Longitudinal time-to-event graph mining pipeline for musculoskeletal injury forecasting

Injury is pervasive to sport and is an important yet challenging forecasting problem with high practical value. Athlete monitoring technologies are deployed for health and wellbeing surveillance, but incidence rates have not decreased in proportion to their widespread use. Existing injury prediction algorithms have deployed off-the-shelf machine learning solutions, leading to mis-formulated approaches that often fail to appreciate the complexities of an athlete’s ever-evolving physiology. This has resulted in a gap between machine learning capabilities and their actual effectiveness in mitigating athletic injuries. The focus of this thesis is to contribute a practical and accurate forecasting framework to mitigate sports-related, non-contact musculoskeletal injuries. Chapter 1 positions my research perspective under the theory of complexity. I formulate injury forecasting as a longitudinal time-to-event problem and specify three specific aims addressed in Chapters 2-4: 1) select an intra-individual dynamic graph construction approach, 2) develop a longitudinal time-to-event forecasting model for dynamic graphs, and 3) design an architecture to identify temporal edges contributing to an athlete’s injury forecast. This thesis ends by threading Chapters 2-4 into an end-to-end graph mining pipeline for sport epidemiology. An athlete monitoring dataset from University of Iowa Department of Intercollegiate Athletics serves as an applied example. Athlete-specific dynamic graphs are constructed from longitudinal force plate data. The longitudinal time-to-event model was trained and tested to forecast musculoskeletal injury from the dynamic graph dataset. Athlete-specific injury forecasts are then explained where novel inferences into the temporal interactions leading to injury are unveiled.