Abstract

This study presents a data-driven framework to uncover patterns of less-studied fatal driver behaviors. We analyzed 84,170 driver behavior reports from the Fatality Analysis Reporting System (FARS) between 2021 and 2023. Using natural language processing methods, we obtained 194 distinct lemmas and quantified their importance in each report. We then applied principal components analysis to obtain 18 underlying behavioral factors. Finally, we clustered the reports via Gaussian Mixture Modeling, obtaining a novel typology of 15 behaviors. Our results reveal spatial disparities in fatal driver behaviors and provide a framework for type-specific crash mitigation efforts.