A deep learning framework for time series prediction of passenger motion sickness based on vehicle dynamics data

Technology development for enhancing passenger experience has been gaining attention as autonomous vehicles (AVs) provide new opportunities for human-vehicle interaction. A new possibility for occupants of AVs is performing productive tasks as they’re relieved from the task of driving. However, it has been shown that non-driving-related tasks cause a conflict between the visual and vestibular sensory systems and can, consequently, lead to Motion Sickness (MS). To develop MS mitigating technologies, it’s advantageous to design a tool that can predict when MS is likely to occur. However, there is currently a lack of in-vehicle systems that are capable of estimating a passenger's MS state in real-time. Furthermore, the lack of real-time physiological data from vehicle occupants constrains the types of sensory data that can be used for estimation under realistic implementations. To address this, a computational model was developed to predict the MS score for passengers solely based on the vehicle's dynamic state. The model leverages self-reported MS and vehicle dynamics time series data acquired by a previous study performed under realistic driving conditions. The data consists of 66 trials (1 trial = 1350s) and includes MS score (1-10 scale), static experimental parameters, and vehicle IMU data. The models proposed are long short-term memory (LSTM) and nonlinear auto-regressive (NARX) deep learning regression frameworks capable of creating a time series mapping between vehicle sensor data, static parameters, and MS score. The predictive models were optimized in terms of their hyperparameters and their results were validated by analyzing the conformance between actual and predicted values of MS Score using statistical measures. NARX outperformed LSTM and yielded a prediction RMSE value below 2.5. The developed model framework is a promising solution for the estimation of MS when only vehicle data is available to the predictor module.