A META-ANALYSIS ON MACHINE LEARNING AS A TRANSFORMATIVE FRAMEWORK FOR PERFORMANCE OPTIMIZATION IN PROFESSIONAL SOCCER

Abstract

The digitization of professional soccer has produced high-dimensional datasets from event-based systems, tracking technologies, and physiological monitoring. Traditional statistical approaches often fail to model the nonlinear and multivariate dynamics of the game. Machine learning (ML), particularly artificial neural networks (ANNs) and decision tree–based models, has emerged as a powerful framework for performance prediction; however, methodological variability limits consolidated interpretation. Objectives: To systematically review and meta-analyse the predictive performance of neural network and decision tree models in competitive soccer across match outcome prediction, injury risk modelling, technical–tactical analysis, and physical workload monitoring. Methods: Following PRISMA 2020 guidelines, databases (Scopus, Web of Science, PubMed, IEEE Xplore, SPORTDiscus) were searched for studies published between January 2000 and March 2025. Twenty-two studies met qualitative criteria, and sixteen were included in a random-effects meta-analysis. Pooled accuracy and Area Under the Curve (AUC) were calculated. Heterogeneity was assessed using I² statistics. Results: The pooled overall accuracy was 86.4% (95% CI: 82.1%–90.7%) with an AUC of 0.90. Neural networks showed higher pooled accuracy (88.2%) than decision tree models (84.7%). In match outcome prediction, ANN models achieved an AUC of 0.92 compared to 0.88 for decision trees. Injury risk modelling reported accuracies of 85.3% (ANN) and 83.9% (decision tree). The overall effect size was moderate-to-large (Cohen’s d = 0.58). Moderate heterogeneity was observed (I² = 41%). Conclusion: Machine learning models demonstrate strong predictive performance in professional soccer analytics. Neural networks excel in modelling nonlinear dynamics, while decision trees provide competitive accuracy with greater interpretability. Standardized validation and explainable AI approaches are recommended to enhance generalizability and practical application.