Answers are grounded in this paper and cite their sources.
BackgroundSports injuries critically constrain athletes' competitive performance and career longevity. Current warning methods struggle with multidimensional data, and complex machine learning models often lack interpretability.ObjectiveThis study aims to construct and validate a highly accurate, interpretable multimodal sports injury risk prediction model, offering a transparent, potential decision-support tool for sports medicine teams.MethodsUtilizing a 5430-sample dataset partitioned at a 7:3 ratio, we applied SMOTE to mitigate class imbalance. A tri-stage feature selection strategy (Boruta, LASSO, MRMR) screened core predictive metrics across physiological, biomechanical, environmental, and training workload dimensions. We compared 12 machine learning models using AUC, accuracy, calibration curves, Brier scores, and DCA. SHAP and LIME frameworks provided interpretability, leading to the exploratory deployment of an R-Shiny interactive warning platform.ResultsFeature engineering established 18 core metrics. The Random Forest (RF) model demonstrated superior overall performance, achieving 0.971 validation accuracy, 0.982 AUC, and favorable internal stability. RF also excelled in probability calibration (Brier: 0.0332) and intervention net benefit. Interpretability analyses identified elevated electromyography amplitude, excessive ground reaction force, acute heart rate spikes, and previous injury history as primary acute injury drivers, exhibiting pronounced nonlinear threshold effects.ConclusionThe RF-based multimodal prediction model effectively balances predictive efficacy, internal stability, and interpretability. The developed online warning system represents a beneficial exploration in translating complex algorithms into practical applications, showing future promise to provide frontline coaching teams with objective support for personalized risk profiling and workload management.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: