2009 Special IssueSelecting features for BCI control based on a covert spatial attention paradigm
Introduction
Brain–computer interfaces (BCIs) depend on the detection of changes in task-related activity as the subject moves from one mental state to the other. This implies that task-related changes must be strong enough and stable over time in order to be useable as a control signal for BCI. It is well-known that a steady-state visual evoked potential (SSVEP), induced by an external stimulus oscillating at a particular frequency, can drive a BCI (Allison et al., 2008, Middendorf et al., 2000, Sutter, 1992). Recently, Kelly, Lalor, Reilly, and Foxe (2005) have shown that the external stimulus might not be required and covert attention to spatial locations in the visual field alone may be sufficient to drive a BCI. They were the first to demonstrate that shifts in covert spatial attention between the left and right visual hemifield can be picked up on the single trial level. This accomplishment is based on the fact that covert shifts in visual attention are paired by alpha-desynchronisation in posterior sites contralateral to the attended position (Sauseng et al., 2005, Thut et al., 2006, Yamagishi et al., 2005) as well as alpha-synchronisation ipsilateral to the attended position (Kelly et al., 2006, Worden et al., 2000).
Kelly et al. (2005) demonstrated that by using alpha power (8–14 Hz) over left and right hemispheres based on 3.52 s windows of EEG data as input to a linear discriminant analysis classifier, a maximum bit rate of 7.5 bits per minute could be achieved. Covert spatial attention is a promising paradigm for BCI control since it is natural for the subject to orient ones attention to the direction of intended control. Furthermore, little training time is required in order to attain acceptable results. However, at present the paradigm remains relatively unexplored.
In this study, we examine the paradigm using data obtained with a 275 channel MEG system for fifteen subjects. To our knowledge, this is the first time covert spatial attention is examined as a paradigm for brain–computer interfacing using MEG as a modality. Our goal is to improve classification performance by using a feature selection approach. We assume that alpha (de)synchronisation over an attention period is indeed the signal of interest, but in contrast to Kelly et al. (2005), the optimal channels are assumed to be subject specific and will be identified using sparse logistic regression (van Gerven, Hesse, Jensen, & Heskes, 2009). Examples of other feature selection approaches in BCI research are Millán, Franzé, Mouriño, Cincotti, and Babiloni (2002), Schröder, Bogdan, Rosenstiel, Hinterberger, and Birbaumer (2003), Lal et al. (2004), Schröder et al. (2005) and Hoffman, Yazdani, Vesin, and Ebrahimi (2008). We compare the results of sparse logistic regression with a method that is analogous to that of Kelley et al. We also examine how classification performance changes as a function of the length of the attention period and as a function of the number of trials. The hope is that the improved classification performance due to the selection of more optimal features and a better insight into the optimal attention period and number of trials leads to a more widespread use of this promising BCI paradigm.
Section snippets
Data collection
Fifteen healthy subjects (mean age 28 9; six females) participated in the experiment. All subjects had (corrected to) normal vision. Four males and two females were left-handed and the remaining subjects were right-handed. The study was approved by the local ethics committee and written informed consent was obtained from the subjects. The subjects viewed a screen with a central fixation cross and four squares at 7.5 degrees of visual angle to the top, right, bottom, and left of the fixation
Lateralisation index
Kelly et al. (2005) obtained the best results using the logarithm of the left hemisphere (EEG electrodes PO7 and O1) alpha power divided by the right hemisphere (EEG electrodes PO8 and O2) alpha power, averaged over the attention period, as input to a linear discriminant analysis algorithm. As a rough approximation to this strategy, we use the average over left and right occipito-parietal MEG channels in order to represent left and right hemisphere alpha power (8–14 Hz). We refer to the log of
Classifier evaluation
In order to determine classification performance, we used the accuracy or classification rate (CR), which measures the proportion of correctly classified trials as a criterion. Classification rate does not take into account the duration of a trial. Therefore, we will also use the information transfer rate (ITR) in order to estimate the amount of information that is conveyed per unit of time (bits per minute). The information transfer rate is computed as where
Experimental results
In this section, we determine the accuracies and information transfer rates that can be achieved by means of the different classification methods and by varying both the length of the attention period and the number of trials. In all experiments, we average power over frequencies in the 8–14 Hz range and over the attention period.
Discussion
In this paper, we have examined how classification performance can be optimised when using covert spatial attention as a paradigm for brain–computer interfacing. Using sparse logistic regression we improved significantly over results that have been obtained using a simple thresholding method. Looking at Fig. 5, we find substantial improvements for a number of subjects. Improvements are explained by the findings that task-related activity can be represented more diffusely, more centrally, and/or
Conclusion
We believe that covert spatial attention is a very promising paradigm for BCI control since it is very natural to orient ones attention to a location of interest. The paradigm therefore requires little subject training. In this paper, we extended the work of Kelly et al. (2005) by showing that the selection of subject-specific channels using sparse logistic regression leads to improved classification performance as compared with thresholding of a lateralisation index, especially when the
Acknowledgements
We thank Sander Berends for his help in data acquisition. The authors gratefully acknowledge the support of the Dutch technology foundation STW (project number 07050), the Netherlands Organisation for Scientific Research NWO (VICI grant number 639.023.604), and the BrainGain Smart Mix Programme of the Netherlands Ministry of Economic Affairs and the Netherlands Ministry of Education, Culture and Science.
References (24)
- et al.
Towards an independent brain–computer interface using steady state visual evoked potentials
Clinical Neurophysiology
(2008) - et al.
Tangential derivative mapping of axial MEG applied to event-related desynchronization research
Clinical Neurophysiology
(2000) The brain response interface: Communication through visually induced electrical brain responses
Journal of Microcomputer Applications
(1992)- et al.
Interpreting single trial data using groupwise regularisation
NeuroImage
(2009) - et al.
Information transfer rate in fMRI experiments measured using mutual information theory
Journal of Neuroscience Methods
(2008) - et al.
Attentional shifts towards and expected visual target alter the level of alpha-band oscillatory activity in the human calcarine cortex
Cognition in Brain Research
(2005) - et al.
Dynamic imaging of coherent sources: Studying neural interactions in the human brain
Proceedings of the National Academy of Sciences
(2001) - Hoffman, U., Yazdani, A., Vesin, J. M., & Ebrahimi, T. (2008). Bayesian feature selection applied in a P300...
- Keerthi, S. S., Shevade, S. K., & Poo, A. N. (2002). A fast dual algorithm for kernel logistic regression. In...
- Kelly, S. P., Lalor, E., Reilly, R. B., & Foxe, J. J. (2005). Independent brain computer interface control using visual...