P300 spellers are among the most popular brain-computer interface paradigms, and they are used for many clinical applications. However, building the classifier for identifying event-related potential (ERP) responses, i.e., calibrating the P300 speller, is still a time-consuming and user-dependent problem. This paper proposes a novel method to reduce calibration times significantly. In the proposed method, a small number of ERP epochs from the current user were used to build a reference epoch. Based on this reference, the Riemannian distance measurement was used to select similar ERP samples from an existing data pool, which contained other-subject ERP responses. Linear discriminant analysis (LDA), support vector machine, and stepwise LDA were trained as ERP classifiers on the selected database and then were used to identify the user-attended character. With only 12 s of EEG data to calibrate, an average character recognition accuracy for 55 subjects of up to 87.82% was obtained. The LDA that built on other-subject samples that were selected by Riemannian distance outperformed the other classifiers. Compared with other state-of-the-art studies, this method significantly reduces P300 speller calibration times, while maintaining the character recognition accuracy.

P300 speller is a famous brain-computer interface (BCI) method, which translates mental attention by identifying the event-related potentials evoked by target stimulus. To improve its efficiency, subject-independent classification models and dynamical stopping strategies have been introduced into P300 speller. However, it has still not been determined whether these methods remain effective when the configurations of visual stimuli are changed. This study investigates whether subject-independent dynamical stopping model (SIDSM) can maintain high efficiency in the case of stimulus onset asynchrony (SOA) change. The SIDSM was built on a 55-subject database, and the classification efficiency was tested online with 14 new subjects. During the online experiment, four SOA conditions were tested, one of which had the same SOA as the modeling data, while the other three had different SOA settings. The SIDSM obtained comparable classification accuracy under different SOA settings. Thus, the efficiency of information transmission can be significantly improved by changing SOA only, without retraining the model. These results suggest that SIDSM has good robustness to changes in stimulus settings and can provide P300 speller with good flexibility for individual optimization.

No abstract