ICHS-SP: An Integrated Chainage-based Hotspot Severity Prediction Framework for Road Accident Risk Assessment Using Ensemble Machine Learning and Spatial Statistics

Abstract

Traffic accidents are a major public health issue worldwide, with  million deaths annually. Small dataset size, imbalance in classes, and the applicability of frequency-based hotspots screening in the national highway corridors of developing economies are challenges to road accidents severity prediction to capture the severity-concentrated risk. The current paper introduces the Integrated Chainage-based Hotspot Severity Prediction (ICHS-SP) system that incorporates multi-method spatial analysis, a new Crash Severity Risk Index (CSRI), and a hybrid ensemble machine learning system in an integrated road safety evaluation system to be applied to the NH-48 corridor (Chitradurga-Davangere, Karnataka, India). Kernal Density Estimation (KDE) and DBSCAN clustering, and the Getis-Ord Gi* statistic of  geo-referenced accident records on five pre-defined chainage zones (CH  km) were run using spatial hotspot analysis. The accident density was the highest at Zone (CH  km) (  accidents/km, ), whereas the proportion of fatal accidents in Zone  (CH  km) was the largest on the corridor, although with only  accidents. In point of rank, Zone  took the top risk priority ( , Critical) even though it has a fourth-place rank in terms of frequency, showing a rank inversion that reveals the systematic misallocation risk of prioritization on a frequency basis. To predict the severity of accidents, a hybrid ensemble of XGBoost, CatBoost   ANN with weighted probability fusion ( ) had the highest overall accuracy of  and a weighted AUC-ROC of  with a -record held-out test set (  split). Cross-validation was repeated five times with the result of , which validated that the model is generalisable and there is no overfitting. The ICHS-SP framework is a computationally manageable severity-sensitive risk assessment device that can be used on data constrained highway corridors, and has direct utility in the evidence-based prioritization of safety interventions.