A machine learning approach to understanding the road and traffic environments of crashes involving driver distraction and inattention (DDI) on rural multilane highways

• Published in: Journal of safety research 2025-02, Vol.92, p.14-26
• Main Authors: Yang, Chenxuan, Liu, Jun, Zhang, Zihe, Kofi Adanu, Emmanuel, Penmetsa, Praveena, Jones, Steven
• Format: Article
• Language: English
• Subjects: Data balancing; Driver distraction and inattention; Machine learning; Road and traffic environment; Rural highways
• ISSN: 0022-4375
• DOI: 10.1016/j.jsr.2024.11.011

•This study investigated crashes involving driver distraction and inattention on rural highways.•Multiple machine learning models were developed to identify factors contributing to crashes.•Two sampling methods (over-sampling and under-sampling) were used to balance the data for machine learning.•Modeling results indicated that the road and traffic environments are strongly linked to crashes in general.•A non-linear relationship between truck volumes and crashes was uncovered. Introduction: Driver distraction and inattention (DDI) are major causes of road crashes, especially on rural highways. However, not all instances of distracted or inattentive driving lead to crashes. Previous studies indicate that DDI-related driving behavior is closely associated with low-traffic and less complex driving environments. Nevertheless, it is unclear if these traffic or road environments also increase the likelihood of crashes involving DDI. Method: This study employed machine learning algorithms to identify the factors contributing to DDI-involved crashes on rural highways. This study applied multiple machine learning models including the Light Gradient Boosting Model (LGBM), Random Forest (RF), and Neural Network (NN) to quantify the correlations of DDI-involved crashes related to road and traffic environments. The study leveraged a statewide crash database with unique roadway data that contains variables for median type (e.g., 4-ft flush medians) and roadside access point density. To deal with the extreme imbalance of data, two sampling methods (over and under-sampling) were used to balance the data for machine learning. Results: Modeling results indicated that the road and traffic environments that are strongly linked to DDI-involved crashes in general overlap with the environments that lead to DDI-related driving behavior, except for the truck volumes in traffic. Crashes that involved DDI were more likely to occur in environments with non-traversable medians (compared to 4-ft flush medians), lower-volume traffic, and greater access spacing on roadsides. With regard to truck volumes, a non-linear relationship with the occurrence of DDI-involved crashes was uncovered. Traffic with about 8 to 10% of trucks is associated with the highest likelihood of DDI-involved crashes. Practical Applications: This study provides valuable information for drivers who need to be careful while driving in certain environments with a risk of DDI-involved crashes and for agencies who need