Unsupervised Online Calibration of a c-VEP Brain-Computer Interface (BCI) Martin Sp¨ uler 1 , Wolfgang Rosenstiel 1 , and Martin Bogdan 1,2 1 Wilhelm-Schickard-Institute for Computer Science, University of T¨ ubingen, Germany 2 Computer Engineering, University of Leipzig, Germany Abstract. Brain-Computer Interfaces (BCIs) can be used to give par- alyzed patients a means for communication. But so far, only supervised methods have been used for calibration of an online BCI. In this paper we present a method that allows to calibrate a BCI online and unsuper- vised. Based on offline data we show that the unsupervised calibration method works and validate the results in an online experiment with 8 subjects, who were able to control the BCI with an average accuracy of 85 %. We thereby have shown for the first time that an online unsu- pervised calibration of a BCI is possible and allows for successful BCI control. Keywords: Brain-Computer interface (BCI), unsupervised learning. 1 Introduction A Brain-Computer Interface (BCI) is a device that enables a user to control a computer by pure brain activity, which is usually recorded by electroencephalog- raphy (EEG). The main application for BCIs is to give paralyzed people a means to communicate, but so far, there are no reports for successful BCI control in complete locked-in patients [1]. Recently, we could show a BCI based on code-modulated visual evoked poten- tials (c-VEPs) to achieve very high communication speeds that made it possible for subjects to spell an average of 21.3 error-free letters per minute [2]. While this BCI used an unsupervised online adaptation, it still depended on a super- vised calibration, for which labeled data is needed to calibrate the BCI on the users brain activity. When looking at BCIs that use other paradigms like motor imagery or P300, there are also different unsupervised adaptation methods [3], but they all depend on a supervised calibration, which needs labeled data. So far, Eren et al. [4] are the only ones, who have shown that a BCI can be calibrated completely unsupervised without the need for labeled training data. Using Gaussian Mixture Models, they have shown in an offline analysis of motor imagery BCI data that their method works for 3 out of 6 subjects. In this paper, we present a method for completely unsupervised calibration of a c-VEP BCI and show it to work for all our subjects in an online study. We further discuss how unsupervised calibration might be useful for complete locked-in patients, for whom supervised calibration does not work [1]. V. Mladenov et al. (Eds.): ICANN 2013, LNCS 8131, pp. 224–231, 2013. c  Springer-Verlag Berlin Heidelberg 2013