Selective Transfer Learning for EEG-Based Drowsiness Detection

On the pathway from laboratory settings to real world environment, a major challenge on the development of a robust electroencephalogram (EEG)-based brain-computer interface (BCI) is to collect a significant amount of informative training data from each individual, which is labor intensive and time-consuming and thereby significantly hinders the applications of BCIs in real-world settings. A possible remedy for this problem is to leverage existing data from other subjects. However, substantial inter-subject variability of human EEG data could deteriorate more than improve the BCI performance. This study proposes a new transfer learning (TL)-based method that exploits a subject´s pilot data to select auxiliary data from other subjects to enhance the performance of an EEG-based BCI for drowsiness detection. This method is based on our previous findings that the EEG correlates of drowsiness were stable within individuals across sessions and an individual´s pilot data could be used as calibration/training data to build a robust drowsiness detector. Empirical results of this study suggested that the feasibility of leveraging existing BCI models built by other subjects´ data and a relatively small amount of subject-specific pilot data to develop a BCI that can outperform the BCI based solely on the pilot data of the subject.