24 | Prediction of carsickness using multisensor data and machine learning approaches

Henry E, Hosni N, Belkacem R, Fenaux E, Bringoux L, Bougard C | Aix Marseille Univ, CNRS, ISM, Stellantis & Telecom Paris, France

Background: Carsickness is a complex phenomenon influenced by internal and external factors, with significant individual variability in symptom onset and progression. Traditional group-level analyses often overlook these differences. This study aimed to identify objective indicators of carsickness and develop a classification model using machine learning (ML) based on vehicle dynamics, physiological, and postural data.

Materials and Methods: Data were collected from multiple experimental driving sessions involving repeated slalom maneuvers (~300 m each). Each session was segmented and time-indexed to track symptom progression. Participants self-reported symptoms on a 5-point Likert scale, later binarized into “carsick” and “non-carsick.” Feature selection was performed using the Gram-Schmidt process. Six ML models—Naïve Bayes, Decision Tree, Random Forest, Support Vector Machine (SVM), Explainable Boosting Machine (EBM), and K-Nearest Neighbors (KNN)—were trained using default scikit-learn settings. Model performance was assessed with ROC curves and AUC metrics.

Results: With vehicle dynamics data alone, the EBM model achieved the highest AUC (0.79). Adding postural data improved accuracy, with EBM and Random Forest models reaching an AUC of 0.85. The best results came from combining vehicle, physiological, and postural inputs, achieving AUCs up to 0.93. These findings highlight the potential of multisensory data and ML techniques for detecting carsickness retrospectively. They also support future development of personalized prediction models based on individual exposure history.