Comparison of video-based and sensor-based head impact exposure

Abstract

Previous research has sought to quantify head impact exposure using wearable kinematic sensors. However, many sensors suffer from poor accuracy in estimating impact kinematics and count, motivating the need for additional independent impact exposure quantification for comparison. Here, we equipped seven collegiate American football players with instrumented mouthguards, and video recorded practices and games to compare video-based and sensor-based exposure rates and impact location distributions. Over 50 player-hours, we identified 271 helmet contact periods in video, while the instrumented mouthguard sensor recorded 2,032 discrete head impacts. Matching video and mouthguard real-time stamps yielded 193 video-identified helmet contact periods and 217 sensor-recorded impacts. To compare impact locations, we binned matched impacts into frontal, rear, side, oblique, and top locations based on video observations and sensor kinematics. While both video-based and sensor-based methods found similar location distributions, our best method utilizing integrated linear and angular position only correctly predicted 81 of 217 impacts. Finally, based on the activity timeline from video assessment, we also developed a new exposure metric unique to American football quantifying number of cross-verified sensor impacts per player-play. We found significantly higher exposure during games (0.35, 95% CI: 0.29-0.42) than practices (0.20, 95% CI: 0.17-0.23) (p<0.05). In the traditional impacts per player-hour metric, we observed higher exposure during practices (4.7) than games (3.7) due to increased player activity in practices. Thus, our exposure metric accounts for variability in on-field participation. While both video-based and sensor-based exposure datasets have limitations, they can complement one another to provide more confidence in exposure statistics.

Fig 1. Instrumented mouthguard for measuring impact severity.

(A) Sensor board containing tri-axial linear accelerometer, tri-axial angular gyroscope, and infrared proximity sensor are embedded (B) inside a custom-formed instrumented mouthguard.

Fig 2. Overview of tiered video assessment for collecting head impact exposure dataset.

(A) Multiple video angles were collected for each practice and game, with at least one camera capturing an end-zone view and one camera capturing a sideline view. (B) Video was trimmed by technicians to only include play footage. (C) Trained raters performed a first round of video assessment, tracking specific players and labeling their activity. (D) A second round of video assessment performed by one of the authors confirmed Helmet Contact activities.

Fig 3. Activity classifications for tracking player activity and identifying helmet contact activities with high sensitivity.

Tracked player marked with a red arrow. (A) Raters identified Helmet Contact activities whenever the tracked player’s head overlapped with an opposing player. (B) Body Contact activities when there was contact not involving the head. (C) No Contact activities when player was in play, but not actively in contact. (D) Obstructed View activities when there was no clear view of the player’s head. (E) Idle activities when players were observed on the sideline, or otherwise not in play. Finally, (F) Not in Video activities when tracked player was not in the video.

Fig 4. Second round video assessment for specific helmet contact activity identification.

Multiple videos were used to confirm Helmet Contact activities. Red arrows mark the tracked player, with blue arrows marking other players. End-zone videos show helmet overlap, but sideline video showed (A) definitive head contact and (B) no helmet contact.

Fig 5. Impact location vectors and mouthguard kinematics processing.

(A) Locations are binned into front, front oblique, side, rear oblique, rear, and top impacts. (B) Video-based helmet contact periods were qualitatively binned into impact locations during second round video assessment by the rating author. For sensor-based head impacts, kinematics were processed by first integrating or differentiating sensor linear acceleration and angular velocity signals to obtain linear velocity, linear position, angular acceleration, and angular position (represented with XYZ Euler angles). Peak motion (angular or linear acceleration, velocity, or position) vectors were found by identifying the peak magnitude and determining the 3 degree-of-freedom components. Peak linear acceleration, velocity, and position vectors were binned directly. We also incorporated peak angular motion vectors to correct respective peak linear motion vectors.

Fig 6. Comparison of video-based and sensor-based head impact exposure.

(A) Exposure rates collected from independent (A) video-based and (B) sensor-based methods differed drastically. Our instrumented mouthguard identified an order of magnitude more discrete head impacts than video-based helmet contact periods. Cross-verifying head impacts with helmet contact periods yields a more consistent 217 discrete head impacts within 193 helmet contact periods. Delineating by event type, we found that there was greater head impact exposure in practices than in games.

Fig 7. Comparison of video-based and sensor-based impact location distributions.

For both video-based and sensor-based distributions, the majority of impacts are to the front, front oblique, and sides. Methods for processing sensor kinematics to obtain location generally did not match well with video locations (number of matches in parentheses). Methods using integrated (position) kinematics and incorporating angular motions (corrected) had the best match.