Evaluating Human Trust States in Human-AI Collaboration During Driving Using EEG: A Machine Learning Approach

Trust plays a pivotal role in determining whether drivers accept autonomous driving during human-AI collaboration driving. Calibrated trust enables the optimal performance of automated systems, thereby fostering harmonious collaboration between driver and autonomous driving. The cornerstone for such calibration lies in real-time evaluation of trust, an area that has recently drawn increased attention through neuromeric-based assessments. This study aims to employ Electroencephalography (EEG) technology to evaluate the real-time trust states of drivers engaging with autonomous driving systems. Leveraging a simulated driving platform, EEG data is collected from drivers by designing a human-AI collaboration obstacle avoidance task. Trust is operationally defined as a binary state, based on whether the driver intervened to take control or not. Utilized 8 distinct machine learning algorithms are utilized to construct a real-time trust evaluation model, subsequently comparing their respective F1 score. Furthermore, feature analysis is conducted on the top-performing models to identify those features that exert a significant influence on trust levels. The findings revealed that Random Forest (RF) model performed best on our dataset, with the F1-score 0.997. Multiple features in the time domain and frequency domain contribute significantly to the trust evaluation model.