Background: Road accident severity estimation is a critical
aspect of road safety analysis and traffic management. Accurate severity
estimation contributes to the formulation of effective road safety policies.
Knowledge of the potential consequences of certain behaviors or conditions can
contribute to safer driving practices. Identifying patterns of high-severity
accidents allows for targeted improvements in terms of overall road safety.
Objective: This study focuses on analyzing road accidents
by utilizing real data, i.e., US road accidents open database called “CRSS.” It
employs advanced machine learning models such as boosting algorithms such as
LGBM, XGBoost, and CatBoost to predict accident severity classification based on
various parameters. The study also aims to contribute to road safety by
providing predictive insights for stakeholders, functional safety engineering
community, and policymakers using KABCO classification systems. The article
includes sections covering theoretical methodology, data analysis, model
development, evaluation, performance metrics, and implications for improving
road safety measures by comparing the performance of different boosting
algorithms on the CRSS dataset. This study aims to identify the most effective
machine learning algorithm to integrate into our product line in the near
future, enabling accurate prediction of both accident severity and occurrence.
Results and Conclusions: This study addresses challenges in
evaluating performance metrics for different severity classes within unbalanced
datasets, emphasizing the impact of dominant classes like Class O (O = no
apparent injury) on overall accuracy. The investigation reveals the limitations
and conservatism associated with imbalanced data in boosting models, hinting at
a potential ceiling in their performance around 80%. Comparative analysis of
algorithms, including CatBoost, XGBoost, and LGBM, demonstrates comparable
performance even in the case of applying KNN algorithm for pre-processing, based
on various metrics, especially accuracy, F
1-score, ROC-AUC, and PR-AUC for all severity classes. XGBoost with
KNN algorithm did not show any significant performance improvement compared to
the XGBoost without KNN algorithm. The study includes performance metrics, such
as F
1-score, CM upper triangle, ROC-AUC, and PR-AUC applied to an
accident analysis case study. Future work directions involve extending the
application of CatBoost, XGBoost, and other algorithms to diverse datasets,
exploring the capabilities of deep neural networks, refining dataset preparation
for accuracy improvement, and creating unified tools for hazard analysis and
risk assessment.
