A Machine Learning-Based Performance Monitoring Approach for Athlete Performance Attenuation Prediction

Assessment of athletes' performance, including pre-game physical measurements and pre/post-game rankings, is challenging due to various physiological, external, and health factors. Established techniques for performance attenuation assessment are vital for mitigating it. Thus, this study aims to improve performance monitoring via predictive models while developing an understanding of performance-attenuation-related metrics through rankings predictions. The study included data from forty-one male Gaelic football players, including physical, anthropometrical, and subjective markers such as Countermovement Jump, Drop Jump, Reactive Strength Index, and perceptual responses, which were captured to evaluate neuromuscular, perceptual, and biochemical responses. The best machine learning models predicted the rankings and were assessed using precision, F1-score, and recall, while principal component analysis (PCA) and balancing techniques (SMOTE, ROSE, and ADASYN) addressed data dimensionality and class imbalance. Our results emphasize the benefits of combining different models and balancing techniques. The alignment between PCA and SHAP values (feature impacts: ±0.7) reinforces factors such as strength, lower body power, and speed as relevant predictors of performance attenuation-related outcomes. This machine learning approach aims to predict performance attenuation in similar scenarios, emphasizing the need for an accurate model and preprocessing technique selection for predictive performance monitoring. We discuss how balancing techniques offer potential and limitations regarding results' generalizability while calling for developing more advanced protocols to address similar prediction requirements.