Bone stress injuries (BSI) are overuse injuries that commonly occur in runners. BSI risk is multifactorial and not well understood. Unsupervised machine learning approaches can potentially elucidate risk factors for BSI by looking for groups of similar runners within a population that differ in BSI incidence. Here, a hierarchical clustering approach is used to identify groups of collegiate cross country runners (32 females, 21 males) based on healthy pre-season running (4.47 m·s-1) gait data which were aggregated and dimensionally reduced by principal component analysis. Five distinct groups were identified using the cluster tree. Visual inspection revealed clear differences between groups in kinematics and kinetics, and linear mixed effects models showed between-group differences in metrics potentially related to BSI risk. The groups also differed in BSI incidence during the subsequent academic year (Rand index = 0.49; adjusted Rand index = -0.02). Groups ranged from those including runners spending less time contacting the ground and generating higher peak ground reaction forces and joint moments to those including runners spending more time on the ground with lower loads. The former groups showed higher BSI incidence, indicating that short stance phases and high peak loads may be risk factors for BSI. Since ground contact duration may itself account for differences in peak loading metrics, we hypothesize that the percentage of time a runner is in contact with the ground may be a useful metric to include in machine learning models for predicting BSI risk.
Keywords: Gait mechanics; Machine learning; Principal component analysis; Stress fracture.
Copyright © 2022 Elsevier Ltd. All rights reserved.