Adaptation to a brain-computer interface of transfer learning using event-related potentials obtained under different measurement conditions

Event-related potentials (ERPs) are widely used in electroencephalogram (EEG)-based brain-computer interfaces (BCIs). Because ERPs vary with age, individual differences, and task difficulty, identifying useful features is challenging. In recent years, deep learning, in which features can be acquired via learning, has been used to identify such features. However, a large amount of BCI data is required for deep learning, and data collection can be difficult. Therefore, we attempted to verify the applicability of BCI to transfer learning based on a considerable amount of data obtained from clinical examinations. The components of an ERP vary depending on measurement conditions; therefore, they degrade the performance of transfer learning. Thus, we proposed to improve performance using latency correction to shift waveforms along the time axis. We trained EEGNet with 5600 responses obtained from a clinical auditory oddball paradigm. In the BCI experiment, 12 subjects input one of the four characters presented on a display. The use of deep learning resulted in an approximately 10% higher performance than the conventional method based on the P300 amplitude. Furthermore, an improvement of approximately 5% in accuracy was observed when latency correction was used.