In L2 automated driving, monotonous tasks induce passive fatigue, yet most prediction models ignore the key psychological variable of human-machine trust. This study incorporates dynamic driver trust as a core feature to build a more accurate fatigue prediction model. A naturalistic driving experiment was conducted to collect multimodal physiological and subjective (fatigue and real-time trust) data. A Variational Autoencoder (VAE) was employed for data augmentation to address limited sample sizes. The results demonstrated that among several machine learning models, LightGBM performed best after integrating the trust feature. The F1_score significantly increased from 0.840 (without trust) to 0.879, achieving 90.9% classification accuracy. SHAP analysis confirmed trust as the fifth most important predictive feature. This research validates the necessity and effectiveness of including dynamic trust in fatigue prediction, significantly enhancing model accuracy and supporting the development of next-generation driver monitoring systems.