Comparison of session-to-session transfer between old and recent session data in motor imagery BCI
Hohyun Cho1, Minkyu Ahn2, Sung Chan Jun1*
1School of Information and Communications, Gwangju Institute of Science and Technology, Gwangju, Korea;
2Department of Neuroscience, Brown University, Providence, RI 02912, USA;
*123 Cheomdangwagi-ro, Buk-gu, Gwangju 61005, Korea. E-mail: scjun@gist.ac.kr
Introduction: Zero training is an important issue in brain-computer interface (BCI), as it minimizes the time- consuming calibration phase in a user-oriented system. Typical approaches transfer pre-existing session data to new session data to reduce the difference between sessions [1-2]. In previous work [3], we proposed a new strategy that used on-site background noise and outperformed existing feature extractors. We showed improved session-to-session transfer using a regularized spatio-temporal filter (RSTF) and a bias correction (BC) without any new session data; however, we did not investigate fully when BC is significantly effective. In this paper, a comparative study was performed on session-to-session transfer between very old pre-existing data (over 3 months old) and data collected recently (within a month).
Materials, Methods and Results: We compared classification accuracies and classifier outputs for pre-existing (more than three months) and recent (within a month) session data using RSTF with offline (background noise from pre-existing data) and on-site noise, and RSTF with on-site noise and BC. For the pre-existing data condition, we tested 6 multi-session data from 3 subjects, and for the recent data condition, 14 multi-session data from 12 subjects were tested. Results showed that RSTF with on-site noise suppression was useful for classification accuracy in both the pre-existing and recent conditions (Figure 1-A), and output results (Figure 1- B). RSTF with on-site noise and BC showed significant improvement in performance in the pre-existing data condition only.
Pre-existing RSTF with on-site + BC Recent
50 100
50
100
*
50 100
50 100
50 100
50
100
**
RSTF with on-site + BC
RSTF with on-site
RSTF with offline
50 100
50
100
**
-100 0 100
19.8 44.1 51.3 15 2.1 39.2 28.6
-100 0 100
9 0.3 2.3 11.9 3 4.1 5.1
-100 0 100
3.7 6.4 2.2 9.9 1.5 2.2 4.3
-100 0 100
38 44.8 20.7 2.5 2.9 23.7 46.9 91 19.6 1.2 15.3 2.1 12.117.1
-100 0 100
7.4 0.9 3.5 15.9 3.7 2.1 8.6 0.1 3 17.2 1.4 5.5 2.2 5.15.5
-100 0 100
5.2 1.2 6.4 5.7 3.4 2.3 14.1 1.6 2.616.9 1.2 6.9 2.2 6.55.4
Pre-existing Recent
RSTF with offline
RSTF with on-site
RSTF with on-site + BC
(A) (B)
4.3
Figure 1. Comparison of three approaches to classification accuracy (A) and classifier outputs (B) with pre- existing (more than three months) and recent (within a month) session data. (A) Statistically significant pairs are marked with * (p<0.10) and ** (p<0.05). (B) Red and green dots indicate classifier outputs of different classes.
Blue-shaded values are degrees of bias and red-shaded values indicate the mean of the degrees of bias defined in the equation.
Discussion: The bias correction method considers the Kullback Leibler distance between two different sources of background noise [3] and showed improved performance in the pre-existing data condition. It is likely that these interval sessions have a different spatial structure, while recent interval sessions do not.
Significance: Our proposed method showed improved session-to-session transfer for pre-existing session data more than three months old simply by using on-site background noise acquisition without new session data.
Acknowledgements: This work was supported by NRF of Korea (2013R1A1A2009029) and MCST/KOCCA in the CT Research & Development Program 2015.
References
[1] Krauledat M, Tangermann M, Blankertz B and Müller KR. Towards zero training for brain–computer interfacing. PLoS One, 3 e2967, 2008.
[2] Li X, Guan C, Zhang H, Ang KK and Ong SH. Adaptation of motor imagery EEG classification model based on tensor decomposition.
Journal of Neural Engineering, 11 056020, 2014.
[3] Cho H, Ahn M, and Jun SC. Increasing session-to-session transfer in a brain–computer interface with on-site background noise acquisition. Journal of Neural Engineering, 12 066009, 2015.
DOI: 10.3217/978-3-85125-467-9-148 Proceedings of the 6th International Brain-Computer Interface Meeting, organized by the BCI Society
Published by Verlag der TU Graz, Graz University of Technology, sponsored by g.tec medical engineering GmbH 148
