Predicting Severity of Expressway Traffic Crashes Using XGBoost: Evidence from a Cold-Region Highway in China

This study aims to predict crash severity on expressways in Heilongjiang Province, China, using the XGBoost algorithm and official crash records from the G1011 Harbin-Tongjiang Expressway (2011-2023). After rigorous data curation, a final dataset of 328 crashes was obtained, with severity categorized as property loss (58.2%), injuries (31.7%), and fatalities (10.1%). Given the severe class imbalance and limited sample size, this study employed stratified 5-fold cross-validation to provide a robust and unbiased estimate of model performance. The model achieved a mean overall accuracy of 0.821, macro-Averaged F 1-score of 0.764, and Cohen's kappa coefficient of 0.698. The recall for fatal crashes was 0.720, demonstrating the model's ability to identify most actual severe crashes despite their rarity. Feature importance revealed that dynamic environmental and human factors play a dominant role in escalating crash severity. Icy or snow-covered road surfaces, low visibility, and unlit nighttime conditions, which are characteristic of the harsh winter environment in Heilongjiang Province, underscore the heightened vulnerability of expressway safety during cold periods and provide a data-driven foundation for developing targeted safety management strategies.