Predictive analytics in autonomous vehicles safety: crash outcome modeling

Traffic-related fatalities, often caused by human error, present a global challenge. Autonomous vehicles (AVs), designed to reduce these fatalities by eliminating human error, still experience crashes. Prior research, limited by the scope of the data set and analytical depth, has not fully explored these factors, resulting in a lingering lack of clarity about the cause of the crashes. The study utilized a comprehensive dataset spanning from 2014 to May 2024 and implemented a stratified sampling alongside class weighting techniques to rectify imbalances within the crash data. A variety of machine learning algorithms such as logistic regression, decision trees, random forests, gradient boosting, bagging classifiers, and easy ensemble methods were utilized to perform the data analysis. The random forest model stood out for its balance in sensitivity and specificity, reducing false negatives. The study revealed that factors such as the extent of vehicle damage, the AV manufacturer, and type of collision were critical predictors of crash injuries. The research provides valuable insights for various stakeholders, including transportation professionals, AV manufacturers, and policymakers. The findings can guide the development of enhanced safety protocols, improved vehicle designs, and more informed policy frameworks aimed at reducing AV-related injuries.