Despite the high severity of pickup truck crashes, research focused on this fleet in Iran remains limited. This study introduces an integrated framework for modeling crash severity, utilizing comprehensive preprocessing, supervised discretization, and a hybrid feature selection method to manage data dimensionality—an approach rarely reported in previous studies. To identify key determinants, the study analyzed 162,745 national crash reports (2014–2018). Following data cleaning, missing value imputation, and supervised discretization, 191 binary variables were generated. Recursive feature elimination was performed using a Random Forest model, removing the three least important features (based on Gini impurity) in each iteration. A Wilcoxon signed-rank test with Holm correction confirmed that model performance with 50 features was not significantly different from the 191-feature model. The study then evaluated several models: Support Vector Machine (SVM), Decision Tree, Extreme Gradient Boosting (XGBoost), Random Forest, Balanced Random Forest, K-Nearest Neighbors (KNN), and PCA-based KNN. XGBoost demonstrated the best performance across Matthews Correlation Coefficient, geometric mean, accuracy, and balanced accuracy metrics. Shapley analysis indicated that collisions with other vehicles (rear-end) and fixed objects are associated with lower severity, while pedestrian, motorcycle, and head-on collisions correlate with higher severity. Crash severity is elevated in Khorasan Razavi, at night, on main streets, and in residential areas. Furthermore, the Paykan vehicle brand, B1 licenses, and missing license status are linked to increased severity, whereas the Zamyad brand and Class 3 licenses are associated with reduced severity.