Impact of feature engineering on predicting non-contact football injuries

Non-contact injuries in professional football impose significant performance and economic burdens, yet the influence of workload feature representation on injury risk modelling remains insufficiently characterised. Traditional monitoring approaches, including the acute-to-chronic workload ratio (ACWR), may inadequately capture the temporal dynamics and instability of training load that underlie injury aetiology. This study systematically compared four complementary temporal feature engineering strategies-rolling workload aggregates, workload balance and exponential smoothing metrics, stability and stress indicators, and polynomial regression residuals-to evaluate their relative discriminative contribution to non-contact injury risk prediction in professional football. GPS-derived external load data from 69 professional male football players across two clubs were analysed over one full competitive season. A total of 23 non-contact injury events were recorded; under a 7-day pre-injury risk window labelling scheme, these generated 109 positive athlete-day observations across 10,134 total daily observations (1.08% positive prevalence). Decision Tree (DT), Random Forest (RF), and XGBoost models were evaluated using stratified group k-fold cross-validation with athlete-level grouping to prevent data leakage. Minimum redundancy-maximum relevance (mRMR) feature selection was applied independently within each fold. Model performance was assessed using Recall, F2-score, ROC-AUC, and Precision-Recall AUC (PR-AUC). Overall sensitivity remained limited across baseline configurations, reflecting the extreme class imbalance of injury data. Polynomial residual features, encoding deviations from expected workload trajectories, produced the most consistent gains in discriminative capacity across models (mean ΔROC-AUC + 0.078; largest absolute improvement: RF ΔROC-AUC =  + 0.131). Compact mRMR-selected subsets (34-42 variables) consistently outperformed full feature spaces. ACWR-based features degraded performance across all classifiers (mean ΔROC-AUC - 0.023). A supplementary optimisation analysis demonstrated that, under calibrated hyperparameters and SMOTE oversampling, RF achieved Recall = 0.667 and ROC-AUC = 0.676 on an independent held-out test set, confirming that the near-zero baseline Recall reflects deliberate methodological conservatism rather than fundamental feature inadequacy. Within this dataset, deviation-based workload representations provided greater discriminative value than traditional ratio-based indicators, suggesting that temporal instability and unexpected departures from established training patterns may carry more predictive information than absolute load magnitude. Given the limited sensitivity achieved under default model configurations, these findings should be interpreted as exploratory methodological evidence rather than a basis for immediate clinical deployment. Future work should integrate larger multi-season datasets, internal load markers, and prospective validation to improve clinical utility.