring may be used to identify a sports-related concussion occurrence with a high level of accuracy and thus reduce the chance of unreported concussion.

Keywords: EEG; EEG monitoring; SVM; electrode networks; non-patient-specific; sports-related concussion.

FIG. 1.

Sets of electrodes mapped on the scalp regions. Associated electrodes of a selected area are highlighted in blue.

FIG. 2.

Feature extraction steps and data-set generation procedures. One-thousand-millisecond EEG segments are pre-processed and transferred into the frequency domain. The features are extracted in a frequency band of interest using the defined feature metrics to generate a (N x N) matrix of the pair-wise coherency of electrodes. N is the number of electrodes in a cluster of electrodes. The (N x N) extracted features from the first and second sessions are stacked together to build a (2N x N) segment of a participant's features, of which the first N^th rows are extracted features of the first EEG session and the second N^th from the second session. These steps are iterated for all 50 segments of the first and second sessions. Fifty segments of each session were used to generate the final data-set for a participant, which generates a rank 3 tensor with 50 segments of (2N x N) extracted features from both sessions. EEG, electroencephalography.

FIG. 3.

Classifier model. Extracted features of a participant are excluded for training the classifier model as the test data set. After training the SVM classifier using extracted features of the other 7 participants, the model is evaluated on the isolated testing dataset (i.e., participant). Model performance reports the accuracy of the modeled classifier on predicting the correct label of the 50 segments of the test data set. SVM, support vector machine.

FIG. 4.

Frequency bands performances. Box plots contain the average performance of modeled classifiers on concussed, control groups, and the total average (average of both groups) in each frequency band using all the defined metrics using the set of electrodes in the All region. Accuracies are in the range of [0 to 1]. Red, blue, and gray boxes are the average performance of the classifiers tested on the concussion, control, and both groups, respectively.

FIG. 5.

Performances of clusters of electrodes in the Theta-Alpha frequency band. Performances of classifiers trained on all the defined feature metrics individually in the Theta-Alpha frequency band are illustrated for groups of electrodes. Accuracies are in the range of [0 to 1]. Red, blue, and gray boxes are the average performance of the classifiers tested on the concussion, control, and both groups, respectively. R-Frontal, right frontal; L-Frontal, left frontal; R-Temporal, right temporal; L-Temporal, left temporal; R-Central, right central; L-Central, left central.

FIG. 6.

Top 10 performances of feature metrics. Average accuracies of the top 10 models with the best performances for each feature metric, regardless of the frequency band and subset of electrodes for training the model, are illustrated for participants in concussion and control groups and overall performance of models for both groups (total average). Accuracies are in the range of [0 to 1]. Red, blue, and gray boxes are the average performance of the classifiers tested on the concussion, control, and both groups, respectively.

FIG. 7.

ROC of models with best performances. ROC curves of the three best models are developed using the Allfeat metric in the Theta-Alpha band in the All (blue), Temporal (red), and Beta band among the Central region (yellow). The AUC of each ROC is provided in the legends. AUC, area under the curve; ROC, receiver operating characteristic.