Machine Learning Applications in Non-Contact Lower Limb Sports Injury Prediction: A Systematic Review

Non-contact Lower limb sports injuries represent some of the most prevalent and impactful conditions within athletic populations, prompting increasing interest in predictive approaches that can inform prevention and rehabilitation strategies. With its capacity to manage high-dimensional and complex datasets, machine learning (ML) has emerged as a promising tool for injury risk prediction. This systematic review, conducted in accordance with PRISMA 2020 guidelines, synthesized evidence from studies retrieved through Web of Science, PubMed, and SPORTDiscus (EBSCO). The literature search was conducted on January 20, 2025. Following independent screening and risk of bias assessment using the PROBAST tool, 15 studies were included from an initial pool of 92. The majority of study populations comprised adult athletes, with basketball and football (soccer) being the most frequently investigated sports. Random Forest and logistic regression were the most commonly applied algorithms, while tree-based approaches yielded the strongest predictive performance in 6 studies. Across 14 studies, area under the curve (AUC) values were reported, with one CHAID-based decision tree achieving the highest performance (AUC = 0.91), and sensitivity values reaching up to 0.92 in eight studies. Importantly, model interpretability was addressed in 87% of included studies, underscoring its emerging importance for clinical translation. Overall, ML exhibits considerable potential in predicting non-contact lower-limb injuries, but its practical value depends on achieving a balance between accuracy, transparency, and reliability. Future research should emphasize the integration of multi-source data and large-scale prospective validation to advance the translation of ML models into precision injury prevention and rehabilitation practice.