de Haan Michel, van der Zwaard Stephan, Sanders Jurrit, Beek Peter J, Jaspers Richard T
Department of Human Movement Sciences, Vrije Universiteit Amsterdam, Amsterdam Movement Sciences, Amsterdam, Netherlands.
Department of Cardiology, Amsterdam University Medical Center, location AMC, University of Amsterdam, Amsterdam, Netherlands.
J Sports Sci Med. 2025 Sep 1;24(3):565-577. doi: 10.52082/jssm.2025.565. eCollection 2025 Sep.
Soccer players are frequently categorized by playing positions, both in the scientific literature and in practice. However, the utility of this approach in evaluating physical match performance and optimizing physical training programs remains unclear. This study compares the effectiveness of categorizing soccer players by their playing position versus using unsupervised machine learning based on match-specific running performance. Match-specific running data were collected from 40 young elite male soccer players over two seasons. Thirty-one of these players completed a 20-meter sprint test and a maximal incremental treadmill test to measure maximal oxygen uptake. Players were categorized both by playing position and by subgroups derived through -means clustering based on match-specific running performance. Differences in sprint capacity, endurance capacity, and match-specific running performance were compared between and within playing positions, as well as between and within clusters. The two categorization methods were further compared for variance within subgroups and standardized differences between subgroups for total distance (TD), low-intensity running (LIR), moderate-intensity running (MIR), high-intensity running (HIR), and sprint distance during matches. Match-specific running performance differed between playing positions, despite notable inter-individual differences in running intensities within playing positions. Clustering based on match-specific running performance revealed less variance within groups (TD: = 0.049, LIR: = 0.032, HIR: = 0.033) and larger standardized differences between groups (LIR: = 0.037, MIR: = 0.041, HIR: = 0.035, Sprint: = 0.018) compared to grouping by playing position. Moreover, 20-meter sprint speed differed between the sprint and high intensity endurance clusters (25.22 vs 23.75 km/h, = 0.012), but not between playing positions. Using unsupervised machine learning to categorize soccer players improves the identification of player groups with similar match-specific running performance, thereby supporting performance evaluation and contributing to the optimization of physical training.
在科学文献和实践中，足球运动员常按比赛位置进行分类。然而，这种方法在评估比赛中的体能表现以及优化体能训练计划方面的效用仍不明确。本研究比较了根据比赛位置对足球运动员进行分类与基于特定比赛跑步表现使用无监督机器学习进行分类的有效性。在两个赛季中收集了40名年轻精英男性足球运动员的特定比赛跑步数据。其中31名运动员完成了20米冲刺测试和最大递增跑步机测试，以测量最大摄氧量。运动员既按比赛位置分类，也按基于特定比赛跑步表现通过K均值聚类得出的亚组分类。比较了不同比赛位置之间和内部、不同聚类之间和内部的冲刺能力、耐力能力以及特定比赛跑步表现的差异。还进一步比较了两种分类方法在亚组内的方差以及亚组之间在总距离（TD）、低强度跑步（LIR）、中等强度跑步（MIR）、高强度跑步（HIR）和比赛中的冲刺距离方面的标准化差异。尽管比赛位置内个体间的跑步强度存在显著差异，但不同比赛位置的特定比赛跑步表现仍有所不同。与按比赛位置分组相比，基于特定比赛跑步表现的聚类显示组内方差较小（TD：F = 0.049，LIR：F = 0.032，HIR：F = 0.033），组间标准化差异较大（LIR：d = 0.037，MIR：d = 0.
