Head impact biomechanics across men’s and women’s contact sports: a comparative and clustering analysis Authors: Zaryan Masood, David Luke, Rebecca Kenny, Daniel Bondi, Adam Clansey, Lyndia C. Wu Published: 2025 ABSTRACT Sports head impacts have been associated with both acute and long-term brain trauma. While wearable sensors provide valuable biomechanics insight, most studies focus on single sports, and the variability in sensor methodologies limits cross-sport comparisons. Our objectives were to conduct a multisport comparison and clustering of head impact biomechanics features implicated in brain injury risk. We uniformly processed a multisport dataset gathered using instrumented mouthguards containing direct head impacts in men’s football, men’s hockey, women’s rugby, and women’s soccer. We statistically compared directional and resultant peak kinematics, impulse durations, and impact directionality metrics. Then, we applied unsupervised k-means and t-distributed stochastic neighbour embedding (t-SNE) models to examine clustering in impact magnitude and frequency features. Statistically significant cross-sport differences were found in all biomechanical features. Men’s football exhibited the highest resultant median peak kinematics, while women’s soccer showed lowest median resultant kinematics. However, directional comparisons revealed unexpected trends such as women’s soccer impacts exhibiting high sagittal kinematics relative to other sports. Clustering analyses grouped impacts into low and high magnitude/frequency clusters that transcended sport boundaries, with only women’s soccer impacts demonstrating tight clustering patterns due to consistent heading biomechanics. We uniquely curated a standardized dataset for multisport head impact biomechanics comparisons. Cross-sport differences in under-investigated biomechanical features such as directional peak kinematics may need to be further examined for potential sport-specific injury risk considerations. Despite substantial gameplay differences, we found interesting shared biomechanical patterns across sports, warranting joint analyses to inform implications in protective equipment design and injury prevention strategies.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-31145-4. Introduction Concussion, or mild traumatic brain injury (mTBI), continues to be a major health concern for many contact sport athletes1 and a leading cause of disability, especially in youth and adolescents2. More recently, there is a growing body of literature suggesting repeated subconcussive events – head impacts that do not lead to clinical concussion diagnosis – can give rise to detectable neurophysiological changes3. To further understand how head impacts during sport are influencing brain health, researchers have applied wearable sensors to collect on-field head impact biomechanics data. Data from these sensors can be leveraged to quantify potential brain injury risk through head kinematics measures such as peak linear and angular accelerations4 as well as simulated brain strains5. Various studies have deployed sensors to gather head impact data in sports, identifying common impact scenarios and mechanisms of head contact. Notably, a number of factors may influence head impact biomechanics across different sports. For example, protective headgear, such as helmets, may reduce peak head acceleration by dissipating impact energy6 and increasing impact duration7. Furthermore, the mechanics of impacting objects can also affect impact biomechanics, such as stiffness and damping differences between common helmet-to-helmet impacts in American football8 and helmet-to-board impacts in ice hockey9. Ice hockey, with typically higher speed on-ice collisions than other field sports, may present greater momentum and higher force transferred during impacts10. These are just a few factors influencing impact biomechanics, with others—including gameplay dynamics and sex—also contributing to differences across sports. However, cross-sport comparison of head impact biomechanics continues to be a challenge given the varying sensor types employed, differences in sensor specifications, and non-standardized data processing techniques. More recently, instrumented mouthguards (iMGs) have become the standard sensor type in recent field studies11, yet differences in sensor specifications and data processing methods still make comparing results across iMG studies a challenge. For example, Jones et al.12 demonstrated substantial differences in both on-field and in-lab peak kinematics across four commonly used iMGs for the same impact scenarios. Additionally, Gellner et al.13 highlighted sensitivity in peak kinematics to varying filter cut-off frequencies. These variations, combined with the tendency for most studies to focus on single-sport data collection and analyses, make cross-sport comparison of head impact biomechanics difficult. Furthermore, most analyses concentrate on peak head kinematics, and there is limited reporting of other biomechanical parameters implicated in brain injury risk, such as impulse durations14, directionality15 and frequency characteristics16. Uniquely, we curated raw iMG impact data across multiple sports and processed all data with a uniform pipeline to enable more controlled comparison of head impact biomechanics across university-level men’s ice hockey, men’s American football, women’s soccer, and women’s rugby. With this dataset, we had two primary objectives: (i) compare biomechanical features of iMG-measured, uniformly processed direct head impacts across sports, and (ii) use clustering methods to further examine if sports distinctly cluster by biomechanical features. We investigated diverse biomechanical features, including impact magnitude, impulse duration, acceleration direction, and frequency features that have been linked to brain injury risk. Methods Methodology overview In our comparison, we included only video-verified direct head impacts and conducted a rigorous data quality review to ensure a high-confidence dataset. As the curated dataset was collected using three different iMGs, we developed the uniform processing pipeline to address sensor setting differences such as varying recording threshold, recording duration and sample rate, as well as ensure standardized filtering methods. Using this standardized multisport dataset, for objective (i), we examined and compared the distributions of biomechanical features across sports, including impact magnitude, impulse duration, and directionality-focused features. For objective (ii), we applied an unsupervised k-means clustering algorithm and a t-distributed stochastic neighbour embedding (t-SNE) model to examine the clustering of impact magnitude and frequency features and determine if sport-specific clusters arise. Dataset descriptions A head impact database was curated from previously collected iMG data from university level men’s American football17,18, women’s soccer19, men’s ice hockey20, and women’s rugby21 teams. Both men’s hockey and women’s rugby teams wore Prevent Biometrics22 iMG’s. The women’s soccer team wore instrumented mouthpieces from Wake Forest Center for Injury Biomechanics23 and men’s football wore Stanford University iMG’s24. All iMG impacts were video-verified and only direct head contacts were included in the analysis (Supplementary Fig. 1). This enables comparison of direct head impact biomechanics separately from inertial head acceleration events due to body contact, which can have very different mechanisms and are also usually more difficult to confidently verify through video analysis. All impact traces were visually inspected to keep only recordings where clearly observable linear and angular acceleration impulses could be identified around the time of trigger (Supplementary Fig. 2). The Prevent iMG, Wake Forest Mouthpiece, and Stanford University iMG all had varying initial recording settings including sampling frequency, trigger thresholds, and impact durations. Coupling of the iMG to the teeth was assessed for Prevent iMG-recorded impacts using proximity sensor data21, where only well-coupled impacts were included for analysis. Part of the Stanford iMG dataset from Wu et al., 201718 used iMGs that included a similar proximity sensor to include only well-coupled impacts. The Wake Forest instrumented mouthpiece was a retainer-style device that was difficult to decouple from teeth without fully removing from the mouth, providing confidence of device coupling. Table 1 includes the initial settings of all iMGs which were then harmonized to follow one uniform processing pipeline. All subjects were consented, and each dataset had research ethics board approval as detailed in the original publications. Table 1Initial sensor specifications and uniform processing settings. Table 1Initial sensor specifications and uniform processing settings. Initial sensor specifications and uniform processing settings. Processing methodology We applied a uniform pre-processing and processing script to raw sensor data to minimize potential sampling and processing differences (Table 1). We down-sampled the Prevent iMG to 1000 Hz using the resample function, and truncated the impact duration window for the Wake Forest mouthpiece and Stanford iMG to match the Prevent iMG (10 ms pre-impact, 40 ms post-impact). To account for trigger differences, we chose a uniform 10 g resultant linear acceleration trigger for all data. We aligned and centered all the impacts to ensure a trigger at 10 g resultant with a 50 ms impact window. Data were rotated to align with anatomical axes. Following rotation, data were filtered using a low-pass 4th order Butterworth filter with cutoff frequencies of 300 Hz for linear acceleration and 180 Hz for angular velocity. Filter cutoff values were selected based on the limits of individual sensor bandwidth ranges. Following filtering, a fourth-order differentiation was applied to angular velocities to determine angular accelerations. Linear accelerations were not transformed to head center of gravity (CoG), as we did not have sufficient information from all datasets to apply subject-specific transforms, and the standard approach of transforming based on 50th percentile male human head dimensions can introduce systematic bias into the analysis. As such, all linear head kinematics analyses are based on kinematics at the upper dentition location, while rotational kinematics are invariant across the head assuming rigid body motion. Feature extraction and analysis To compare head impact biomechanics across sports, we computed key biomechanical features that are commonly used in head impact biomechanics research and have known injury risk implications. These features were calculated for each recorded head impact across all sports in our dataset. Directional magnitude features Peak linear acceleration (PLA) in anterior-posterior (AP), left-right (LR), and inferior-superior (IS) axes, as well as peak angular velocity (PAV), and peak angular acceleration (PAA) in coronal (cor.), sagittal (sag.) and horizontal (hor.) planes were computed for each impact. Resultants (R.) were also computed for PLA, PAV, and PAA. Impulse duration features Linear and angular acceleration impulse durations were determined from the maximum linear and angular acceleration directions for each impact, respectively. Impulse durations were extracted by calculating the width of the impulse at half of the peak acceleration value of the maximum axis. Therefore, each impact consisted of a linear acceleration and an angular acceleration impulse duration. Acceleration direction features Peak acceleration directions were computed by determining the peak resultant acceleration value and finding its corresponding per axis values to create a 3D acceleration vector, for both the linear and angular acceleration individually. The peak acceleration direction vectors were then divided into six direction categories (forward/anterior, backward/posterior, leftward, rightward, upward/superior, downward/inferior), separated by the 45-degree oblique directions in each anatomical plane. Sagittal and coronal ratios During a head impact, linear and rotational head kinematics are intrinsically coupled due to the complex head-neck-torso system’s response to external impact forces. We studied the coupling between linear and rotational kinematics in sagittal and coronal impacts by calculating the linear-angular acceleration ratio. For sagittal impacts (> 80% of peak angular acceleration in sagittal plane), we calculated the sagittal acceleration ratio (rSagittal) by dividing anterior-posterior linear acceleration by sagittal angular acceleration at the moment of peak angular acceleration. Similarly, for coronal impacts (> 80% of peak angular acceleration in coronal plane), we calculated the coronal ratio (rCoronal) by dividing left-right linear acceleration by coronal angular acceleration. Frequency features We extracted wavelet transform (WT) features to examine signal characteristics in the time-frequency domain18. Morlet WT was performed on the linear and angular accelerations in all directions. For each impact, the peak frequency at the peak WT amplitude in the time-domain, was computed. Further, we estimated the bandwidth of each impact by identifying the range of frequencies where the WT amplitude remained above 50% of its peak value (full-width at half maximum). Statistical methods and clustering analysis For objective (i), Kruskal-Wallis test was used to test for significance, followed by a post-hoc pairwise comparison with Bonferroni correction to statistically compare directional magnitude, impulse duration, as well as rSagittal and rCoronal features across sports. Post-hoc analyses were performed in MATLAB using the multcompare function (‘CType’,‘bonferroni’), which conducted all six possible sport-to-sport comparisons within each metric with a Bonferroni-adjusted significance threshold of α = 0.05/6 = 0.0083. Following the Kruskal-Wallis test, we also calculated the epsilon square (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2) to determine effect size, with the classification of 0.01 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\le\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\le\:$$\end{document} 0.06 as a small effect, 0.06 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\le\:\epsilon\:$$\end{document}2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\le\:$$\end{document} 0.14 a medium effect, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\ge\:$$\end{document} 0.14 as large effect25. A Chi-squared test for independence, followed by a post-hoc pairwise comparison with Bonferroni correction, was used to determine if any significant association existed between sports and the distribution of acceleration direction features. For objective (ii), we applied a k-means clustering algorithm for two separate feature groups (directional impact magnitude features and frequency features) to examine clusters in directional magnitude distributions and in impact frequency characteristics. Only impact magnitude and frequency features were used for k-means clustering, as we expected time-frequency domain features to capture the impulse duration information, and acceleration direction features did not form clear clusters (low silhouette scores) in preliminary analysis. Additionally, coupling ratios were only computed for sagittal and coronal impacts, and thus not available for including as a feature for all impacts. After our k-means analysis, we then trained a t-SNE model on all combined magnitude and frequency features to project high-dimensional non-linear features into 2D space, while maintaining local relationships, thus optimizing for cluster identification with the high-dimensional feature space. Before applying the k-means clustering algorithm, we normalized features using the min-max method prior to applying the MATLAB kmeans function (R2023b, The MathWorks, Inc., Natick, MA). We initialized seeds by implementing the MATLAB kmeans++26 algorithm and computed silhouette scores to evaluate the effectiveness of the clustering for each feature group27. To determine the optimal number of clusters, we generated an elbow plot using average silhouette scores. We identified the features contributing most to centroid movement by analyzing within-cluster sum of squares (WCSS)28,29. For the t-SNE algorithm30, following recommended pre-processing techniques30, we reduced the data to two dimensions using a principal component analysis (PCA). Based on the dataset size, we selected a perplexity value of 30 (default value for this data size) to define the effective number of local neighbors for the model. Results Database Our database consisted of 1250 total video-verified direct head impacts, after data quality screening and re-thresholding. Of the 1250 impacts, 146 were from men’s hockey (20 athletes and 80 athlete-events), 232 from men’s football (10 athletes and 22 athlete-events), 306 from women’s rugby (19 athletes and 57 athlete-events), and 566 from women’s soccer (13 athletes and 146 athlete-events), where each athlete-event represents one instrumented athlete participating in a single practice or game. Magnitude features Significant differences in peak kinematics across the four sports were identified using the Kruskal-Wallis test (Fig. 1, Supplementary Tables 1 & 2). Effect sizes indicated medium to large sport-based differences for PLALR (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2 = 0.40), PAVcor (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2 = 0.33), and PAVR (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2 = 0.31), suggesting these metrics vary substantially by sport. In contrast, PLAAP (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2 = 0.04) and PAAR (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}2 = 0.13) showed weaker effect sizes. For PLA, men’s football presented the highest median PLAAP and PLALR (10.8 g and 10.6, respectively) while women’s soccer displayed the highest PLAIS (11.2 g). Additionally, men’s football demonstrated the highest 95th percentile PLAAP (27.5 g) and PLAIS (25.3 g) while men’s hockey had the highest 95th percentile PLALR (30.7 g). Post-hoc analyses highlighted most pairwise comparisons were significant, except those involving men’s hockey vs. women’s rugby. Additionally, comparisons between men’s hockey vs. women’s soccer, and women’s rugby vs. women’s soccer for PLAAP, men’s football vs. women’s rugby for PLALR, and men’s hockey vs. men’s football for PLAIS were not significantly different. For PAV, men’s football displayed the highest median PAVcor (6.8 rad/s) and PAVsag (6.3 rad/s), while men’s hockey displayed the highest median PAVhor (5.7 rad/s). Men’s football demonstrated the highest 95th percentile PAV for all axes. Post-hoc pairwise comparisons consistently showed no significant differences between men’s hockey and women’s rugby, or between men’s hockey and men’s football, across all axes. For PAA, men’s football displayed the highest median PAAcor (1332 rad/s2), PAAsag (966 rad/s2), and PAAhor (749 rad/s2), and highest 95th percentile PAA for all axes. Interestingly, men’s hockey vs. women’s rugby remained the only non-significant comparison across PAA. Across all resultant values, men’s football had the highest median PLAR (17.5 g), PAVR (11.1 rad/s), and PAAR (1789 rad/s2), while women’s soccer had the lowest (13.1 g, 5.5 rad/s, 786 rad/s2). Post-hoc analysis revealed that most pairwise comparisons were significant, with a few exceptions. For PLAR, differences were not significant between men’s hockey and women’s rugby, men’s hockey and women’s soccer, or between men’s football and women’s rugby. For PAVR, no significant differences were found between men’s hockey and women’s rugby, men’s hockey and men’s football, or men’s football and women’s rugby. Lastly, for PAAR, all pairwise comparisons were significant except between men’s hockey and women’s rugby. Fig. 1Per axis median (A) peak linear acceleration (PLA), (B) peak angular velocity (PAV), and (C) peak angular acceleration (PAA) across all sports. Scatter points indicate all impacts for each sport. Boxplots indicate median (horizontal black line) with 25th and 75th percentiles. AP: Anterior-Posterior, LR: Left-Right, IS: Inferior-Superior, Cor: Coronal, Sag: Sagittal, Hor: Horizontal, R: Resultant. Per axis median (A) peak linear acceleration (PLA), (B) peak angular velocity (PAV), and (C) peak angular acceleration (PAA) across all sports. Scatter points indicate all impacts for each sport. Boxplots indicate median (horizontal black line) with 25th and 75th percentiles. AP: Anterior-Posterior, LR: Left-Right, IS: Inferior-Superior, Cor: Coronal, Sag: Sagittal, Hor: Horizontal, R: Resultant. Impulse duration Comparing linear acceleration impulse durations of the maximum axis, men’s hockey and women’s rugby exhibited longer median durations at 11 ms and 10 ms, respectively, while men’s football and women’s soccer were shorter at 9 ms and 8 ms (Fig. 2A). For angular acceleration impulse durations, men’s hockey again displayed the longest median duration of 10.5 ms, closely followed by women’s rugby at 10 ms. Men’s football and women’s soccer were lower at 7 ms and 8 ms, respectively (Fig. 2B). Statistical comparisons revealed significant differences in linear impulse durations across sports using the Kruskal-Wallis test, however post-hoc tests highlighted non-significant findings between men’s hockey vs. women’s rugby, men’s football vs. women’s rugby, and men’s football vs. women’s soccer (Supplementary Table 3). Similarly, angular impulse durations were significantly different between all sports, except for men’s hockey vs. women’s rugby and men’s football vs. women’s soccer. (Supplementary Table 3). Fig. 2Median linear (A) and angular (B) impulse duration across all sports. Scatter points indicate all impacts for each sport. Boxplots indicate median (horizontal black line) with 25th and 75th percentiles. Median linear (A) and angular (B) impulse duration across all sports. Scatter points indicate all impacts for each sport. Boxplots indicate median (horizontal black line) with 25th and 75th percentiles. Acceleration direction Linear directionality analysis showed men’s hockey to display a predominant proportion of leftward (21.9%, 32/146 impacts) and rightward (22.6%, 33/146 impacts) accelerations. Men’s football had primarily backwards impact accelerations (34.3%, 80/232 impacts) while women’s soccer consisted of primarily upwards accelerations (59.4%, 336/556 impacts). Women’s rugby had a more distributed impact classification with most impacts being leftward (32.6%, 99/306 impacts) accelerations (Fig. 3A; Table 2). Chi-squared tests and post-hoc comparisons showed statistically significant differences between all sports for linear acceleration directions. Fig. 3Linear (A) and angular (B) classification of head accelerations for each sport in all axes. Created in BioRender. Wu, L. (2025) https://BioRender.com/kx9ev5m. Linear (A) and angular (B) classification of head accelerations for each sport in all axes. Created in BioRender. Wu, L. (2025) https://BioRender.com/kx9ev5m. Table 2Peak linear acceleration directions for all sports. Table 2Peak linear acceleration directions for all sports. Peak linear acceleration directions for all sports. Table 3Peak angular acceleration direction for all sports. Table 3Peak angular acceleration direction for all sports. Peak angular acceleration direction for all sports. Men’s football displayed the greatest proportion of impacts where the skull rotates in the coronal plane (leftward: 28.4%, rightward: 36.9%, 151/232 impacts). Women’s soccer had a large proportion of impacts where the skull is rotating backwards in the sagittal plane (79.2%, 440/556 impacts) (Fig. 3B; Table 3). Both men’s hockey and women’s rugby had more balanced distributions of sagittal, coronal, and horizontal impacts. Chi-squared tests followed by post-hoc tests revealed statistically significant differences between all sports for angular acceleration directions. Sagittal and coronal ratios Of the 1250 head impacts, 570 were sagittal dominant, with 34 from men’s hockey, 46 from men’s football, 69 from women’s rugby, and 412 from women’s soccer (Fig. 4A). Women’s soccer displayed the highest proportion of sagittal dominant impacts (n = 406/556, 73%) while men’s football had the least (n = 46/232, 20%). Men’s football exhibited the highest median rSagittal length of 15.9 cm while men’s hockey displayed the lowest median with 9.7 cm (Fig. 4B). We also found 299 coronal dominant impacts with 43 from men’s hockey, 124 from men’s football, 104 from women’s rugby, and 28 from women’s soccer (Fig. 4A). Men’s football had the highest percentage of rCoronal impacts (n = 123/232, 53%) while women’s soccer had the least (n = 28/556, 5%). Men’s hockey displayed the highest median rCoronal lengths with 13.8 cm while men’s football had the lowest median rCoronal length of 4.6 cm (Fig. 4B). Kruskal-Wallis tests highlighted significant differences between all sports for rSagittal and rCoronal ratios. Post-hoc tests highlighted significant pairwise comparisons between all comparisons for rSagittal except between men’s football and women’s rugby, and men’s hockey and women’s soccer. Further, rCoronal showed significant differences between all pairwise comparisons except men’s football vs. women’s soccer and men’s hockey vs. women’s rugby (Supplementary Table 4). Fig. 4(A) Distribution of dominant sagittal and coronal impacts for each sport and (B) median of sagittal acceleration ratio (rSagittal) and coronal acceleration ratio (rCoronal) impacts across all sports. Scatter points indicate all impacts for each sport. Boxplots indicate median (horizontal black line) with 25th and 75th percentiles. (A) Distribution of dominant sagittal and coronal impacts for each sport and (B) median of sagittal acceleration ratio (rSagittal) and coronal acceleration ratio (rCoronal) impacts across all sports. Scatter points indicate all impacts for each sport. Boxplots indicate median (horizontal black line) with 25th and 75th percentiles. Magnitude and frequency clusters For both the magnitude and frequency feature groups, k-means clustering identified k = 2 as the most optimal (i.e., two clusters), with silhouette scores of 0.45 and 0.37, respectively. Clustering with magnitude features identified two clusters that may be differentiated into low magnitude (Cluster 1) and high magnitude (Cluster 2) clusters based on resultant peak kinematics (Fig. 5A). They may also be differentiated by directional kinematics, where the low magnitude cluster consisted of primarily sagittal plane kinematics, and the high magnitude cluster showed higher coronal plane kinematics. Of those two clusters, women’s soccer impacts were found to largely represent the low magnitude cluster (N = 415/519, 79.9% of all total impacts) (Cluster 1), whereas the high magnitude cluster comprised of impacts mostly from men’s football, men’s hockey, and women’s rugby (Cluster 2). Frequency feature clustering also comprised of low frequency (Cluster 1) and high frequency (Cluster 2) clusters, mainly differentiated by low vs. high peak frequencies but comparable ranges of bandwidths (Fig. 5B). Similar to the magnitude features, the lower frequency cluster (Cluster 1) was moderately dominated by women’s soccer (N = 470/714, 65.8% of all total impacts), while the higher frequency cluster (Cluster 2) had a relatively even distribution of other sports. For magnitude-based features and clustering, WCSS analysis revealed that PAVcor had the greatest influence on cluster centroid displacement, while PAAsag had the least. Similarly, for frequency-based features, angular frequency in the sagittal direction caused the highest centroid shifts, whereas linear frequencies in the IS direction had the lowest effect. Interactions of feature groups Interactions of magnitude and frequency features highlight some interesting sport specific similarities and differences (Fig. 5C). Both men’s hockey and men’s football impacts were mostly high magnitude with more even distribution in low and high frequency clusters. Men’s football and women’s rugby showed slight dominance in impacts with high magnitude and low frequency (48.9% and 54.4% of all impacts, respectively). Lastly, women’s soccer demonstrated mostly low magnitude and low frequency impacts (62.2% of all impacts). Fig. 5(A) Two optimal magnitude clusters with their respective cluster median peak linear acceleration (PLA), peak angular velocity (PAV), and peak angular acceleration (PAA). (B) Three optimal frequency clusters with their respective median linear and angular peak frequencies and bandwidths. (C) Interaction of both magnitude and frequency clusters with sport-specific distribution of impacts in clusters. AP: Anterior-Posterior, LR: Left-Right, IS: Inferior-Superior, Cor: Coronal, Sag: Sagittal, Hor: Horizontal. (A) Two optimal magnitude clusters with their respective cluster median peak linear acceleration (PLA), peak angular velocity (PAV), and peak angular acceleration (PAA). (B) Three optimal frequency clusters with their respective median linear and angular peak frequencies and bandwidths. (C) Interaction of both magnitude and frequency clusters with sport-specific distribution of impacts in clusters. AP: Anterior-Posterior, LR: Left-Right, IS: Inferior-Superior, Cor: Coronal, Sag: Sagittal, Hor: Horizontal. t-SNE trained on all feature groups To further explore our impact dataset, we trained our t-SNE model on all of our features to further observe any sport-specific clustering, as t-SNE is a useful tool to determine similarities in high dimensional data. We generally observed tighter clustering of women’s soccer impacts, consistent with k-means findings, while other sports impacts showed more overlapping patterns (Fig. 6). Some sport-specific small subclusters were noted and further analyzed to examine if they belonged to specific players or playing positions, yet no conclusive patterns were found. Fig. 6All head impacts visualized using t-SNE, trained on comprehensive biomechanical features, including peak linear acceleration (PLA), peak angular velocity (PAV), peak angular acceleration (PAA), as well as linear and angular frequencies and bandwidths across all directions. All head impacts visualized using t-SNE, trained on comprehensive biomechanical features, including peak linear acceleration (PLA), peak angular velocity (PAV), peak angular acceleration (PAA), as well as linear and angular frequencies and bandwidths across all directions. Discussion We uniformly processed raw iMG head impact data across four sports—men’s football, men’s hockey, women’s rugby, and women’s soccer—and compared their biomechanical features. For objective (i), we identified statistically significant differences across sports in the distributions of peak kinematics, impulse durations, coupling ratios, and resultant acceleration directionality. However, for objective (ii), our clustering results highlighted cross-sport similarities, suggesting that head impacts may exhibit shared biomechanical characteristics across sports. Men’s football exhibited the highest median peak resultant PLA, PAV, and PAA, while women’s soccer had the lowest across all resultant kinematic metrics. Comparing our kinematic magnitudes to previous literature remains challenging due to the wide variability in reported kinematics for similar impact scenarios. A non-comprehensive review of prior studies showed a wide range of reported peak kinematics in the four sports we examined, with mean PLA and PAA ranges of 21 –32 g and 1017 rad/s2 – 2120 rad/s2 for men’s football31–34, 19–40 g and 1750–3500 rad/s2 for men’s hockey10,35, 11 –31 g and 800 rad/s2 − 4291 rad/s2 for women’s rugby35–37, and 5 –39 g and 1750 rad/s2 – 7713 rad/s2 for women’s soccer38–41. Notably, our study reports values on the lower end of these ranges, perhaps reflecting improvements in data quality and standardization using instrumented mouthguard sensors. A cross-sport comparison study by Naunheim et al.42 in 2000, used a helmet-mounted sensor, and found men’s football to have the lowest mean PLA (29.2 g), followed by hockey (35 g) and soccer (54.7 g), which not only show substantially higher kinematics for all sports compared with our study, but also opposite trends in cross-sport impact severity. More recently, using iMG sensors, Gabler et al.43 reported median kinematics for collegiate men’s football at 17 g and 1180 rad/s2, and Miller et al.,44 reported soccer heading with 9.4 g and 689 rad/s2 which align more closely with our current findings. Therefore, we believe as sensor accuracy continues to improve and processing methods become more standardized, cross-sport comparisons will become more consistent and reliable. Notably, men’s football exhibited substantially higher coronal rotational kinematics compared with the other sports, with a relatively large proportion of high rotation impacts (14% impacts over 5,000 rad/s2 of coronal angular acceleration). Interestingly, while women’s soccer impacts had generally lower peak resultant values, the median and 75th percentile inferior-superior peak linear accelerations in soccer impacts were the highest among sports. In addition, women’s soccer median and 75th percentile sagittal angular acceleration were second highest among sports. Since the soccer heading impacts were primarily in the sagittal plane, our directional magnitude analyses revealed that women’s soccer impacts can exhibit higher kinematics relative to other sports in certain directional measures, even though traditional resultant values would indicate lowest severity impacts. As it is known there can be direction-dependence of injury risk15, and some injury risk criteria such as the Brain Injury Criterion (BrIC) have established directional thresholds45. Our findings show that directional kinematics analyses are important when investigating sport head injury risks. Our acceleration direction analyses also revealed some impact directionality differences across sports that may have injury risk implications. From Tables 2 and 3; Fig. 4A, we see that only women’s soccer showed predominantly sagittal impacts during heading (82.6% peak angular accelerations in sagittal plane). Further, one interesting observation was the dominance of upward acceleration observed in women’s soccer impacts (59.4%). Upon reviewing the data, we found that women’s soccer impacts typically exhibited two impulsive peaks in succession, one from the initial ball-head contact (typically to the frontal hairline region), and the second from an opposite acceleration impulse likely due to the reaction force from the head-neck stopping at the end range of motion. The upwards accelerations corresponded to the second peak, which often surpassed the magnitude of the first peak. This observation shows that traditional peak analysis may not fully capture the complexities of impact biomechanics in sports, since one may have originally considered the ball-to-head impact to be more important to analyze for injury considerations, while we are finding that the ‘rebound’ impact can be higher magnitude. Men’s football impacts also showed some dominance in coronal peak angular accelerations (65.3%), and combined with the earlier observation of higher peak kinematics distributions compared with other sports, may prompt further investigation of coronal impacts and injury risk considerations for football. Early primate studies have shown that the coronal plane may have lower tolerance to diffuse axonal injury46. From our analysis of sagittal and coronal ratios, we found that men’s football had substantially lower ratios in coronal impacts, which is consistent with the observation that for a similar amount of linear accelerations, higher angular accelerations were seen in football impacts. Past work investigating the instantaneous head centre-of-rotation has noted the importance of understanding head/neck/torso and spinal constraints in investigating head rotational kinematics47. Interestingly, this prior laboratory-based study showed higher linear-angular ratios in the coronal plane than the sagittal plane47, which is consistent with what we saw in men’s hockey, but opposite to our findings in football impacts, demonstrating potential on-field differences in head-neck-torso engagement during impact. Overall, men’s hockey and women’s rugby showed more balanced impact directionalities, and substantially more horizontal impacts compared with men’s football and women’s soccer. Past literature have also indicated potential horizontal plane vulnerability48 that may be important to consider for these sports. While we included both men’s and women’s data, this unique iMG-only dataset did not include matched sports to enable a sex-based analysis. Despite this, we can make preliminary observations that women’s head impacts had comparable ranges of peak kinematics as men’s head impacts in our study, especially in women’s rugby. This would challenge the theory of weaker neck muscles in women leading to more severe impacts and subsequently higher injury risk49,50. Since our study quantified sagittal and coronal linear-angular ratios, we expected these ratios to approximate the distance between the sensor and the instantaneous center of rotation. If neck length played a significant role, we anticipated higher ratios in men’s impacts compared to women’s. However, we also did not find consistent sex-based differences. Various factors, including anthropometrics, neck muscle strength, and playing style, may all contribute to differences in impact biomechanics and subsequent injury risk. Further studies with matched sports between men’s and women’s teams can further investigate these factors. Our comparison of impulse durations along with clustering of frequency-domain features give insight to the temporal dynamics and frequency characteristics of impact biomechanics across sports. While most studies focus on reporting peak head kinematics metrics, injury biomechanics literature have highlighted the importance of examining not only peak acceleration, but also the duration of the acceleration impulse when defining concussion risk curves14,51. Theoretical structural mechanics analyses and earlier cadaver/animal research52 have derived injury risk curves that predict higher tolerance to peak accelerations at short durations and lower tolerance to peak accelerations at long durations. Our analyses have shown a wide range of impulse durations from 3 ms to 50 ms or longer, with largely overlapping ranges across sports. Interestingly, despite the expectation that helmet padding should lengthen impulse durations and mitigate high frequency content53 we did not observe substantial differences in impulse duration between helmeted (linear median: 9 ms, angular median: 8 ms) and unhelmeted (linear median: 9 ms, angular median: 9 ms) impacts. Additionally, we found the helmeted impacts from men’s football and men’s hockey to have a greater proportion of impacts in the high frequency cluster. It is worth noting the measured frequency-dependence of brain material behaviour54, and the potential existence of low-frequency resonance behaviour of brain tissue55 that may prompt future studies to also consider these biomechanical parameters in conjunction with impact magnitude when determining injury risk implications. Overall, combining the findings from the two objectives, we observe that despite some notable distribution differences in biomechanical parameters, direct head impacts in the sports we examined have largely overlapping biomechanical characteristics. Women’s soccer impacts, consisting solely of heading events, exhibited the tightest biomechanical parameter distributions. Since soccer heading impacts involve largely similar directionalities and the same impacting object, they are highly similar to each other in these biomechanical features, which we observed in the majority of soccer impacts falling into the same magnitude and frequency clusters. With more varying directionality and impact scenarios in other sports, we find less clear sport distinction in the clustering analysis, despite large differences in protective equipment use and gameplay dynamics. Even within soccer heading, a small proportion of impacts have similar characteristics as those found in other sports with some of these impacts found in the high magnitude and high frequency clusters. With the T-SNE analysis that optimized for cluster formation by preserving local structures in the high dimensional feature space, there did not seem to be clear sport-specific clustering of features. Overall, we show that sport head impacts span a continuous spectrum of biomechanical parameters, and combining multisport impact data along with epidemiological considerations may give a bigger picture view of the relationship between biomechanical parameters and brain injury risks. This study has some limitations that should be considered. Our analysis focused solely on direct head impacts, excluding indirect impacts that frequently occur in sports. These indirect impacts may also contribute to brain injury risk and could alter the biomechanical comparisons across sports. Furthermore, the soccer dataset had only headers and no other recorded impact scenario (e.g., elbow-to-head, head-to-head, etc.,), which may limit the generalizability of all impacts experienced in soccer. Additionally, while we aimed to standardize sensor settings and data processing techniques, we may not be able to fully address inherent sensor hardware differences in post-processing. Also, the standardization process of down sampling and truncating impacts from some sensors may have resulted in loss of frequency- and time-domain information that we are unable to compare (e.g., normalized filter cutoffs to lowest available sensor bandwidth, reducing 100–150 ms impacts to only 50 ms). Additionally, estimating one impulse duration for each impact can be challenging for real-world impacts. Our approach of duration calculation using the maximum axis may not fully capture the complexities of impacts that are multidirectional and/or not perfectly unimodal. Future work may explore different ways of duration calculations in each direction of interest to further understand potential complexities in impulse duration. Moreover, our clustering model was trained on a select set of biomechanical features rather than an exhaustive list of all possible parameters, to focus on key parameters with injury risk implications, reduce redundancy, and improve the interpretability of findings. Including additional features may alter the observed clustering patterns. Additionally, a sport-matched comparison of sex differences was not feasible with our dataset, despite some preliminary observations of similarities in the men’s and women’s head impacts. A more controlled prospective study design could provide further insights into sex-based similarities or differences in head impact biomechanics. Conclusions This study examined head impact biomechanics across multiple sports through a comprehensive analysis of various biomechanical parameters that have been implicated in brain injury risk. We uniquely curated a multisport dataset collected using iMG-type sensors. After controlling for sensor setting differences and performing uniform processing, we note interesting differences and similarities in biomechanical characteristics across men’s hockey, men’s football, women’s rugby, and women’s soccer head impacts. While men’s football showed highest peak resultant kinematics and women’s soccer showed the lowest, consistent with past observations of relative impact magnitudes, directional kinematics analyses revealed more nuanced patterns. For example, women’s soccer impacts exhibited relatively high sagittal peak linear and rotational kinematics. From our clustering analyses, we show that aside from women’s soccer demonstrating biomechanically similar impacts from heading, there may be shared biomechanical signatures across sports head impacts. These findings suggest that standardized multi-sport analyses may offer greater insights into brain injury mechanisms than isolated sport-specific studies. Future research should leverage similar standardized methodologies across expanded datasets to further elucidate the relationships between the full spectrum of impact characteristics and brain injury risk, ultimately informing more effective protective strategies across sports. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4