Machine learning models for injury risk prediction in football players: A systematic review of predictors, performance, practical applications, and limitations.

ObjectiveFootball is associated with a high incidence of musculoskeletal injuries because of its dynamic, high-intensity, and contact nature. Machine learning methods offer the potential to capture nonlinear relationships among diverse risk factors. This systematic review aimed to evaluate the application, accuracy, and practical utility of machine learning models for predicting injuries in football players.MethodsThis systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 guidelines. A systematic search was performed in Web of Science, PubMed, and Scopus, along with manual searches, through October 2025. Cohort and cross-sectional studies applying machine learning methods for injury prediction in football players were included (n = 22). Study quality and risk of bias were assessed using the Prediction model Risk of Bias Assessment Tool.ResultsWe performed a narrative synthesis because of substantial heterogeneity in study design, machine learning models, injury definitions, and predictor variables. Common predictors included training load metrics, previous injury history, biomechanical and neuromuscular characteristics, anthropometric variables, and psychological and biological indicators. Most studies reported moderate-to-high discriminative performance based on area under the curve, sensitivity, specificity, and precision metrics. Reported accuracy values ranged from approximately 55% to 96%. However, several studies relied only on accuracy as the primary metric, which is susceptible to misinterpretation in the presence of class imbalance between injury and noninjury observations; metrics such as sensitivity, specificity, F1-score, or area under the curve-ROC would provide a more balanced evaluation. Models generally demonstrated better performance for noncontact and soft tissue injuries than for contact injuries.ConclusionsMachine learning models may have exploratory value for injury risk stratification in football. However, the current evidence remains heterogeneous and methodologically limited. Further studies with large-scale designs, standardized methodologies, external validation, appropriate calibration, and clinical interpretability are required before these models can be recommended for routine injury-prevention practice.Review protocol was registered in the International Prospective Register of Systematic Reviews database (ID: 1279575).