Capturing head impacts in boxing: a video-based comparison of three wearable sensors

Wearable sensors are used to quantify head impacts in athletes, but recent work has shown 2 that the number of events recorded may not be accurate. This study aimed to compare the 3 number of head acceleration events recorded by three wearable sensors during boxing and 4 assess how impact type and location affect the triggering of acceleration events. Seven boxers 5 were equipped with an instrumented mouthguard, a skin patch, and a headgear patch. 6 Contacts to participants’ head s were identified via three video cameras over 115 sparring 7 rounds. The resulting 5168 video-identified events were used as reference to quantify the 8 sensitivity, specificity, and positive predictive value (PPV) of the sensors. The mouthguard, 9 skin patch, and headgear patch recorded 695, 1579, and 1690 events, respectively, yielding 10 sensitivities of 35%, 86% and 78%, respectively, and specificities of 90%, 76%, and 75%, 11 respectively. The mouthguard, skin patch, and headgear patch yielded 693, 1571, and 1681 12 true positive events, respectively, leading to PPVs for head impacts over 96%. All three 13 sensors were more likely to be triggered by punches landing near the sensor and cleanly on 14 the head, although t he mouthguard’s sensitivity to impact location varied less than the 15 patches’ . While the use of head impact sensors for assessing injury risks remains uncertain, 16 this study provides valuable insights into the capabilities and limitations of these sensors in 17 capturing video-verified head impact events. 18


Introduction
Wearable devices are being used to record and quantify impacts to athletes' heads during sports participation.The information gained from these devices has been used to improve rules, 25,27 adapt practices, 2,3 and influence the design and use of protective equipment, 34 all aimed at increasing athlete safety.Injury risk curves have also been proposed to describe the relationship between concussion and head impact dynamics, 31 and several studies have evaluated the association between head impacts and physiological, neurological and cognitive detriments. 26,32 ,21,24,30 Among the first reports of spurious recordings -or false positive (FP) sensor events, i.e., when a sensor records an acceleration event, but the video review does not show an impact occurring -was the American football study by Hernandez et al. in 2015.13 The authors reported that 99% of all mouthguard measurements collected were spurious recordings.
Cortes et al. later found that 25 to 64% of events occurring during play were not associated with an impact observed on video. 7Across several studies, the proportion of FPs and events that could not be video-verified (e.g., player not in view) ranged from 31% to 98%. 1,5,7,20,21,29 There aso exist several studies reporting false negative events, where an impact to the head is visible on video but the sensor did not record anything.Identifying all true contact periods and/or individual impacts from video is time-consuming and has rarely been completed. 5,11,20,24,30 In uch studies, the false negative rate varied between 26 and 59%, and events of particular interest, such as impacts leading to an athlete displaying signs of concussion, have been missed by sensors.18 The combination of all these studies indicates that two-way video verification, where sensor recordings are matched to video events and video events are matched to sensor recordings, is an important step in the full assessment of a sensor's performance.20,24 Head impact devices record an event when one or more of the sensor's signals, typically a linear acceleration signal, reaches a pre-set threshold.However, actions such as jumping, running, or sitting down, have been shown to trigger spurious recordings in sensors attached to the skin or to headbands.21,29,30 When loosely coupled to the skull via soft tissue (the skin) or an external object (headband, helmet), sensors can move independently from the skull, 15,38 which affects the magnitude of the acceleration signal, triggering spurious recordings and resulting in overestimates of exposure.33,38 Instrumented mouthguards have been shown to be more tightly coupled to the skull 38 but still record spurious events, with spitting or chewing as possible causes.2 Therefore, it remains unclear whether there is a type of head impact sensor that is more accurate than others.
Most in-vivo studies involving one sensor have been hindered by a high proportion of events that could not be verified because the athlete was out of view of the cameras, the quality of the video was low, or the view of the head was obstructed. 11,37 n imprecise quantification of exposure to head impacts may hinder the advancement of our understanding of concussive injury. 10,28,37 I order to help develop more trustworthy devices, the factors that can result in false positives or false negatives need to be better understood.Thus, this study was designed to record head acceleration events using three sensors simultaneously, representing three types of coupling to the skull.The objectives were to: (1) compare the number of acceleration events recorded by the three sensors relative to events observed on video; (2) assess whether impact type, location, or participant affects the capacity of each sensor to be triggered; (3)    evaluate the agreement of the sensors to record events of particular interest for injury risks.

Materials and Methods
This observational cohort study took place from September to December 2020 in Auckland, New Zealand.Competitive amateur boxers were equipped with an instrumented mouthguard, a skin patch, and a headgear patch during sparring sessions.All sessions were recorded by multiple video cameras and videos were reviewed to identify every contact to the participants' head and body.Acceleration events recorded by the sensors were matched to video events and the number of true/false positives were analyzed.Additionally, a certified boxing judge reviewed a sample of the video data, identifying scoring punches, i.e., punches landing with good technique and a certain amount of force.This study was approved by the Auckland University of Technology Ethical Committee (AUTEC 20/153).
Experience in boxing ranged from 1 to 8 years; five participants had competed at the national level or higher, and one had no previous fighting experience.All regularly took part in weekly sparring sessions, which were organized in 3-minute rounds with 30-s breaks and lasted 40 to 60 minutes.Boxers were matched in weight and ability and wore headgear.No participant reported a history of diagnosed concussion nor sought medical attention for a head injury over the course of the study.
The participants sparred either in a boxing ring or a defined area on the floor.Using five cameras (GoPro Hero 3+ to Hero 7 Black, GoPro, Inc., San Mateo, CA, USA), three angles of video were always available for each area.The cameras recorded at 60 fps with a shutter speed of 1/120 s to minimize motion blur, and at a resolution of 1080p.To align the video and sensors in time, all sensors were placed on a stiff board that was dropped once in view of the primary camera, which was kept on for the duration of the session.Later, all videos were synchronized to the primary camera and the timestamps of the sensors were re-calculated to match the video timeline.

Instrumentation
The Prevent Biometrics Hybrid mouthguards (Prevent Biometrics Inc., Edina, MN) are boiland-bite mouthpieces that were molded to the upper jaw dentition according to the manufacturer's guidelines.They incorporate a triaxial linear accelerometer (ADXL372, Analog Devices, Boston MA, range ± 200 g) and a triaxial angular rate sensor (BMG250, Bosch, Gerlingen Germany, range ± 35 rad/s), both sampling at 3200 Hz (FIGURE 1A).The sensors are located in the vicinity of the first left lateral incisor. 23During use, any time the linear acceleration signal reached 10 g on any axis, the device recorded a 50-ms acceleration event, that was timestamped using the iPad clock with a 1-ms resolution.The Hybrid mouthguards incorporate the same hardware and software as Prevent Biometrics' other devices, which have been shown to identify head impacts with a positive predictive value (PPV) of 81.6% for the previous boil-and-bite version in American football and 96.4% for the custom mouthguard in rugby. 17Sx patches (CSx Systems Ltd, Auckland, New Zealand) used in this study measure 25 x 21 x 6 mm and weigh 4.5 g (FIGURE 1A).One CSx patch was taped behind the right ear (FIGURE 1B) using sports adhesive spray and wig tape (Walker Tape Ultra Hold, West Jordan, UT) and a second patch was taped to the back of the headgear using standard double-sided tape (FIGURE 1C).Each patch houses a triaxial linear accelerometer (ADXL375, Analog Devices, Norwood, MA, range ±200 g, sampling frequency 3,200 Hz) and a triaxial angular rate sensor (ITG-3701, Invensense, TDK, San Jose, CA; range ±4000°.s-1 , sampling frequency 8,000 Hz).A 10-g threshold on any linear acceleration axis was also used and each 50-ms recording was timestamped with a resolution of 1 second using the device's GPS-synched clock.The Prevent Biometrics mouthguards and the CSx patches both incorporate a proprietary classification algorithm that determines whether the acceleration event is a true head impact or a spurious recording.All recordings were included in this study, independent of the manufacturer's classification.Sparring rounds where a participant's sensors collected data from start to end were used.The start and end times of each round were identified from the primary video, and acceleration events that occurred outside of a round were excluded for all sensors.Sensor data were processed using custom MatLab scripts (R2019a, MathWorks, Natick, MA).

Video processing
Video coding and verification were carried out using Nacsport Elite 6.0.0 software (Nacsport, Canary Islands, Spain).For each round, all video angles were synchronized in Nacsport and a single coder (author ELF) carried out the identification and characterization of all contact events, watching the videos at 30% of their normal speed, and reviewing each event several times from all angles.Contact events were characterized based on definitions agreed upon between the primary coder (ELF) and a combat sports expert (author SL), and included definite head impacts, probable head impacts, definite body impacts, and probable body impacts (TABLE 1).Definite head impacts were classified as prolonged head impacts if the glove lingered against the head.All head-related events were further defined as direct or indirect 35 (TABLE 1, FIGURE 2).A highly inclusive approach to capturing impacts was followed, where every contact event was coded, independent of its observed magnitude.The coder used objective criteria to identify contact events, such as the deformation of the glove or a sudden change in the head and/or the punching glove's velocity.Sixteen head impact areas were identified on an image of a helmeted head (FIGURE 3).The vicinity of a sensor was defined as a circular area around each sensor's theoretical location, seized to mimic the area of the knuckle part of a boxing glove (10-15 cm in diameter).Similarly, the face area was aposteriori defined as a rectangular area, such that in most cases, a punch landing in this area would not hit the headgear.Based on a test-retest protocol using the first minute of 10 randomly-chosen rounds, the primary coder demonstrated a total agreement in sequence 6 of 90.2% (95%CI: 74.7-96.7%),or a difference of 1.1 ± 2.3 events per minute of sparring.

Type of event Definition
Video coding events Definite head impact Every time the head or headgear is visibly hit or touched.
Headprolonged contact When an arm or glove lingers against the head of the opponent, potentially pushing it and rubbing against the headgear; also called "framing" in boxing.
Probable head impact Every time a head impact is suspected but the video does not allow to be certain (e.g., no good angle, obstruction).

Body impact
Every time the body is visibly hit.
Probable body impact Every time a body impact is suspected but the video does not allow to be certain (e.g., no good angle, obstruction, loose clothing).

Clinch
Event where a boxer holds the opponent's body and/or arms with one or both of their arms to prevent or hinder the opponent's punches or movements. 35This may cause the heads to hit or rub against each other.

Artefact
Any event where the sensor may record something not related to a head impact (e.g., the boxer pushes the headgear back in place, or touches the headgear patch at the back of their head).

Head impacts only
Direct impact A hit that visibly lands on the head with the knuckle part of the glove without any obstruction and bounces off in the opposite direction.

Indirect impact
A hit that is partially or fully blocked (contact to the head with either boxer's glove or arm), partly deflected, or glancing.

Certified official's events
Scoring punch An impact that lands effectively, i.e., with good technique. 14,35 coring punches contact cleanly with the head and may therefore transmit a higher force.
Hard punch Any punch that seems to "shake" the boxer and that a referee would be looking for to assess whether the boxer was fit to continue.

Video verification
The sensors' recordings were first manually aligned to each other, then to the video events, accounting for a ±2 s windows around the mouthguard's timestamp due to the patch's lower time resolution.The video events served as reference and if a sensor recording could be matched with a video event, it was marked as a true positive (TP).If a sensor recording existed but could not be matched with a video event, it was marked as a false positive (FP).The video events not associated with an acceleration event were marked as false negative (FN).The criteria for video-verified impacts previously established by Cortes et al. 7 were adapted for this study: (1) the video setup ensured that the participants were visible on video at all times; (2) we differentiated between definite impacts and probable impacts when we couldn't be certain of the contact to the head; and (3) because boxers can receive several punches to the head in one second (i.e., the patches' timestamp resolution), we marked the sensors recordings where we couldn't identify with certainty which video event they were associated with.

Analyses
The video and sensors datasets were described using mean and standard deviation (SD), and median and interquartile range (IQR).For each sensor, the sensitivity and positive predictive values (PPV) to record events observed on video were calculated with their 95% confidence intervals (CI).Sensitivity values were considered different if their 95% CI did not overlap.The analyses focused on definite head impacts and prolonged contacts to the head, as they were the only types of punches that were identified as scoring punches by a boxing judge (see below).The sensitivities and their 95% CI were compared with respect to the type of head impact (direct vs. indirect), the impact location (in bins and respective to the sensor), and by participant.
To evaluate whether the sensors were accurately capturing events that may be relevant from a scoring or refereeing perspective, 14 an active IBA-certified boxing judge and referee independently reviewed 30 randomly-selected 1-minute-long video segments of sparring (9% of the total dataset).The official identified any punch that would influence a judge's assessment (e.g., "scoring punches") 14 and a referee's assessment (see TABLE 1 for definitions).The official also looked for foul punches, which have a higher injury potential.
Eight of the video segments were presented to the official twice, without their knowledge, to determine the judging's reliability.The assessments were in total agreement 6 for 50% of the At the end of each sparring session, the participants were asked whether they had sustained "one (or more) heavy blow to the head that "rocked" or "fazed" them, left them feeling dizzy, off-balanced or with blurry vision, even for a short time".When possible, heavy blows were identified on video using the description reported by the participant, matched to sensor data, and qualitatively described.

Results
A total of 115 rounds of sparring were analyzed, from which 5168 video events were coded.
Definite head impacts and prolonged contacts to the head comprised 57% of the video events (TABLE 2), for an average of 8.4 ± 5.1 contacts per minute (median: 7.7 [IQR: 5.1 -10.3], range: 1.0 -32.9)(seeFIGURE 4 for an example round).Definite head impacts and prolonged contacts to the front of the head totaled 69%, followed by the right (16%) and left (13%) sides of the head.Direct contact represented 30% of the definite head impacts and 11% of the prolonged contacts (TABLE 2).Totals of 27%, 6%, and 1% of observed head impacts landed in the vicinity of the mouthguard, skin patch, and headgear patch, respectively, as defined in FIGURE 3. Rec+: sensor recording present, Rec-: sensor recording absent; the percentages represent the proportion of the sensor dataset that is associated with the type of impact, e.g., 87% of mouthguard recordings were associated with a definite impact to the head.Artefacts and false positives included headgear contact from a body punch landing on the bottom of the participant's guard, punches being thrown, and headgear being adjusted.We verified 695 mouthguard, 1579 skin patch, and 1690 headgear patch events (TABLE 2).
We matched 693 mouthguard events, 1571 skin-patch events, and 1681 headgear-patch events to video events, equaling TP rates over 99.5% for each sensor.Participants often sustained multiple impacts within a couple of seconds, therefore, 14-17% of the TPs could not be matched with one specific video event (they were marked as TPs, but the type and impact location were uncertain).Both patches recorded twice as many events as the mouthguard and all sensors were triggered more frequently from definite head impacts than other types of impacts, resulting in higher sensitivities (TABLE 2, FIGURE 5).The PPV for a sensor recording to be associated with a head impact (definite head impacts, prolonged contacts, and probable head impacts combined) was 97.1% for the mouthguard, 96.3% for the skin patch, and 96.1% for the headgear patch.
All three sensors were more sensitive to direct than indirect head impacts, but similarly sensitive to impacts to the face area (FIGURE 5).All three sensors also showed differences in sensitivity related to the location of impact (FIGURE 5, FIGURE 6).Specifically, the patches were significantly more sensitive to impacts occurring in their vicinity (skin patch: 72.2%  6).Inter-participant sensitivity differences existed and were consistent across all three sensors (e.g., all three sensors seem more sensitive for participant 03 than for participant 04)(FIGURE 7).The certified official identified 50 scoring punches (44 head impacts, 5 prolonged contacts, 1 body punch).There was no foul punch or punch that visibly shook any of the participants.The number of scoring punches to the head averaged 1.6 ± 1.8 punches per minute (range: 0 -8, median: 1 [IQR: 0 -2.25]).The mouthguard, skin patch, and headgear patch showed sensitivities of 35%, 86% and 78%, respectively, and specificities of 90%, 76%, and 75%, respectively (TABLE 3).Five heavy blows to the head were reported by three participants over 18 sparring sessions; only two could be identified on video.Both were sustained by the same participant, and resulted from rear hooks to the left ear area that triggered all three sensors.MG: mouthguards, SK: skin patch, HG: headgear patch, Rec+: recording present, Rec-: recording absent.a Video events include every type of events: head impacts, prolonged contacts and probable head impacts, clinches, definite and probable body impacts.b The "No scoring punches" are all events that were coded by the primary coder, but that were not identified by the certified official as a scoring punch.

Discussion
The goals of this study were to: (1) compare the number of acceleration events recorded by a mouthguard, a skin patch, and a headgear patch used simultaneously; (2) assess whether various factors affect the recording of an acceleration event; and (3) evaluate the agreement of the sensors to record scoring punches or heavy blows.Overall, the patches recorded twice as many events as the mouthguard.When compared to the video, all three sensors recorded very few FPs, leading to PPVs of over 96% for head impacts.In contrast, a large proportion of head contacts identified on video did not generate a sensor recording (78% for the mouthguard, 53% for the skin patch and 49% for the headgear patch).All three sensors were more likely to be triggered by definite head impacts and direct head impacts and showed difference sensitivities to impact locations.While this was true for all three sensors, the mouthguard's sensitivity across impact types varied less than the patches'.With respect to scoring punches, the skin and headgear patches were more sensitive than the mouthguard (86 and 78% vs 35%) but less specific (76 and 75% vs 90%).

Numbers of acceleration events
The proportion of FPs recorded by the sensors was low (0.3-0.7%).In the conservative scenario where all acceleration events associated with body impacts and clinches were artefacts, this proportion was still lower than 4%.These are amongst the lowest numbers reported, as FP proportions have varied between 3.6 and 89% with instrumented mouthguards in rugby, 17 soccer, 30 and American football, 11,17,20 and 15 to 90% with skin patches in Australian rules football 24 and soccer. 29The differences may have arisen from our definition of true/false positives and our inclusive video criterion.A more restrictive criterion, e.g., classifying as TP only the impacts that visibly modify the trajectory of the head, 11 would likely have resulted in a higher proportion of FPs.The low proportion of FP events may also be a consequence of the sport we studied.First, there is no running, and only limited jumping during boxing, two common causes of FPs from patches. 29,30 econd, a round of sparring means three minutes of intense activity, with limited occasions to manipulate the mouthguard or headgear.Third, sparring was dictated by a bell announcing the start and end of each round, and we used this to exclude all acceleration events that occurred outside of sparring time.
Verified impacts from the mouthguard, skin, and headgear patches represented 22%, 47%, and 51% of the video events, respectively.Other studies in team sports have found mouthguards and skin patches sensitivities around 71% with a 5-10 g triggering threshold (range: 41 to 98%). 5,11,20,21,30 In ur approach, the inclusion of all contact events independent of magnitude did not match the methods used by the sensors, which are designed to ignore low-magnitude impacts.A less inclusive video coding approach may have resulted in higher sensitivity numbers.Given that the sensors and video coding approaches were not equally inclusive, the overall sensitivity cannot be used alone to evaluate the performance of the sensors.Nonetheless, the differences in sensitivity between the sensors can help understand how various factors affect the sensors' capacity to be triggered.
Both patches were triggered more than twice as often as the mouthguard, despite all sensors using a 10-g threshold.The only studies comparing mouthguard, skin patch and headgearbased device were conducted in the laboratory with semi-controlled impacts, 19,38 making a direct comparison with our study difficult.Yet, these studies and others have shown that skin patches, skull caps and helmets are not rigidly coupled to the skull. 16,38 hen a sensor decouple from the skull during or because of an impact, it sustains larger accelerations, 16,19,33,38 which might trigger the recording of acceleration events while the skull's motion is below threshold.In this study, we often observed the headgear sliding back on the head from a punch and the skin patch sometimes loosening due to intense sweating.Additionally, we found an increased sensitivity to impacts in the vicinity of the sensors.If the sensors were rigidly fastened to the skull, the location of the impact (excluding a direct contact) should not influence their capacity to be triggered to that extent.Furthermore, while the patches were twice as sensitive to definite head impacts as the mouthguard, they were three-to-four times more sensitive to prolonged contacts and probable head impacts.These types of impacts may have made the patches move enough to trigger the sensors, while the acceleration at the mouthguard stayed below threshold.While decoupling is likely the main cause of the differences in event numbers between the mouthguards and patches, we cannot exclude the possibility that differences in sensor design or technology may also have contributed to this disparity.Nonetheless, our findings raise an important point: while most of the sensor recordings were video-verified, it was impossible to know if a recording was triggered by true head motion or by skull/sensor decoupling.Ensuring proper coupling should therefore be a priority.

Implications for impact monitoring in boxing
The IBA-certified official's identification of scoring punches enabled us to assess the sensors' accuracy in recording head impacts.The PPV for all three sensors was similar (28-31%), indicating they were all moderately effective in distinguishing between scoring and nonscoring punches, provided the event was recorded.The patches were more sensitive (78-86%) than the mouthguard (35%) in detecting scoring punches, making them more useful for quantifying the number of punches classified as scoring.However, due to their moderate specificity, patches would not be well suited for scoring fights or assessing an athlete's actual exposure.Although both patches were sensitive to impact location, the back of the head was less frequently impacted than the side (2% vs. 16%), reducing the issue of an overpowered proportion of events recorded to the patch's side of the head.The headgear patch was also more convenient to use, adhering better than a skin patch and easier to check or replace without removing the headgear, and would therefore be the most suitable sensor for independent impact counting.However, our findings do not allow us to conclude whether any of the sensors can predict injury risk.The number of events recorded by a sensor may overestimate exposure if an acceleration event is an artefact resulting from skull/sensor decoupling rather than a valid measurement of head motion.The analysis of kinematic signals is required to determine the prevalence of artefacts and the sensors' potential to comprehend exposure and injury risks.

Strengths and limitations of the study
Boxing sparring provided a suitable environment for the verification of head impacts.The multiple cameras covering only two athletes in a small defined space and the limited clinching and grappling made identifying contacts to the head easier than in many other sports.As a result, less than 0.2% of sensor events could not be verified, much less than the 15-20% proportion of unverifiable sensor events reported in previous studies. 1,36 onetheless, the number of probable head impacts, i.e., when contact to the head could not be asserted (vs. definite head impacts or prolonged contacts), represented 20% of all our video events.It is possible that additional cameras or better positioning may have reduced this proportion, but we were restricted by the number of cameras available, the need to capture two sparring areas at the same time, and the space available in the gym.
Our highly inclusive approach to identifying contact events on video did not match the sensor's processes and included many low-magnitude events that may be of little interest in the study of concussive injury risks.Our approach likely underestimated the sensors' sensitivity, but also over-estimated the PPV, as it was common for the skin and/or the headgear patches to record an event while the punch appeared to have little effect on the boxer's head motion.We did not attempt to visually estimate the severity of each head impact from the video, as has been done previously, 11,20 because the reliability of such an approach relies on definitions and criteria that we did not believe could be accurately applied to video data.A related limitation was the suboptimal test-retest reliability of the primary coder and IBA-certified official, especially for the latter who reviewed the videos at full-speed, making the already complex and controversial task more prone to errors. 4,9,14 Immary, this study investigated the capability of three head impact sensors to record video-observed head impact events.Our setup and procedures during boxing sparring allowed us to verify the large majority of events recorded by the sensors.As a result, the number of spurious events was one of the lowest reported in head impact research and the positive predictive value was high for all sensors.Our approach to identifying head contacts on video, which differed from the sensor's threshold-based triggering, provided a detailed comparison of the sensitivities of the mouthguards and patches.Notably, our findings indicated that patches recorded more acceleration events than the mouthguard and were more sensitive to impacts occurring in their proximity.These observations suggest that all sensors may present varying degrees of decoupling from the skull upon impact, highlighting the importance of ensuring proper coupling in future research.Furthermore, the headgear patch demonstrated ease of use and higher sensitivity to punches, which could make it a valuable tool as an impact counter.While the use of head impact sensors for assessing injury risks remains unclear, this study provides valuable insights into the capabilities and limitations of these sensors in capturing head impact events during boxing sparring.

Conflicts of interest
Author GPS is a shareholder and director of MEA Forensic Engineers & Scientists, a forensic consulting company that may derive benefit from being involved in this study.

FIGURE 1 .
FIGURE 1. A) A CSx patch (bottom-left) and a new, unmolded Prevent Biometrics Hybrid mouthguard (topright).B) Placement of the CSx patch on the flat aspect of the right mastoid process (indicated by the yellow arrow).C) Placement of the CSx patch at the back of the participant's headgear.

FIGURE 2 .
FIGURE 2. Examples of head impacts observed on video.A) Direct head impact, landing cleanly with the knuckle part of the glove, without obstruction.B) Indirect head impact, with contact to the head made by the participant's own glove.C) Indirect head impact from the palm side of the glove, with the punch glancing about the face.D) Indirect head impact, with the punch being deflected by the participant's arm before making contact.

FIGURE 3 .
FIGURE 3. Image used on Nacsport to manually point at the location of contact to the participants' head (the same image was used for all participants).Pointing at the image would automatically categorize the impact in a location bin (based on the bins defined by Greenwald et al. 12 ) and record the coordinates of the point.The blue, orange, and red circles represent the area around the mouthguard, skin patch, and headgear patch, respectively, with an arbitrarily defined radius to mimic the size of the knuckle part of a boxing glove.The green rectangle represents the face area, where for most participants, a punch would not hit the headgear.

FIGURE 4 .
FIGURE 4. Representation of the events over time for one round of sparring (3 minutes), after video verification.The scoring punches were only coded for the first minute of the round, shaded in grey.Each vertical line represents one event, and all lines are the same width; lines that appear thicker mean there were two or more events in close proximity.Video events include head impacts and prolonged contacts only.

FIGURE 5 .FIGURE 6 .FIGURE 7 .
FIGURE 5. Sensitivity of the sensors to definite head impacts and prolonged head contacts relative to various parameters.The event numbers correspond to the video-coded events for each category.The diamonds represent the sensitivity and the lateral bars the 95% confidence intervals.The colored dashed vertical lines represent the overall sensitivity for each sensor.MG: mouthguard, SK: skin patch, HG: headgear patch.The sensitivity between two conditions was considered different if the 95% confidence intervals did not overlap.