LABORATORY PC AND MOBILE POCKET PC BRAIN-COMPUTER INTERFACE ARCHITECTURES G. Edlinger, G. Krausz, F. Laundl, I. Niedermayer, C. Guger g.tec – medical engineering GmbH Herbersteinstrasse 60, 8020 Graz, Austria e-mail:
[email protected] Abstract- An EEG-based brain computer interface (BCI) system converts brain activity into control signals. BCI systems have been developed for people with severe disabilities to improve their quality of life. A BCI system has to satisfy different demands depending on the application area. A laboratory PC based system allows the flexible design of multiple/single channel feature extraction, classification methods and experimental paradigms. The key advantage of a Pocket PC BCI approach is its small dimension and battery supply. Hence a mobile BCI system e.g. mounted on a wheelchair can be realized. This study compares and discusses thoroughly the two mentioned approaches.
I.
INTRODUCTION
EEG based brain computer interface (BCI) systems have been developed during the last years for people with severe disabilities to improve their quality of life. Applications of BCI systems comprise the restoration of movements, communication and environmental control [1]. However, recently BCI applications have been also used in different research areas e.g. in the field of virtual reality [2]. General used parameters to quantify the performance of BCI systems are the accuracy and speed. Furthermore, a BCI approach should ensure that the users learn to control the system within a few training sessions. The level of control should be stable after an initial learning phase and should improve over time [1, 3, 4]. Family members must be able to help in operation of the BCI system on a daily basis. Therefore, the systems must be robust and easy to use. System appearance and how the users look like while using the device are also important issues when realizing the BCI system. BCI systems have been successfully realized based on different EEG phenomena: (i) oscillatory EEG components in the mu and beta range, (ii) slow cortical potentials and (iii) evoked potentials. Depending on the BCI concept (synchronous or asynchronous), the experimental strategy (movement imagery, visual attention, …) and feedback method (continuous/discrete, abstract/realistic, …) different electrode montages are used for measuring EEG. Thereafter, feature extraction and classification of EEG data is performed resulting in the control signal. After some training sessions the BCI accuracy enhances to a certain degree, meaning the
BCI system and the subject have adapted to each other for a better general system performance [1, 3]. Up to now BCI research has primarily comprised demonstrations and studies on individual persons based on dedicated hardware and software. A more general working environment is desired to optimize BCI systems. Such an environment optimally enables comparing results, using different types of brain signals and also using various signal processing techniques [5]. Hence it is necessary to have a very flexible and extendable but still user friendly system architecture. These demands can be satisfied by a lab PC based BCI system supporting single/multi channel EEG recordings, comparing different feature extraction and classification methods and allowing the development of different applications [1,3,6]. However, for the general usage outside the research lab new key features must additionally be realized: the BCI system must be as small as possible and easy to use. Therefore, both a flexible research system and an embedded Pocket PC system are discussed in this paper. II.
SYSTEMS
A.. Laboratory System Fig. 1 displays the main components necessary implementing a BCI laboratory system: The subject is connected via electrodes measuring the brain signals to a biosignal amplifier (g.BSamp or g.USBamp). Depending on the type of EEG signal acquired (scalp EEG signal or ECoG signal the amplifier must meet different safety conditions according to normative IEC 60601-1 and its variants. A data acquisition board (DAQ board) for the g.BSamp or a 24 bit integrated analog to digital converter for the g.USBamp performs the digital conversion at a user defined sampling rate. Other signals are also measured in synchrony to the biosignal data including timing information, trigger signals or signals from other external devices. In order to control all hardware parameters, it is important to have a powerful software package that also includes the driver for the DAQ-boards and supports the USB interface, respectively.
Fig. 1. Components of a laboratory PC based BCI system. Explanation of components is given in the text.
It is also important to select a programming language that enables an easy setup or adaptation of the programs for the experimental paradigm, data acquisition, analysis and final application. The rapid prototyping environment discussed in this paper is based on MATLAB (MathWorks, Inc., Natick, USA), in combination with the signal flow oriented Simulink Toolbox. Simulink is used for the real-time and online calculation of different parameters describing the current state of the EEG (g.RTanalyze). The selection of the parameter estimation algorithms is mainly dependent on the type of brain signals used for the BCI approach. A BCI device based on the event-related synchronization and desynchronization EEG phenomenon can e.g. be realized using band power feature extraction methods. The robust FFT based calculation method includes typically two frequency bands in the alpha band (between 8-12 Hz) and beta band (between 16-24 Hz) for the feature extraction. However, adaptive autoregressive (AAR) methods can also be used for power spectrum estimation. AAR methods have the advantage that no subject specific frequency band selections for feature extraction must be done. A disadvantage of these methods is the higher sensitivity to a low signal to noise ratio compared to FFT based methods. The BCI system supports in general the implementation and integration of different user specific algorithms for comparison. Feature extraction is then followed by data set classification converting the computed parameters into specific commands. Common classifications methods are based on linear discriminant analysis, neural networks or hidden Markov models. Beside the real-time parameter extraction and classification MATLAB handles the data acquisition, timing and presentation of the experimental paradigm. Thus, the system can be programmed graphically and it is also running in real-time under Windows [3].
Fig. 2. Components of a Pocket PC based BCI system.
Digital and analog outputs allow controlling external devices such as a hand orthosis or a stimulation unit (g.STIMunit). Such devices are used to present different experimental paradigms to the subject and enhance the BCI performance with visual and/or acoustic feedback of the classification result. A personal area network is used to remotely control the BCI system for operation control, algorithm updates as well as BCI data transfer [3]. B.
Mobile System The rapid prototyping environment described in section A is well suited for experimental purposes in a laboratory. Such a system configuration can be used for developing a language supporting system where the patient can select letters or words on the display of a notebook. However, for assistive applications like TV channel selection or a wheelchair mounted language supporting system an embedded solution including the processor and DAQ board without mechanical disks and extra display is more suitable. Size, robustness and usability are major considerations for assistive communication devices. The hardware must be also fully portable, supplied by a battery and cheap [4]. For the embedded BCI a standard Pocket PC is used as a portable host. The Pocket PC is connected via a serial cable to an embedded target computer system g.MOBIlab (see Fig. 2 and Fig. 3). The serial interface has a data transfer rate of 115 kBaud. The embedded system consists of a µC operating at 12 MHz to optimize the power consumption. A 16 bit analog to digital converter samples 8 analog channels. Each channel is sampled at 256 Hz. The amplifier module is equipped with 4 EEG type channels, 2 ECG type channels and 2 analog inputs for external sensors. Two digital inputs and 2 digital outputs allow controlling external devices. Two batteries of type “AA” power the embedded system. The Pocket PC operating system is Windows Mobile and the BCI system is
Fig. 4. Training phase displayed on the Pocket PC: Red arrows indicate that the subject should image a left hand (left panel) or right hand (right panel) movement imagery. Fig. 3.
Pocket PC based BCI system. Explanation of components is given in the text.
programmed in eMbedded Visual C++. The integrated Wireless LAN (WLAN) module of the Pocket PC can be used for wireless data transmission. Data are stored on the internal 64 MByte storage or streamed to a Compact Flash card (512 MByte) for later analysis. An application programming interface allows accessing the hardware components and data buffers. Hence BCI applications can be adapted to optimally meet user specific needs or novel applications can be developed. The following paragraphs give one example for a typical BCI experiment based on oscillatory brain activity measured over electrode positions C3 and C4 [6]. Training phase For BCI training two different experimental paradigms are implemented. In order to acquire EEG data in the training phase the first experiment is performed without feedback. Therefore, an arrow pointing to the left or right side of the computer monitor is shown (Figure 4). Depending on the direction of the arrow the subject has to imagine a specific kind of movement. If the arrow is pointing to the left hand side the subject should imagine a left hand movement, if the arrow is pointing to the right side the subject should imagine a right hand movement. EEG data for a total of 160 trials (80 right and 80 left hand movement imageries) are acquired. Specific EEG parameters are then extracted from the data and the trials are classified into two classes yielding a subject specific classifier. Application phase After computing the weight vectors the application phase can be started. The classifier weights the extracted features calculated from the EEG data in such a way that the thoughts are converted into cursor movements in real-time. A classification result of a right hand movement extends the bar to the right side. A classification result of a left hand movement class extends the bar to the left side. The length of the bar represents the reliability of the classification.
Fig. 5. Application phase displayed on the Pocket PC: The direction of the bar indicates the classification in either the right hand movement class (right panel) or the left hand movement class (left panel). The length of the bar represents the reliability of the classification result.
III.
DISCUSSION
The PC based BCI system allows a very flexible design of BCI experiments concerning electrode configuration, utilized algorithms and applications. The rapid prototyping environment speeds up the development cycle significantly. The embedded BCI system with its compact dimension allows the usage of the BCI outside the research lab for patient training and (as Pocket PC CPUs are getting more and more powerful) also for implementing sophisticated applications. The system can be mounted easily on a wheelchair or beside the bed and is fully battery powered. A big advantage is that the Pocket PC BCI operates immediately after switching it on without booting of the operating system. Hence the Pocket PC BCI approach supports bridging the gap between laboratory research and practical application. Furthermore the Pocket PC concept supports BCI applications in other research areas as well. The PC based BCI system was used for example in a Virtual Environment to walk around in virtual city by thoughts [2]. A limitation in this project was the size of the
system which required that the subject was sitting throughout the experiment. The Pocket PC BCI system would allow the subject to move more flexible through the virtual world. ACKNOWLEDGMENT This project is funded by the European Union project PRESENCIA, IST-2001-37927. REFERENCES [1] G. Pfurtscheller, C. Guger, G. Müller, G. Krausz, and C. Neuper, “Brain oscillations control hand orthosis in a tetraplegic,” Neuroscience Letters, 292, pp.. 211-214, 2000.
[2] R. Leeb and G. Pfurtscheller, “Walking through a Virtual City by Thought,” Proc. 26th IEEE Engineering in Medicine and Biology Society (EMBS), San Francisco, CA, pp. 4503-4506, 2004. [3] C. Guger, “Real-time data processing under Windows for an EEG-based brain-computer interface,” Dissertation, University of Technology Graz, 1999. [4] T.M. Vaughan, J.R. Wolpaw, and E. Donchin, “EEG-based co mmunication: prospects and problems,” IEEE Trans. Rehab. Engng, 4s(4), pp. 25-430, 1996. [5] G. Schalk, DJ McFarland, T. Hinterberger, N. Birbaumer, JR. Wolpaw, „BCI2000: A general-purpose brain-computer interface (BCI) system,“ IEEE Trans. Biomed Eng. 2004: 51:1044-1051. [6] C. Guger, et al, How many people are able to operate an EEG-based brain-computer interface (BCI) ?, IEEE Trans. Rehab. Engng, 11(2), pp. 145-147, 2003.
