Вы находитесь на странице: 1из 4

Final Paper Page 1 Nonlinear Dynamic Analysis of ECoG for seizure prediction

Vivek Menon ESE 412

Epilepsy is a neurological disorder that is characterized by the recurrence of spontaneous seizures. Physiologically, seizures can be generally defined as sudden disruptions in the electrical activity of the brain that are usually synchronous and/or excessive and can also cause an alteration in mental state of the epileptic. These episodes of total loss of control can be dangerous for the epileptic and can result in serious injury, in addition to a loss of memory. Treatment for seizures rely on calming down the electrical activity of the brain; benzodiazepine (BZD) is a common medication for epilepsy patients since it enhances the effect of GABA, the primary inhibitory neurotransmitter in the central nervous system. Unfortunately, it can be difficult to provide someone undergoing a seizure the proper medication. As a result, pre-emptive action would provide the best treatment to an epileptic, which motivates the study of seizure prediction. Scientists who research epilepsy and seizures rely on the electroencephalogram (EEG) to learn more about the physiological nature of the seizure. EEGs record real time overall electrical activity in voltages of the brain using electrodes that are either placed on the surface of the scalp or directly on the surface of the brain. The latter method is known as intercranial-EEG (iEEG) or electrocortigography (ECoG). Scientists and engineers have been looking for patterns in EEG recordings of seizure activity to help predict oncoming seizure activity in the brain. One such proposed pattern stems from dynamic analysis of EEGs. Iasemidis et al. demonstrated the existence of chaotic dynamics in the brain.1 Mathematical analysis of seizures is done by continuous EEG recordings of the pre-ictal (before seizure), ictal (during seizure) and post-ictal brain. This provides a high-quality time-series of voltage readings. Typically, dynamical analysis is conducted when we can define the dynamics of a system at the interested states; however, time-series data does not directly provide us with this information. Because of this, many of the concepts and types of analysis learned in this course could not be applied to seizure prediction. Therefore, my goal for this final paper was to 1) learn how to characterize the dynamics of EEG data 2) learn how to measure chaoticity/complexity on time series data 3) verify that ictal-phase EEG is characterized by a loss of complexity. The remainder of this paper will focus on the motivation behind these goals and what I achieved. Casual observation of typical EEG* recordings will show that the EEG signal is chaotic and unpredictable in nature. Additionally, there is hardly any difference in the graphs generated during preictal and the ictal stage, though, upon literature, I learned that there is one notable difference between normal and (pre)-ictal EEG stages: as the seizure approaches, the waveforms start to take a more regular and repetitive shape.2 This is intuitively indicative of a loss of complexity or chaoticity in the EEG signal. As a result, engineers found reason to measure the chaoticity and complexity of the EEG signal before, during, and after the seizure.

a)

Figure 1: Normal EEG

b) Seizure EEG

*Calculations and graphs were generated in MATLAB using my own code. All data was acquired in .txt format from the 4th reference.

Final Paper Page 2

Vivek Menon ESE 412

One of my goals was to characterize the dynamics of EEG signals and determine whether or not there was a reason to believe that the dynamics were nonlinear / chaotic in nature. Takens demonstrated a method by which one could construct a phase space with similar topological properties as the actual underlying phase portrait with a time-series and converges to the same attractors a set of physical properties to which a system tends to evolve.3 This method is known as the Figure 2: a+b+c (from top to bottom). method of delays or embedding phase Graphical Representations of <x(t), x(t-2)> space construction. If is the data set of voltages from the EEG with respect to time, then the vector where is referred to as the embedding dimension. Graphing with embedding dimension 2 allows for a graphical (approximate) representation of the phase space of the system. Figure 1 shows my results* of phase space construction from EEG recordings4 of a) non-seizure b) seizure and c) a time series of a sinusoid: a non-chaotic periodic function. Figure 2c was included to demonstrate that the phase space of a known function can be constructed to look like what the actual portrait of the vector field would be (in the case of sin(t), a circle/oval shape). Clearly, the vector fields generated from the EEG data is nonlinear. The attractors generated from the EEG data (2a+b) are known as strange attractors. These attractors are chaotic in nature two orbits with nearby initial conditions diverge very quickly. After having demonstrated visually that EEG signals are chaotic in nature, my next goal was to be able to measure both the chaoticity and complexity of different timeseries data sets. One method of quantifying the complexity of the attractor in the phase space is to calculate the dimension of the attractor. The correlation integral is proportional to the total number of pairs of points that are less than a specified distance apart.6 When , . is referred to as the correlation dimension which essentially Figure 2: a+b+c (from top to bottom). measures the extent to which the presence of Correlation Dimension Estimation (from slope). a point X affects the other points in the phase space. When (degrees of freedom of the data set), then by definition there exists an attractor and gives its dimension.1 Figure 2 shows my attempt to calculate the correlation dimension of pre-ictal data, ictal data and the sine wave. My computer was not able to process calculations beyond an embedding dimension of 1, so these graphs provide a very crude estimate of the correlation dimensions [ ] of the attractors from Figure 1. is estimated by finding the slope of . As is apparent from the graph, the sinusoid has the lowest slope whereas the ictal data has the highest slope. The most common measure of the chaoticity of the graph the divergence of two trajectories that with very close initial conditions is the largest Lyapunov exponent (L). The Lyapunov exponent measures the average rate of exponential divergence between two adjacent trajectories in the attractor.

*Calculations and graphs were generated in MATLAB using my own code. All data was acquired in .txt format from the 4th reference.

Final Paper Page 3

Vivek Menon ESE 412

If L>0, the system creates information and is chaotic, whereas if L<0 the system destroys information and is predictable in nature. I was not able to create an accurate program by which the maximum Lyapunov exponent can be calculated because of 1) a lack of knowledge of computing / MATLAB and 2) the inability of my computer to handle an embedding dimension of higher than 1 (or 2 in small intervals). As a result, I shall detail the mathematical formulation and calculation of L and the approach for calculating L for a time series. Exponential divergence from one state to another can be expressed at where is the Lyapunov exponent. This relationship shows that a positive corresponds to divergence, implying chaos. In order to test this measure against a non-chaotic waveform, I will use the time series . is the initial difference between two points on the vector field, in this case, and . is then the difference between the two points on the vector field for each . In the case of , the vector field associated with this system is a circle and over time, two corresponding points would travel around the circle and remain a set distance from each other. Thus, and , therefore as expected from a periodic non-chaotic function. For a time series, the formula calculating the maximum lyapunov exponent resembles the earlier formula for a time series is rewritten as
| |

( .

). The classic lyapunov exponent calculation is the distance between two close points in the

attractor and is the difference between these two points as time evolves. In this case, is the number of samples for which the lyapunov exponent is being calculated. The embedding dimension for in order to properly estimate L is 7 as given by Kolmogorov-Smirnov test, since p=7 allows for the rejection of H0: that the vectors are too similar. When this algorithm is applied to EEG data, it has been shown that the Lyapunov exponent of the time series is positive and decreases significantly on the onset of a seizure at the ictal stage of the EEG. This significant drop in complexity and chaoticity of trajectories is seen universally across various scientific studies4,5. Measurements of the correlation dimension and the Lyapunov exponent can help to anticipate epileptic episodes up to 20 minutes before seizure onset.7 Despite the consistent loss of complexity and chaoticity during seizure, these measurements have their flaws. Reduction in correlation dimension and lyapunov exponents occur within normal EEG recordings as well this change is not always constant even upon seizure onset. Recent studies have shown that application of these methods produce a high number of false positives.7 Even so, analysis on the dynamics of EEG can be helpful in predicting seizures and can perhaps be used as one tool or one perspective by which the prediction process would take place. Perhaps, if I could redo this project or change my goals, I would learn more about linear dynamic analysis techniques for time series data and would try to find more robust statistics that can better predict seizures from EEG waveforms.

*Calculations and graphs were generated in MATLAB using my own code. All data was acquired in .txt format from the 4th reference.

Final Paper Page 4

Vivek Menon ESE 412

References 1. Iasemidis, L.D., Zaveri, H.P., Sackellares, J.C. and Williams, W.J. Modelling of ECoG in temporal lobe epilepsy. 25th Ann. Rocky Mountain Bioeng. Symposium, 1988(a), 24:187-193. 2. Lyapunov Exponent of EEG Signal for Epileptic Seizure Characterization - Bartosz Swiderski 3. Takens, F. Detecting strange attractors in turbulence. In: D.A. Rand and L.S. Young (Eds.), Dynamical Systems and Turbulence. Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1981:366-381. 4. Andrzejak RG, Lehnertz K, Rieke C, Mormann F, David P, Elger CE (2001) Indications of nonlinear deterministic and finite dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state, Phys. Rev. E, 64, 061907 a. Data: Exemplary recordings as defined by the authors from sets A and E. 5. Iasemidis, L.D., Zaveri, H.P., Sackellares, J.C. and Williams, W.J. Phase space analysis of EEG in temporal lobe epilepsy. IEEE Eng. in Medicine and Biology, 10th Ann. Int. Conf., 1988(b):12011203. 6. Hirsch M., Smale S., Devaney R.L., Differential Equations, Dynamical Systems, and an Introduction to Chaos 2012, 0123820103 7. H. Korn, P. Faure, Is there chaos in the brain? II. Experimental evidence and related model C. R. Biologies 326 (2003)

*Calculations and graphs were generated in MATLAB using my own code. All data was acquired in .txt format from the 4th reference.

Вам также может понравиться