Proceedings of the Fifth International Brain-Computer Interface Meeting 2013 DOI:10.3217/978-3-85125-260-6-122 Published by Graz University of Technology Publishing House, sponsored by medical engineering GmbH Article ID: 122 Unsupervised BCI Calibration as Possibility for Communication in CLIS Patients? M. Spüler 1 , W. Rosenstiel 1 , M. Bogdan 1,2 1 Wilhelm-Schickard-Institute for Computer Science, University of Tübingen, Sand 13, 72076 Tübingen, Germany 2 Computer Engineering, University of Leipzig, Johannisgasse 26, 04103 Leipzig, Germany Correspondence: M. Spüler, Sand 13, 72076 Tübingen, Germany. spueler@informatik.uni-tuebingen.de Abstract. In this paper we present first results that a Brain-Computer Interface (BCI) can be calibrated online in a completely unsupervised manner. Thereby it is possible to provide a user with contingent feedback without the need for any goal-directed action. Since the extinction of goal-directed thinking is the assumed cause, why there are no reports of successful BCI communication in patients suffering from complete locked-in syndrome (CLIS), we discuss if unsupervised calibration could be used to enable communication in those patients. Keywords: Brain-Computer Interface (BCI), complete locked-in syndrome (CLIS), goal-directed thinking, unsupervised learning 1. Introduction The main motivation for Brain-Computer Interface (BCI) research is to give paralyzed patients a new means for communication. Although BCI communication works for patients who have some remaining muscle-control, using other communication methods that utilize the remaining muscle control are more effective in most cases. Therefore, the use of BCI technology would be especially interesting for people suffering from complete locked-in syndrome (CLIS), who have no remaining muscle control. Despite the extensive research in this area, so far there are no reports of successful BCI communication in CLIS patients. One reason that BCI did not work for CLIS patients yet, is the assumed extinction of goal-directed thinking [Kübler and Birbaumer, 2008]. Due to the CLIS, the patient lacks voluntarily motor control and thereby feedback. Since cognitive activity does not result in any feedback, the missing feedback might be responsible for the extinction of goal-directed thinking. Disregarding the recent introduction of classical conditioning for BCI control, there are two approaches, how a person is enabled to control a BCI: (1) In the neurofeedback approach the user is given feedback of certain parameters of his brain activity and the user learns to voluntarily control his brain activity to alter the feedback. (2) In the supervised machine learning approach the computer learns to decode the user’s brain activity. Therefore, a supervised calibration phase is needed in which certain instructions are given to the user which actions to perform or which stimuli to attend. To ensure that the machine learning is working properly, the user is required to voluntarily follow these instructions. Taking under consideration the assumed extinction of goal-directed thinking in CLIS patients, both approaches do not work, because: (1) the patient is not able to voluntarily alter his brain activity and (2) the patient is not able to voluntarily follow any instructions and perform voluntary actions. In this paper we show first results of a BCI based on code-modulated visual evoked potentials (c-VEPs) that can be calibrated with unsupervised machine learning and we discuss that an unsupervised BCI calibration may provide the user with contingent feedback without the need for the user to voluntarily alter his brain activity or follow any instructions. 2. Material and Methods Based on our recent developments for improved classification [Spüler et al., 2012] in a c-VEP BCI and a method for unsupervised online adaptation [Spüler et al., 2012a], we developed a method that allows a completely unsupervised calibration of the c-VEP BCI. This method will be presented elsewhere in more detail. Although the c- VEP BCI can be used with 32 stimuli, only 2 stimuli were used during the unsupervised calibration. To show that an unsupervised calibration is possible, the method was evaluated offline on data from 18 sessions (9 healthy subjects, 2 sessions each) with each session consisting of 640 trials. The first 64 trials were used for unsupervised calibration, while the remaining 576 trials were used for accuracy estimation.