Beyond threshold-based monitoring: A deep temporal learning framework with dual-path ACWR feature encoding for sports injury risk prediction

Sports injury is a core issue constraining professional athletes’ competitive performance and health, and the scientific management of training load is a key intervention for injury prevention. However, existing machine learning injury prediction models generally neglect the temporal dynamic characteristics and domain-specific semantics of the Acute:Chronic Workload Ratio (ACWR), resulting in limited prediction accuracy and clinical interpretability. This study proposes the ACWR-LSTM framework, which explicitly integrates dual-pathway ACWR feature engineering based on Rolling Average (RA) and Exponentially Weighted Moving Average (EWMA), four-level risk interval one-hot encoding, and a temporal attention mechanism into a two-layer stacked LSTM architecture, and employs Focal Loss to handle sample class imbalance. The primary methodological contribution lies in the systematic integration of sports-science domain knowledge—specifically the dual-path ACWR computation paradigm and evidence-based risk-zone encoding—into the input representation of a temporal deep learning model, a design that has not been previously validated in multi-sport injury prediction settings. Based on a multi-center dataset integrating 22 cohort studies (921 athletes, 657 injuries, covering four sports: soccer, tennis, rugby, and field hockey), the model was systematically evaluated via 5-fold temporal cross-validation and compared against six baseline methods: logistic regression, random forest, XGBoost, SVM, standard LSTM, and CNN-LSTM. ACWR-LSTM achieved AUC = 0.847 (95% CI: 0.831–0.863), sensitivity 78.1%, specificity 80.8%, F1 score 0.783, and Brier score 0.087 on the overall test set, significantly outperforming all baseline methods ($$p < 0.001$$); AUC ranged from 0.794 to 0.862 across sports, and the inter-fold standard deviation of 5-fold cross-validation was below 0.022. Ablation studies confirmed that removing ACWR features caused an AUC drop of 0.084, making it the most critical source of model performance gain; temporal attention weight heatmaps revealed that the model autonomously focused on the high injury-risk window 7–10 days after workload spikes, highly consistent with sports medicine clinical consensus. Extensive new robustness analyses—including a 2× 2 double-ablation matrix revealing a significant synergistic interaction between ACWR features and temporal attention ($$+0.021$$ AUC, $$p < 0.05$$), feature subset sensitivity, 10-seed data split sensitivity (bootstrap 95% CI [0.825, 0.868]), three-dimensional heterogeneity stratification (max ΔAUC $$= 0.018$$), and a failure mode analysis—confirm that model performance is stable across operational variability and not contingent on specific design choices. By explicitly embedding sports science domain prior knowledge into a deep temporal modeling framework, ACWR-LSTM achieves substantial improvements in injury prediction accuracy, cross-sport generalization, and clinical interpretability, while important limitations remain regarding external validation and data heterogeneity across the 22 integrated cohort studies. The framework provides a replicable methodological paradigm for domain-knowledge-guided temporal deep learning in multi-sport injury prevention scenarios.