<p>Autonomous vehicles (AVs) have demonstrated significant potential to enhance traffic safety and efficiency. However, their performance in high-risk situations remains underexplored. This study aims to identify the determinants of AV crash risks in high-risk testing scenarios through a novel analysis framework. The framework comprises three key components: scenario construction, risk modeling, and model interpretation. Data for this analysis are obtained from virtual simulation safety tests of AVs under high-risk scenarios generated by a Time-Series Generative Adversarial Network using in-depth crash data. To overcome the limitations of traditional statistical methods in analyzing high-dimensional data and the opacity of machine learning algorithms, this study employs interpretable machine learning techniques. Key contributions include a focus on analyzing crash risks in high-risk virtual testing scenarios, the generation of extensive AV safety test data through crash reconstruction and virtual simulation, and the application of interpretable machine learning to identify critical crash determinants. The findings highlight the superior predictive performance of advanced machine learning models, particularly eXtreme Gradient Boosting with feature interaction constraints, in identifying crash risks and their contributing factors. Key determinants, such as the initial velocity of the ego vehicle, weather conditions, and the relative distance between the ego and objective vehicles, are identified. These insights are instrumental in optimizing sensor configurations and refining autonomous driving algorithms to enhance AV safety in high-risk environments.</p>