Injuries are a common occurrence in team sports and can have significant financial, physical and psychological consequences for athletes and their sporting organisations. As such, an abundance of research has attempted to identify factors associated with the risk of injury, which is important when developing injury prevention and risk mitigation strategies. Traditionally, research has implemented reductionist approaches to identify injury risk factors. These reductionist methodologies assume that all the parts of a system (in this case, injury aetiology) can be broken down and examined individually and then summed together to represent the system as a whole. Reductionist approaches are useful in establishing associations between specific factors and the risk of injury. However, in order to predict the occurrence of injuries at an individual level, complex approaches should be implemented. In light of this, machine learning has been suggested as an appropriate method of applying complex approaches to the prediction of injuries in sport. Machine learning is a field of computer science which involves building algorithms to learn from data and make predictions without being programmed what to look for or where to look for it. Whilst machine learning cannot be used to establish causal relationships between specific factors and the occurrence of injuries, it differs from reductionist methodologies in that it has the ability to identify the complex, non-linear interactions that occur amongst risk factors. Study 1 (Chapter 4) aimed to utilise machine learning methods to predict the occurrence of hamstring strain injuries (HSIs) in elite Australian footballers. Hamstring strain injury is the most common injury in elite Australian football and three of the most consistently identified risk factors for HSI are increasing age, prior HSI and low levels of eccentric knee flexor strength. While some iterations of the predictive models achieved near perfect performance (maximum area under the curve [AUC] = 0.92), others performed worse than random chance (minimum AUC = 0.24). It was concluded that age, previous HSI and eccentric knee flexor strength data could not be used to identify Australian footballers at an increased risk of HSI with any consistency, despite these factors being highly associated with the risk of HSI. It is suggested that more observed injuries, in addition to more frequent measures of the variables included in the models, may have improved the performance of the predictive models in Study 1. To overcome the limitations acknowledged in Study 1, Study 2 (Chapter 5) investigated whether more frequent measures of the impact of prior injury (in the form of session availability), in addition to a greater number of observed injuries (albeit non-specific pathologies), improved the ability to identify injury risk. It was observed that greater session availability in the previous 7 days increased injury probabilities by up to 4.4%. Additionally, lesser session availability in the previous 84 days increased injury probabilities by up to 14.1%, only when coupled with greater availability in the previous 7 days. It was concluded that session availability may provide an informative marker of the impact of prior injury on subsequent injury risk and can be used by practitioners to guide the progression of training, particularly for athletes that are returning from long periods of injury. Study 1 and Study 2 implemented complex approaches in an attempt to improve injury risk identification at an individual level. Despite the findings of Study 1 and Study 2, quantifying injury risk on a daily basis remains a complex and challenging task for practitioners working in Australian football. Commonly implemented tools such as self-reported wellness questionnaires provide a much more accessible measure of athletes’ wellbeing and how they are responding to the demands of training/competition. Whilst improving the ability to estimate injury risk at an individual level is an important focus area, it may also be important to determine the level of information that more accessible and more frequently measured variables (such as self-reported wellness) provide regarding injury risk. To make this determination, however, it is also necessary to understand the factors that directly influence self-reported wellness. Accordingly, Study 3 (Chapter 6) aimed to investigate the factors that impact wellness in elite Australian footballers. Measures of external load examined on their own were able to explain changes in wellness to a large degree (root mean square error = 1.55, 95% confidence intervals = 1.52 to 1.57). However, there was a proportion of wellness that could not be explained by external loads. It is suggested that examining the interaction between external training loads and self-reported wellness may assist practitioners in their load management strategies. However, there is limited research investigating the interaction between external loads and wellness and the impact this information may have on subsequent injury risk. Accordingly, Study 4 (Chapter 7) aimed to investigate the ability of external load data, session availability data and self-reported wellness data, as well as the interaction between the three, to identify the risk of lower limb non-contact injuries in elite Australian footballers. The model with the least input variables (athlete ID and session type) displayed the highest predictive ability (AUC = 0.76, Akaike information criterion [AIC] = 479, Brier score = 0.009). The models built using external load, session availability and wellness data all displayed similar predictive ability (AUCs = 0.72 to 0.75, AICs = 477 to 478, Brier scores = 0.009 to 0.009). Despite observing higher predictive performance compared to previous research, the addition of external load, session availability and wellness data, as well as demographic and pre-season external load data, did not improve the ability to predict lower limb non-contact injuries in Study 4. Overall, this program of research displayed a limited ability to predict injuries in elite Australian football. The findings of this thesis highlight a need for a larger number of observed injuries when implementing predictive modelling strategies to identify injury risk at an individual level. Despite this, the predictive modelling strategies implemented in this thesis may assist researchers and practitioners in better understanding the relationships that exist between variables that are commonly collected, analysed and interpreted. Whilst the efficacy of complex approaches and their application in sports research may warrant further investigation, researchers and practitioners alike need to strongly consider the limitations of input data and the predictive modelling strategies used to analyse these data when conducting (as well as interpreting) future research.