Non-invasive beat-to-beat cardiac output monitoring by an improved method of transthoracic bioimpedance measurement.

The report describes a method of impedance cardiography using an improved estimate of thoracic volume. The formulas and their implementation in hardware and software are explained and new shortband electrodes are described which generate a good homogeneous thoracic field. Examples of stroke volume and cardiac output curves underline the capabilities of the monitoring system "Task Force Monitor". In several experiments, results are compared to thermodilution as well as to BioZ measurements: the new method excels in comparison with thermodilution and is comparable to the BioZ device. Compared to traditional electrodes, the new shortband electrodes are shown to provide better reproducibility.

Feature extraction for on-line EEG classification using principal components and linear discriminants.

The study focuses on the problems of dimensionality reduction by means of principal component analysis (PCA) in the context of single-trial EEG data classification (i.e. discriminating between imagined left- and right-hand movement). The principal components with the highest variance, however, do not necessarily carry the greatest information to enable a discrimination between classes. An EEG data set is presented where principal components with high variance cannot be used for discrimination. In addition, a method based on linear discriminant analysis (LDA), is introduced that detects principal components which can be used for discrimination, leading to data sets of reduced dimensionality but similar classification accuracy.

Adaptive autoregressive modeling used for single-trial EEG classification.

An adaptive autoregressive (AAR) model is used for analyzing event-related EEG changes. Such an AAR model is applied to single EEG trials of three subjects, recorded over both sensorimotor areas during imagination of left and right hand movements. It is found that discrimination between both types of motor-imagery is possible using linear discriminant analysis, but the time point for optimal classification is different in each subject. For the estimation of the AAR parameters, the Least-mean-squares and the Recursive-least-squares algorithms are compared. In both methods, the update coefficient plays a key role: it determines the adaptation ratio as well as the estimation accuracy. A new method, based on minimizing the prediction error, is introduced for determining the update coefficient.

Timing of EEG-based cursor control.

Recent studies show that humans can learn to control the amplitude of electroencephalography (EEG) activity in specific frequency bands over sensorimotor cortex and use it to move a cursor to a target on a computer screen. EEG-based communication could be a valuable new communication and control option for those with severe motor disabilities. Realization of this potential requires detailed knowledge of the characteristic features of EEG control. This study examined the course of EEG control after presentation of a target. At the beginning of each trial, a target appeared at the top or bottom edge of the subject's video screen and 1 sec later a cursor began to move vertically as a function of EEG amplitude in a specific frequency band. In well-trained subjects, this amplitude was high at the time the target appeared and then either remained high (i.e., for a top target) or fell rapidly (i.e., for a bottom target). Target-specific EEG amplitude control began 0.5 sec after the target appeared and appeared to wax and wane with a period of approximately 1 sec until the cursor reached the target (i.e., a hit) or the opposite edge of the screen (i.e., a miss). Accuracy was 90% or greater for each subject. Top-target errors usually occurred later in the trial because of failure to reach and/or maintain sufficiently high amplitude, whereas bottom-target errors usually occurred immediately because of failure to reduce an initially high amplitude quickly enough. The results suggest modifications that could improve performance. These include lengthening the intertrial period, shortening the delay between target appearance and cursor movement, and including time within the trial as a variable in the equation that translates EEG into cursor movement.

EEG-based discrimination between imagination of right and left hand movement.

Three subjects were asked to imagine either right or left hand movement depending on a visual cue stimulus. The interval between two consecutive imagination tasks was > 10 s. Each subject imagined a total of 160 hand movements in each of 3-4 sessions (training) without feedback and 7-8 sessions with feedback. The EEG was recorded bipolarly from left and right central and parietal regions and was sampled at 128 Hz. In the feedback sessions, the EEG from both central channels was classified on-line with a neural network classifier, and the success of the discrimination between left and right movement imagination was given within 1.5 s by means of a visual feedback. For each subject, different frequency components in the alpha and beta band were found which provided best discrimination between left and right hand movement imagination. These frequency bands varied between 9 and 14 Hz and between 18 and 26 Hz. The accuracy of on-line classification was approximately 80% in all 3 subjects and did not improve with increasing number of sessions. By averaging over all training and over all feedback sessions, the EEG data revealed a significant desynchronisation (ERD) over the contralateral central area and synchronisation (ERS) over the ipsilateral side. The ERD/ERS patterns over all sessions displayed a relatively small intra-subject variability with slight differences between sessions with and without feedback.

On-line EEG classification during externally-paced hand movements using a neural network-based classifier.

EEGs of 6 normal subjects were recorded during sequences of periodic left or right hand movement. Left or right was indicated by a visual cue. The question posed was: 'Is it possible to move a cursor on a monitor to the right or left side using the EEG signals for cursor control?' For this purpose the EEG during performance of hand movement was analyzed and classified on-line. A neural network in form of a learning vector quantizertion (LVQ) with an input dimension of 16 was trained to classify EEG patterns from two electrodes and two time windows. After two training sessions on 2 different days, 4 subjects showed a classification accuracy of 89-100%. For two subjects classification was not possible. These results show that in general movement specific EEG-patterns can be found, classified in real time and used to move a cursor on a monitor to the left or right. On-line EEG classification is necessary when the EEG is used as input signal to a brain computer interface (BCI). Such a BCI can be a help for handicapped people.

Graz brain-computer interface II: towards communication between humans and computers based on online classification of three different EEG patterns.

The paper describes work on the brain--computer interface (BCI). The BCI is designed to help patients with severe motor impairment (e.g. amyotropic lateral sclerosis) to communicate with their environment through wilful modification of their EEG. To establish such a communication channel, two major prerequisites have to be fulfilled: features that reliably describe several distinctive brain states have to be available, and these features must be classified on-line, i.e. on a single-trial basis. The prototype Graz BCI II, which is based on the distinction of three different types of EEG pattern, is described, and results of online and offline classification performance of four subjects are reported. The online results suggest that, in the best case, a classification accuracy of about 60% is reached after only three training sessions. The online results show how selection of specific frequency bands influences the classification performance in single-trial data.

Sleep classification in infants by decision tree-based neural networks.

This paper presents an AI-based approach to automatic sleep stage scoring. The system TBNN (Tree-Based Neural Network) uses a decision-tree generator to provide knowledge that defines the architecture of a backpropagation neural network, including feature selection and initialisation of the weights. The case study reports a successful application to the data from polygraphic all-night sleep of 8 babies aged 6 months. The teaching input was provided by a medical expert in accordance with the rules of Guilleminault and Souquet. The performance of TBNN is compared with 5 other methods and the results are discussed.

EEG-based brain computer interface (BCI). Search for optimal electrode positions and frequency components.

Several laboratories around the world have recently started to investigate EEG-based brain computer interface (BCI) systems in order to create a new communication channel for subjects with severe motor impairments. The present paper describes an initial evaluation of 64-channel EEG data recorded while subjects used one EEG channel over the left sensorimotor area to control on-line vertical cursor movement. Targets were given at the top or bottom of a computer screen. Data from 3 subjects in the early stages of training were analyzed by calculating band power time courses and maps for top and bottom targets separately. In addition, the Distinction Sensitive Learning Vector Quantizer (DSLVQ) was applied to single-trial EEG data. It was found that for each subject there exist optimal electrode positions and frequency components for on-line EEG-based cursor control.

Selection of electrode positions for an EEG-based brain computer interface (BCI).

One major question in designing an EEG-based Brain Computer Interface to bypass the normal motor pathways is the selection of proper electrode positions. This study investigates electrode selection with a Distinction Sensitive Learning Vector Quantizer (DSLVQ). DSLVQ is an extended Learning Vector Quantizer (LVQ) which employs a weighted distance function for dynamical scaling and feature selection. The data analysed and classified were 56-channel EEG recordings over sensorimotor areas during preparation for discrete left or right index finger flexions. Data from 3 subjects are reported. It was found by DSLVQ that the most important electrode positions for differentiation between planning of left and right finger movement overlie cortical finger/hand areas over both hemispheres.

Differentiation between finger, toe and tongue movement in man based on 40 Hz EEG.

Movements of right and left index fingers, right toe and tongue were studied by EEG measurement in the alpha and gamma (30-40 Hz) bands. The EEG was recorded with a 56-electrode array over pre- and postcentral areas. For each movement the average power decrease, as a measurement of the event-related desynchronization or power increase in narrow frequency bands, was calculated. Single-trial data from 8 electrodes, 3 frequency bands and 4 time points within a 1 sec window were subject to a classification task. It was found that, based on single EEG trials, the data from the 4 movements could be differentiated with an accuracy of 70% when alpha and gamma band activity were used but only with 58% in the case of the alpha band activity alone. This shows that the gamma band activity or 40 Hz EEG is strongly related to planning of a specific movement and therefore, improves the accuracy of classification significantly.

AI-based approach to automatic sleep classification.

The primary goal of this paper is to introduce the potential of artificial intelligence (AI) methods to researchers in sleep classification. AI provides learning procedures for the construction of a sleep classifier, prescribing how to combine the observed parameters and how to derive the corresponding decision thresholds. A case study reporting a successful application of an automatic induction of decision trees and of a learning vector quantizer to this domain is presented.

Towards automated sleep classification in infants using symbolic and subsymbolic approaches.

The paper addresses the problem of automatic sleep classification. A special effort is made to find a method of extracting reasonable descriptions of the individual sleep stages from sample measurements of EGG, EMG, EOG, etc., and from a classification of these measurements provided by an expert. The method should satisfy three requirements: classification accuracy, interpretability of the results, and the ability to select the relevant and discard the irrelevant variables. The solution suggested in this paper consists of a combination of the subsymbolic algorithm LVQ with the symbolic decision tree generator ID3. Results demonstrating the feasibility and utility of our approach are also presented.

EEG classification by learning vector quantization.

EEG classification using Learning Vector Quantization (LVQ) is introduced on the basis of a Brain-Computer Interface (BCI) built in Graz, where a subject controlled a cursor in one dimension on a monitor using potentials recorded from the intact scalp. The method of classification with LVQ is described in detail along with first results on a subject who participated in four on-line cursor control sessions. Using this data, extensive off-line experiments were performed to show the influence of the various parameters of the classifier and the extracted features of the EEG on the classification results.

Sleep classification in infants based on artificial neural networks.

The study reports on the possibility of classifying sleep stages in infants using an artificial neural network. The polygraphic data from 4 babies aged 6 weeks, 6 months and 1 year recorded over 8 hours were available for classification. From each baby 22 signals were recorded, digitized and stored on an optical disc. Subsets of these signals and additional calculated parameters were used to obtain data vectors, each of which represents an interval of 30 sec. For classification, two types of neural networks were used, a Multilayer Perceptron and a Learning Vector Quantizer. The teaching input for both networks was provided by a human expert. For the 6 sleep classes in babies aged 6 months, a 65% to 80% rate of correct classification (4 babies) was obtained for the testing data not previously seen.

Prediction of the side of hand movements from single-trial multi-channel EEG data using neural networks.

Thirty channels of EEG data were recorded prior to voluntary right or left hand movements. Event-related desynchronization (ERD) was quantified in the 8-10 Hz and 10-12 Hz bands in single-trial data and used as training input for a neural network comprised of a learning vector quantizer (LVQ). After a training period, the network was able to predict the side of hand movement from single-trial EEG data recorded prior to movement onset.