<p>Most current experimental studies on motion sickness prediction are conducted in laboratories, where the constant excitation frequency differs significantly from real-world dynamic conditions. In this paper, we break this paradigm by adopting a mountain city natural driving route and executing a comprehensive set of motion sickness test procedures. These procedures aim to leverage vehicle data and capture occupants actual motion sickness levels as authentic data inputs to predict the passengers motion sickness in real-world conditions. Based on LSTM neural network and attention mechanism, a new model which named as LSTM-A is integrated into the 6DOF SVC fundamental model. Experimental results indicate that the proposed LSTM-A model outperforms both the LR models and GRU models. Specifically, the RMSE is improved by 6.12% compared to the LR model and by 5.78% compared to the GRU model. The MAE is enhanced by 9.52 and 7.94%, while the MCC is improved by 13.64 and 11.25%. These significant improvements suggest that the LSTM-A model enabling it to capture multi-directional motion sickness dynamics more effectively. The LSTM-A model delivers exceptional accuracy in predicting individual motion sickness levels, establishing a robust foundation for innovative theoretical and methodological advancements in the field of motion sickness prediction.</p>