Published: Građevinar 77 (2025) 12
Paper type: Original scientific paper
Download article (Croatian): PDF
Download article (English): PDF
Predicting unsafe road sections using machine learning
Abstract
This paper presents an ML methodology to predict hazardous road segments, using the weighted accident index (Wi). The analysis covers 161 road segments in North Macedonia (~1,300 km)—with 23+1 variables categorized into Road, Traffic, Environmental, and Accident data. Feature influence is evaluated using six models with an 80/20 training/testing split. Weighted SHAP is applied to obtain a single variable ranking; XGBoost with 15 inputs is the final predictor. The model achieves a validated performance (R² = 0.65), while operational prioritization yields AUROC = 0.69 at Wi ≥ 10.13, enabling timely identification of hazardous segments and interventions by relevant authorities.
Keywordsroad safety, machine learning, prediction, SHAP, weighted accident index, traffic analysis
