Interpretable Feature Selection for Truck Collision Injury Severity: A SHAP-RFE Approach

Sana Chodri, Hossein Zolfagharinia, Aida Haghighi

Keywords: interpretable machine learning, feature selection, SHAP, truck collisions, multiclass classification, injury severity prediction

TL;DR: We propose an interpretable feature selection framework to improve injury severity prediction in truck collisions and enhance supply chain safety and resilience.

Abstract:

Truck-involved collisions pose a significant safety and operational risk within supply chains, often resulting in costly disruptions, injuries, and delays. Accurate and interpretable prediction of injury severity is critical for supporting proactive safety interventions and risk mitigation strategies. This study presents a SHAP-guided Recursive Feature Elimination (SHAP-RFE) framework for identifying the most informative features related to injury severity in truck crashes, using data from the 2022 Fatality Analysis Reporting System (FARS).

We compare SHAP-RFE against two benchmark feature selection methods: Principal Component Analysis (PCA) and a literature informed feature set synthesized from 58 prior studies. Our approach achieves the highest adjusted macro F1-score, while selecting a compact set of 26 interpretable features. Notably, 20 of these overlap with domain-validated risk factors, confirming strong alignment with existing research.

The results highlight SHAP-RFE’s ability to balance performance and interpretability in imbalanced multiclass classification tasks. This interpretable framework offers practical value for transportation safety planners and logistics decision-makers seeking to reduce crash impact and enhance supply chain resilience.

Submission Number: 23