Interpretable machine learning for risk analysis of truck driver injury severity in fatal collisions

This study develops an interpretable machine learning framework to analyze injury severity outcomes among truck drivers involved in fatal motor vehicle collisions. Although every analyzed collision involved at least one fatality, truck-driver injury outcomes spanned the full range from no injury to fatal injury, highlighting substantial heterogeneity in driver injury severity despite the shared fatal-collision context Using data from the Fatality Analysis Reporting System (FARS) for 2022–2023, the study identifies key factors associated with variation across the three truck-driver injury severity classes. To address the limitations of traditional econometric models and the opacity of black-box machine learning approaches, the proposed framework integrates multiple feature selection strategies, including literature-guided subsets and recursive feature elimination (RFE), with a SHAP-based interpretation. In particular, SHAP-guided RFE is adapted for multiclass injury severity prediction using class-specific SHAP attribution and macro-SHAP feature ranking, supporting parsimonious and interpretable model development. The results show that machine learning models effectively capture nonlinear predictive relationships among driver, roadway, environmental, vehicle, and collision-related factors. The SHAP-based analysis provides both global and class-specific explanations, revealing the highest-ranked predictors and how their SHAP-based importance varies across injury severity classes. These findings highlight the importance of combining predictive modeling with interpretable feature selection to derive safety-relevant insights from complex crash data. This study contributes to the development of interpretable, data-driven approaches for safety assessment in transportation systems. The proposed framework supports risk-informed prioritization by identifying the factors most predictive of truck-driver injury severity and highlighting areas for further safety evaluation.