Examining the Epileptic Brain with Nonlinear Mathematics

Correlation Dimension of Brain Hemisphere Activity in Patient EEG

 

 

 

 

 

 

By

 

 

 

 

Miron Derchansky

 

 

 

MAT335

 

 

 

PART III

 

 

 

 

 

 

 

 

 

 

 

 

2002

 

Summary

 

Epilepsy is caused by repetitive and entrained seizure activity, which affects 0.8% of the human population. Clinical interventions include drug treatments, surgery and electrical stimulation. The underlying physiological arrangement of the neurons in the brain allows for nonlinear behavior, which results from complex circuitry. The application of nonlinear dynamics in the analysis of this organ has resulted in profound discoveries into the innate workings of this system. Nonlinear mathematics allows the extraction of crucial information and serves as a vital tool in epileptology, both in detecting and predicting a seizure. 

            Seizures can be experienced in the generalized form, which affect most of the brain, or in the focal form, which constrains the seizure to one locality. The complexity of neuronal activity can be modeled by obtaining the correlation dimension values of the system, which can be obtained from the EEG analysis. In patients who are suffering from focal epilepsy, the dimension of one hemisphere is greater than the other as such a hemisphere contains the driving focus, or foci, of the seizure. In primary generalized epilepsy, the correlation dimension is equivalent in both hemispheres. Furthermore, it has been shown that a dimensional drop in rats occurs before seizure onset, as the brain undergoes a transformation from its normal, chaotic activity, to the periodicity found in seizures.

            In this study, patient EEG was analyzed and the correlation dimension of both hemispheres was extracted. These values were compared both to each other and to surrogates, which are randomly generated data sets. During this analysis, a novel method for the automatic extraction of the correlation dimension was addressed and tested for precision. Although the patient was clinically diagnosed as having primary generalized epilepsy, it was observed that some differences in hemispheric dimension occurred, and that a large drop in dimension prior to seizure activity was present. This suggests that correlation dimension can be utilized as a real-time method for seizure detection and possesses predictability power.

 

Examining the Epileptic Brain with Nonlinear Mathematics

Correlation Dimension of Brain Hemisphere Activity in Patient EEG

 

1. Introduction

1.1 Epilepsy and nonlinear mathematics.   Epilepsy, a condition of chronic and recurring seizures, is a brain dysfunction that affects approximately 0.8% of humans. Of these, 20% of patients are resistant to drug treatment, which emphasizes the importance of anticipating seizure onset (Lehnetz et al., 1998) and implementing some form of clinical intervention (Jerger et al., 2001). Unlike other neurological diseases, such as Parkinson’s or Alzheimer’s, where the disease process is ever present, epileptic patients both appear and function normally most of the time. However, seizure onset, an intermittent pathologic condition, disturbs normality and illustrates the duality of this disease (Schwatzkroin, 1997).

With the application of the theory of nonlinear deterministic dynamics, new principles, as well as powerful algorithms were devised to analyze the behavior of the seizure state (Schiff, 1998). These principles, when applied on the time series domain, can yield measures of dimensions, Lyapunov exponents, entropies, or other fundamental information about complex brain dynamics (Elger et al., 2000). In rats, for example, it has been illustrated that a dimensional drop in brain activity occurs before seizure onset, and the decline in dimensionality is now thought to possess predictive powers (Babloyantz, 1986). The brain, with its intrinsic complex connectivity of neurons and their regulation makes neuroscience the ideal realm for chaos and nonlinear mathematics. Furthermore, the well investigated nonlinear behavior of individual neurons, and the expectation that neural networks will mimic this behavior, makes the application of nonlinear mathematics to epileptic models a seamless marriage.

 

1.2 Focal versus primary generalized epilepsy.  The disorder known as epilepsy can be classified into various categories, which may have such causes as perinatal injuries, metabolic abnormalities, tumors, brain lesions and infections. Seizures may occur as a generalized form, by affecting all or most of the brain, or as a partial form, affecting only a certain region of the brain (Getz et al., 2002). In patients with generalized epilepsy, both hemispheres are affected and share the same complexity in brain activity (Bullmore et al., 1994). Consequently, the treatment of this type of epilepsy consists mainly of drugs, which have various side effects varying from mild to sever. In patients with partial, or focal epilepsy, only one hemisphere of the brain is affected and the complexity of electrical activity in the hemispheres is more pronounced at the lobe where the focus is found (Widman et al.,1999). Oftentimes, such patients suffer intractable seizures, which cannot be ameliorated by drugs. These patients may be candidates for surgery, where the area generating the seizures is removed. 

The treatment of epilepsy depends greatly on its nature and it is vital for a physician to differentiate between generalized or focal epilepsy. The primary method for diagnosis of a patient  utilizes the electroencephalograph, which measures the electrical activity of the brain, and may tell where the seizure focus is located. New methodologies are emerging, such as Functional Magnetic Resonance Imagine, or fMRIs to detect the foci of seizures and to give the neurologist a clearer picture of cerebral activity. However, detection and epileptic diagnosis can be difficult at times, since the patient sometimes exhibits various symptoms of the disease, which makes the qualitative nature of EEG interpretation esoteric in nature.

 

1.3 Methods of capturing epilepsy in the laboratory.    The two major methodologies for studying epilepsy in a lab setting are the in vitro study of rat hippocampal brain slices, and the study of human data, obtained from an electroencephalograph recording, or EEG. Both pathways provide information about the system that can be analyzed using nonlinear mathematics. In this thesis project, the human data was utilized to extract information for analysis.

The human data was obtained via scalp EEG from a patient that was diagnosed with primary generalized epilepsy that could not be treated pharmacologically. Hence, as the seizure commences, both sides of the brain fire simultaneously and are active. Although the patient’s EEG data, evaluated by a neurologist, showed generalized epilepsy, she exhibited signs of focal epilepsy. Such symptoms included the rotation of her body to one side during seizure. Hence, the technique of correlation dimension, a nonlinear technique, was employed to investigate the EEG traces. The purpose of the project was to analyze the patient EEG and extract a dimensional value from each of the brain hemispheres in time, and draw a comparison between the two. Therefore, if the patient did indeed have full brain epilepsy, then the dimensions of the two lobes should be approximately equal (Drake et al., 1995).

 

1.4 Summary. Over the past decade, nonlinear dynamics has been increasingly applied to the problem of epilepsy. The major classes of epilepsy include temporal and focal epilepsy, and in the laboratory, human data can be obtained for analysis via EEG. The diagnosis of the seizure type is crucial for the proper treatment of the patient, as the effectiveness of surgery is greater than that of drugs, especially in intractable epilepsy.

This thesis examines these issues and will address the following hypotheses:

1.     If the patient’s pathology supports focal epilepsy, there will be a difference in dimension between both brain hemispheres.

2.     There will be a sizeable drop in dimension before the patient begins going into a seizure state.

 

2. Background on the general methods. The techniques required for data acquisition and its consequent analysis are a mixture of electrophysiological and mathematical approaches. The physics behind the recording techniques is crucial to the understanding of the results, as is the mathematics behind the data analysis techniques. It is important to understand the tools utilized in the experiment, from obtaining the patient EEG to the correlation dimension algorithm.

 

2.1  Electroencephalograph on human subjects.  The technique of EEG has been widely utilized

to measure brain electrical activity. In this technique, small, noninvasive electrode are placed on the patient’s head and voltage traces ranging from 5 to 500 microvolts are captured (Lehnertz, 1999). These are fed into an amplifier, which amplifies the voltages and produces either a digital trace (digital EEG, or dEEG), or a polygraphic strip chart to be read and interpreted by a neurologist. Figure 1 shows an EEG system with 19 electrodes implanted on the patient, each recording the distinct activity of a certain brain region. The EEG works in such a manner that it reads voltage differences on the head, relative to a given point. Hence, if the electrical activity between two hemispheres is to be ascertained, then one will need to place 3 electrodes, one on each hemisphere, and another in the center, connected to both electrodes. This will give an absolute difference between hemispheric brain activity.

The physics behind the EEG employs various fundamental concepts of electricity and magnetism. To find the voltage on a specific point on the head (f), a radius ‘r’ from the source, one must:

·       Create a primary current, denoted as Jp, which will produce an electric field leading to the secondary current, Js.

·       Solve Ñ(sÑf) = Ñ(Jp), where s represents the conductivity of the scalp. This is known as the Poisson Equation for the electrical potential.

·       For the Poisson equation to hold true, there must be a boundary conditions stating that there is no potential outside the scalp, or mathematically, f/n = 0.

 

2.2  Inter-Peak-Interval analysis. From the EEG recording, the inter-peak-intervals, or IPIs, are

calculated and are utilized to extract valuable information from the system. Figure 2 shows a clearer diagram of how the technique of IPI extraction is performed. This algorithm creates data based on the peaks of the traces and the distances between consecutive peaks detected. It is this distance, or the interval between the peaks, which is recorded and analyzed by the correlation dimension algorithm. The peak detection algorithm is one that chooses the peaks based on width and amplitude parameters of the system being recorded. IPI analysis was performed rather than utilizing the original voltage traces since this calculation reduces the number of the points analyzed by the correlation dimension algorithm, and has been proved to contain as much information as the voltage. Hence, there is no loss of information, and the file sizes are smaller, which increase the efficacy and speed of the correlation dimension calculation.

 

2.3  Phase space and the dimension of a system. To understand the mathematics behind the

method of correlation dimension, it is vital to comprehend the abstract notion of phase space, since the correlation dimension algorithm is performed in this space.

 

2.31 Phase space.  Differential equations are often utilized to model and describe the behavior of ordinary systems. However, when dealing with the complex system represented by the brain, there are yet to be found equations that succinctly characterize the organ. One way to comprehend phase space is to imagine a multidimensional space in which a moving point constructs a curved line. Hence, the location of the point in any moment contains complete information about its state, as determined by the various axes representing the system variables (Gallez et al., 1991). The phase space defines all the states of the variables of the system, and as the points are connected in time, the evolution of the system can be seen in the production of a unique curve. In this study, the variables depicted on the axes are the IPIs of the system.

An illustrative example of the utility of phase space and its analytic and predictive power may be found in the simple pendulum. A pendulum constrained in the plane from p to -p may oscillated ad infinitum, unless constrained by friction. When friction is applied to the system, the pendulum will stop, or be attracted to the position on the plane with coordinate p/2. In two dimensions (Figure 3A), an elliptical orbit may be seen, with the point p/2 as the attractor. When time is added as a dimension (Figure 3B), the phase space becomes three dimensional and the trajectory of the system is a paraboloid.

The phase space of the EEG data was computed using the IPI data obtained from the raw voltage trace. This three dimensional phase space (Figures 3C and D) can be analyzed graphically and the dimension of the space may be calculated based on the dimensional algorithm, which reconstructs the phase space.

 

 2.32 Correlation dimension algorithm. This algorithm is performed in phase space and employs the box - counting technique (Elbert et al., 1994). The correlation dimension is defined mathematically as:

n = lim             [ln(C) / ln(r)]

             (r ® ¥) 

 

Here, r is the radius of the box and C is the number of points that land within the box. Figure 4A illustrates a system whose correlation dimension is calculated from its phase space. When the number of points that land within a certain radius, r, of the box are plotted against the radius, on a log-log plot (natural units), the slope of the linear region of the curve produces the correlation dimension (Figure 4B).  The correlation dimension can be thought of as the 3D fractal dimension, but performed in phase space, and not in the 2D plane.

The correlation dimension algorithm expands the system into its phase space in m dimensions. From each of these dimensions, the correlation dimension is calculated by finding the linear region on the log-log curve and taking its respective slope (Babloyantz et al., 1996). Hence, m distinct slopes will be found. It is vital to comprehend that the dimension represented by m is not the correlation dimension, but rather the phase space extracted dimension. The true dimensional value of the system may be found where the plot of m versus the slope of the mth dimension plateaus, as illustrated in Figure 4D. This is done due to the fact that an optimal dimension of the phase space reconstruction of the brain (m) is not yet known and a range of dimensional values must be analyzed.

 

2.4  The linearity approximation. When the correlation dimension calculation is performed on a

system,  the experimenter tries to isolate the linear region of the log-log curve. However, when the system being analyzed produces a large magnitude of files, it is useful to automate this procedure. As illustrated in Figure 4C, the linearity approximation (LA) automatically isolates the linear region of the curve based on a graphical technique making use of the maximal y value (y_max) and the slope of the nonlinear curve. As the computer moves from point to point on the log-log curve, a straight line is constructed, based on the slope of the curve. The LA calculation will be performed until the y_max value of the line exceeds that of the curve. At this point, the linear region is isolated and its slope is calculated. Such a novel approximation, once coded, allows thousands of files to be processed in mere minutes, with a relative high level of precision.

 

2.5  Surrogate time series. Surrogate data is generated by randomizing the real data,

according to a  randomization algorithm. When this algorithm is applied, randomized data is obtained with the same frequency spectrum as the original data. When the surrogate data is compared to the real data, its purpose is to illustrate that random, gaussian processes could not have generated the observations (Elger et al., 1998). The magnitude by which an observation deviates from a random process is given as:

s* = |Cs - Xd| / STDs

Where Cs is the surrogate mean and Xd is the mean of the obtained data. STDs is the standard deviation of the surrogates. It has been established in the literature that if s* is greater than 2, then there is substantial drift in the real data from randomness and that the results are statistically significant. The surrogated data can also be utilized as a comparative trace, as two mutually exclusive traces can be compared to the same random noise, or surrogate. Hence, by the comparison of the two traces to a common trace, the two traces can be relatively compared.

 

3. Methods

3.1 EEG recordings. Twenty hours of extracranial EEG recordings were obtained from a patient that was diagnosed with temporal lobe epilepsy. The patient was in subclinical status epilepticus for 3 hours, where there was no visible sign of seizure, yet the EEG showed clear seizure activity. Four observable seizures occurred, each approximately 2 minutes in duration. Figure 5.0 illustrates the position of the electrodes of interest on the head. As both hemispheres were compared, the channels f3-pz and f4-pz were analyzed. All EEG data was digitized at 200Hz, and notch filtered to eradicate any 60Hz noise. 

 

3.2 Window size and baseline. As suggested by the literature, a sliding window of 40000 points, or 200s was shifted along the EEG with an 80% overlap (Babloyantz et al., 1996). The baseline of the curves was then removed using the ESAF software provided by Khosravani. This was performed and coded in Visual Basic for both channels of interest. Taking a window size is necessary for the purpose of capturing the evolution of the system, and hence the large overlap. The brain is a dynamical system and changes very rapidly and in lieu of this fact, comparing the first 200s to the consequent 200s will be extremely audacious, since overlap is needed to capture neuronal changes.

 

3.3 IPI generation. Using a peak detection algorithm, IPIs were acquired from the segmented files. The peak detection criterion was a function of both the amplitude and the width of the signal trace. The standard deviation of the voltage trace was computed (s) for each window size and an optimal value of 2s was obtained. The amplitudes of such a trace were compared to this value both below and above the baseline and peaks surpassing this value were taken as statistically significant. These calculations were done in ESAF, an analysis software developed by Khosravani.

 

3.4 Correlation dimension algorithm. The source code for the algorithm was obtained from the nonlinear analysis TISEAN package developed by Kantz and Schreiber. A batch file with all the IPI file names was created and read by the correlation dimension code, which generated 4 files per IPI file. The file of interest contained a list of radii and point numbers that landed within the given radii.

 

3.5 Testing the linearity approximation. Before the LA was applied in analyzing the log-log plots, this approximation was tested to see how well it estimated the dimensions of the known  Rossler attractor (Grassberger, et al., 1983). The Rossler attractor was created to explain chemical kinetics and contains 3 coupled, nonlinear differential equations. These equations can be solved analytically and do not require numerical solutions.  It is important to note that this system does not result in a limit cycle and is an illustrative example for a system exhibiting deterministic chaos. To test the LA, the system was modeled using the following differential equations:

x¢ = -(y + z)

y¢ = x + ay

z¢ = b +xz - cz, where a, b, c are defined constants.

 

Maple code was generated to solve these equations analytically and their trajectories were analyzed by the linearity approximation and compared to the data in the literature (Ashkenazy, 1999).

 

3.6 Computing the system’s dimension. The log-log plots were plotted using ESAF, a visual basic program coded by Khosravani. The LA was incorporated into the body of the code and an automated technique for the analysis of the dimension of thousands of files was created.

3.7 Surrogate data testing. After utilizing the TISEAN package to generate surrogate data, the same analysis was performed on the surrogate data as on the IPI files. For each segmented file, 19 surrogates were generated, as suggested by the literature (Elger et al., 1998). The s* was then computed to verify that the data generated from the correlation dimension of the system was different than random noise.

 

3.8 File handling and computational efficacy. To date, correlation dimension algorithms have never been performed on such prolific amounts of data, as exemplified by 20 hours of recorded EEG, sampled at 200Hz (approximately 1.4 billion points, excluding 80% overlap). At the completion of the analysis, 57000 files were created with the aid of Visual Basic, Origin Pro and Maple code. Due to the large file-handling aspect of the project, a large portion of the methodology was spent on automation and computational efficacy of the code. This included array and data-storage architecture as well as code debugging. Figure 6.0 has a diagrammatic representation of the experimental methodology, including the compilers generated for the code and the consequent number of files.

 

4. Results

4.1 Linearity Approximation. The Rossler attractor was generated in Maple using the aforementioned differential equations as can be visualized by Figure 7. This system was generated using 10000 points and the Linearity Approximation was applied to acquire the correlation dimension. This value was compared with the known values of the system (Table 1), which were obtained by the standard GCD method for capturing the correlation dimension of an attractor (Ashkenazy, 1999).  The error generated with the LA was 6%, and the approximation was utilized to generate the dimensional values of the system.

 

4.2 Correlation dimension values. Figure 8.0 shows the correlation dimension values of the left and right brain hemispheres. In the normal state, the patient exhibited an average dimension of 1 for both hemispheres. A large drop in the system dimension can be seen 2 hours before status epilepticus, which commenced at 22:00 and lasted until 03:00. Clinical seizure occurred at 23:53, 00:44, 02:00 and 02:30 hours. During the four seizures, the dimension deviated from the mean dimensional value of 1, as lower values were observed.

 

Figure 9.0 shows the superposition of the two traces, as differences in the dimension are observed.  Although the general geometry of the dimension curve is sustained in both hemispheres, closer examination of the trace shows noticeable differences. Three distinct time intervals were selected to illustrate such dimensional differences: pre-seizure, beginning of status epilepticus, and end of status epilepticus. The differences fluctuate mainly before and at the beginning of seizure activity. During seizure, there seems to be similar dimensional values present in both hemispheres, as the difference is minimal.  Dimensional drops can also be observed before each of the 4 seizures occurred. The decay of such drops was faster than the drop that occurred before status epileptics at 20:00, while maintaining a comparitive magnitude, especially for the 3rd seizure, at 26:00.

Time windows of 10s were obtained in the voltage domain and plotted in correlation with four periods in the dimension trace. Figure 10 illustrates the complexity of the raw voltage traces as they correspond to the dimension. During seizure activity (Figure 10D), the dimension takes on low values and the voltage trace shows a less complex waveform. At the peak before the large dimensional drop, (20:00), the dimensional value is at 1.5, and the waveform of the voltage trace (Figure 10B) can be seen as more complex, or noisy, than other compartive traces of lower dimensional value.

 

4.3 Surrogate data. With 19 surrogate files generated, a tester surrogate file was obtained by averaging these files. Both hemispheres were compared to the averaged surrogate file and the deviation from randomness (s*) was plotted. Figure 11 illustrates the hemispherical deviation from random noise. The average deviation for the left hemisphere, was observed at 2.5s*, and the right hemisphere had a deviation of 3s*. In both hemispheres, the largest deviation from noise occurred at the interval corresponding to the dimensional drop, at 20:00 to 22:00 respectfully. The majority of the traces was found at above the 2s* threshold value.

5. Discussion

The correlation dimension of the left and right hemispheres, when analyzed with the Linearity Approximation, was shown to be different and deviating from random noise artifacts. There was a corresponding large drop in dimensional values in both the analyzed hemispheres, approximately 2 hours before seizure. The Linearity Approximation that aided in selecting the linear region of the log-log plots had an error of 6%, a realism that was mitigated by its utility on both brain hemispheres, which presented a constant and steady error. Hence, the values of the system dimension are relative and are not absolute numbers. The voltage waveforms of the high dimensionality epochs showed a greater complexity than those of low dimensionality. Seizure periods retained a relatively low complexity voltage signal, corresponding to a low dimensional value.

Before the start of a seizure, there is a transitional period, know as the inter-ictal period (Drake, 1998; Fisher, 1992), which provides the bridge between normality and seizure. In this period, there is a transformation from normal, chaotic brain activity to periodicity, which is characteristic of seizures. Hence, the voltage signal shows the emergence of a more organized pattern in the brain before seizure onset (Figure 10C). During periods of seizures, the neuronal activity of the brain becomes synchronized and a large number of neurons fire periodically. From this pathology, an observed low-dimensional signal emerges, corresponding to brain synchrony. Hence, the dimensional drop in both brain hemisphere occurs due to either transition into a seizure, or during a seizure and is a function of the physiology of the disease.

As the dimensional traces from both hemispheres are compared, differences in hemispherical activity are observed. During the seizure period, the difference in the patient’s hemispheres was minimal and this confirms the diagnosis that the patient suffers from primary generalized epilepsy. Hence, the correlation dimension data as well as the EEG trace can systematically rule out the theory that there is a rhythm generator in one of the patient’s hemispheres. However, subtle dimensional differences can be observed in the correlation dimension calculation which qualitative EEG analysis would not pronounce. It is here suggested that for borderline patients, whose clinical diagnosis is challenging to make, the hemispherical dimensional values can be utilized as aids to facilitate the diagnostic process.  

            When the results were compared to random noise, it was shown that there was a statistically significant deviation. Hence, the data was not stochastic and bears physical significance. Amplifying the fact that the hemispherical differences can be observed is the fact that the deviation signals (Figure 11) from both hemispheres are different from each other when compared to a common trace. Furthermore, the most statistically significant period of the dimensional traces, where the deviation from noise was the greatest, occurred in the dimensional drop. This further illustrates the statistical significance of the inter-ictal period where the brain is on a trajectory to synchrony.

            The investigation of the correlation dimension of a system can aid in determining the activity of the system. Probing the brain with this nonlinear technique illustrated a strong correspondence with neurophysiological theory. The conceptual information gained from this investigation support the idea of seizure detection, via a drop in dimension before seizure. Such is a vital application of this nonlinear tool, since in the realm of epileptology, detection is the first step to the prevention of seizure. Current work on prevention includes electrical stimulation of the brain at a crucial transition period into epileptiform activity. Future work involves application of the correlation dimension algorithm in real time both in vivo and in vitro studies (Gotman, 1982), as the correlation dimension calculation provides a crude method of seizure detection and possess some predictability power.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

References

 

Ashkenazy Y. The use of generalized information dimension in measuring fractal dimension in      time series. Physica 1999;271:427-447.

 

Babloyantz A, Destexhe A. Low-dimensional chaos in an instance of epilepsy. Proc Natl Acad          Sci U S A 1986;83:3153-7

 

Babloyantz A, Lourenco C. Brain chaos and computation. Int J Neural Syst 1996;7(4):461-71

 

Bullmore, E.T, Brammer, Bourlon, Alarcon, Polkey, Elwes, Binnie, C.D. Fractal analysis of

electroencephalographic signals intracerebrally recorded during 35 epileptic seizures: evaluation of a new method for synoptic visualisation of ictal events. Electroencephalography and Clinical Neurophysiology. 1994;91:337-345

 

Drake Jr., M.E. EEG correlation dimension in partial and generalized epilepsy.

Electroencephalography and Clinical Neurophysiology 1995;9:12

 

Drake ME, Padamadan H, Newell SA. Interictal quantitative EEG in epilepsy. Seizure 1998;7:39-        42.

 

Elbert-T; Ray-WJ; Kowalik-ZJ; Skinner-JE; Graf-KE; Birbaumer-N. Chaos and physiology:

deterministic chaos in excitable cell assemblies. Physiol-Rev. 1994; 74: 1-47

 

Elger C, Lehnertz K. Seizure prediction  by non-linear time series analysis of brain electrical           activity. Europ. J of Neuroscience 1998; 10:786-9.

 

Elger CE, Widman G, Andrzejak R, Arnhold J, David P, Lehnertz K. Nonlinear EEG analysis

and its potential role in epileptology. Epilepsia 2000;41:S34-8

 

Fisher RS, Webber WR, Lesser RP, Arroyo S, Uematsu S. High-frequency EEG activity at the    start of seizures. J Clin Neurophysiol 1992;9:441-8.

 

Gallez D, Babloyantz A. Biol Cybern. Predictability of human EEG: a dynamical approach.

1991;64(5):381-91.

 

Getz K, Hermann B, Seidenberg M, Bell B, Dow C, Jones J, Woodard A, Rutecki P, Sheth R,

O'Leary D, Magnotta V. Negative symptoms in temporal lobe epilepsy. Am J Psychiatry 2002;159:644-51

 

Gotman J. Automatic recognition of epileptic seizures in the EEG. Electroencephalogr clin             Neurophysiol 1982;54:530-40.

 

Grassberger P, Procaccia I. Characterization of strange attractors. Phys Rev Letters 1983;50:346-     9.

 

Jerger KK, Netoff TI, Francis JT, et al. Early seizure detection. Clin Neurophysiol 2001; 18:259-    268.

 

Lehnertz K. Non-linear time series analysis of intracranial EEG recordings in patients with

epilepsy--an overview. Int J Psychophysiol 1999;34(1):45-52

 

Lehnetz K, Elger C. Can epileptic seizures be predicted? Evidence from nonlinear time series        analysis of brain electrical activity. Phys Rev Lett 1998;80:5019-26.

 

Schiff SJ. Forecasting brain storms [news; commnet]. Nat Med 1998;4:1117-8.

 

Schwatzkroin. Origins of the Epileptic State. Epilepsia 1997;38:853-858

 

Widman, G, Bingmann, Lehnertz, Elger,  Reduced signal complexity of intracellular recordings:

a precursor for epileptiform activity? Brain Research 1999; 836:156-163.