Predicting Field-Sport Distances Without Global Positioning Systems in Indoor Play: A Comparative Study of Machine-Learning Techniques.

PURPOSE: Accurately predicting the distance covered by athletes during indoor sport activities without the use of GPS (global positioning systems) presents a significant challenge. This study evaluates the effectiveness of various machine-learning techniques in predicting total distance, sprinting distance, and running distance for athletes in men's and women's soccer and lacrosse at the University of Notre Dame.

METHODS

The techniques assessed include XGBoost Regressor, ElasticNet Regression, Ridge Regression, and Lasso Regression. Key performance metrics such as root-mean-square error, SDs, means, and 95% CIs are analyzed to provide insights into the relative performance of each method.

RESULTS

XGBoost provided the lowest root-mean-square error for total distance (97.962 [12.973]), sprint distance (91.616 [4.234]), and running distance (137.103 [2.789]).

CONCLUSION

The results demonstrate varying levels of accuracy and precision across different sports, genders, and contexts (game vs practice), highlighting the importance of selecting the appropriate model for specific applications in optimizing team performance, injury prevention, and player conditioning.