Predicting Traffic-Accident Severity in Sarajevo (2019–2025): An Imbalance-Aware, Interpretable Machine-Learning Study
Systems · 07.11.2026 · 14 (11) · 1425
DOI: 10.3390/systems14111425 ↗Kurz gefasst
Fachartikel in der Zeitschrift Systems, online erschienen am 07.11.2026. Verfasst von Muhamed Begović und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Reliable classification of the severity of traffic accidents requires appropriate treatment of class imbalance. This paper applies machine-learning methods to a verified and standardised database of police-recorded injury accidents from Sarajevo Canton, Bosnia and Herzegovina: 1517 accidents from August 2019 to December 2025, with 28 predictors. Severity was defined as a binary Killed or Seriously Injured (KSI) outcome: 410 Major (27.0%) and 1107 Minor (73.0%) accidents. Seventeen classifiers and a soft-voting ensemble were compared under 10 × 10 repeated stratified cross-validation, with Major-class recall as the primary measure. Precision, F1, balanced accuracy, the areas under the receiver operating characteristic and precision–recall curves (ROC-AUC, PR-AUC), and calibration were also reported. Classifiers were compared with the Nadeau–Bengio corrected resampled t-test. Random Data Partitioning with Voting Rule (RDPVR; recall 0.822 ± 0.056) and Balanced Random Forest (0.821 ± 0.063) recognised the most KSI accidents and did not differ significantly. The soft-voting ensemble and cost-weighted eXtreme Gradient Boosting (XGBoost) offered a better precision–recall balance (F1 0.749 and 0.733) and better-calibrated scores than these two ensembles. Leak-free threshold tuning moved the models along the precision–recall trade-off without materially changing F1. SHapley Additive exPlanations (SHAP) and permutation importance identified pedestrian involvement, location, accident year, primary-road category, and vertical geometry as the most important predictive features. Because police coordinates are approximate, a supportive analysis without them was also run: recall of the two leading ensembles fell only slightly, to 0.79–0.80. Performance was largely retained under spatial block validation. Forecasting from earlier years worked for 2022 (ROC-AUC 0.77–0.82) and moderately for 2023 (0.61–0.78), but collapsed for 2024–2025 (0.34–0.57). The models are post-crash severity classifiers, not prospective crash-location predictors, and must be revalidated before being applied to later years.
Zitieren
APA
Begović, M., Hassanat, A. B., Kosovac, A., Omerhodžić, A., & Kovačević, T. (2026). Predicting Traffic-Accident Severity in Sarajevo (2019–2025): An Imbalance-Aware, Interpretable Machine-Learning Study. Systems, 14(11), 1425. https://doi.org/10.3390/systems14111425
Vancouver
Begović M, Hassanat AB, Kosovac A, Omerhodžić A, Kovačević T. Predicting Traffic-Accident Severity in Sarajevo (2019–2025): An Imbalance-Aware, Interpretable Machine-Learning Study. Systems. 2026;14(11):1425. doi:10.3390/systems14111425
BibTeX
@article{Begovi2026,
author = {Begović, Muhamed and Hassanat, Ahmad B. and Kosovac, Amel and Omerhodžić, Adnan and Kovačević, Teodora},
title = {Predicting Traffic-Accident Severity in Sarajevo (2019–2025): An Imbalance-Aware, Interpretable Machine-Learning Study},
journal = {Systems},
year = {2026},
volume = {14},
number = {11},
pages = {1425},
doi = {10.3390/systems14111425}
}
EUVIMED