Hybrid Machine Learning Algorithm for Identifying Feature Levels Associated with Safety Critical Events

Abstract

According to the National Highway Traffic Safety Administration (NHTSA), about seven million traffic accidents claimed more than 36,560 human lives in the U.S. in 2018 . These statistics have prompted researchers to investigate the driver characteristics associated with safety-critical events (SCE). This paper presents a hybrid CatBoost algorithm for identifying the feature levels associated with SCE. The model accounts for numerous difficulties and drawbacks reported in the literature. The model was trained and validated using the entire set of the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP2-NDS) events (crash/near-crash and normal/baseline). Results indicate that secondary tasks (interacting with object in-vehicle, reaching for objects in the vehicle, pet interaction, cellphone/tablet use, and writing/texting), intersection influence (parking lots/driveway/entrance/exit, uncontrolled intersections, traffic signals, interchanges, and stop signs), income (under $29,000 and $100,000–$149,000), age (16–19 and 20–24), traffic density (level of service C, D, and E/F), high sensation-seeking tendency (scoring 18–35 on a scale of 35), low driving knowledge (scoring 0–8.9 on a 19-point scoring system questionnaire), and gender = female are the feature levels having an association with SCE with a probability varying between 51% and 87%. Results also revealed that passenger interactions, eating/drinking, driving away from intersections or interchanges, being age 70—79, or driving in traffic density = A are more related to safe driving. Consideration of these results can contribute to reducing roadway crashes and improve traffic safety.