Machine learning insights into AV crash dynamics: predicting injury outcomes

Road traffic-related fatalities, most of which are caused by human error, are a major problem globally. Autonomous vehicles (AVs) eliminate the potential for human error and thus hold promise for reducing the number of fatalities and improving overall road safety. Existing literature lacks comprehensive studies on predicting injury outcomes of AV-related crashes using advanced machine learning techniques. Most research has been constrained by limited datasets and traditional statistical methods, and often fails to address class imbalances in crash data, which can skew analysis and predictions. This study aims to fill this gap by providing new insights into predicting the injury outcomes of collisions involving AVs. The research employed advanced machine learning and a comprehensive dataset derived from crash reports assembled by the California Department of Motor Vehicles, covering the period from 2014 to July 2023. Imbalances in the crash data were addressed by class weighting techniques used in conjunction with stratified sampling. The data was analyzed using a range of machine learning algorithms, including logistic regression, random forests, bagging classifiers, decision trees, gradient boosting, and easy ensemble methods. The results showed that the bagging classifier was the best model, as it minimized false negatives while striking a balance between sensitivity and specificity. This research will benefit stakeholders such as transportation experts, AV manufacturers, and urban planners to improve road safety, advance AV technology, and inform a wide range of stakeholders about the implications of autonomous driving. Future research directions could involve expanding the geographic scope of data collection to include diverse driving environments, integrating detailed telematics and sensor data for a more comprehensive analysis, and exploring the impact of different regulatory frameworks on AV performance and safety.