Efficient Mind-wandering Detection System with GSR Signals on MM-SART Database

Mind-wandering (MW) is a ubiquitous phenomenon where the attention involuntary shifts from task-related to task-unrelated thoughts, and thus MW has negative impacts on task performance during learning. In this paper, we propose a MW detection system with galvanic skin response (GSR) signals on the multi-modal for Sustained Attention to Response Task (MM-SART) database. To explore the relationships between GSR and MW, we extract total 119 features including time, frequency, entropy, and wavelet domain. By using XGBoost as the classifier, we can achieve 0.713 AUC on the MM-SART database. However, large number of features may cause high training complexity and long inference latency. To reduce the number of features and find the most dominant features related to MW, we apply Pearson’s correlation coefficients and the importance scores given by extreme gradient boosting (XGBoost) classifier. Experiment results show that by using 10 dominant features we can achieve 0.706 AUC, 70.3% accuracy, 70.8% weighted F1 score and 0.294 Cohen’s kappa score on the MM-SART database. Moreover, the latency of training and inference are significantly reduced by 5x and 184x respectively. In conclusion, we have proposed an efficient MW detection system with GSR signals on the MM-SART database.