Efficient Histogram-Based Gradient Boosting Approach for Accident Severity Prediction With Multisource Data

Abstract

Many people lose their lives in road accidents because they do not receive timely treatment after the accident from emergency medical services; providing timely emergency services can decrease the fatality rate as well as the severity of accidents. In this study, we predicted the severity of car accidents for use by trauma centers and hospitals for emergency response management. The predictions of our model could be used to decide whether an ambulance unit should be dispatched to the crash site or not. This study used histogram-based gradient boosting (HistGBDT), a modification of the gradient boosting (GBDT) classifier that accelerates the learning process and increases a model’s prediction power. The HistGBDT model was compared with seven state-of-the-art machine learning models: logistic regression, multilayer perceptron, random forest, extremely randomized trees, bagging, AdaBoost, and GBDT. The experiments were conducted on French accident data from 2005 to 2018. The HistGBDT model, with an overall accuracy of 82.5%, recall of 76.7%, and precision of 81.9%, outperformed other models. An analysis of feature importance indicated that safety equipment was the most important feature and vehicle category, department, localization, and region were other significant features. The Fβ measure (i.e., the weighted harmonic mean of recall and precision) was optimized with different weights on recall for the four best performing models to compare the tradeoff between the two crucial performance measures.