Abstract details
| Abstract-ID: | 1607 |
| Title of the paper: | Cracking the Code: Predicting Player Positions in Handball with Machine Learning |
| Authors: | Venzke, J., Schäfer, R., Manchado, C., Platen, P. |
| Institution: | Ruhr-University Bochum |
| Department: | Department of Sports Medicine and Sports Nutrition |
| Country: | Germany |
| Abstract text | INTRODUCTION: In team sports, an increasing amount of data is being collected to assess tactical behavior and monitor physical performance. Strong physical performance is a cornerstone of success in professional team sports. An individual or position-specific performance profile is highly desired by coaches for player monitoring. However, many analyses focus on selected parameters to describe position specific exercise load. A comprehensive understanding of player performance requires the consideration of multiple parameters. However, coaches are faced with an overwhelming number of parameters and require a more detailed view of the most meaningful parameters. Our aim is to identify the most influential predictor of locomotion data across different positions in handball. METHODS: A supervised machine learning model was used to classify positions with 10-fold cross validation. Our analysis included several conventional parameters (distance, mean velocity, speed/acceleration zones) and parameters derived from the metabolic power approach (energy cost, metabolic power). Local positioning system data (Kinexon Precision Technologies) from all 65 EHF EURO 2020 matches were used. 414 elite male handball players were included, resulting in 1596 datasets. We analyzed net playing time with durations longer than 1 min. Goalkeepers were excluded. RESULTS: 1437 datasets were used for training and validation and 159 for testing. The test accuracy of our model was determined to be 78.6%. Wings were the most accurate position to predict, with 80 out of 81 players correctly predicted. Center backs were often confused with outer backs. Pivots were also often predicted as outer backs. Average speed had the greatest relative influence on predicting position (47.4%) and weight the second greatest (31.2%), followed by average metabolic power (5.2%) and height (4.9%). CONCLUSION: Our study highlights the importance of using multiple parameters to accurately classify player positions in elite handball. Our model demonstrated robust performance, as evidenced by high accuracy on the test dataset. Wings consistently stood out as the most accurately predicted positions, reflecting their typically lower body weight. Average speed being the best predictor is consistent with recent research showing positional differences in average speed (Manchado et al., 2022). For example, wing players cover the most distance and spend the most time in the high-speed categories. These findings highlight the potential of machine learning techniques to improve our understanding of player performance dynamics in team sports such as handball. |
| Topic: | Statistics and Analyses |
| Keyword I: | |
| Keyword II: | |
| Keyword III: | |