Spatial filtering based on Riemannian distance to improve the generalization of ErrP classification

Due to the inherent non-stationarity of EEG signals, before each experimental session, BCI is usually calibrated to build the classification models, thus avoiding performance decay. This tedious re-calibration procedure is a limiting factor for real-world applications. Therefore, single-calibration...

Full description

Saved in:
Bibliographic Details
Published inNeurocomputing (Amsterdam) Vol. 470; pp. 236 - 246
Main Authors Cruz, Aniana, Pires, Gabriel, Nunes, Urbano J.
Format Journal Article
LanguageEnglish
Published Elsevier B.V 22.01.2022
Subjects
Brain-computer interface (BCI)
Error-related potentials (ErrP)
Generalization
Riemannian geometry
Single-calibration
Spatial filter
Zero-calibration
Riemannian geometry
Single-calibration
Brain-computer interface (BCI)
Zero-calibration
Error-related potentials (ErrP)
Spatial filter
Generalization
Online AccessGet full text

Cover

More Information
Summary:Due to the inherent non-stationarity of EEG signals, before each experimental session, BCI is usually calibrated to build the classification models, thus avoiding performance decay. This tedious re-calibration procedure is a limiting factor for real-world applications. Therefore, single-calibration or zero-calibration plays a crucial role in the use of BCIs in real contexts, outside the laboratory. Here, we propose and validate a statistical spatial filter, Riemannian Fisher criterion beamformer, based on Riemannian geometry able to use the invariance properties of Riemannian distance to handle cross-session and cross-subject generalization. The proposed method is validated with two datasets publicly available, consisting of error-related potentials. The results show that the proposed filter improves the generalization across sessions and across subjects and that it is robust to the amount of error samples used to train the classification model.
ISSN:0925-2312
1872-8286
DOI:10.1016/j.neucom.2021.10.078