5B 127 MACHINE LEARNING PREDICTING RUNNING INJURIES INTRODUCTION We used the context and data of [1]. In [1], using statistics was the fundament of modelling and knowledge about the risk of sustaining an injury in running. In the following, we investigate the possibility of applying machine learning to predict sustaining an injury. We showed the influence of the Acute:Chronic Workload Ratio (ACWR) change on the risk of sustaining an injury. The knowledge of the relationship between workload and the effect on injuries might also be improved by using machine learning techniques. However, machine learning techniques are pointed at enabling prediction instead of fully understanding the system [2]. Although, the input of the knowledge about the influence of change in relative workload on sustaining an injury indicates the variables in the machine learning model. The identified load variables will be used to develop an machine learning model for predicting injuries. Statistically proven relations improve the quality of the machine learning models[3], [4]. The aim is to predict the risk of sustaining an injury using the ACWR of competitive runners using machine learning. To our knowledge, no study in running has investigated the combination of ACWR and machine learning techniques to predict injuries supporting the trainer and runner on intervention in training. MATERIALS AND METHODS The participants, the definition of an injury, quantifying workloads, the definition of ACWR, and data analysis, are the same as in [1]. Machine Learning A machine learning model was constructed to predict the occurrence of an injury. The differences in load between runners were eliminated by using the ACWR. Therefore, we used the ACWR calculated [1] as the machine learning model features. To construct the ACWR as features, the ACWR was split between the four weeks preceding the injury and the four weeks not preceding an injury, labelled as preceding an injury or not. Also, the fortnightly difference between the ACWR preceding an injury or not was used. The constructed features are represented in Table 1.
RkJQdWJsaXNoZXIy MjY0ODMw