Instrumented mouthguards (iMGs) capture thousands of head acceleration events in
contact sports, though processing the data from these into meaningful
information remains a challenge. To bridge this gap, we developed a principled
clustering pipeline to extract concise kinematics that are representative of
real-world impact scenarios. We evaluated six multivariate time-series distance
metrics using k-medoids clustering and a dataset of 5149 impacts across three
sports: mixed martial arts, boxing, and rugby union. Model performance was
evaluated using a Coverage Quality Score (CQS), prioritizing reconstruction
coverage and cluster separation. Joint Shape-Based Distance (SBD) significantly
outperformed all other metrics, including Dynamic Time Warping (DTW), driven
primarily by superior reconstruction coverage. The methodology yielded two
complementary archetypes. First, a 36-impact canonical set provides distinct
medoids representing core kinematic mechanisms; however, distributional checks
confirm these medoids alone do not reproduce cohort severity distributions. To
resolve this, we generated a second, extended impact set sampled proportionally
within clusters along specific severity axes (e.g., peak acceleration and
whole-brain strain). This targeted sampling recovers full severity-distribution
fidelity using fewer than 150 impacts per dataset. Ultimately, this framework
successfully distills large iMG datasets into manageable cohorts providing a
scalable solution for assessment, experimental targets, and finite element
simulations.
