100

Electroencephalography and clinical Neurophysiology , 86 (1993) 100-109 1993 Elsevier Scientific Publishers Ireland, Ltd. 0013-4649/93/$06.00

EEG92108

Pattern recognition of the electroencephalogram by artificial neural networks

Gfibor Jand6 *, Ralph M. Siegel, Zsolt Horvfith* and Gy6rgy Buzsgtki
Center for Molecular and Behavioral Neuroscience, Rutgers, State University of New Jersey, Newark, NJ 07102 (USA)

(Accepted for publication: 16 September 1992)

Summary A back-propagation network was trained to recognize high voltage spike-and-wave spindle (HVS) patterns in the rat, a rodent model of human petit real epilepsy. The spontaneously occurring HVSs were examined in 137 rats of the Fisher 344 and Brown Norway strains and their F1, F2 and backcross hybrids. Neocortical EEG and movement of the rat were recorded for 12 night hours in each animal and analog data were filtered (low cut: 1 Hz; high cut: 50 Hz) and sampled at 100 Hz with 12 bit precision. A training data set was generated by manually marking durations of HVS epochs in 16 representative animals selected from each group. Training data were presented to back-propagation networks with variable numbers of input, hidden and output cells. The performance of different types of networks was first examined with the training samples and then the best configuration was tested on novel sets of the EEG data. FFT transformation of EEG significantly improved the pattern recognition ability of the network. With the most effective configuration (16 input; 19 hidden; 1 output cells) the summed squared error dropped by 80% as compared with that of the initial random weights. When testing the network with new patterns the manual and automatic evaluations were compared quantitatively. HVSs which were detected properly by the network reached 93-99% of the manually marked HVS patterns, while falsely detected events (non-HVS, artifacts) varied between 18% and 40%. These findings demonstrate the utility of back-propagation networks in automatic recognition of EEG patterns.

Key words: EEG analysis; Neural nets; Parallel processing; Back-propagation; Epilepsy; Spike-and-wave

With the introduction of digital computers various mathematical methods have been applied to quantify the EEG signal and thereby ease the clinical diagnosis. These mostly linear methods include Fourier analysis, brain mapping, feature extraction and pattern recognition (Gevins 1980; Gotman 1986, 1990; Gevins and R6mond 1987; Ktonas 1987; Lopes da Silva and Mars 1987; John et al. 1988). To date, pattern recognition has remained the most difficult problem of EEG analysis (Gevins 1980; Gotman 1990). Several reviews of applications of computer methods to the analysis of the EEG and computer analysis of automatic detection in epilepsy have been published in the last few years (Gotman 1986; Gevins and REmond 1987; Lopes da Silva and Mars 1987). In this paper we present a novel

Correspondence to: Gy6rgy Buzs~iki, Center for Molecular and Behavioral Neuroscience, Rutgers University, 197 University Avenue, Newark, NJ 07102 (USA). * Permanent address:-Department of Physiology, Medical School, 7643 P6cs, Hungary.

method of EEG analysis by using neural nets for the recognition and quantification of various EEG patterns. The development of the 3-layer neural net by Rummelhart and McClelland (1986) permitted the discovery of non-linear relationships between complex input and output data. Their neural network consisted of simple non-linear input-output elements. The input propagates through the network in a feed-forward fashion. Back-propagation, a gradient descent method, is used to minimize the output error by altering the strengths of the connections (weights) between the layers by a non-linear least square algorithm. In our approach, we designed and tested the performance of a back-propagation neuronal network for recognizing spike-and-wave patterns of inbred rat strains. HVSs in these rat strains differ in their probability of occurrence, amplitude and intraburst frequency. Spike-and-wave EEG patterns were chosen because they are relatively easy to recognize visually and we have accumulated a large database (approximately 2000 h) of neocortical EEG from 137 rats with genetically inherited petit mal epilepsy (Buzsfiki et al. 1990; Carpi et al. 1992).

P A T T E R N R E C O G N I T I O N O F E E G BY N E U R A L N E T W O R K S

101

Material and methods

Animals and surgery One hundred and thirty-seven male and female rats of Fischer 344 and Brown Norway inbred strains and their F1, F2 and backcross descendents were used. All rats were between 6 and 8 months old at the time of testing. Previous studies have established that these strains express genetically mediated neocortical high voltage spike-and-wave patterns (HVS). Epidural recording electrodes were used with the following coordinates: L = 2.5 mm, AP = 2.5, relative to bregma; and L = 5.0, AP = 0.0, bilaterally. These coordinates were chosen because previous mapping experiments showed amplitude maxima of HVS at these locations (Buzsfiki et al. 1988). Two additional screws were placed in the midline above the cerebellum and served as indifferent and ground electrodes. Recording and data processing The animals were kept under normal dark-light cycle throughout the experiments. During the recording session the rat was placed into a sound-insulated cage. E E G activity was recorded to a computer hard disk after digitization of the traces at 100 Hz with an IBM PC. The E E G was band-pass filtered at 1-50 Hz

(Grass E E G amplifier). A recording session lasted for 12 h (8.00 p.m-8.00 a.m.). For analysis, E E G derived from the right frontal cortical electrode was selected. Neocortically recorded HVSs were detected by visual inspection and the duration of the event was marked manually by a mouse-guided editor. These epochs, obtained from 16 rats along with HVS-free periods, were used as training signals for the network. HVS-free periods were randomly selected from interictal E E G periods. The total length of HVS-free periods was approximately 3 times longer than the total length of the HVS epochs.

Back-propagation network The non-linear back-propagation algorithm according to Rummelhart and McClelland (1986) was used (Fig. 1). The input of a particular element was calculated as the sum of the input values multiplied by the connection strength (synaptic weights).
rrl i-1

where H i is a hidden layer cell, |i is the ith input cell, m is the number of input cells, wji is the synaptic strength between ith input cells and jth hidden layer cell.

f0
m

Nii~i

Input Hidden

Of.r__
..~...

llllllll, , f4 - fol
@ijJ, foi

Output

Fig. 1. A: schematic diagram of the artificial neuronal network. It consists of an input layer, a hidden layer and an output layer. The n u m b e r of cells in each layer varies and the optimal net configuration is determined by feedback training. Each cell is connected to each cell from the neighboring layer. (Only some connections are shown here.) T h e n u m b e r of input cells is equal to the size of the sliding window in pixels (boxed area of a 12 sec E E G trace). Each different input cell receives contiguous samples of the raw EEG. W h e n the calculation cycle is done the window moves one pixel (10 msec) further. B: the E E G sample in the sliding window (boxed area in A) is transformed into the frequency domain by an F F T algorithm. Both the real and imaginary parts are used as input data for the network. The differently textured input cells (f0 to fT) receive different f r e q u e n c y / a m p l i t u d e and f r e q u e n c y / p h a s e components of the E E G epoch. Bottom trace: analog output of the network.

102
The value for each cell is passed through a non-linear sigmoidal function g(x): this function is used because it permits the network to perform a high-dimensional, non-linear approximation to the data. Indeed, this allows the construction of quite arbitrary, non-pretermined functions by the network.
1

G. JAND() ET AL.
sive time slices were generated by shifting 10 msec forward in time (i.e., a single data point). The same sliding window procedure was used for analyzing the F F T of the E E G signal. However, instead of the time function of the EEG, both the real and imaginary parts of F F T were presented to the network, which was called FNET. Prior to searching for the best weights, the network was initialized by setting the weight values randomly between 0.5 and - 0 . 5 . A single output cell of the network was used to predict the presence or absence of HVS in the recorded EEG. The output point was arbitrarily chosen to lie at the center of the input window. The summed squared error between the predicted and pre-evaluated output was used as a measure of the performance of the network in predicting HVS epochs. Various configurations of the network were tested by systematically changing the number of inputs cells, hidden layer cells and output cells (see Results). By taking into consideration network performance (false positive: non-HVS events, which were falsely marked by the network; false negative: HVS events, which were missed by the network) and running time efficacy, the most effective configuration (16 input cells, 19 hidden layer cells and 1 output cell; see Results) was then tested with a large E E G data base derived from 137 rats. The events selected by the network in each rat were visually edited. The initiation and termination (length) of each event was corrected manually. False positive events were deleted (e.g., Fig. 6). The overall performance of the network was calculated by comparing the manually determined and network selected events of the E E G database.

g(x)

1+e x

(2)

The output of a hidden layer element was the input to the cells located in the final output layer of the network. Since the cell in the final output layer received inputs from all the previous layer cells, the output of the whole network at a given p pattern O p can be described as follows:

O~ = g (j~lw~ig (i~lWjil P + Oj) )

(3)

where m is the number of input cells, n is the number of hidden cells, wkj are the weights between the hidden and output layer, w~i are the weights between the input and hidden layer cells; Oj is the threshold of the jth hidden layer cell; k is always 1 in this study. The error of the network for identifying a given pattern p is calculated as:

E = (0 p-Op) 2

(4)

where d o is the desired output, O p is the actual output of the network. The weights were corrected according to the delta rule: wJ~ew= w?td+ Yawj,

0E

(5)

where y is a learning coefficient, 0E/0wji is the derivative of the error function with respect to the given weight wji. Computations were done on an IBM RS/6000 Model 540 with 32-bit precision. The manually selected HVS and explicitly non-HVS epochs were marked on the E E G records of 16 representative animals selected from each group. Values of 0 and 1 were assigned to non-HVS and HVS epochs, respectively. These files served as a template during the training period and for initial testing. For developing the network, various tests were carried out by varying the number of input and hidden layer cells and output cells. In most of the tests the number of output cells was set to one. The performance of the network was assessed by the summed squared error. Identical network configurations were presented with two fundamentally different data sets: (A) the raw E E G records and (B) the Fast Fourier Transform (FFT) of the E E G (Fig. 1). The F F T was computed using the methods described in Press et al. (1988). For training with the raw EEG, contiguously shifting time slices of 40-640 msec (4-64 data points) were analyzed by the network. This network was called NET. Succes-

Results

Most HVSs occurred during awake immobility and slow wave sleep. The behavioral correlation, genetic control and other aspects of HVS are the subject of a different report (in preparation).

Training by feedback
A typical example of the HVS together with the network-evaluated outputs is shown in Fig. 2. The training set consisted of 1007 visually recognized HVS (total duration: 5629 sec) derived from 16 rats of the Brown Norway and Fischer 344 strains and their F I and F2 descendents together with explicitly non-HVS epochs (13,595 sec) of similar average duration. With the initial random weights the network produced an analog output similar to the input signal in NET. During training, the output gradually approximated the desired square wave pattern. For extensive training, the "best" network configuration (see below) was used

P A T T E R N R E C O G N I T I O N O F E E G BY N E U R A L N E T W O R K S

103

-.u.b,u- .i,,b,.lil.,~blilllil,.

D
I

'

'

'

'

'

'

'

'

'

'

6 time (sec)

10

12

D
,

~ F f ' l ' ~ ' ' l ' ' ' ' l ' ' ' ' l ' ' ' ' l ' ' ~ ' l ' ' ' ' l ' ' ' ' l ' ' ' ' l ' ' ' ' l

10 20

30 40 50 60 70 interations (x 106 )

80

90 100

Fig. 2. Performance of the network at different stages of feedback training. A - D : output of the network for the evaluation of an HVS episode (top trace). The E E G trace was converted into the frequency domain and both the real and imaginary parts of F F T were presented to the network. The straight line in D (arrow) is the manually determined HVS (the ideal network output). The square of the difference between the manually detected event and network output is the squared error. Bottom: s u m m e d squared error as a function of the n u m b e r of interations (symbolic time). Dots on the learning curve correspond to network outputs A - D above. RND: random weights (maximum error). 0 = ideal output (minimum error). Learning coefficient 3' was set to 0.0001 (see Eq. 5).

after about 6 million iterations the network reliably recognized HVS when they occurred in the awake animal with a desynchronized background. However, slow waves occurring in the sleeping rat were sometimes falsely classified as HVS. With further training this discrimination substantially improved. Performance of the network with the FFT-transformed E E G data reached an asymptote after about 50 million iterations by which time the summed squared error decreased to 20% of the initial random weights. The magnitude of the final summed squared error depended on the variability of the HVSs presented to the network. For example, when only a single 12 h long record from a single animal was used as a training template, the summed squared error decreased by 90% or more. However, the general performance of the network trained on data from a single animal was significantly inferior in detecting HVSs obtained from other animals than networks that were trained by more variable patterns derived from several animals. Using the same network configuration, performance did not change noticeably until after 80 million iterations when the raw E E G was used as an input. The "learning" curve was sigmoid and performance did not reach an asymptote even after 160 million iterations. Even with such an extensive training the summed squared error did not decrease below 40% of the initial random weights. A caveat is in order here, however. If the training had been continued for longer, it is possible that the summed squared error of the network with the raw E E G patterns may have eventually decreased below the error computed for the FFT-transformed data. However, such calculations are computationally very extensive and the necessary calculations would have had required several more months of computer time even on a superfast (60 MFlop) workstation. Training the network with the raw E E G was approximately 10 times slower than training with the FFTtransformed data. For the above reasons we tried to extrapolate the most likely result of further training. The learning curve of the raw input nework (NET) was best fit with a single sigmoidal function. We used the Iogit function:
365,000 E = 1 + e 1)'169i 8.77 + 184,000 (6)

(input = 16; hidden = 19; output = 1; y = 0.0001; see Eq. 5), and the summed squared error was plotted as a function of the number of iterations. Performance increased rapidly with the FFT-transformed data and

where E is the summed squared error, i refers to the number of iterations. Levenberg-Marquardt's method for determining the best parameters of the function was used (Press et al. 1988). A single sigmoidal function was a poor fit to the Fourier input network (FNET) learning curve as assessed with the variance accounted for. We found that the data was better fit with the sum of two sigmoidal functions:
E 337,000 435,000 l+el.0ti 3.o3 + l+el).llsi+I.03 1-96,000 (7)

104 The F N E T and N E T curve fits were extrapolated to large number of iterations and did not cross. While one must be careful in extrapolating data these results are indicative that the F N E T was always better than N E T regardless of the number of iterations.

G. JANDO ET AL. The summary of all possible interactions between various numbers of input and hidden layer cells is presented in Fig. 4. Both raw E E G and FFT-transformed data were tested. On one axis the number of input cells is given, on the other is the number of hidden cells. The error is color coded. The network with the best performance is indicated by the + and is located at 16 input and 19 hidden cells. This is referred to elsewhere as the "best configuration". It is evident from these experiments that the efficiency of the network does not increase with a larger number of input and hidden layer ceils and, in fact, a large number of cells in either layer deteriorate performance. The "best" configurations were different for analyzing raw E E G and FFT-transformed data. We also changed the number of output ceils. However, choosing a value other than 1 dramatically increased the normalized summed squared error. Therefore, in all subsequent tests the number of input cells was kept at 1.

Network configuration
In order to determine the most effective network configuration, various parameters of the network were systematically varied. Since the back-propagation network consists of 3 ~ayers (i.e., number of input, hidden and output cells), one of the parameters was varied at a time, while leaving the remaining two constant. For these computations 2 million iterations were used in all cases and the 3' was fixed to 0.001 (see Eq. 5). The number of input cells was varied between 4 and 64. The summed squared error showed a minimum at 8 input cells using the raw E E G signal (NET) and at 16 input ceils using the FFT data (FNET). Increasing the number of input cells did not improve the performance of the networks. In fact, the performance for the evaluation of raw E E G signal deteriorated significantly when the number of input cells was increased to 16 or more (Fig. 3A). The effect of varying the number of hidden cells on the summed squared error is illustrated in Fig. 3B. The number of input and output cells was kept constant (16 and 1, respectively) and the number of hidden layer cells was varied between 2 and 97. By increasing the number of hidden cells the summed squared error gradually decreased for both sets of data (raw E E G and FFT data). Again, the performance was better when FFT data were used as the input (FNET) and the best performance was obtained with 18-25 hidden layer cells. Increasing the number of hidden cells further slightly deteriorated the efficiency.

Comparison of network performance with manually edited data

The continuous output of the network was converted to give an unambiguous discrete signal (1 or 0) by a threshold function (Fig. 5). Fig. 6 illustrates exampies of true and false detections of HVS by the "best" configuration of the network. The effect of threshold on the precision of HVS detection by the network is illustrated in Fig. 7A and B. In these tests the threshold was gradually increased from 0.85 to 0.99. Both missed HVS (false negativity) and spuriously detected non-HVS events (false positivity) were plotted as a function of the increasing threshold values. These tests were carried out for both the training and novel sets of data. By increasing the threshold, detection of spurious

hidden=27

input=16 D = -NET

FNET

....

10

i ....

20

r ....

30

i ....

40
cells

i . . . . . . . . . . . .

5'o

6'0

70

. . . . . . . .

i ....

i ....

i,,

,i

. . . . . . . . . .

110

20

30

40

50

6J0 cells

710 ' 8 r0 ~' I ' 9 '0 ~

. . . .

1 0I0

input

hidden

Fig. 3. Effect of network configuration on performance. A: summed squared error as a function of input cells. The number of hidden layer cells was constant (hidden = 27). The curves for raw EEG (NET) and its FFT (FNET) were plotted out after equal number of iterations. Learning coefficient 3' was set to 0.001 (see Eq. 5). B: summed squared error as a function of hidden layer cells. The number of input cells was constant (input = 16). Other parameters were the same as in A.

PATTERN RECOGNITION OF EEG BY NEURAL NETWORKS

105

Fig. 4. Effect of network configuration on performance. Interaction of independent variables (horizontal: number of input cells; vertical: number of hidden cells) as a function summed squared error (color). Dark blue: randomly initialized weights (max). Red: 20% of the dark blue value (min). The best network configuration is marked by " + " on FNET. The number of input cells was 4, 8, 16, 32, 64 and the number of hidden cells was varied between 2 and 97 in steps of 5. The summed squared error was calculated at each node after 2 million iterations. The maps were smoothed by interpolations using a cubic spline algorithm.

events was d e c r e a s e d at t h e e x p e n s e o f missing s o m e H V S events. N e v e r t h e l e s s , at t h r e s h o l d levels w h e r e less t h a n 7 % o f all m a n u a l l y d e t e c t e d H V S w e r e det e c t e d by t h e n e t w o r k , only 40% s p u r i o u s events w e r e classified as H V S . A f r e q u e n t s o u r c e o f t h e falsely d e t e c t e d events w e r e s c r a t c h i n g a n d g r o o m i n g artifacts. It is r e l e v a n t to m e n t i o n h e r e t h a t the f r e q u e n c y o f h i n d leg s c r a t c h i n g in the rat is in t h e r a n g e of t h e H V S ( 6 - 1 0 Hz). Such events w e r e f r e q u e n t a n d o f t e n of l a r g e a m p l i t u d e . A n o t h e r s o u r c e o f false positive events was m e c h a n i c a l artifacts d u e to b u m p i n g o f t h e h e a d s t a g e p r e a m p l i fiers into t h e ceiling o f t h e c a g e at t i m e s w h e n the rat

t r i e d to climb o u t o f t h e r e c o r d i n g cylinder. T h e final source of s p u r i o u s l y d e t e c t e d events was s l e e p s p i n d l e s a n d t h e i r r e g u l a r m i x t u r e of slow waves a n d o c c a s i o n a l spike-like events d u r i n g slow wave sleep. T h e m i s s e d H V S events w e r e always " a t y p i c a l " , e i t h e r b e c a u s e the spike c o m p o n e n t s w e r e very small o r b e c a u s e t h e entire H V S event was o f small a m p l i t u d e .

Discussion
This study i n t r o d u c e s a novel a p p r o a c h for the a u t o matic recognition and discrimination of EEG patterns

106

G. JANDO ET AL.

~ ' - ' i

threshold level

time(sec) Fig. 5, Setting the threshold of the analog network output. The straight line superimposed on the analog output is the manually determined HVS ("ideal" network output). The threshold (horizonal line) is set at 0,85. At this threshold level greater than 99% of HVS were detected, The HVS occurred during sleep. Note large amplitude slow waves preceding the HVS.

using non-linear methods. Various combinations of a 3-layer back-propagation network were tested to derive the optimal configuration for the efficient and reliable detection of HVS patterns in various rat strains. The most difficult aspect of EEG analysis is pattern recognition (Gevins 1980; Niedermeyer and Lopes da Silva 1987). To date, various linear methods have been successfully used for detecting both normal and pathological patterns in the EEG (cf., John et al. 1988). A major disadvantage of the linear methods is the rapid increase of false positivity and false negativity when new and more variable patterns are presented. This is due to the fact that most linear methods use a priori selection criteria. For the recognition of highly variable patterns the method of choice is parallel distributed processing as it is designed to reveal the non-linear relationships between complex input and output (Rummelhart and McCleUand 1986). In this study we have tested the utility of the back-propagation model on HVS of various strains of rats. It must be emphasized

10

.......

auto

,. ~ manual ls Fig. 6. Examples of truly recognized, falsely recognized and missed HVS events by the network. Interrupted lines: network detected events. Continuous lines: experimenter detected events. Note that falsely recognized non-HVS events (1-4) during sleep contain large amplitude spikes and rhythmic spike-and-wave pattern at 2/see. Since HVS were defined as 7-9/see rhythmic events, these were omitted by the experimenter. Missed events (traces 5-8) are usually atypical HVSs, with either unusually small amplitude wave components (trace 5) or missing spike components (traces 6-8). Missed events comprised less than 1% of all HVSs. Traces 9 and 10 indicate events detected by both experimenter and network but with some discrepancy regarding the beginning and end of the events.

125 AJVI

P A T T E R N R E C O G N I T I O N O F E E G BY N E U R A L N E T W O R K S

107

A
140

100o~ 80-

~ 60'

20

n .=gative ~

False
negal tves ~r

0
0.85 0.87 0.89 0 . 9 1 0.93 threshold 0.95

~
0.97 0.99 0,85 0.87 0.89 0.91 0.93 threshold 0.95 0.97 0.99

Fig. 7. Effect of threshold on network performance. Percentage of spuriously detected events (false positives) and missed HVS events (false negatives) are plotted as a function of threshold. A: training set (192 h of E E G from 16 rats). B: novel data set (1021 h = 3,675,600 sec of E E G from 137 rats). Lines represent 2 S.E.M.

that the intraburst frequency, the amplitude ratio of the spike and wave components and other parameters of HVS are significantly different in Fischer 344 and Brown Norway strains and in their F1 and F2 descendents (Carpi et al. 1992). The magnitude of strain-dependent variability is comparable to the variability seen in various forms of human generalized spike-and-wave epilepsy (Gloor and Fariello 1988). To date, various parallel-distributed processing models are available for such complex input-output analyses (Gevins and Morgan 1988), and there is no reason to believe that the back-propagation network is the best or most suitable for EEG pattern recognition. Our reason for choosing the back-propagation method was that its mathematical background is well understood and sufficient experience is available regarding its efficiency in various pattern recognition/completion problems (Sejnowski et al. 1988; Lytton and Sejnowski 1992). Our experiments with various number of input and hidden layer cells indicates the need for an extensive testing of the back-propagation network before it can be used efficiently for data analysis. An important finding of the present experiments is that increasing the number of the input ceils or the hidden layer cells actually lead to the deterioration of the network. Only after extensive testing of the various input cell hidden layer cell combinations could we determine the most effective network configuration. When equally good configurations were found as indicated by the minima of summed squared error (cf., Fig. 4), combinations with the least number of cells were chosen to increase the speed of analysis. With the FFT-transformed EEG data, the best configuration for the most efficient recognition of HVS episodes consisted of 16 input ceils, 19 hidden cells and 1 output cell. Other EEG patterns may require substantially different combinations. A consistent finding was that using the FFT

transformation of the EEG as the input, significantly shorter training was required to attain asymptote performance as compared to the analysis of the raw EEG. Determination of the "best" cell configuration of the network and asymptote training are computationally expensive and require computers with tens of MFlop speed or preferably massively parallel computers. However, it must be emphasized that once the optimal weights among the interconnected cells had been obtained, the chosen configuration can easily be implemented in low-priced computers for efficient analysis of similar data in many EEG laboratories. An important issue in the evaluation of any new method is its reliability. Using the lowest threshold value (0.85) of the network output, none of the manually detected HVS events were missed by the network. This was true with both the original and novel data sets. At this threshold value the incidence of false positive events was relatively high and almost every second detected event was false. These spuriously detected events were substantially reduced by increasing the threshold. However, increasing the threshold of the network output also increased the incidence of the missed events. Most of the falsely detected events were due to extremely large amplitude artifacts that saturated the amplifier and to rhythmic scratching and grooming artifacts in the frequency range of the HVS. The percentage of falsely recognized events can be further decreased by retraining the network with a data set in which artifacts are explicitely noted. The network configurations in this study were trained with relatively few explicitely non-HVS epochs (twice the number of HVS). It must be emphasized that the total duration of real (HVS) and spuriously detected events represented only 0.1-1% of the total 12 h EEG record and our major goal was to positively identify all HVS events. Thus, instead of a 12 h record the experimenter had to screen and edit only a very small portion of the already

108

G. JANDO ET AL.

presorted EEG events and without missing a single HVS epoch. During the past year, we have extensively examined the performance of the back-propagation network by screening more than 3000 h of EEG records obtained from various rat strains. Visual analysis of such a large data base would have been prohibitively time consuming. Another obvious way of increasing performance is to simultaneously analyze several channels. The topographic information inherent in multichannel data would predictively enhance the detection capability of the network because artifacts and HVS have different topographical relationship. Neural networks in EEG diagnosis The advantages and possible disadvantages of neural networks in aiding the clinical diagnosis have yet to be explored (Gevins and Morgan 1988; Mamelak et al. 1991; Vaselis et al. 1991). Nevertheless, a few candidate applications and future directions can be outlined. For example, neural nets may be extremely useful in differential diagnosis. First, massively parallel networks could be trained with large data sets derived from patients with clearcut, but clinically different diseases. The weights obtained thereby can then be distributed for use in any laboratory with affordable computers. Since usually only l - 5 % of the EEG record is of interest for the clinical diagnosis, neural networks can become useful for on-line monitoring of pathological events. Automatic detection methods would then result in considerable data reduction. Thus, a 24 h monitoring session can be reduced several-fold and the detected events can be stored by at low cost. Furthermore, technicians can easily be trained for the manual selection of the already detected events, whereas recognition of abnormal patterns in the background of ongoing EEG requires substantial experience. Training of back-propagation networks require explicit and a priori judgments of the human observer and therefore the network is useful only for the automatization of the mental process. More recent non-linear methods, however, are capable of clustering the input information with hitherto unsuspected precision and reliability (Hopfield and Tank 1986; Farmer and Sidorowich 1988; Casdagli 1989; Doutriaux and Zipser 1991). For these analyses multichannel recordings will substantially increase the power of the analysis. Multichannel data can be analyzed in a truly parallel fashion and the network can provide evaluations corresponding to identifiable physiological states of the brain, such as sleep/awake stages and seizure episodes.
We thank Drs. P. Wong, A.S. Gevins and O. Gl6ckler for comments and discussions. This research was supported by N I N D S / N I H (NS-27058, 28121, 02383), the Whitehall Foundation and ADRDA.

References
Buzs~iki, G., Bickford, R.G., Armstrong, D.M., Ponomareff, G., Chen, K.S., Ruiz, R., Thai, L.J. and Gage, F.H. Electrical activity in the neocortex of freely moving young and aged rats. Neuroscience, 1988, 26: 735-744. Buzs~iki, G., Smith, A., Berger, S., Fisher, L.J. and Gage, F.H. Petit mal epilepsy and parkinsonian tremor: hypothesis of a common pacemaker. Neuroscience, 1990, 36: 1-14. Carpi, D., Jand6, G., Pierre, E., Vadi, D., Fleischer, A., Lajtha, A., Vad~.sz, C. and Buzs~ki, G. Inheritance of neocortical high voltage spindle spike-and-wave (HVS) patterns in rats. Soc. Neurosci. Abst., 1992, 18: in press. Casdagli, M. Nonlinear prediction of chaotic time series. Physica D, 1989, 35: 335-330. Doutriaux, A. and Zipser, D. Unsupervised discovery of speech segments using recurrent networks. In: D.S. Touretzky, J.L. Elman, T.J. Sejnowski and G.E. Hinton (Eds.), Connectionist Models: Proc. 1990 Summer School. Morgan Caufmann, San Mateo, 1991: 303-309. Farmer, D.J. and Sidorowich, J.J. Exploiting chaos to predict the future and reduce noise. In: Y.C. Lee (Ed.), Evolution, Learning and Cognition. World Scientific, Singapore, 1988. Gevins, A.S. Pattern recognition of brain electrical potentials. IEEE Trans. Patt. Anal. Mach.lntell., 1980, 2: 383-404. Gevins, A.S. and Morgan, N.H. Applications of neural-network (NN) signal processing in brain reseach. Proc. IEEE Acoust. Speech Signal Proc., 1988, 36: 1152-1161. Gevins, A.S. and R6mond, A. (Eds.). Methods of Analysis of Brain Electrical and Magnetic Signals. EEG Handbook, Vol. 1. Elsevier, Amsterdam, 1987. Gloor, P. and Fariello, R.G. Generalized epilepsy: some of its cellular mechanisms differ from those of focal epilepsy. Trends Neurosci., 1988, 11: 63-68. Gotman, J. Computer analysis of EEG in epilepsy. In: F.H. Lopes da Silva, W. Storm van Leeuwen and A. R6mond (Eds.), Handbook of Electroencephalography and Clinical Neurophysiology, Vol. 2. Elsevier, Amsterdam, 1986: 171-204. Gotman, J. Automatic seizure detection: improvements and evaluations. Electroenceph. clin. Neurophysiol., 1990, 76: 317-324. Hopfield, J. and Tank, D. Computing with neural circuits: a model. Science, 1986, 233: 625-633. John, E.R., Prichep, L., Fridman, J. and Easton, P. Neurometrics: computer-assisted differential diagnosis of brain dysfunctions. Science, 1988, 239: 162-169. Ktonas, P.Y. Automated spike and sharp wave (SSW) detection. In: A.S. Gevins and A. R6mond (Eds.), Handbook of Electroencephalography and Clinical Neurophysiology, Vol. 1. Elsevier, Amsterdam, 1987: 211-241. Lopes da Silva, F.H. and Mars, N.J.I. Parametric methods in EEG analysis. In: A.S. Gevins and A. R6mond (Eds.), Handbook of Electroencephalography and Clinical Neurophysiology, Vol. 1. Elsevier, Amsterdam, 1987. Lytton, W.W. and Sejnowski, T.F. Computational neuroscience. In: A.K. Asbury, G.M. McKhann and W.I. McDonald (Eds.), Diseases of the Nervous System: Clinical Neurobiology. Saunders, Orlando, FL, 1992. Mamelak, A.N., Quattrochi, J.J. and Hobson, J.A. Automated staging of sleep in cats using neural networks. Electroenceph. clin. Neurophysiol., 1991, 79: 52-61. Niedermeyer, E. and Lopes da Silva, F.H. Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Urban and Schwarzenberg, Baltimore, MD, 1987. Press, W.H., Flannery, B.P. Teukolsky, S.A. and Vetterling, W.T. Numerical Recipes of the Art of Scientific Computing. Cambridge University Press, Cambridge, 1988.

PATTERN RECOGNITION OF EEG BY NEURAL NETWORKS Rummelhart, D.E. and McClelland, J.L. Parallel Distributed Processing: Explorations in the Microstructure of Cognition. MIT Press, Cambridge, MA, 1986. Sejnowski, T.J., Koch, C. and Churchland, P. Computational neuroscience. Science, 1988, 241: 1299-1306.

109 Vaselis, R.A., Reinsel, R., Sommed, S. and Carlon, G. Use of neural network analysis to classify electroencephalograpbic patterns against depth of midazolam sedation in intensive care unit patients. J. Clin. Monit., 1991, 7: 259-267.