Machine learning-based sports injury prediction models demonstrated substantial predictive efficacy with a pooled diagnostic odds ratio of 9.02, a pooled sensitivity of 0.79, and a pooled specificity of 0.71.
Meta-Analysis (n=12,249)
Do machine learning-based prediction models accurately predict sports injuries in athletes?
Machine learning models demonstrate moderate to good overall predictive value for sports injuries, though their clinical utility is currently limited by significant methodological heterogeneity and a lack of external validation.
Odds Ratio: 9.02 (95% CI 4.31–18.88)
Predicting sports injuries is highly challenging, and current machine learning-based models for predicting sports injuries hold promise for addressing the difficulties in sports injury prediction. This study aims to assess the overall performance of existing prediction models in forecasting sports injuries. A comprehensive search was conducted in five databases: PubMed, Web of Science, Scopus, Cochrane Library, and IEEE Xplore. The search covered studies from the inception of each database until December 2025, and the performance of sports injury prediction models was reported. Subgroup and meta-regression analyses were performed based on potential influencing variables. Results: Ten predictive models across ten independent studies were ultimately included in the present analysis. The bivariate random-effects model yielded a pooled sensitivity of 0.79 (95% CI: 0.66–0.87) and a pooled specificity of 0.71 (95% CI: 0.61–0.80) for the included predictive models. The overall diagnostic odds ratio (DOR) was 9.02 (95% CI: 4.31–18.88), demonstrating substantial predictive efficacy. Furthermore, the pooled positive and negative likelihood ratios (PLR and NLR) were 2.72 (95% CI: 1.93–3.83) and 0.30 (95% CI: 0.18–0.49), respectively. Assessment of heterogeneity revealed no significant threshold effect, as evidenced by a weak correlation between logit sensitivity and logit specificity ( r = -0.11) and a non-significant asymmetry parameter in the HSROC model (beta= -0.30, P = 0.372). Potential factors such as sample size, model type, validation method, and injury type may influence model performance. Conclusion: Existing machine learning-based sports injury prediction models exhibit considerable methodological differences, and these methodological barriers affect the improvement and enhancement of model performance. It is recommended to apply imbalance handling to datasets with small sample sizes, use k-fold cross-validation to prevent overly optimistic prediction performance, and combine time-series data with other data for feature extraction. These suggestions need to be further validated through larger-scale studies to improve the performance of sports injury prediction models.
Lou et al. (Wed,) conducted a meta-analysis in Sports injuries (n=12,249). Machine learning prediction models was evaluated on Diagnostic odds ratio for predicting sports injuries (DOR 9.02, 95% CI 4.31-18.88). Machine learning-based sports injury prediction models demonstrated substantial predictive efficacy with a pooled diagnostic odds ratio of 9.02, a pooled sensitivity of 0.79, and a pooled specificity of 0.71.