$231 Evoked Potential Studies (J. Clin. Neurophysiol., 1984, 1: 3). The frequency of OPs is in the region of 100 to 160Hz, whereas the primary components are considerably slower, from 0.01 to 25Hz. This fortunate frequency content allows us to separate the components of the ERG by using linear zero-phase shift filters. We have designed a finite impulse response (FIR) filter for this purpose. The coefficients of this filter are 1, 1, 0, - 7 , - 1 3 , 36, - 1 3 , - 7 , 0, 1, 1 and when they are spaced at intervals of 2.5 msec the filter attenuates a- and b- wave amplitudes practically to zero and passes OPs unchanged. Our filter gives almost equal curves when applied to two consecutive averages of five stimuli. This indicates that the signal-to-noise ratio of the filtered curves is high and they can be used for quantitative analysis of oscillatory potentials.
DPW2.10 E N H A N C E D AVERAGE EVENT-RELATED POTENTIAL ESTIMATES USING STATISTICAL PATTERN RECOGNITION.
A.S. Germs, N.H. Morgan, J.C. Doyle and S.L Bressler (San Francisco, CA, USA) Many attempts have been made to form improved estimates of the average event-related potential (ERP). Most methods have inherent unrealistic assumptions about the statistical properties of signal and noise, and further assume that task-related signals are present in every trial. Here we present a simple new method without either type of assumption. The procedure is as follows: 1) Filter the single trials to remove alpha and higher frequency components; 2) Down-sample the resulting time series; 3) Choose a contiguous set of time points from an interval of interest (e.g. around the P300 peak) to produce features representing one class of a linear discriminant function, (the features are simply the voltages at the chosen time points). Form the other class using time points from an interval with low average energy (such as pre-stimulus); 4) Train the Fisher discriminant function to separate these two classes using features from two thirds of the set of single trials; 5) Test the discrimininant function with the remaining one-third of the trials. If the classification accuracy is significant, make a list of the correctly-classified trials; 6) Repeat steps 4 & 5 with the data divided into different, non-overlapping training and testing data sets; 7) Compute the average of the unfiltered trials corresponding to the filtered trials which were correctly classified in (5). This is the enhanced ERP estimate. The method has been applied several hundred times to different types of ERPs. In the averages of the correctly classified trials, the ERP peaks are greatly enhanced in comparison with the original averages. The averages of the incorrectly classified trials resemble the background EEG. These results suggest that the method is successful in identifying trials without discriminable event-related signals, and provides a simple method of average ERP estimation without a priori assumptions about the characteristics of the (unknown) event-related signal made in techniques such as Wiener or minimum mean-square error filtering.
DPW2.11 M U L T I D I M E N S I O N A L FACTORIAL METHO D S FOR EEG DATA.
J.P. Banquet, W. Guenther and D. Breitling (Paris, France) In the past decade, EEG data have been submitted to different kinds of multidimensional factorial analyses (MFA). An overview of the literature shows that the rationale of the use of these methods is totally different according to the type of data analysed: spontaneous EEG versus evoked/event-related potentials (ERPs). In the latter, the approach is very similar to that of the early cognitive psychologists who tended to test well established hypotheses on cognitive factors. The factors have been replaced by ERP components. The corresponding type of MFA is a genuine factor analysis with factor rotation. In the geometrical representation of the results, the elements of interest are the undimensional geometrical factors affected by a physiological significance. Conversely, in the case of the spontaneous EEG, the approach is one of exploration in the absence of any predetermined hypothesis. The corresponding method can be labelled Inertial Analysis. It is implemented by the diagonalisation of either covariance correlation matrices for the correlational approach or disjunctive correspondence tables in the distributional approach. The elements of interest are not only the factors but the subspaces determined by several factors whereupon elements and variables are projected. These different techniques are described and implemented on sleep and E R P s / E E G data.
DPW2.12 RULE-BASED AUTOMATED ANALYSIS OF SENSORY EVOKED POTENTIALS.
J.R. Boston (Pittsburgh, PA, USA) Sensory evoked potentials (EP) provide a means to evaluate central nervous system function in anaesthetized or comatose patients, and they are being increasingly utilized for monitoring applications in the operating room and the intensive care unit. A major problem, however, is the need for continuous interpretation of the results. The development of automated techniques to screen or interpret the responses would greatly facilitate monitoring applications. The steps in response interpretation include: determination of whether a response is present; identification of peaks that correspond to normal responses; measurement of latencies and amplitudes. Several previous studies have addressed specific steps using analytical programming techniques. These techniques have generally been correct for 80-90 percent of responses, but it is the incorrectly analyzed responses that are often of the greatest clinical interest. Examination of the failures suggests that heuristic rules will be needed for clinically useful automated interpretation. These rules should incorporate the subjective criteria used by human
$232 interpreters that are specific to particular response patterns. We are implementing a rule-based system to analyze the brain-stem auditory EP, using the OPS5 production programming system. The features are obtained using analytical techniques, and rules are used to assign appropriate peak numbers to the input peaks. I thank Harry Pople and the Decision Systems Laboratory, University of Pittsburgh, for their assistance and for the use of computer facilities.
abnormal a m o u n t of beta activity and borderline abnormal. For the derivations F4-C 4, I~-C 3, Fs-A 2, FT-A 1, T4-T6, T~-Ts, P4-O2 and P3-O1 the most discriminating variables were found in the frequency intervals from about 5 to 7.5 Hz, around 10 Hz and from 0 to 1 Hz. The scores on the two discriminant functions of an EEG spectrum characterized the corresponding EEG in a continuous way. The application of this concept of analysis to work in a routine laboratory will be reported.
EEG DATA P R O C E S S I N G . DPPI.01 A U T O M A T I C D E T E C T I O N O F ICTAL AND INT E R I C T A L EPILEPTIC ACTIVITY IN A M B U L A T O R Y CASSE'FTE R E C O R D I N G S .
DPPI.03 C O M P A R I S O N OF S O M E Q U A N T I T A T I V E M E T H O D S IN D I S C R I M I N A T I O N BETWEEN EEG DESYNCRHONIZING AND SYNCHRONIZING REACTIONS.
J. Gotman and D. Koffler
T. J~rdhnh?tzy, Z. Ori and 1. Gyori
(Montreal, Canada)
(Szeged. Hungary)
Twenty-four hour EEG recordings from ambulatory cassette equipment were subjected to computer analysis for the concurrent detection of three patterns: seizures, bursts of spike-andwave and isolated spikes. EEGs were first played back at 20 times recording speed, digitized and written on a computer disk. Analysis then proceeded off-line. Spikes and seizures were recognized by existing and formally evaluated methods. Bursts of spike-and-wave were detected by a new method based on the recognition of groups of spikes and sharp waves occurring either in rapid succession in one channel or with a relatively loose temporal pattern in two or three channels. This method allowed detection of classical 3 / s e c bursts as well as short and irregular bursts seen during sleep and secondary generalized epilepsy. The performance was quantitatively evaluated and it was found that only very poorly defined bursts were missed~ False detections due to artefacts could be frequent during periods of physical activity, despite special artefact-rejection procedures. Results were presented in the form of a discontinuous paper tracing including valid and false detections, for visual interpretation. Original recordings were typically reduced by factors of 20 to 50. Quantification of true epileptic events could be obtained after interactive editing of false detections.
As mentioned in our other presentation, (J~rdanhSzy and Ori, Electroenceph. clin. Neurophysiol., 1985; this w)lume) the automatic recognition of synchronizing and desynchronizing reactions is of great importance for computerised scoring of the depth of the light thiobarbiturate narcosis. In this second step, the so called period and amplitude analysis was used on the same EEG activity as it was in the spectral analysis study. Because of the results of a pilot investigation, only the parameters related to the number of waves and the sum of the area under them in the 'traditional' EEG bands were used. It was found that the per cent changes in the sum of the areas under the waves in the alpha, theta and beta bands were the best in the discrimination. In a 'probit' discriminant analysis it was possible to get an about 80% agreement with visual scoring. In the last step, the three parameters described by Hjorth were calculated for analysis, It was found that the best parameters were the per cent changes in those related to Activity and Mobility and their 'coefficients of form' after stimulation. In a 'probit' discriminant analysis an 87% agreement with visual scoring was achieved which result is about the same as that obtained by spectral analysis.
DPPI.02 C H A R A C T E R I Z A T I O N OF EEG P O W E R SPECTRA IN CLINICAL PRACTICE.
DPPI.04 EEG M O N I T O R I N G W I T H P O W E R S P E C T R A L DENSITY BAND ANALYSIS. N.J. Etherington, D.S.L. Lloyd and C.J. Watkins
G.H. Wieneke, T. Jfirdb.nhSzy and W. Storm van Leeuwen (Royston and London, U K ) (Utrecht, The Netherlands) The frequency bands which discriminate between normal and classes of abnormal EEG were determined using a linear discriminant analysis with, as variables, the power densities in 0.6 Hz bands of 100 second spontaneous EEG (eyes closed). The abnormal classes were: abnormal amount of slow activity,
A simple quantitative electroencephalographic (EEG) analysis method, based on r.m.s, power spectral density, and employing broad-band (8th order) Chebyshev analogue filters, has been developed to monitor patients undergoing surgical, anaesthetic, and intensive care treatment. This system is based on the concept that clinicians require quantitative EEG analysis which