IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 51, NO. 6, JUNE 2004 979 An Asynchronously Controlled EEG-Based Virtual Keyboard: Improvement of the Spelling Rate Reinhold Scherer*, Gernot R. Müller, Member, IEEE, Christa Neuper, Bernhard Graimann, and Gert Pfurtscheller, Member, IEEE Abstract—An improvement of the information transfer rate of brain-computer communication is necessary for the creation of more powerful and convenient applications. This paper presents an asynchronously controlled three-class brain-computer inter- face-based spelling device [virtual keyboard (VK)], operated by spontaneous electroencephalogram and modulated by motor im- agery. Of the first results of three able-bodied subjects operating the VK, two were successful, showing an improvement of the spelling rate , the number of correctly spelled letters/min, up to (average ). Index Terms—Asynchronous control, brain-computer interface (BCI), motor imagery, virtual keyboard (VK). I. INTRODUCTION C OMMUNICATION and the ability to interact with the en- vironment are basic needs for human relationships. For people who suffer from severe physical disabilities or palsy, the ability to comply with this need is limited or even impossible. In contrast to impaired motor activity, the sensory and cognitive functions are usually almost intact (locked-in state). Bioelec- trical brain signals, such as those reflected by electroencephalo- gram (EEG) or electrocorticogram (ECoG), have been proved to provide an alternative communication channel. Intellectual activity can modify the bioelectrical brain activity without any motor action. A brain-computer interface (BCI) is able to recog- nize voluntary changes in the ongoing electrophysiological sig- nals and to map different brain states to appropriate commands in order to operate communication aids [1]–[5]. Patients suffering from amyotrophic lateral sclerosis (ALS) learned to operate an electronic spelling device [1]. With a bi- nary decision, which requires the discrimination of two dif- ferent brain states (classes), the German alphabet was split it- eratively, following a fixed procedure, into two halves until the desired letter was isolated. The communication performance, Manuscript received June 23, 2003; revised February 6, 2004. This work was supported by the “Land Steiermark” under Project FA6A-10V4-00/1 and the “Fonds zur Förderung der Wissenschaftlichen Forschung,” Austria, under Project P16326-B02. Asterisk indicates corresponding author. *R. Scherer is with the Institute of Human-Computer Interfaces, Graz Uni- versity of Technology, 8010 Graz, Austria (e-mail: reinhold.scherer@tugraz.at). G. R. Müller and B. Graimann are with the Institute of Human-Computer Interfaces, Graz University of Technology, 8010 Graz, Austria. C. Neuper is with the Ludwig Boltzmann Institute of Medical Informatics and Neuroinformatics, Graz University of Technology, 8010 Graz, Austria. G. Pfurtscheller is with the Institute of Human-Computer Interfaces, Graz University of Technology, 8010 Graz, Austria and also with the Ludwig Boltz- mann Institute of Medical Informatics and Neuroinformatics, Graz University of Technology, 8010 Graz, Austria. Digital Object Identifier 10.1109/TBME.2004.827062 given by the spelling rate and measured in correct selected letters/min, reached values of about 0.5 letters/min . Using the same letter selection strategy, a patient suffering from severe cerebral palsy achieved spelling rates of approximately one letter/min controlling the Graz-BCI [6]. To im- prove the spelling rate or, more generally, to increase the information transfer rate is a main goal in BCI research. The aim of this paper is to introduce a new Graz-BCI based spelling application designed to provide an increase in the in- formation transfer rate. The basic principle of the Graz-BCI is the classification of sensorimotor EEG patterns generated by the imagination of motor activity (e.g., left hand, right hand, foot, or tongue) [4], [7]. The setup of the classifier is done by performing a cue-based repetitive training of mental motor imagery. During the following feedback training, the real-time classification re- sult of the ongoing EEG is presented to the subject (e.g., moving cursor). By repeating this training and updating the classifier, the subject and the BCI can mutually adapt to one another. The following points were taken into account for the imple- mentation and design of the virtual keyboard (VK). 1) Improvement of the classification accuracy: Since bio- logical signals show a large inherent variability, the relia- bility of a classifier is very important. The adaptation and optimization of the parameters of the selected information processing methods should lead to a better generalization and, consequently, to a reduction of misclassification. 2) Increase of the number of discriminable brain patterns: An increase of the number of brain patterns that can be equally reliably detected may increase the communica- tion speed. If the two-class process described above is di- vided in three instead of two parts, less selection steps are necessary. 3) Noncue-based (asynchronous) information transfer: The real-time Graz-BCI operates in a cue-based or synchronous communication mode [4], [7]. If the BCI (receiver) is prepared to handle an input, a ready signal (cue) is sent to the user (transmitter). Therefore, cue-based communication requires an additional signal to enable a proper information transfer. A side effect is an idling period, in which the user and the BCI are waiting. Besides this synchronous transfer mode, asynchronous transfer is also possible. In the latter case, no additional signal is required, since all the information is already contained in the conveyed signal. The BCI processes the incoming physiological signals and reacts properly if a known input pattern is found [8], [9]. From an engi- neering point of view, the asynchronous mode is more 0018-9294/04$20.00 © 2004 IEEE