3 54 CHAPTER 3 and decision support using machine learning is a novelty that has not yet been realized for team sports. In order to build an in-match physical performance prediction and decision support, models have to be based on derived time-motion data variables. These derived timemotion variables can be divided into type-1 or type-2 [14]. Type-1 variables include external load measures such as distance covered and distance covered in speed category. Type-2 variables include load measures related to changes in velocity such as accelerations, decelerations, and summarized variables like metabolic power and energy expenditure. Researchers have tried to quantify physical performance decline as a decrease in various type-1 variables. It turns out that during the match, the distance covered, and the distance in the speed category decreases [2], [3], [15]. However, type2 variables are more sensitive to identifying in-match physical performance decline than type-1 variables [14]. Furthermore, condense variables like metabolic power are specially equipped for identifying in-match performance decline. They hold a more linear relationship with fatigue and include accelerations and decelerations in their calculation [16]. These findings highlight the sensitivity of type-2 variables for physical performance decline. Therefore, we include both the more common type-1 and the more sensitive type-2 variables in our prediction models. Contextual factors like home or away, rankposition, and score show a difference in the overall distance covered [17]. Although we acknowledge the contextual factors such as home or away, rank position and score, we excluded these contextual factors in this proof-of-concept study. Instead, we focused on the individual player in-match motion data. The study’s goal is to predict the in-match physical performance decline of the individual soccer player using machine learning. To our knowledge, no prior study in professional soccer has investigated the in-match physical performance prediction using machine learning techniques enabling decision support for the coach on substitutes. We aim to prove: (1) if physical performance decline can be identified using both type-1 and type2 variables; (2) if substitutes perform better than entire match players on both type-1 and type-2 variables and (3) if the degree of physical performance of a player can be predicted in an early stage of the match using machine learning models for type-1 and type-2 variables. 2. MATERIALS AND METHODS 2.1 Experimental Approach to the problem For our study, we retrospectively collected the in-match position tracking data from 302 competitive professional soccer matches between 18 teams during the Dutch ‘Eredivisie’
RkJQdWJsaXNoZXIy MjY0ODMw