Modelling Training Loads and Injuries in Australian Football

Submission note: Submitted in total fulfilment of the requirements of a Doctor of Philosophy to the La Trobe Sport and Exercise Medicine Research Centre, School of Allied Health, Human Services and Sport, College of Science, Health and Engineering, La Trobe University, Victoria, Australia.

Injuries are a frequent burden in Australian football and can negatively impact on individual player health and team performance. Wearable sensor technologies such as global positioning systems, accelerometers, and heart rate monitors are increasingly being used to collect data on the physical demands of training and matches in elite Australian footballers. This thesis aimed to investigate how longitudinal load monitoring data can be utilised to improve decision making for injury risk reduction in Australian football. A literature review was conducted to identify the training load variables associated with injury in professional footballers, and to critique the techniques used to model relationships between load and injury data. Studies used global positioning system (GPS) devices, heart-rate monitors, accelerometers, ratings of exertion and match frequency counts to quantify the load placed on athletes. The amount of training load accumulated by players over consecutive days or weeks, as well as the relative changes training load were reported to be associated with injury risk in a range of football codes. Few prior studies evaluated the ability to predict future injuries using statistical models of training load data. Additionally, many studies applied discretisation transformations to training load data, potentially increasing the risk of false positive and false negative findings. The first study examined whether relative training loads, quantified using the ratio of short-term (acute) to longer term (chronic) load, were associated with injury in two seasons of Australian football player data. It extended previous research by considering a range of periods of short (2- 9 days) and long-term (2-5 weeks) training loads. The ratio of moderate speed running (18-24 km/h) in a 3 or 6-day acute period and 21-day chronic time-window had the strongest association with injuries. These time periods reflected typical training and match schedules in the studied AFL team, and provide some indication that the sport-specific schedule of competition may be important when choosing acute and chronic load monitoring periods. The second study investigated the most influential predictors of athlete ratings of perceived exertion in Australian football training sessions. Predictive modelling approaches were compared for the ability to predict how hard an athlete would subjectively rate a training session using objective training data such as running distances, speeds, and accelerations as inputs to the model. Non-linear modelling and machine learning methods outperformed linear models for predicting subjective athlete responses to training, achieving a root mean square error of approximately 1 unit on a 10-point scale. The model constructed could be used to inform training session design that targets a specific athlete response, or to longitudinally track changes in perceived exertion in response to similar training sessions. xvii The third investigation considered the task of designing training plans that incorporated evidence-based information from training load and injury risk models. Mathematical optimisation was used to generate pre-season training plans for Australian football that objectively maximised performance goals whilst being constrained to not exceed injury risk thresholds. Simulations showed the method enabled the creation of feasible training plans that allowed for individualised training plan design and the ability to adapt to changing training objectives and different training load metrics. The approach detailed in this study has applications for cost-benefit scenario modelling such as estimating the increases or decreases in training that could be expected if an athlete or team was willing to accept a higher or lower level or injury risk. Three seasons of professional Australian football training load monitoring and injury data were used in the fourth study to investigate the ability of machine learning and statistical models to predict injuries. Models were constructed for non-contact, non-contact time-loss and hamstring specific injuries using two seasons of data. Then tested on a held-out sample of data from a third season. Predictive performance was only marginally better than chance for models for noncontact and non-contact time-loss injuries. Targeting the models to only hamstring injuries improved predictive performance. Injury prediction models built on data from a single team had limited ability to predict injury occurrence on out-of-sample data. Results of this study indicated that focusing the modelling approach on specific injury types and increasing the amount of modelling data may improve predictive models for injury prevention. Finally, a critical examination of the prevalent modelling methods in training load and injury risk studies was performed using simulated data. Injuries were simulated in training load data using three risk profiles (flat, U-shaped, and S-shaped). One-hundred datasets were simulated with sample sizes of 1000 and 5000 observations. Discrete modelling methods were compared to continuous methods (fractional polynomials and spline regression) for their capacity to fit the specified risk profiles. Models were evaluated using metrics of discrimination (area under ROC curve) and calibration (Brier score). Continuous methods were found to be better suited to modelling the relationship between training load and injury than discrete methods. Comparing injury risk models using ROC curves led to inferior model selection. Measures of calibration were better able to judge the utility of injury risk models. This thesis contributes new evidence on the relationships between training loads and injuries in Australian football. It gives recommendations on how to model non-linear relationships between training loads and injuries and describes a framework for incorporating injury risk models into training load planning. The thesis provides practitioners with evidence-based tools for using training load monitoring data for injury risk decision making in Australian football.