Detection of American Football Head Impacts Using Biomechanical Features and Support Vector Machine Classification

Abstract

Accumulation of head impacts may contribute to acute and long-term brain trauma. Wearable sensors can measure impact exposure, yet current sensors do not have validated impact detection methods for accurate exposure monitoring. Here we demonstrate a head impact detection method that can be implemented on a wearable sensor for detecting field football head impacts. Our method incorporates a support vector machine classifier that uses biomechanical features from the time domain and frequency domain, as well as model predictions of head-neck motions. The classifier was trained and validated using instrumented mouthguard data from collegiate football games and practices, with ground truth data labels established from video review. We found that low frequency power spectral density and wavelet transform features (10~30 Hz) were the best performing features. From forward feature selection, fewer than ten features optimized classifier performance, achieving 87.2% sensitivity and 93.2% precision in cross-validation on the collegiate dataset (n = 387), and over 90% sensitivity and precision on an independent youth dataset (n = 32). Accurate head impact detection is essential for studying and monitoring head impact exposure on the field, and the approach in the current paper may help to improve impact detection performance on wearable sensors.

Figure 1

Study Overview. We deployed instrumented mouthguards (a) to football players during games and practices (b). Videos of the events were recorded to generate ground truth event labels for mouthguard recordings (c). Using the labelled dataset containing head impacts and nonimpact events, we trained a classifier to distinguish between these two classes of events (d).

Figure 2

Extracted Training and Validation Dataset. In the first round of video review, raters identified periods of possible helmet contact, player being idle, or in activity without contact. To extract the impact set, a second round of video review was performed to establish confidence on the helmet contact label. Instrumented mouthguard recordings within verified helmet contact periods were further matched with video observed directionality and checked for device placement through infrared sensing before inclusion into the impact dataset. High IR mouthguard recordings from idle and no contact periods were included in the nonimpact dataset.

Figure 3

Distributions of Kinematics Measures for Impacts and Nonimpacts. Shown here are high IR impact (n = 156) and high IR nonimpact (n = 231) distributions of common kinematic measures that could be used for impact registration or classification. The subplots are anterior-posterior (AP), left-right (LR), inferior-superior (IS) linear acceleration, and coronal (Cor), sagittal (Sag), horizontal (Hor) plane angular velocity. It is shown that impacts and nonimpacts have similar linear acceleration distributions. Comparing linear acceleration magnitude, impacts had values ranging from 10.1 g to 65.6 g (mean 22.0 g, median 19.1 g), while nonimpacts ranged from 10.0 g to 104.3 g (mean 18.5 g, median 15.8 g). For change in angular velocity magnitude, impacts had values ranging from 1.6 rad/s to 26.1 rad/s (mean 10.5 rad/s, median 9.7 rad/s), while nonimpacts ranged from 0.1 rad/s to 55.7 rad/s (mean 5.4 rad/s, median 3.3 rad/s). Although there are more low-angular velocity recordings in nonimpacts compared to impacts, there are overlaps between the two classes. With the overlaps in distributions, the kinematics features may not be the most predictive features for classification.

Figure 4

Kinematics, PSD, and WT of an Example Impact and Nonimpact. Example impact (a) and nonimpact (b) kinematics show qualitative differences between these two events, with the nonimpact exhibiting higher frequency impulses and oscillations. Such frequency-domain differences are reflected in the Fourier transform power spectral density (PSD) plots (c and d) and wavelet transform (WT) plots (e and f), where color represents amplitude.

Figure 5

Distinguishing Features between Impacts and Nonimpacts. Here we show distributions of 8 features with the lowest p-value in the Wilcoxon Rank-Sum test between impacts and nonimpacts. These features include linear acceleration and angular acceleration PSD and WT features at low frequencies (10–20 Hz), as well as estimated AP torso displacement from the kinematic measurements. With less overlap between impacts and nonimpacts in these distributions, the features tend to have higher predictive value in distinguishing the two classes.

Figure 6

PCA and Correlation Analysis of Features. Principal component analysis of the feature space shows that 90% of the variance could be explained by the first principal component, indicating low dimensionality of the feature space (a). Among the first five principal components, the top contributing features are mainly PSD and WT features, and the time-domain linear acceleration features have a relatively significant contribution only in the first principal component (b). Pearson’s correlation analyses of features show that many features are highly correlated (c).

Figure 7

Best Performing Features and Classification Decision Boundary. The top three features from forward feature selection optimizing for AUC are shown in (a) with distributions of impacts and nonimpacts in the space of these three features. The ROC and PR curves of the AUC classifier and F-measure classifier show areas under curves of close to 1, while a classifier based on acceleration thresholding has similar performance as random guessing (b,c).

Figure 8

Classifier Performance in Classifying All Events Recorded During a Practice. Since the classifier was only trained on a small fraction of data that could be verified through video, we tested the performance of the classifier in classifying all events recorded by the mouthguard over a sample event (total recordings = 1219). This figure shows the timeline of a practice for an offensive lineman. Over the recorded video duration, the first round of video analysis identified periods of helmet or body contact, as well as when contact likely occurred but the type of contact could not be confidently judged due to obstructed video view. 46 events recorded by the mouthguard were classified as helmet contact by the trained classifier, and they tend to fall close to periods of observed or obstructed contact from video analysis.