Académique Documents
Professionnel Documents
Culture Documents
EEG EEG+NIRS
Subject
Acc. Sen. Spec. AUC Acc. Sen. Spec. AUC
VPeaa 53.0 20.7 63.9 0.56 58.5 74.9 54.6 0.76
VPeab 77.8 72.6 80.2 0.79 77.1 72.7 79.0 0.78
VPeac 31.0 84.4 17.3 0.56 34.2 86.5 21.0 0.59
VPead 71.3 60.1 75.3 0.76 71.5 64.0 74.3 0.77
VPeae 61.7 56.6 63.4 0.69 56.6 64.8 53.9 0.68
VPeaf 36.6 83.5 25.0 0.60 38.1 80.3 27.5 0.61
VPeag 66.7 59.4 69.5 0.71 67.5 70.9 66.3 0.75
VPeah 66.2 55.4 70.9 0.67 67.4 58.1 71.3 0.69
VPeai 36.3 46.9 33.1 0.50 48.5 45.1 49.6 0.58
VPeaj 63.5 38.3 75.5 0.59 63.0 65.8 61.8 0.66
VPeak 27.0 80.8 13.1 0.52 61.4 87.7 55.2 0.80
VPeal 57.4 40.0 64.5 0.59 57.4 44.5 62.4 0.61
VPeam 45.9 55.9 42.9 0.60 57.0 63.7 55.1 0.70
VPean 75.3 64.4 80.2 0.76 76.1 74.1 77.0 0.80
Mean 55.0 58.5 55.3 0.64 59.6 68.1* 57.8 0.70*
For separation of the three classes ('left hand imagery', 'right IV. DISCUSSION AND CONCLUSION
hand imagery' and 'idle') we considered multi-classification To stabilize the performance of asynchronous BCIs, we
strategy namely One Versus the Rest (OVR) [3]. Three types of introduced a multi-modal hybrid BCI. Our results outline that
OVR-based linear discriminant analysis (LDA) classifiers were for some subjects a single modality is not capable to estimate
trained for EEG: left vs others, right vs others, idle vs rest (all mental states reliably enough for the robust control of
asynchronous BCIs. By employing a hybrid-BCI approach, the
OVR). Furthermore, one OVR classifier was estimated from
overall performance was significantly enhanced and some
NIRS (HbO) data: idle vs rest. To emulate classification in subjects gained remarkably improved accuracies. Furthermore,
real-time, the classifiers were fed with data from a moving subjects (and patients) who are not able to operate a BCI (so-
window (width=1s, step size=0.5s). This procedure emulates a called 'BCI illiteracy' [5]) will find a viable alternative in the
pseudo-online paradigm [4]. presented hybrid asynchronous BCI approach. However, the
3) Subject-dependent Classifier Combination: To estimate analyzing routines presented here, have been solely validated in
the three mental states (right, left and idle state), we propose an offline fashion using pseudo-online technique. Our future
subjects-dependent strategy based on EEG and NIRS work will extend our research to a real-time asynchronous
classifiers. At first, the current mental state was estimated by hybrid BCI with visual feedback or for controlling external
OVR-based EEG classifiers, then the output of classifiers were devices.
adjusted (or weighted) by the output of the NIRS-based
classifier. In order to find the optimal weighting parameter for ACKNOWLEDGMENT
each subject, a grid search was performed on the validation This work was supported by National Research Foundation
data set using Area under the Receiver Operating Curve of Korea (NRF) funded by the Ministry of Science, ICT and
(AUC) as a loss function. For the performance evaluation of Future Planning (No. 2012-005741).
an asynchronous BCIs, the accuracy, sensitivity, specificity,
and AUC scores were calculated for stand-alone EEG and the REFERENCES
proposed hybrid EEG-NIRS system.
[1] J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G. Pfurtscheller and, T.
III. RESULTS M. Vaughan, “Brain–Computer Interfaces for Communication and
Control,” Clinical neurophysiology, vol. 113, 2002, pp.767–791.
Our first goal is to show that the proposed subject-
dependent classifier strategy can improve the classification [2] S. Fazli, J. Mehnert, J. Steinbrink, G. Curio, A. Villringer, K. R. Müller,
and B. Blankertz, “Enhanced Performance by a Hybrid NIRS-EEG
accuracy of asynchronous BCI systems. Table 1 shows the Brain Computer Interface,” Neuroimage, vol. 59, 2012, pp. 519-529.
classification results for stand-alone EEG and the proposed [3] G. Dornhege, B. Blankertz, G. Curio, and K.-R. Müller, “Boosting Bit
multi-modal hybrid EEG-NIRS system. The hybrid asyn- Rates in Noninvasive EEG Single-trial Classfications by Feature
chronous BCI outperformed the EEG stand-alone system. The Combination and Multiclass Paradigms,” Biomedical Engineering, vol.
results showed that all the evaluation parameters (accuracy, 51, 2004, pp. 993–1002.
sensitivity, specificity, and AUC score) enhance the [4] B. Blankertz, G. Curio, and K.-R. Müller, “Classifying Single Trial
performance on average across all subjects. Paired t-tests EEG: Towards Brain Computer Interfacing,” Advances in neural
information processing systems, vol. 1, 2002, pp. 157–164.
yielded significant performance increases for the sensitivity
[5] C. Guger, G. Edlinger, W. Harkam, I. Niedermayer, and G. Pfurtscheller,
and AUC scores (p = 0.03 and p = 0.01, respectively). However, “How many people are able to operate an EEG-based brain-computer
accuracy and specificity did not reach significant levels interface (BCI)?,” Neural Systems and Rehabilitation Engineering, vol.
(p = 0.10 and p = 0.52, respectively). 11, 2003, pp. 145–147.