Driver proficiency assessment using multidimensional driving simulator data and machine learning

This study proposed a driver proficiency classification framework based on multi-modal behavioral data collected from a desktop driving simulator, including gaze behavior, operational behavior, and vehicle dynamics. Multi-modal features were aggregated at the segment and lap levels, and five machine learning models were evaluated under subject-wise validation to examine participant-independent driver proficiency classification. The primary nested cross-validation results showed that the overall classification performance varied across models and evaluation metrics under the current dataset. The chance-level tests further showed that only SVM significantly exceeded random classification in segment-level accuracy, with a mean value of 52.3%, while Transformer showed a marginally significant AUC tendency at the segment level, with a mean value of 61.7%, and a significant AUC result at the lap level, with a mean value of 69.4%. In contrast, recall and f1-score more frequently exceeded the chance level, especially for Transformer, which showed segment-level recall of 81.5% and f1-score of 64.1%, and lap-level recall of 86.4% and f1-score of 64.9%. Additional analyses further examined the classification results from multiple perspectives, including differences between intersection and straight segments, comparison with traditional assessment methods, feature differences by driver proficiency, and feature importance analysis. Overall, this study provides a data-driven method for simulator-based driving behavior analysis by integrating gaze behavior, control input, and vehicle dynamics with machine learning models.