Open Access

Felix Faude

• In collaboration with the professional football club VfB Stuttgart 1893 AG, this thesis develops machine learning models aimed at predicting non-contact injuries in youth players. The sample includes players (Age: M = 18.6, SD = 3.4) from the club’s U16 to U21 teams. The primary objective is to create models that exhibit high performance while ensuring prediction interpretability. This study utilizes various load metrics, including internal measures such as the Rating of Perceived Exertion, and external GPS-derived data. The entire process involves data preprocessing, feature engineering, data analysis, and model training using Random Forest, XGBoost, and Neural Networks. The models are evaluated using recall, precision, balanced accuracy, and Brier score. The findings reveal that Random Forest and XGBoost models achieve moderate recall and balanced accuracy. However, their low precision limits their practical application due to a high number of false positives. In contrast, the Neural Network model demonstrates better probabilistic predictions and precision, which were still not sufficient for practical utility. Given the moderate performance of these models, the study recommends that the models should not be implemented for practical injury risk assessments, particularly due to the potential risks to player’s health from false predictions. Nonetheless, further development and adjustments of these models are recommended, considering the given methodology and limitations, to build more reliable machine learning models for injury prevention.