3 68 CHAPTER 3 Following these arguments, the main finding of our study is that our machine learning models could reliably identify and predict the physical performance of a player after 15 minutes in the match. The early prediction of physical performance can support a decision support system as advocated by Robertson [12] and further illustrates the opportunities provided by machine learning in player monitoring during the match. A limitation of the study is the exclusion of contextual factors like home or away, rankposition, position system, and score show a difference in the overall distance covered [17]. Although these contextual factors on their own influence the overall distance of the team. To generate a machine learning model on individual physical performance, every combination of the contextual factors needs to be sufficiently present in the data. Not every combination of an individual player, home or away, rank-position, position system, and score will be present in one season. A coach will need to use his or her insight and knowledge to judge the prediction of physical underperformance on its merits. The use of a machine learning approach also goes in hand with some limitations. To conduct a reliable model for an individual player, there must be enough entire match data available. We did not identify any literature in soccer to refer to the amount of data needing to be available. In the literature on fitness trackers, it is found that three days of repeated measures is necessary to represent adults’ normal activity levels with an 80% confidence [30]. In parallel, three entire matches for a player may be sufficient to identify his average physical performance. A method to conduct a reliable model is to retrain models frequently and monitor precision to identify the optimum amount of data [31]. Another limitation is that physical underperformance is just one of several reasons for a coach to substitute a player. Substitutions can also be initiated by a player’s injury, necessary tactical changes (e.g., because of being behind in a match), or tactical underperformance of a player [7]. In our study, the data was limited to the individual player’s speed, acceleration, and distance measures. Next to contextual influences [17], other physiological markers of fatigue such as individual measures like heart rate, breathing, and body temperature were not included. Including contextual influences and physiological markers of fatigue in the machine learning model could enable a more informative system. Finally, the thresholds of physical underperformance were randomly chosen, and the 90% threshold is relatively rarely seen. 5. CONCLUSIONS Our study confirmed that the identification of the physical performance could be based on type-1 and type-2 variables calculated from the position tracking systems. Also, substitutes perform better than entire match players on both type-1 and type-2
RkJQdWJsaXNoZXIy MjY0ODMw