Enhancing CNN-Based Motion Sickness Detection Using EEG Signals Weighted by Motion Sickness Dose Value

초록

With the advancement of autonomous driving technologies, passengers increasingly engage in nondriving activities. However, these activities are often limited by motion sickness (MS), which worsens with vehicle vibration characterized by longer exposure duration and greater magnitude. Although recent studies have detected MS using EEG-based artificial intelligence models, they have often overlooked the characteristics of MS in vehicles. In this study, we propose a method to enhance the performance of EEGbased convolutional neural network (CNN) models by incorporating the motion sickness dose value (MSDV) defined in ISO 2631-1, which considers both exposure duration and vibration magnitude. For the CNN models, we selected ShallowConvNet, EEGNet, TSception, and EEG-Conformer, which are widely used in brain-computer interface (BCI) and emotion recognition. The proposed method assigns MSDV-based weighting factors to electroencephalogram (EEG) segments, with MSDV computed from triaxial head acceleration. To validate the effectiveness of the approach, EEG signals were recorded from 42 participants who were exposed to various simulated road conditions with a motion simulator. Results demonstrated that the MSDV weighted models exhibited statistically significant improvements in classification performance compared to the models without MSDV. Furthermore, classification accuracy was enhanced by applying a cascade adaptive filter to eliminate movement artifacts induced by whole-body vibration (WBV). The proposed method achieved within-subject (WS) classification accuracies of 94.55%-95.54% and cross-subject (CS) accuracies of 67.92%-74.77%, outperforming existing MS detection models. These findings suggest that integrating EEG signals with MSDV weighting provides a promising pathway for accurate real-world MS detection.

키워드