LETTERS International Journal of Recent Trends in Engineering, Vol 2, No.

2, November 2009

Dissimilarity Index in the Study of Patterns in EEG Using Fractal Dimension

Kiran.V, lecturer, kv_bee@yahoo.co.in
Department of Electronics and Communication Engineering , R.V. College of Engineering, Bangalore
Abstract Past two decades has seen a tremendous progress in the theory of nonlinear dynamics and this has given a new perspective to the nonlinear modeling of physiological systems like brain. We have used the notions of complexity and self-similarity in the electroencephalogram (EEG) signal to calculate fractal dimension which is used to detect the difference in structure or organization rather than amplitude or shape. Running fractal dimension curves have been obtained on several channels of EEG signals. The changes in the background activity due to eye closure have been clearly differentiated by the fractal method. The running fractal dimension curves are shown to be effective in detecting changing pattern in speech utterances also. The effects of different window lengths and shifts on the running fractal dimension have also been studied. The utility of fractal method for EEG data compression is highlighted.Next the use of dissimilarity index in the analysis of EEG signals has been explored. We have calculated dissimilarity index for the eyes open/eyes closed multichannel EEG data and shown how it can be used to quantify the dissimilarity between patterns of EEG. Key words: Electroencephalogram, Nonlinear dynamics, Fractal dimension, Self-similarity, Dissimilarity index, Changing EEG patterns, Data compression. 1 INTRODUCTION Progress in the theory of nonlinear dynamics in the past two decades has made it possible to think of modeling the brain as a continuous, spontaneously changing, nonlinear dynamical system [1]. Grassberger and Procacia [2] have further developed the tools to model experimental data as the output of a nonlinear dynamical system. The study involving the EEG falls in this category. These new concepts for the investigation of microscopic properties of brain activity offer a fresh way to provide new explanations. In fact, it has been reported [3] that the complex brain dynamics involve only a few degrees of freedom. Thus the dimensionality analysis provides a prime step for exploring the complexity of a nonlinear dynamical system. We have used the notions of complexity and selfsimilarity in the EEG signal, and the index that draws together these two notions and at the same time quantifies them is the fractal dimension. The fractal dimension, as presented by Mandelbrot [4], is an individual number fundamental to the quantification of most complex shapes. The pattern features quantified by this method do not depend upon the assumption that the EEG is generated by a periodic phenomenon. The fractal dimension of a pattern is also a measure of its complexity and is intimately related to the shape and dimension of the pattern. Fractal dimension measurements detect the difference in structure or organization rather than amplitude or shape. Therefore, this method may lend itself to a better understanding of patterns and their transitions in the EEG. Recently, fractal dimensions have been applied for feature detection in various waveforms. The fundamentals of fractal geometry and characterization of speech waveforms by fractal dimension measurements have been described by Pickover and Khorasani [5]. The application of fractal dimensions to the detection of transients in the EEG has been reported [6]. Many objects and patterns in the natural world possess the quality of self similarity-no matter what scale is used to view the pattern, the magnified portion of the shape looks (qualitatively) like the original pattern. Such objects include mountains and coastlines, as well as several classes of patterns derived purely from mathematics. This scale invariance has been

253

2009 ACADEMY PUBLISHER

LETTERS International Journal of Recent Trends in Engineering, Vol 2, No. 2, November 2009
studied extensively by Mandelbrot, who coined the word fractal to describe such irregular shapes. Mandelbrots famous and pioneering work with fractal geometry and his introduction of many new concepts like fractal dimension and self-similarity have affected most branches of science [4]. The attributes of complexity and self-similarity determine whether shapes are fractals. For example, suppose we magnify a segment of coastline, consisting of bays, capes and a peninsula, for its variability. In that magnified segment there will be yet another layer of complexity or variability, corresponding to coves and estuaries. If any part of that segment is amplified, then that too will exhibit complexity of similar intricacy. Self-similarity is a property of many natural sets as well as a property of all mathematically generated fractal sets. The Self-Dissimilarity Many of the systems characterized as complex the patterns exhibited on different scales differ markedly from one another [7]. For example, the biomass distribution in a human body looks very different depending on the scale at which one examines it. Conversely, the patterns at different scales in simple systems (e.g., gases, mountains, crystals) vary little from one scale to another. Accordingly, the degrees of selfdissimilarity between the patterns of a system at various scales constitute a complexity signature of that system [8]. Whatever distance measure is chosen, our quantification of selfdissimilarity can be measured for many kinds of real-world data. This allows comparisons of the complexity signatures of wholly different kinds of systems (e.g., systems involving information density in a digital computer vs. capital density in an economy, etc.). The measure that is quantifying the intuitive notion of the complexity of systems has a long history. To measure the complexity one can use Dissimilarity index (SD) in addition to the fractal dimension discussed earlier. If the spatiotemporal patterns exhibited on different scales differ markedly from one another then that system is called as complex or self-dissimilar. Conversely, for systems commonly characterized as simple the patterns are quite similar, those are called as self-similar. We discuss here what Dissimilarity index is and how it quantifies the dissimilarity between patterns. The human body is a familiar example of such self-dissimilarity; as one changes the scale of the spatiotemporal microscope with which one observes the body, the pattern one sees varies tremendously. The (out of equilibrium) terrestrial climate system is another excellent illustration, having very different dynamic processes operating at all spatiotemporal scales and typically being viewed as quite complex. Conversely, the patterns at different scales in simple systems such as gases and crystals do not vary significantly from one another. Similarly, once it has fossilized, a dead organism is static across time, i.e., completely self-similar along the time axis. In this paper, we have explored the use of fractal dimension in better understanding of patterns and their transitions in the EEG. The study has explored the possibility of using running fractal dimension as an effective method of presenting large-scale EEG data in a comprehensive single line plot. The study has also explored the role played by the length of the data window in the calculation of the fractal value. The fractal merits further investigation as a method of data compression, and may have advantages in certain applications. Next we have studied the use of dissimilarity index in the analysis of EEG signals. In a selfdissimilar system, the spatio-temporal patterns exhibited on different scales differ markedly from one another. We have calculated dissimilarity index for the eyes open/eyes closed multichannel EEG data and explored whether it can be used to quantify the dissimilarity between patterns. These techniques are found to be efficient methods of data compression, and they could potentially be used to compress EEG data, while complexity and transitions are maintained. Such a method may find possible applications in sleep monitoring and in monitoring stages of anesthesia during surgery. 2. FRACTAL DIMENSION AND DISSIMILARITY INDEX The index that draws together the notions of complexity and self-similarity and at the same time quantifies them is the fractal dimension. To grasp the notion of fractal dimension, let us consider a drawing of a snowflake under simple

254

2009 ACADEMY PUBLISHER

LETTERS ernational Journal of Recent Trends in Engineering, Vol 2, No. 2, November 2009
proceeding forwards monotonically, have natural magnification. As the picture is magnified, it becomes less complex; the features are blown up into the available viewing field. The ratio of complexity to space available has not changed under magnification. It is true that there is less detail, but there is also less space. Furthermore, the amount of detail present is less than the amount of detail at lower magnification. In contradistinction, fractal shapes do not obey this principle. As the space available decreases in a fractal, the complexity does not. The ratio of logarithm of space to logarithm of complexity is the fractal dimension. It should be noted that the variable representing space available would have a direct relation to the dimension of Euclidean space in which the fractal set is embedded [9]. For example, the variable for space will naturally have an exponent of one if the space is a line, two if the space is a plane and three if the space is a three-dimensional space. For fractals, this exponent of space term becomes implicated in the ratio of logarithms; the value will fall between the dimension of the Euclidean containing space of the fractal and the next lower dimension, the Euclidean space from which the fractal erupted. The containing space in our EEG example has a dimension 2, the dimension of a plane, and the simpler space has the dimension 1, the Euclidean dimension of a line. Thus we can expect the EEG fractal dimension to lie always between 1 and 2. The more that the EEG line fluctuates, the more of the plane it appears to cover, although it cannot cover the plane because the line from which it is made has no thickness. Nevertheless, our intuition tells us that rapid high-frequency fluctuations of the EEG waveform are somehow more complicated than straight lines. Indeed, the fractal dimension increases with greater fluctuations, provided that each fluctuation has self-similarity. In general, the fractal dimension, D, of a planar curve is given by: D=log(L)/log(d) 1) where L is the total length of the curve, and d is the diameter (the planar extent) of the curve [27]. For waveforms, which are ordered sets of (x,y) point pairs, the total length L is simply the sum of the distance between successive points, i.e. L=Sum(dist(i,i+l)) where dist (a,b) means the distance between two point pairs a = (x,y) and b = (u, v). Waveforms, starting points. Thus, the planar extent or the diameter of a waveform can be considered to be the farthest distance between the starting point (point 1) and any other point (point i) of the waveform, i.e. d=max(dist(1,i)) Fractal dimensions have the property that they cannot easily be compared when measured in arbitrary units. This is because fractal characterizations presume that each shape is fabricated as a collection of discrete units of similar size [4]. Thus the space must be discretized. One general convention for discretizing space and for normalizing fractal dimensions defines the standard unit to be the average step. For waveforms, this average step is s, the average distance between successive points. Using s, the formula for the fractal dimension, D, becomes: D=log(L/s)log(d/s), i.e., D=log(n)/{log(n)+log(d/L)} where n = L/s is the number of steps in the curve [10]. We have adopted this convention for estimating the fractal dimension of waveforms. 3.1 Running Fractal Dimension The technique of running fractal dimension has been used in this work. A window is applied to the EEG data in which the fractal dimension of the contents is calculated. The window is moved along the EEG. By sliding the window it is possible to generate a curve whose points not only correspond to the time axis of the time series, but reflect the fractal dimension at any given time. In this study, we present the continuous fractal dimension of speech and biomedical signals to demonstrate the utility of (fractal method for data compression. They also demonstrate its effectiveness in the analysis of changing patterns in these signals. 3.2 Dissimilarity Index The simplest version of such a signature is to reduce all of the patterns to a single number measuring their aggregate self-dissimilarity. This (2) would be analogous to conventional measures that quantify a systems complexity as a single number. We can use richer signatures however. (4) (3)

(5)

255

2009 ACADEMY PUBLISHER

LETTERS International Journal of Recent Trends in Engineering, Vol 2, No. 2, November 2009
waveforms. It clearly differentiates the signal in One is the symmetric matrix of the dissimilarity values between all pairs of patterns at different scales. More generally, say we have a dissimilarity measure that can be used to quantify how spread out a set of more than two patterns is. We define the dissimilarity index following expression: by the eyes open and eyes closed condition. The fractal values are more in eyes closed condition compared to the eyes open condition. It may also be observed that there are only 40 fractal dimension values as compared to the original

(6) If two nearest-neighboring points i and j belong to the same community, then the average distance dikfrom i to any another point k (k~=i, j) will be quite similar to the average distance djk from j to k, therefore the two perspectives ~based on i and j, respectively, will be quite similar. Consequently, dissimilarity index will be small if i and j belong to the same community and large if they belong to different communities [11]. 4. RESULTS AND DISCUSSION In what follows, we present the results obtained on the various studies made on the application of fractal dimension to speech and biomedical signals. In particular, we have shown the utility of running fractal dimension in analyzing changing patterns of waveforms. For studies on EEG, we have considered a multichannel record made on a normal subject under eyes open and eyes closed conditions . It has been observed that the front channels show more changes than the back channels during the transition stage of eyes open to eyes closed state. Fig.1 shows the EEG waveform for channel 1 (i.e. position Fp1). Fig. 2 shows the corresponding running fractal dimension. A window length of 100 samples is considered. Results are shown for both overlapping (with overlap of 50) and non-overlapping cases. It may be observed that the running fractal dimension curves are able to detect the differences in the structure of the EEG

Fig.1 EEG waveform for channel 1

Fig.2 Running fractal dimension curves for channel 1 4000 data values and hence a compression of 100 has been achieved. However, it should be noted that we couldnt obtain the original signal from the fractal dimension values.

256

2009 ACADEMY PUBLISHER

LETTERS International Journal of Recent Trends in Engineering, Vol 2, No. 2, November 2009

Fig. 3 shows the EEG signal from channel 16. Fig. 4 shows the corresponding running fractal dimension.

50 40 30 20 10 0 -10 -20 -30 -40 -50 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Fig. 4a Plot of EEG signal with eyes open and eyes closed

Fig. 3

EEG waveform for channel 1 (i.e. Fp1)

1.16 1.14 1.12 1.1 1.08 1.06 1.04 1.02 0 50 100 150 200 250 300 350 400

Fig. 4b Running Fractal Dimension with W=300 and S=10 1.14 Fig.4 Running fractal dimension curves for channel 16 It may be observed that there is not much change observed in the signal during transition stage and this is also clearly reflected in the fractal dimension curve. The effects of changing the window size and the overlap size have also been studied. Fig 4 shows the effect of changing the overlap size from S=10 to S=300 keeping the window size W=300 constant. 1.12 1.1 1.08 1.06 1.04 1.02 0 5 10 15 20 25 30 35 40

Fig. 4c Running Fractal Dimension with W=300 and S=100

257

2009 ACADEMY PUBLISHER