A Random Forest model was trained on a large real-world road-crash dataset (328,407 cases) to predict the occurrence of physical injury. A total of 19 explanatory variables were considered, covering driver, vehicle, environmental, and contextual characteristics. To address the significant class imbalance, eight resampling techniques were tested, including SMOTE, undersampling, and Tomek Links. The model was optimized through an exhaustive grid search combined with 10-fold cross-validation, using F1-score as the optimization criterion. The best model achieved an F1-score of 58.4, an AUC of 85.1, and a G-Mean of 75.4 on the hold-out test set. The feature importance analysis confirmed several well-established findings in road safety research, notably the strong influence of vehicle category (with motorcycles standing out), accident type, and the greater vulnerability of teenage drivers. At the same time, the model revealed less intuitive patterns: run-off-road crashes were identified as the most likely to result in injuries, whereas collisions with obstacles within the roadway were associated with the lowest injury rates. These results demonstrate the potential of Random Forest for identifying relevant risk factors, combining predictive performance with interpretability, and supporting more effective injury prevention strategies. Furthermore, the model validated known risk factors while uncovering less intuitive patterns, reinforcing its value for evidence-based road safety analysis.