Intelligent Road Accident Risk Forecasting using Spatio-Temporal Data and Machine Learning
Abstract
Road traffic crashes cause serious injuries, deaths and economic losses annually, and are a primary public health problem. Accurate prediction of accident risk can help authorities implement effective preventive measures and improve road safety. This study presents a machine-learning-based analysis of road accident severity using a linked accident–vehicle–casualty dataset of 59,998 recorded accidents. Three classifiers – Decision Tree, Random Forest, and Logistic Regression – are trained and compared, together with a Gradient Boosting baseline, under a documented preprocessing and feature pipeline. Because the target variable (accident severity) is heavily imbalanced (86.6% Slight, 12.3% Serious, 1.1% Fatal), models are evaluated using macro-averaged precision, recall, and F1-score, confusion matrices, and multiclass ROC-AUC, in addition to overall accuracy, and are validated with 5-fold stratified cross-validation, a chronological (forecast-style) holdout, and a geographic holdout across unseen districts. The results show that although ensemble methods reach approximately 86–87% raw accuracy, this is close to the accuracy of a naive majority-class classifier, and macro-averaged metrics reveal substantially weaker detection of Serious and Fatal accidents. Class-weighted Logistic Regression trades overall accuracy for more balanced recall across severity classes. These findings indicate that accuracy alone is an insufficient metric for this task, and that severity-class imbalance must be explicitly addressed and reported when evaluating accident-risk models.