Investigating running styles and fatigue: a machine learning approach to explore whether stereotypical behaviours exist

: (Alternative format thesis)

Abstract

Sports biomechanists have extensively studied the relationship between running technique, economy and fatigue to optimise performance and minimise injuries. However, the literature remains inconclusive on the key factors of technique that runners should focus on to enhance economy and fatigue resilience. The lack of clear findings may be explained by the high individuality of running technique. Identifying groups of runners who share similarities in their technique could enable the development of more tailored interventions, potentially enhancing training success. The purpose of this PhD thesis was to assess whether stereotypical running technique patterns and responses associated with fatigue exist. The first study addressed the challenge of detecting contact events during running, which are key for technique analysis but difficult to identify without force data. A novel algorithm using marker-based kinematics and deep learning was introduced, improving the accuracy of previous methods across different running conditions. The second study featured a comprehensive comparison of different methods for dimensionality reduction, a key step for the efficacy and efficiency of clustering to group runners based on their technique. Principal Component Analysis (PCA) achieved the greatest data compression and led to the most distinct clusters. The third study used PCA and hierarchical clustering to group runners based on their technique at various speeds. Two distinct clusters were identified: the \enquote{neutral pelvis} and \enquote{tilted pelvis} clusters. The tilted pelvic cluster exhibited a more tilted pelvis, greater trunk to pelvis extension and hip flexion throughout the gait cycle as well as smaller duty factor. Despite their differing techniques, clusters showed no significant variations in running economy, demographics, anthropometrics, physiology, training, or performance metrics. The fourth study examined whether the technique modifications associated with fatigue during a run to exhaustion were specific to the identified clusters. Both clusters showed small adjustments in technique, but no distinct modifications specific to each cluster were observed. The fifth study introduced a classification algorithm to allocate new runners to the proposed clusters without the need for a camera-based motion capture system. Several consumer-based wearable sensor networks (2-3 sensors) were identified and achieved reasonable accuracy ($\geq$82\%). These results enhance the reach of our research to the wider running community. The findings of the thesis enable the design of targeted interventions which may enhance the success of running training. Furthermore, a framework was presented leveraging camera- and wearable-based motion capture with machine learning methods to bridge the gap between laboratory research and in-field application, which may be easily applied to other sporting actions.

Date of Award	13 Nov 2024
Original language	English
Awarding Institution	University of Bath
Sponsors	NURVV Ltd
Supervisor	Ezio Preatoni (Supervisor), Grant Trewartha (Supervisor), Xi Chen (Supervisor) & Dario Cazzola (Supervisor)