Annals of Biomedical Engineering
February 8, 2025
In response to growing evidence that repetitive head impact exposure and concussions can lead to long-term health consequences, many research studies are attempting to quantify the frequency and severity of head impacts incurred in various sports and occupations.
The most popular apparatus for doing so is the instrumented mouthguard (iMG). While these devices hold greater promise of head kinematic accuracy than their helmet-mounted predecessors, data artifacts related to iMG decoupling still plague results. We recreated iMG decoupling artifacts in a laboratory test series using an iMG fit to a dentition mounted in a NOCSAE headform.
With these data, we identified time, frequency, and time-frequency features of decoupled head impacts that we used in a machine learning classification algorithm to predict decoupling in six-degree-of-freedom iMG signals. We compared our machine learning algorithm predictions on the laboratory series and 80 video-verified field head acceleration events to several other proprietary and published methods for predicting iMG decoupling. We also present a salvaging method to remove decoupling artifacts from signals and reduce peak resultant error when decoupling is detected. Future researchers should expand these methods using on-field data to further refine and enable prediction of iMG decoupling during live volunteer use. Combining the presented machine learning model and salvaging technique with other published methods, such as infrared proximity sensing, advanced triggering thresholds, and video review, may enable researchers to identify and salvage data with decoupling artifacts that previously would have had to be discarded. …
Methods
Development of a Machine Learning Model for Decoupling Detection
We used data from a laboratory study that explored the effects of decoupling on iMG accuracy to develop a machine learning algorithm that reliably detected head impacts with decoupling event features [14]. We then applied this algorithm to field data from video-verified head acceleration events collected from female adolescent ice hockey players. We compared our algorithm’s outputs to three other impact classification methods—the iMG manufacturer’s change-in-proximity measurement (Delta-prox), the iMG manufacturer’s impact quality label (Quality), and a published method from Luke et al. that uses iMG-specific proximity measurement thresholds. [19]
Our previous study defining decoupling events in the laboratory used a National Operating Committee on Standards for Athletic Equipment (NOCSAE) headform mounted to a Hybrid-III 50th percentile male neck [24] and modified with a mouth area cut out to fit 3D-printed upper teeth and a lower aluminum plate [25] to study the effects of instrumented mouthguard decoupling. This surrogate headform represents the properties of a 50th percentile male [26] and was mounted on a linear slide table with 5 degrees of freedom. A boil-and-bite iMG (Prevent Biometrics, Edina, MN, USA) was fitted to the upper teeth. The fit was confirmed to pass the open-mouth test by pressing the instrumented mouthguard onto the upper dentition and ensuring it did not fall off the teeth. We placed the lower aluminum plate at three distances (0 mm—control, 1.6 mm—small gap, and 4.8 mm—large gap) from the iMG after firmly pressing the iMG onto the upper teeth. The lower plate is attached to the head via a bolt at the midpoint between the rear molars. The lower plate was securely fastened during all tests by tightening it against either the instrumented mouthguard in the coupled conditions or against a series of washers in the open-mouth conditions. These washers effectively tightened the lower plate to the roof of the headform’s mouth without touching the instrumented mouthguard. The lower plate was not free to move during any of the tests.
We conducted tests on a pendulum impactor at various impact durations and severities (Fig. 1). Two impactor faces (padded (VN600, Dertex Corporation) and rigid nylon), four target peak linear acceleration severities (25, 50, 75, and 100 g), four locations (front, front boss, rear, and rear boss) [25, 27], and two trials at each condition generated 192 total tests. Both impactor faces were 127 mm in diameter; the padded impactor face was 40 mm thick, while the rigid nylon was 25 mm thick. The faces generated linear impact durations in the 3–5 ms range (rigid) and 10–12 ms range (padded). The pendulum arm was 1.905 m long and the anvil mass was 15.5 kg.
Figure 1
The pendulum impactor A setup with four impact locations. B–E. A boil-and-bite iMG was fit to 3D-printed teeth in a medium NOCSAE headform with measured gaps to the fixed lower dentition (F). We used data from this test series to train the machine learning classification algorithm. …
Discussion
We have presented a machine learning model for detecting decoupling in instrumented mouthguard six-degree-of-freedom measurement signals. Our final model is intentionally low in feature count and complexity to avoid over-fitting the training dataset derived from laboratory head impact events. We valued sensitivity to detecting an actual decoupling event because the subsequent salvaging process minimally influenced peak kinematic accuracy in coupled events but considerably increased accuracy in decoupled events.
Relative to other decoupling prediction methods, our model had the highest sensitivity to decoupling in the training dataset. Although other methods had higher specificities, our model was the most balanced between sensitivity and specificity. In field head acceleration events, the Delta-prox method had the most similar decoupling prediction rate to our model. When considering all head acceleration events, however, the iMG manufacturer’s Quality prediction score was most similar to our model regarding agreement in both coupling status categories.
Our model differed from the other tested methods in that other methods use proximity sensor readings from the iMG, while our machine learning model does not. Our model uses information about known decoupling events [16] to detect decoupling. We used features from the kinematic signals rather than lower sampling rate proximity readings because coupling status can change throughout an impact event [14] and lower sampling rates may miss these changes. Therefore, when salvaging data is the goal, it is essential to classify coupling status with data from within the impact duration rather than only before and after. Additionally, infrared proximity sensors are fundamentally measuring different physical phenomena than the kinematic sensors upon which our presented algorithms rely. Proximity sensors measure the distance to a surface by calculating the intensity of reflected infrared light; these reflections can be influenced by factors other than true iMG motion relative to the dentition. The kinematic sensors, on the other hand, are measuring iMG motion directly, but here the challenge is to discern iMG motion from head impact characteristics, as we have attempted to train our model to do.
The supplementary coupling prediction methods we included had varying levels of agreement with one another. The modified Luke et al. method agreed well with the Delta-prox method in the laboratory training data. This agreement was likely due to both methods using only proximity readings as inputs and the controlled laboratory setting, which resulted in no iMG motion relative to the teeth except during impact. On the other hand, the Delta-prox method did not always agree with the Quality method. This lack of agreement likely comes from the manufacturer’s use of kinematic signal features in their Quality label assignment.
Advanced published post-processing methods such as HEADSport and AI-denoising offer similar promise to the methods presented: the ability to remove artifacts from iMG field data while keeping relevant head kinematic information. In those methods and ours, signal frequency content is used to decide how to move forward with a given impact most appropriately. Although the other advanced methods focus on signal-to-noise ratio and artifacts in general, we have chosen to focus on decoupling artifacts in particular as a problem with poor or loosely fitting iMGs. Issues with poor-fitting instrumented iMGs have been noted in sports, such as ice hockey [19], where mouthguard abuse is common, and helmet cages make ejecting an iMG mid-play inconsequential to players. Each of these advanced methods, including our presented algorithm and salvaging process, performs reasonably well in a laboratory environment, generally improving accuracy compared to a reference and removing egregious outliers. Testing the methods’ efficacy in the field is difficult, as coupling status is typically unknown, especially during an impact event.
Previously reported machine learning studies decided to reject data classified as non-impact or poor-quality [20, 21, 23]. To date, no study has reported methods for salvaging iMG data initially deemed to have decoupling artifacts. Our salvaging method improved accuracy in all three kinematics in laboratory decoupled impacts compared to the pre-salvaged iMG data. We applied our salvaging procedure to coupled laboratory head impacts to understand the potential effect this procedure could have should our algorithm misclassify in the field. The influence on peak kinematics was minimal for coupled data. This was expected because the salvaging procedure removes noise associated with decoupling, such as late, high-frequency peaks—these artifacts should not be present in coupled impacts. Peak angular acceleration was affected the most, with a higher percent difference in mean after the salvaging process. Despite the minimal effects, we do not recommend applying our salvaging methods to any impact data unless decoupling is suspected to be the primary noise source in the signal to ensure the greatest possible kinematic accuracy.
We expect salvaging field head acceleration events to have results similar to our laboratory results. Our preliminary sweep of the available field data reinforced this. We observed similar patterns of change from pre- to post-salvaging in our predicted-decoupled field data as we saw in our laboratory decoupled head impacts, with salvaged data having lower peak kinematics. After re-running the salvaged head acceleration events through our classification model, we found that 38% were unsalvageable. This likely means that decoupling was not the dominant source of noise in those 8 head acceleration events. Field data will inherently include more noise sources (e.g., chewing or jaw clack), but when decoupling is the primary source of noise in the signal, the presented salvaging methods can reduce kinematic error. Additional differences between laboratory head impacts and field head acceleration events include differences in rigidity between the NOCSAE headform surrogate and living volunteers and differences in neck response during head impact between surrogates and volunteers. Each of these factors may influence the accuracy and capability of our algorithms to successfully detect and remove decoupling artifacts. …