Machine Learning-Based Prediction of Road Accident Risk Using Driver Behavior and Environmental Factors
Abstract
Predicting whether a road accident will be fatal is not as straightforward as published accuracy figures might suggest. Many models that perform well have been developed using very large datasets from Europe or North america. When similar approaches were applied to a 3000 record Indian dataset, none of the features had a pearson correlation with accident severity above 0.046. Three classifiers-Logistic Regression, Decision Tree, and Random Forest-were tested using 22 accident related features covering driver details, road and weather conditions, and vehicle type. Since the binary target was imbalanced, SMOTE was applied to the training data. Five-fold stratified cross-validation was then used to assess model performance given the small dataset size. On the test set, the Decision Tree produced the highest recall for fatal cases, with a recall of 0.188 and an F1 score of 0.261. Random Forest, on the other hand, had the most stable cross-validation performance, with an F1 score of 0.734±0.023. A continuous risk Score was also developed from the Random Forest output. It is calculated by multiplying the model's predicted probability of a fatal accident by 100. This allows traffic authorities to rank accident scenarios without having to choose a fixed cutoff. Also, Speed limit, month of the year, and driver age turned out to be the three most important features based on Mean Decrease in Impurity. The primary finding demonstrates that fatal accident risk depends on many factors together, not any single one.