ENGLISHIEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. XX, NO. Y, 2006 1The Berlin Brain-Computer Interface: EEG-basedcommunication without subject trainingBenjamin Blankertz, Guido Dornhege, Matthias Krauledat, Klaus-Robert Müller,Volker Kunzmann, Florian Losch, Gabriel CurioAbstract—The Berlin Brain-Computer Interface (BBCI)project develops a non-invasive BCI system whose key featuresare (1) the use of well-established motor competences as controlparadigms, (2) high-dimensional features from 128-channel EEGand (3) advanced machine learning techniques. As reportedearlier, our experiments demonstrate that very high informationtransfer rates can be achieved using the readiness potential (RP)when predicting the laterality of upcoming left vs. right handmovements in healthy subjects. A more recent study showedthat the RP similarily accompanies phantom movements in armamputees, but the signal strength decreases with longer loss of thelimb. In a complementary approach oscillatory features are usedto discriminate imagined movements (left hand vs. right handvs. foot). In a recent feedback study with 6 healthy subjectswith no or very little experience with BCI control, 3 subjectsachieved an information transfer rate above 35 bits per minute(bpm), and further two subjects above 24 and 15 bpm, whileone subject could not achieve any BCI control. These results areencouraging for an EEG-based BCI system in untrained subjectsthat is independent of peripheral nervous system activity and doesnot rely on evoked potentials even when compared to results withvery well-trained subjects operating other BCI systems.Index Terms—Brain-Computer Interface, Classification, Com-mon Spatial Patterns, EEG, ERD, Event-Related Desynchro-nization, Information Transfer Rate, Readiness Potential, RP,Machine Learning, Single-Trial AnalysisI. INTRODUCTIONThe aim of Brain-Computer Interface (BCI) research is toestablish a new augmented communication system that trans-lates human intentions—reflected by suitable brain signals—into a control signal for an output device such as a com-puter application or a neuroprosthesis [1]. According to thedefinition put forth at the first international meeting for BCItechnology in 1999, a BCI “must not depend on the brain’snormal output pathways of peripheral nerves and muscles”[2]. This viewpoint is certainly for research purpose in orderto have clear evidence of what information a systems uses andwhere it comes from. Nevertheless there seems to be consensusin the BCI community that in specific BCI applications (e.g.,for paralyzed patients) it may be reasonable to get all signalsthat provide useful information regardless of their origin.This work was supported in part by grants of the Bundesministeriumfür Bildung und Forschung (BMBF), FKZ 01IBE01A/B, by the DeutscheForschungsgemeinschaft (DFG), FOR 375/B1, and by the IST Programme ofthe European Community, under the PASCAL Network of Excellence, IST-2002-506778. This publication only reflects the authors’ views.BB, GD, MK and KRM are with Fraunhofer FIRST (IDA), Berlin, Ger-many, E-mail: [email protected]. KRMand MK are also with University of Potsdam, Potsdam, Germany.VK, FL and GC are with the Dept. of Neurology, Campus BenjaminFranklin, Charité University Medicine Berlin, Berlin, Germany.There is a huge variety of BCI systems, see [1], [3], [4] for abroad overview. Our Berlin Brain-Computer Interface (BBCI)is a non-invasive, EEG-based system, which does not useevoked potentials. BCI systems relying on evoked potentialscan typically achieve higher information transfer rates (ITRs)in contrast to systems working on unstimulated brain signals,cf. [5]. On the other hand with evoked potential BCIs theuser is constantly confronted with stimuli, which can becomeexhaustive after longer usage.Here we present two aspects of the main approach taken inthe BBCI project. The first is based on the discriminabilityof premovement potentials in self-paced movements. Ourinitial studies ([6]) show that high information transfer ratescan be obtained from single-trial classification of fast-pacedmotor commands. Additional investigations point out waysof improving bit rates further, e.g., by extending the class ofdetectable movement related brain signals to the ones encoun-tered when moving single fingers within one hand. A morerecent study showed that it is indeed possible to transfer theresults obtained with regard to movement intentions in healthysubjects to phantom movements in patients with traumaticamputations.Taking another approach, we established a BCI systembased on motor imagery that works without subject training.Using general, complex features derived from 128-channelEEG recordings the system automatically adapts to the specificbrain signals of each user by using advanced techniques ofmachine learning and signal processing [7], [8], [9]. Thisapproach contrasts with the operant conditioning variant ofBCI, in which the subject has to learn to control a specific EEGfeature which is hard-wired in the BCI system. According tothe motto ’let the machines learn’ our approach minimizesthe need for subject training and copes with one of the majorchallenges in BCI research: the huge inter-subject variabilitywith respect to patterns and characteristics of brain signals.II. APPROACH 1: SELF-PACED AND PHANTOM FINGERMOVEMENTSA. Exploiting the limits of the refractory behavior in fast-paced motor commandsThe main goal of BCI is to improve autonomy of peoplewith severe motor disabilities by new communication andcontrol options. These persons cannot move but can thinkabout moving their limbs and produce in this way stablemotor-related signals like the readiness potential (RP, orBereitschaftspotential, BP) and event-related desyncroniza-tion (ERD). The RP is a transient postsynaptic response2 ENGLISHIEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. XX, NO. Y, 2006of main pyramidal peri-central neurons, see [10]. It leadsto a negativation of the EEG over primary motor corticesduring motor preparation that peaks about movement onset. Inhand movements it is focused contralateral to the performinghand, cf. [11], [12] and references therein for an overview.The preparation of movements is reflected also by an ERD,i.e., an attenuation of pericentral µ- and β -rhythms in thecorresponding motor areas. With respect to unilateral handmovements these blocking effects are