Brain 1
Brain 1
Brain 1
6, NOVEMBER 2007
1005
I. I NTRODUCTION HE HUMAN brain is among the most complex systems that are known to mankind. Neuroscientists seek to understand brain functions through detailed analysis of neuronal excitability and synaptic transmission. However, the dynamical transitions to neurological dysfunctions of brain disorders are not well understood in current neuroscience research. Epilepsy is the second most common brain disorder after stroke yet the most devastating one. The most disabling aspect of epilepsy is the chronic condition of diverse etiologies with the common symptom of spontaneous recurrent seizures, which can be characterized by a chronic medical condition that is produced
by temporary changes in the electrical function of the brain. These electrical changes can be captured by electroencephalograms (EEGs), which is a common tool for evaluating the physiological state of the brain. While EEGs offer excellent spatial and temporal resolution to characterize rapidly changing electrical activity of brain activation, neuroscientists understand very little about the seizure development process from EEG data. The unpredictable occurrence of seizures has presented special difculties regarding the ability to investigate the factors by which the initiation of seizures occurs in humans. If seizures could be predicted, it will revolutionize neuroscience research and provide greater understanding of abnormal intermittent changes of neuronal cell networks that are driven by the seizure development. Recent advances in data mining and signal processing research for excavating hidden patterns or relationships in massive data (such as EEGs) offer a possibility to better understand brain functions (as well as other complex systems) from a system perspective, which will generally be very useful in medical diagnosis. Thus, the vital step to advance research in seizure prediction is to develop novel techniques that are capable of recognizing and capturing epileptic activity in EEGs before a seizure occurs. As an answer to this question, the discriminant ability to differentiate and classify the preseizure (abnormal) EEG signal is logically a prerequisite of the seizure prediction/ warning process. Thus far, none of the current epilepsy studies in the literature is undertaken to develop quantitative classication techniques that can be used to differentiate normal and preseizure EEG signals. The goal of this study is to develop and present a novel EEG classication technique as a theoretical foundation and host of methodologies to enhance the ability to differentiate normal and abnormal EEG signals. Such technique is based on the integration of the brain dynamics, k-nearest neighbor (KNN), and statistical time series analysis. A. Background: Epilepsy At least two million Americans and another 40 million people worldwide (1% of population) currently suffer from epilepsy, which is the second most common brain disorder after stroke [1], [2]. In about half of all cases, no cause can be found. In the other half, head injuries, brain tumors, lead poisoning, problems in brain development before birth, and certain genetic and infectious illnesses are the cause of epilepsy. The mainstay of contemporary treatment for epilepsy is pharmacological, which involves chronic administration of drugs, attempting to keep the concentration of those drugs within a therapeutic window. Antiepileptic drugs are commonly used to reduce
Manuscript received April 7, 2006; revised January 27, 2007. This work was supported in part by the National Science Foundation under CAREER Grant CCF 0546574 and in part by Rutgers Research Council under Grant 202018. This paper was recommended by Associate Editor J. Lee. W. A. Chaovalitwongse and Y.-J. Fan are with the Department of Industrial and Systems Engineering, Rutgers University, Piscataway, NJ 08854 USA (e-mail: [email protected]; [email protected]). R. C. Sachdeo is with the Department of Pediatrics, Jersey Shore University Medical Center, Neptune, NJ 07754 USA (e-mail: [email protected]). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TSMCA.2007.897589
1006
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 37, NO. 6, NOVEMBER 2007
Fig. 1. Inferior transverse views of the brain, illustrating approximate depth and subdural electrode placement for EEG recordings, are depicted. Subdural electrode strips are placed over the left orbitofrontal (LOF), right orbitofrontal (ROF), left subtemporal (LST), and right subtemporal (RST) cortex. Depth electrodes are placed in the left temporal depth (LTD) and right temporal depth (RTD) to record hippocampal activity.
Fig. 2. The 20-s EEG recordings of (a) normal activity, (b) preseizure activity, (c) seizure onset activity, and (d) postseizure activity from patient 1 that were obtained from 32 electrodes. Each horizontal trace represents the voltage that is recorded from the electrode sites that are listed in the left column (see Fig. 1 for anatomical location of electrodes).
seizure onset, and postseizure [5][7]. The 20-s EEG proles during these stages are illustrated in Fig. 2(a)(d), respectively.
the frequency of seizure episodes. However, the limitation of this approach is that drugs have a great deal of side effects, and 25%30% of epileptic patients remain unresponsive to antiepileptic drug treatment. Alternative treatment is surgical approaches, which attempt to remove the epileptic region of the brain. Nevertheless, surgery is not always feasible and involves the risk of craniotomy. In addition, only 60% of surgical cases are successful, the mean length of hospital stay for patients with intractable epilepsy admitted for invasive EEG monitoring for presurgical candidates ranged from 4.7 to 5.8 days, and the total aggregate costs exceeded $200 million each year [3]. The cost per patient ranged from $4, 272 for persons with remission after initial diagnosis and treatment to $138, 602 for persons with intractable and frequent seizures [4]. Uncontrolled epilepsy poses a signicant burden to society due to associated healthcare cost. The diagnosis and treatment of epilepsy is complicated by the disabling aspect that seizures occur spontaneously and unpredictably due to the nature of the chaotic disorder. The brain activities (both normal and abnormal) are postulated to result from complex systems that follow dynamical transitions and of which the statistical properties depend on both time and space (electrode location). It has been shown that the development of the epileptic state can be considered as changes in the network circuitry of neurons in the brain that produce changes in voltage potential, which can be captured by an EEG. These changes are reected by wriggling lines along the time axis in a typical EEG recording. For this reason, EEGs have been the main tool for neurologists and neuroscientists that are used to study the epileptogenetic processes and other neurological disorders. A typical electrode montage for the intracranial EEG recordings that are used in our study is shown in Fig. 1. In the emerging view in recent epilepsy research, it is believed that there are four stages that evolved in the seizure process: normal, preseizure,
B. Epilepsy Research: Seizure Prediction During the past decade, there has been an explosion of interests in applying quantitative methods to seizure prediction research including a measure of the brain dynamics [8], [9], a measure of the relative complexity [10][12], a measure of signal similarity in the correlation dimension [13], [14], increasing incidence of energy bursts [15], [16], and multivariate linear discrimination [17]. The studies that are reported in [9], [18][22] suggest that seizures are deterministic rather than random, and it may be possible to predict the onset of epileptic seizures based on quantitative analysis of the brain electrical activity through EEGs. Subsequently, the seizure predictability has also been conrmed by several other groups [10][17], [23]. The results of these studies indicate that it may be possible to detect state transitions to epileptic seizures based on analysis of the brain electrical activity through EEG signals. It is also believed that a seizure is essentially a reecting transition of progressive changes of hidden dynamical patterns in EEG. Such a transition that precedes seizures for periods on the order of minutes to hours is detectable in the EEG by the convergence in value of chaos measures among critical electrode sites on the neocortex and hippocampus [21]. This transition has been shown to be detectable through quantitative analysis of the brain dynamics [9], [20][24]. The results from our previous studies demonstrated that the spatiotemporal dynamical properties of EEGs manifest patterns corresponding to specic clinical states [9], [20], [25][27]. These detectable patterns represent an episode of seizure precursors. These seizure precursors are reected from the convergence of the brain dynamics from a group of electrode sites during the hour preceding seizures [18], [19], [28], [29].
1007
C. Motivation Research in seizure prediction is still far from complete in spite of promising signs of the seizures predictability. While the previous studies demonstrate that preseizure transitions are detectable, the existence of preseizure transitions remains to be further investigated with respect to its specicity and accuracy, i.e., if it only reects epileptic activity or it also occurs with other brain activity. In addition, the development of a model for the mechanism of generation of epileptic seizures remains a difcult task, and the mechanisms of epileptogenesis are not well understood. Essentially, there is a need for an answer to a fundamental, but extremely crucial, question of whether the brains normal and preseizure epileptic activities are distinctive or differentiable. Thus far, this question still remains unanswered. Specically, one needs to demonstrate that normal EEGs differ from preseizure EEGs. In order for one to verify that seizures are predictable, one would have to demonstrate substantial evidence that the brains normal activity differs from the brains preseizure epileptic activity. If they differ, the next question is Are different brains states classiable? Thus far, visual inspection of multiple time series of EEG signals in their unprocessed form is still the predominant way of discriminating and classifying EEG patterns in the medical community. This task requires highly trained medical professionals to continuously eyeball the EEG data. None of the current epilepsy studies in the literature is undertaken to develop classication techniques that can be used to systematically classify/differentiate normal and abnormal EEGs. In order to answer these crucial questions, we are motivated to develop such classication technique. D. Paper Organization The organization of the succeeding sections of this paper is given as follows. In Section II, the quantication of the brain dynamics and statistical similarity measures of time series for KNN classication are described. The EEG data description and experimental design are presented in Section III. The classication results and the performance characteristics of the proposed technique are addressed in Section IV. The concluding remarks are given in Section V. II. M ETHOD The classication technique that is developed in this study is comprised of three key features. The rst feature is quantitative measures of the brain dynamics. We employ the estimation of the short-term maximum Lyapunov exponent, which has been previously shown to be capable of contemplating dynamical mechanisms of the brain network from EEG signals [20]. The study of the brain dynamics is motivated by the proof concept of chaos theory, i.e., understanding brain dynamics is capable of providing insights about the different states of brain activities that are reected from pathological dynamical interactions of the brain network [8], [30]. Based on the quantication of the brain dynamics, the subsequent second feature is the statistical similarity measures for similarity/dissimilarity of classiable
features of different brains physiological states. We herein employ three time series similarity measures to the analysis of brain dynamics in EEG time series. Those measures include Euclidean, T-Statistical, and Dynamic Time Warping (DTW) distances. The third feature is the classication algorithm, which is based on the KNN rule. Integration of the second and third features results into a novel time series classication technique. We will use this integrated technique to classify normal and preseizure EEGs. A. Quantication of Brain Dynamics Quantication of the brain dynamics from EEGs in this study is suitable to the investigation of a nonstationary system such as the brain because it is capable of automatically identifying and appropriately weighing existing transients in the data. This technique is motivated by mathematical models from chaos theory that are used to characterize multidimensional complex systems and reduce the dimensionality of EEGs [31][35]. To quantify the brain dynamics, we divide EEG signals into sequential 10.24-s epochs (nonoverlapping windows) to properly account for possible nonstationarities in the epileptic EEG. For each epoch of each channel of the EEG signals, we estimate the measure of chaos, which is known as short-term maximum Lyapunov exponent, to quantify the chaoticity of the attractor. A chaotic-system-like human brain is a system in which orbits that originate from similar initial conditions or nearby points in the phase space diverge exponentially in expansion process. The rate of divergence is an important aspect of the dynamical system and is reected in the value of the Lyapunov exponents. In other words, the Lyapunov exponents measure the average uncertainty along the local eigenvectors of an attractor in the phase space. Next, we will give a short overview of mathematical models that are used in the estimation of the short-term maximum Lyapunov exponent from the EEG signals. 1) EEG Time Series Embedding: In the study of brain dynamics, the initial step in analyzing the dynamical properties of EEG signals is to embed it in a higher dimensional space of dimension p, which enables us to capture the behavior in time of the p variables that are primarily responsible for the dynamics of the EEG. We can now construct p-dimensional vectors X(t), whose components consist of the values of the recorded EEG signal x(t) at p points in time that is separated by a time delay. Construction of the embedding phase space from a data segment x(t) of duration T is made with the method of delays. The vectors Xi in the phase space are constructed as Xi = (x(ti ), x(ti + ), . . . , x (ti + (p 1) )) (1)
where is the selected time lag between the components of each vector in the phase space, p is the selected dimension of the embedding phase space, and ti [1, T (p 1) ]. The vectors Xi in the phase space are illustrated in Fig. 3. 2) Estimation of Short-Term Maximum Lyapunov Exponent ST Lmax : The method for estimation of ST Lmax for nonstationary data (e.g., EEG time series) is previously explained in [8] and [36]. In this section, we will only give a short description and basic notation of our mathematical models that
1008
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 37, NO. 6, NOVEMBER 2007
Fig. 3. Diagram illustrating an EEG epoch that is embedded in phase space for the quantication of brain dynamics, assuming that p = 4. The ducial trajectory, i.e., the rst three local Lyapunov exponents (L1, L2, L3), is shown.
are used to estimate ST Lmax . First, let us dene the notation that is given here. Na is the number of local ST Lmax s that will be estimated within a duration T data segment. Therefore, if Dt is the sampling period of the time domain data, T = (N 1)Dt = Na t + (p 1) . X(ti ) is the point of the ducial trajectory t (X(t0 )) with t = ti , X(t0 ) = (x(t0 ), . . . , x(t0 + (p 1) )), and X(tj ) is a properly chosen vector that is adjacent to X(ti ) in the phase space. Xi,j (0) = X(ti ) X(tj ) is the displacement vector at ti , i.e., a perturbation of the ducial orbit at ti , and Xi,j (t) = X(ti + t) X(tj + t) is the evolution of this perturbation after time t. ti = t0 + (i 1) t and tj = t0 + (j 1) t, where i [1, Na ] and j [1, N ] with j = i. t is the evolution time for Xi,j , i.e., the time one allows Xi,j to evolve in the phase space. If evolution time t is given in seconds, then L is given in bits per second. t0 is the initial time point of the ducial trajectory and coincides with the time point of the rst data in the data segment of analysis. In the estimation of ST Lmax , for a complete scan of the attractor, t0 should move within [0, t]. ST Lmax is dened as the average of local Lyapunov exponents in the state space and can be calculated by the following equation: ST Lmax 1 = Na t
Na
log2
i=1
porating statistical time series similarity measures for differentiating EEG signals from normal and preseizure states. We will also apply and compare the performance of those similarity measures, including Euclidean, T-Statistical (index), and DTW distances. Each of these measures can give us different aspects and insights about the temporal and spatial characteristics of EEG signals. In this section, we will represent the time series of measures of chaos in the notation of ST Lmax as time series X to simplify the mathematical models. 1) DTW Distance: Given two time series (or vector sequences) X and Y of equal length |X| = |Y | = n, pattern similarity is determined by aligning time series X with time series Y with the distortion of alignment Dalign (X, Y ). DTW is used to compute the best possible alignment warp between two time series by selecting the one with the minimum distortion. In other words, the DTW distance is a distance measure (or similarity measure) between two time series by computing the best possible alignment or the minimum mapping (aligning) distance between two time series. In this study, all our EEG data samples are equal in length; however, the DTW can be extended to the case where the lengths of the two time series are not equal. DTW has been widely used in many contexts including data mining [37], [38], gesture recognition [39], robotics [40], speech processing [41][43], and medicine [44]. The problem of calculating the DTW distance can be solved by a dynamic programming approach. The basic concept can be described as follows. First, construct an alignment of every time series vector (data point) in X to match with every corresponding vector in Y . The n n alignment matrix will then be constructed by using the Euclidean distance as the local distance between two vectors, i.e., d(xi , yj ) = (xi yj )2 , where the (i, j)th element of the matrix is distance d(xi , yj ) between the ith point of time series X and the jth point of time series Y . Subsequently, we construct a warp path, i.e., W = w1 , . . . , wK , starting at the beginning of each time series w1 = (1, 1) and nishing at the end of both time series wK = (n, n). Note that K is the length of the warp path and max(|X|, |Y |) K < |X| + |Y |. The kth element of the warp path represents the matching point of two time series, i.e., wk = (i, j), where (i, j) corresponds to index i from time series X and index j from time series Y (shown in Fig. 4). The warp path can actually be calculated in reverse order, starting at the end of both time series. There is also another constraint on the warp path on the monotonically increasing indices of a warp path. Specically, if wk = (i, j) and wk+1 = (i , j ), then a warp path must satisfy i i i + 1 and j j j + 1. Note that there can be an exponential number of warping paths that satisfy the preceding conditions. However, the optimal warp path is the one with a minimum warping (distortion) cost dened by 1 K
K
B. Similarity Measures In this study, the choice of similarity measures is a very important step in achieving accurate classication results. To our knowledge, we are the rst to develop a framework incor-
In a dynamic programming approach, the warp path must either be incremented by one unit (adjacent) or stay at the same i-axis or j-axis. Therefore, we only need to evaluate the
1009
where |xy| is the sample standard deviation of the absolute difference between time series X and Y estimated over a window with length n. Note that the t-index follows a t-distribution with n 1 degrees of freedom. 3) Euclidean Distance: The Euclidean distance is the most commonly used similarity measure. It is easy to understand and bear certain success in many classication problems. It measures the degree of similarity in terms of intensity of the data. In short, the Euclidean distance tells us an average of the difference in intensity of two time series. The Euclidean distance ED between time series X and Y is dened as EDxy = ( n (xi yi )2 )/n. i=1 C. KNN Classication KNN is a very intuitive method in which the classier labels samples (time series) based on their similarity between samples in the training data. In other words, the classier makes a decision by comparing a new unlabeled sample with the baseline data. This technique represents a bridge between the parametric techniques that require a priori knowledge of the distributions underlying the data and nonparametric techniques, which presuppose the functional form of the discriminant surfaces separating the different pattern classes. Applications of KNN have been successful in many areas including handwritten digit recognition [45], gene expression classication [46], and text mining [47]. In general, for a given unlabeled time series X, the KNN rule nds the k closest (neighborhood) labeled time series in the training data set and assigns X to the class that appears most frequently in the neighborhood of k time series. Besides the training data, the KNN rule only requires three input parameters that are used for classifying a new unlabeled time series, i.e., the size of the neighborhood k and a similarity function that is used as a measure of closeness. There are two common rules for classifying the new unlabeled data: 1) majority voting and 2) similarity degree summing. In majority voting, a class (category) gets one vote for each instance of that class in the set of k neighborhood samples. Then, the new data sample is classied to the class with the highest amount of votes. In similarity score summing, each class gets a score that is equal to the summation of the similarity degrees of the instances of that class in the set of k neighborhood samples. Then, the new data sample is classied to the class with the highest similarity degree sum. Note that majority voting is more commonly used because it is less sensitive to outliers. Thus, we shall use majority voting in this study. This is illustrated in Fig. 5, which shows data records, each with two attributes (2-D plot) that are representations of three classes of data (red, blue, and green). In this case, k = 5. Thus, the unknown-labeled test point would be labeled by the category of the red points. However, note that we will train the classier to nd the best value of k to maximize the classication performance. III. M ATERIALS AND E XPERIMENTAL D ESIGN The underlying hypothesis is that the measure of chaos from EEG time series can be used as features to discriminate different stages of the brain dynamics, and these features from the
Fig. 4. Warping matrix with the minimum-distance warp path of two time series X and Y .
recurrence of the cumulative distance that is found in the adjacent elements, i.e., D(i, j 1) D(i, j) = d(xi , yj ) + min D(i 1, j) D(i 1, j 1). This results into the optimality condition for the warping path: DTW (wk ) = d(xi(k) , yj(k) , wk )r(wk ) + DTW (wk1 ). xy xy 2) T-Statistical (Index) Distance: The t-statistics is a statistical analysis that is used to examine if two time series are statistically different from each other. In particular, it assesses whether the means of two time series are statistically different from each other. In other words, we commonly use the t-test to determine if two time series differ from each other in a signicant way under the assumptions that the paired differences are independent and identically normally distributed. Note that the t-test deals with the problems that are associated with inference based on small samples (epochs of time series), which implies that the calculated mean and standard deviation may deviate from the real mean and standard deviation. The t-index is a similarity degree of the t-statistics from the paired t-test for comparisons of means of paired-dependent observations. In this study, we use the t-index as a measure of statistical distance between two time series. Specically, it is used to estimate the difference of the EEG signals from different brain states. The t-index can be seen as a ratio of the difference between the two means or averages and a measure of the variability or dispersion of the scores. The property of t-index is essentially another example of the signal-to-noise metaphor in research: The difference between the means is the signal, and a measure of variability is essentially noise that may make it harder to see the group difference. The t-index Txy between the time series X = x1 , . . . , xn and Y = y1 , . . . , yn is then dened as Txy =
n i=1
|x yi | i n|xy|
(3)
1010
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 37, NO. 6, NOVEMBER 2007
sequent ofine analysis. These EEG recordings were viewed by two independent electroencephalographers to determine the number and type of recorded seizures, seizure onset and end times, and seizure onset zones. B. Data Selection and Sampling Procedure In this study, the classication will be performed separately for each subject. For consistency, we analyze and investigate EEG time series only from 26 standard electrodes from every patient. First, we randomly sample two groups (normal and preseizure) of 5-min EEG epochs from the continuous recordings in each patient. Per seizure, ve EEG epochs from each of the normal and preseizure states are randomly and uniformly sampled. Normal EEG samples are selected from EEG recordings that are more than 8 h apart from a seizure. Preseizure EEG epochs are selected from EEG recordings during the 30-min interval before. For example, patient 1 had seven seizures; therefore, 70 EEG epochs (35 normal and 35 preseizure) will be sampled from patient 1s data set. C. Classication Procedure After two groups of EEG epochs are collected, we rst calculate the measures of chaos (ST Lmax proles) from EEG signals using the method that was described in the previous section. Each measure was calculated continuously for each nonoverlapping 10.24-s segment of EEG data. After we calculate an ST Lmax prole for each EEG channel, we then perform the KNN classication by categorizing an unknown sample using the KNN rules based on the average statistical distance over 26 electrodes. There are a few parameters in the KNN classication that need to be trained. This can be done by dividing the EEG data into a training set (baseline) and a test set. There are many alternatives on how to divide the data into the training and test sets. We describe the technique that is used to perform this operation in the next section. For each training set, we use different k values to train for the optimal k value that gives the best result in classication. Then, we use EEG epochs from the test set to evaluate the performance of this k value. In this study, we will vary the value of k from 3 to 13 (only odd numbers). The optimal k value will be determined in the training phase (discussed in next section). Note that we apply the KNN using all three time series similarity measures that are discussed in this paper. Note that, although the Euclidean distance is commonly used in a standard KNN algorithm, the KNN is very sensitive to the choice of similarity measure that is used, and the EEG characteristics are much more complicated than the Euclidean distance can capture. We anticipate that more sophisticated similarity measures such as t-index and DTW will outperform the standard Euclidean distance. D. Training and Testing: Cross Validation There are many choices of how to divide the data into training and test sets. In order to reduce the bias of training and test data, we propose to employ an n-fold cross-validation technique
Fig. 5. KNN query starts at the test point and grows a spherical region until it encloses k training samples, and it labels the test point by a majority voting of these samples. In this case where k = 5, the test point would be labeled by the category of the red points. TABLE I EEG DATASET CHARACTERISTICS
same state should be more similar than the one from different states. In other words, the characteristics of the brain dynamics during the normal state should be more similar to each other than those during the preseizure state and vice versa. This study is undertaken to demonstrate that the classication technique that was described in the previous section can differentiate the EEG signals from the normal and preseizure states. The classiability of different brain states will be demonstrated through the quantitative analysis of the brain dynamics ST Lmax by using the time series KNN classication that is proposed here. A. EEG Data Acquisition The data sets consisted of continuous long-term (313 days) multichannel intracranial EEG recordings from bilaterally surgically implanted macroelectrodes in the hippocampus, temporal, and frontal lobe cortexes of ten epileptic patients with medically intractable temporal lobe epilepsy (outlined in Table I). The recordings were obtained as part of a presurgical clinical evaluation, using a Nicolet BMSI 4000 recording system with ampliers of an input range of 0.6 mV, sampling rate of 200 Hz, and lters with a frequency range of 0.570 Hz. Each recording included a total of 2632 intracranial electrodes (eight subdural and six hippocampal depth electrodes for each cerebral hemisphere, and a strip of four additional electrodes if deemed necessary by the neurologist). The recorded EEG signals were digitized and stored on magnetic media for sub-
1011
F. Selection of the Best Classication Schemes To quantitatively select the best classication scheme as well as the best parameter setting (i.e., the best k value and similarity measure), we will implement a statistical method that is called receiver operating characteristics (ROCs) and derived from the detection theory to identify the optimal parameter setting. The ROC analysis is used to indicate an appropriate tradeoff that one can achieve between the detection rate (sensitivity, which is plotted on the Y -axis) that is desired to be maximized and the false alarm rate (1 specicity, which is plotted on the X-axis) that is desirable to be minimized. Basically, the sensitivity and the specicity of each classication scheme for each parameter setting are calculated. By denition of ROC analysis, the best scheme is selected such that it is closest to the ideal classier (best performance). For each individual patient, the best classication scheme can be identied by selecting the scheme whose performance is closest to the ideal point (sensitivity = 1 and 1 specicity = 0), i.e., the scheme that is closest to the top left-hand corner on the ROC plot will be selected. The performance evaluation of the classication schemes and the results for each of classication schemes will be discussed in more detail in the next sections. IV. R ESULT Among all ten patients, we sample and preprocess the total of 1530 EEG epochs from the normal and preseizure states, as described in the previous section. Each epoch contains 780 data points (26 electrodes, 5-min epoch = 30 points). The next step is to train our classier for the best value of k-nearest neighbors. In this experiment, we try different k values ranging from 3 to 13 (note that k should be an odd number). As mentioned earlier, threefold and vefold cross validations are employed in this study. We will test the KNN algorithm using three time series similarity measures (i.e., Euclidean, t-index, and DTW). The sensitivity and specicity of each parameter setting (value of k and choice of similarity measure) are calculated. It is worth noting that we are studying two-class classication in this study. Thus, the specicity of classifying preseizure EEGs is equivalent to the sensitivity of classifying normal EEGs. In this section, we will use these terms interchangeably. Finally, based on the classication performances from all combinations of parameter settings, the best parameter setting for each patient is selected from the ROC analysis. A. ROC Analysis In order to achieve the most appropriate tradeoff to maximize the detection rate (sensitivity, which is plotted on the Y -axis) and minimize the false alarm rate (1 specicity, which is plotted on the X-axis), the best classication scheme for each patient is selected such that it is closest to the ideal classier (best possible performance) at the top left-hand corner of the ROC plot. Per individual patient, we apply ROC analysis to indicate the best parameter setting in the KNN classication. Fig. 7 illustrates an ROC plot of the KNN classication using DTW applied to patient 10 using vefold cross validation with
Fig. 6. Evaluation concept of classication results. Note that we dene preseizure as abnormal.
to train for the best value of k in this study. The crossvalidation technique is extensively used as a method to estimate the generalization error based on resampling. In other words, these techniques will be implemented during the training phases to estimate how well the classiers learning from the training EEG data is going to perform on future as-yet-unseen data in the testing phase. Generally, with n-fold cross validation, EEGs data will be divided into n subsets of (approximately) equal size. The proposed KNN will be trained and tested n times, in which one of the subsets from training is left out each time, and tested on the omitted subset [22], [48]. As mentioned in [49], the result from one n-fold cross validation may not be reliable. In order to have low mean square error and bias, we perform ten repetitions of cross validations. After ten cross-validation replications for each value of n (folds), we compare the average of the classication performance on different values of k (nearest neighbors). In this study, we select the values of n to be 3 and 5 as the EEG sample sizes in some patients are very small. It is important to note that the classication techniques will be trained and tested individually for each subject. E. Performance Evaluation of Classication Schemes In general, to evaluate the classier, we categorize the classication into two classes: 1) positive (preseizure) and 2) negative (normal). Then, we consider four subsets of classication results: 1) true positives (TPs), which denote correct classications of positive cases; 2) true negatives (TNs), which denote correct classications of negative cases; 3) false positives (FPs), which denote incorrect classications of negative cases into class positive; and 4) false negatives (FNs), which denote incorrect classications of positive cases into class negative. To better explain the concept of the evaluation of classiers, let us consider in the case of the detection of preseizure EEG data (see Fig. 6). Sensitivity and specicity are widely used in the medical domain as classication performance measures. Sensitivity measures the fraction of positive cases that are classied as positive. Specicity measures the fraction of negative cases that are classied as negative. In fact, the sensitivity can be considered as a probability of accurately classifying EEG samples in the preseizure case. The specicity can be considered as a probability of accurately classifying EEG samples in the normal case. The sensitivity and specicity are dened as follows: TP TP + FN TN . specicity = TN + FP sensitivity =
1012
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 37, NO. 6, NOVEMBER 2007
Fig. 7. ROC plot for DTW in patient 10 using vefold cross validation to train and test the algorithm.
Fig. 9. ROC plot for t-statistics in patient 10 using vefold cross validation to train and test the algorithm.
using the T-Statistical distance outperforms all other time series similarity measures. B. Classication Results We apply the ROC analysis, as described in the previous section, to identify the best classication scheme (value of k and choice of similarity measure) per individual patient. The performance characteristics of the best scheme for each individual patient in threefold and vefold cross validations are listed in Table II. From the table, we can see the consistent result of the best performance from the T-Statistical distance. The results follow the intuitive sense in the classication theory that the performance on vefold cross validation should be better the one on threefold cross validation, with the reason being that the training sample size is larger. It is clear that the KNN classication is sensitive to the size of training samples, and we expect the performance to be better with a higher value of n-fold. With vefold cross validation, the KNN classication achieved about 81% sensitivity and 73% specicity, on average, over ten patients. With threefold cross validation, the performance is slightly worse as the KNN classication achieved about 79% sensitivity and 71% specicity, on average, over 10 patients. As observed in Table II, the optimal value of k-nearest neighbors for the T-Statistical distance ranges from 3 to 13. While k = 3 tends to be the best global setting, this implies that the T-Statistical distance is capable of capturing specic patterns that are manifested in EEG time series at some specic location at a specic time. This will lead to a computational complexity advantage. As the value of k is lower, the algorithm requires less computational time to search for the nearest neighbors. Thus, the algorithm will run much faster. Figs. 10 and 11 illustrate the overall classication results that are tested in ten patients with optimal scheme (k value and statistical distance) per individual patient to train and test the algorithm using vefold cross validation and threefold cross validation, respectively. In Fig. 10, employing the vefold cross-validation technique, with the optimal k value using the
Fig. 8. ROC plot for EDs in patient 10 using vefold cross validation to train and test the algorithm.
ten replications to train and test the KNN algorithm. Note that each point in the plot represents a value of k-nearest neighbors (ranging from 3 to 13). In Fig. 7, we observe that the KNN classication with DTW achieves its best performance when k = 3, which yields a sensitivity of about 84% and a specicity of about 75%. Figs. 8 and 9 illustrate the ROC plots of the KNN classication using Euclidean and T-Statistical distances, respectively, in patient 10 using vefold cross validation with ten replications. In Fig. 8, we observe that the KNN classication with Euclidean distance achieves its best performance when k = 11, which yields a sensitivity of about 84% and a specicity of about 65%. In Fig. 9, we observe that the KNN classication with T-Statistical distance achieves its best performance when k = 3, which yields a sensitivity of about 89% and a specicity of about 91%. From these ROC plots, in the case of patient 10, it is clear that the KNN classication
1013
TABLE II PERFORMANCE CHARACTERISTICS OF THE BEST CLASSIFICATION SCHEME FOR EACH INDIVIDUAL PATIENT USING THREEFOLD AND FIVEFOLD CROSS VALIDATION ON ONLY ST Lmax VALUES
Fig. 10. Overall classication results among all similarity measures using vefold cross validation to train and test the algorithm.
about 79% and 71%, respectively. With the optimal k value using the DTW distance, the KNN classication can achieve a sensitivity and specicity of about 68% and 65%, respectively. With the optimal k value using the Euclidean distance, the KNN classication can achieve a sensitivity and specicity of about 73% and 66%, respectively. It can also be observed that the KNN classication achieved lower specicity than all other cases. One speculation is that there are more variabilities in the EEG patterns during the normal period than during the preseizure period. In other words, because the preseizure EEG patterns are more similar as they evolve in the same state transition (seizure development), the KNN classication is able to better capture similarity in preseizure patterns. Note that the results in this study are consistent with the results from our previous study in classifying EEGs [22]. Surprisingly, the use of the DTW distance did not improve the performance of the KNN classication to achieve better classication results. Nevertheless, with a choice of any similarity measures in general, the KNN algorithm shows a potential to be a good time series classication as it preserves both the spatial and temporal properties of the data. These results conrm our hypothesis that the brains states are classiable based on quantitative analyses of EEG. The framework of the classiers that are proposed in this study can be extended to the development of an abnormal brain activity classier, which will be described in the next section.
T-Statistical distance, the KNN classication can achieve a sensitivity and specicity of about 81% and 73%, respectively. With the optimal k value using the DTW distance, the KNN classication can achieve a sensitivity and specicity of about 68% and 66%, respectively. With the optimal k value using the Euclidean distance, the KNN classication can achieve a sensitivity and specicity of about 75% and 67%, respectively. In Fig. 11, employing the threefold cross-validation technique, with the optimal k value using the T-Statistical distance, the KNN classication can achieve a sensitivity and specicity of
This study addresses the open question of the classiability of the brains preseizure and normal states. The experimental results on the performance of the time series KNN are very encouraging. These results represent a proof concept of the successful brain dynamics quantication through EEGs. This concept can provide insights and characterize different states of brain activities that are reected from pathological dynamical interactions of the brain network. In addition, these results also conrm our hypothesis that it is possible to differentiate and classify the brains preseizure and normal activities based on optimization, data mining, and dynamical system approaches in multichannel intracranial EEG recordings. The reliable classication is conceivable because, for the vast majority of seizures, the spatiotemporal dynamical features of the preseizure state
1014
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 37, NO. 6, NOVEMBER 2007
Fig. 12. Flowchart of an automated system, which can be used as a base module of ABAC or ASPS.
sufciently differ from those of the normal state. It can be easily observed that the brain dynamics from different states of the same patient are qualitatively and quantitatively differentiable. In addition, we notice that the performance of the KNN algorithm is irrelevant to the characteristic of the patients. Surprisingly, a choice of smaller k value for all time series similarity measures tends to be favorable. In addition, the proposed technique is very fast and scalable, and it does not require any assumption of distribution of the EEG data. All of the programming was done in Matlab environment on a desktop computer with Pentium Xeon 2.8 GHz with 2 GB of random access memory. The running time for the statistical cross-validation technique is less than 5 min on average. The bottleneck of the algorithm is twofold. One is the calculation of the measure of chaos, and another is the calculation of the DTW and T-Statistical distances. Calculating the DTW distance, we are required to solve multiple dynamic programs in addition to the calculation of the Euclidean distance. Calculating the T-Statistical distance, we are required to estimate the standard deviation of the difference of two time series. Both of these steps can be improved when we implement an online real-time analysis in our future study. The results of this study provide a framework of methodology that can be used to elucidate the dynamical mechanisms of the brain mechanisms underlying the epileptogenic process. In addition, the results can be extended to the development of an online brain monitoring system. For instance, the classication algorithm can be embedded into the EEG monitoring unit. The unit can collect normal and preseizure EEG samples as a baseline database and subsequently perform the EEG classication by periodically sampling the ongoing EEG signals. If the EEG samples are classied as preseizure, one can trigger the warning of an impending seizure. Success of this study is a necessary rst step in the development of an abnormal brain activity classier (ABAC) and an automated seizure prediction system (ASPS), which are computer-based systems that are used to predict seizures or classify abnormal EEGs (see, e.g., Fig. 12). In order to make another step closer to the ABAC, in our future study, we would like to extend the proposed classication tech-
nique to scalp EEG data as it is more practical and feasible to obtain scalp EEG data. In this study, the intracranial EEG data that are studied here have good characteristics for classication because they are clean of artifacts that are caused by the EEG machines as well as the patients movements (e.g., coughing and swallowing). As opposed to intracranial EEGs, the most challenging issue is that scalp EEGs are much more sensitive and susceptible to artifacts. These will make it very hard for our algorithm to classify the real physiological states of the brain. An analysis of our algorithm on scalp EEGs will be studied in the future. Such classication algorithms for scalp EEGs can also be used to develop a quick screening procedure to classify normal and epilepsy patients. In addition, the feature selection study will be possible in the future. This study will help us to select electrodes that show prominent changes and give us the best classication results, which might lead us to the solution to the epileptogenic zone localization problem. Another prospective of this study is to understand exactly how biology makes certain people vulnerable to seizures. The reigning theory holds that abnormal activities/scenarios affect the risk of seizures being developed. Success of this project will help neurologists to nd better faster ways of gauging the risk of seizures being developed in epileptic patients. We believe that the lessons that were learned in controlling abnormal heart rhythms could be applied to controlling brain arrhythmias. Traditionally, cardiac arrhythmias were treated with medicines. Today, many are treated with automatic debrillators, which are devices that are surgically placed in patients chests. The devices monitor heart rhythms, and when the rhythms become abnormal, they will give a very small jolt of electricity to put the heart back in its normal rhythm. In the same view, epilepsy may respond to a similar kind of closed-loop feedback warning system. The device would predict the onset of a seizure and prevent the seizure before it occurs. These results are a very important proof of concept investigation of the development of implantable seizure feedback control devices to directly regulate therapeutic pharmacological or physiological intervention to prevent impending seizures or other brain disorders. For example, such an intervention might be achieved by electrical or magnetic stimulation (e.g., vagal nerve stimulation) or by a timely release of an anticonvulsant drug. In the future, a new outlook for epilepsy treatment would be to trigger a pulse intervention long (ranging from minutes to hours) before the seizure actually begins. It is possible that an electrical stimulus that is delivered during this preseizure period could abort an impending seizure before it actually occurs. A warning of several minutes in duration could even trigger the release of an anticonvulsant drug in time to abort an impending seizure. The outcome of this study will form a bridge between seizure prediction research and the implementation of biofeedback seizure control devices, which will be a revolutionary approach for handling epileptic seizures (i.e., the brain pacemaker). ACKNOWLEDGMENT The authors would like to thank Prof. E. Keogh, Prof. P. M. Pardalos, Prof. L. D. Iasemidis, and Prof. J. C. Sackellares for their fruitful comments and discussion.
1015
R EFERENCES
[1] O. Cockerell, I. Eckle, D. Goodridge, J. Sander, and S. Shorvon, Epilepsy in a population of 6000 re-examined: Secular trends in rst attendance rates, prevalence, and prognosis, J. Neurol., Neurosurg., Psychiatr., vol. 58, no. 5, pp. 570576, 1996. [2] W. H. O. (WHO), Epilepsy: Historical Overview, 2004. [Online]. Available: http://www.who.int/mediacentre/factsheets/fs168/en/ [3] C. Begley, M. Famulari, J. Annegers, D. Lairson, T. Reynolds, S. Coan, S. Dubinsky, M. Newmark, C. Leibson, E. So, and W. Rocca, The cost of epilepsy in the United States: An estimate from population-based clinical and survey data, Epilepsia, vol. 41, no. 3, pp. 342351, 2000. [4] C. Begley, J. Annegers, D. Lairson, T. Reynolds, and W. Hauser, Cost of epilepsy in the United States: A model based on incidence and prognosis, Epilepsia, vol. 35, no. 6, pp. 12301243, 1994. [5] L. Iasemidis, Epileptic seizure prediction and control, IEEE Trans. Biomed. Eng., vol. 50, no. 5, pp. 549558, May 2003. [6] K. Lehnertz and B. Litt, The rst international collaborative workshop on seizure prediction: Summary and data description, J. Clin. Neurophysiol., vol. 116, no. 3, pp. 493505, 2005. [7] B. Litt and J. Echauz, Prediction of epileptic seizures, Lancet Neurol., vol. 1, no. 1, pp. 2230, May 2002. [8] L. Iasemidis, On the dynamics of the human brain in temporal lobe epilepsy, Ph.D. dissertation, Univ. Michigan, Ann Arbor, MI, 1991. [9] P. Pardalos, W. Chaovalitwongse, L. Iasemidis, J. Sackellares, D.-S. Shiau, P. Carney, O. Prokopyev, and V. Yatsenko, Seizure warning algorithm based on spatiotemporal dynamics of intracranial EEG, Math. Program., vol. 101, no. 2, pp. 365385, 2004. [10] C. Elger and K. Lehnertz, Seizure prediction by non-linear time series analysis of brain electrical activity, Eur. J. Neurosci., vol. 10, no. 2, pp. 786789, Feb. 1998. [11] K. Lehnertz and C. Elger, Can epileptic seizures be predicted? Evidence from nonlinear time series analysis of brain electrical activity, Phys. Rev. Lett., vol. 80, no. 22, pp. 50195022, Jun. 1998. [12] F. Mormann, T. Kreuz, C. Rieke, R. Andrzejak, A. Kraskov, P. David, C. Elger, and K. Lehnertz, On the predictability of epileptic seizures, J. Clin. Neurophysiol., vol. 116, no. 3, pp. 569587, 2005. [13] M. L. V. Quyen, J. Martinerie, M. Baulac, and F. Varela, Anticipating epileptic seizures in real time by non-linear analysis of similarity between EEG recordings, NeuroReport, vol. 10, no. 10, pp. 21492155, Jul. 1999. [14] J. Martinerie, C. V. Adam, and M. L. V. Quyen, Epileptic seizures can be anticipated by non-linear analysis, Nat. Med., vol. 4, no. 10, pp. 11731176, Oct. 1998. [15] M. Harrison, M. Frei, and I. Osorio, Accumulated energy revisited, J. Clin. Neurophysiol., vol. 116, no. 3, pp. 527531, 2005. [16] B. Litt, R. Esteller, J. Echauz, D. Maryann, R. Shor, T. Henry, P. Pennell, C. Epstein, R. Bakay, M. Dichter, and G. Vachtservanos, Epileptic seizures may begin hours in advance of clinical onset: A report of ve patients, Neuron, vol. 30, no. 1, pp. 5164, Apr. 2001. [17] K. Jerger, S. Weinstein, T. Sauer, and S. Schiff, Multivariate linear discrimination of seizures, J. Clin. Neurophysiol., vol. 116, no. 3, pp. 545551, 2005. [18] W. Chaovalitwongse, L. Iasemidis, P. Pardalos, P. Carney, D.-S. Shiau, and J. Sackellares, Performance of a seizure warning algorithm based on the dynamics of intracranial EEG, Epilepsy Res., vol. 64, no. 3, pp. 93133, May 2005. [19] W. Chaovalitwongse, P. Pardalos, L. Iasemidis, J. Sackellares, and D.-S. Shiau, Optimization of spatio-temporal pattern processing for seizure warning and prediction, U.S. Patent application led August 2004, Attorney Docket No. 028724-150, 2004. [20] W. Chaovalitwongse, P. Pardalos, L. Iasemidis, D.-S. Shiau, and J. Sackellares, Applications of global optimization and dynamical systems to prediction of epileptic seizures, in Quantitative Neuroscience, P. Pardalos, J. Sackellares, L. Iasemidis, and P. Carney, Eds. Norwell, MA: Kluwer, 2003, pp. 136. [21] W. Chaovalitwongse, P. Pardalos, L. Iasemidis, D. Shiau, and J. Sackellares, Dynamical approaches and multi-quadratic integer programming for seizure prediction, Optim. Methods Softw., vol. 20, no. 2/3, pp. 383394, 2005. [22] W. Chaovalitwongse, P. Pardalos, and O. Prokoyev, Electroencephalogram (EEG) time series classication: Applications in epilepsy, Ann. Oper. Res., vol. 148, no. 1, pp. 227250, Nov. 2006. [23] R. Esteller, J. Echauz, M. DAlessandro, G. Worrell, S. Cranstoun, G. Vachtsevanos, and B. Litt, Continuous energy variation during the seizure cycle: Towards an on-line accumulated energy, J. Clin. Neurophysiol., vol. 116, no. 3, pp. 517526, 2005.
[24] W. Chaovalitwongse, P. Pardalos, and O. Prokoyev, Reduction of multiquadratic 01 programming problems to linear mixed 01 programming problems, Oper. Res. Lett., vol. 32, no. 6, pp. 517522, 2004. [25] W. Chaovalitwongse, Optimization and dynamical approaches in nonlinear time series analysis with applications in bioengineering, Ph.D. dissertation, Univ. Florida, Gainesville, FL, 2003. [26] P. Pardalos and J. Principe, Biocomputing. Norwell, MA: Kluwer, 2003. [27] L. Iasemidis, D.-S. Shiau, W. Chaovalitwongse, J. Sackellares, P. Pardalos, P. Carney, J. Principe, A. Prasad, B. Veeramani, and K. Tsakalis, Adaptive epileptic seizure prediction system, IEEE Trans. Biomed. Eng., vol. 50, no. 5, pp. 616627, May 2003. [28] J. Sackellares, L. Iasemidis, D.-S. Shiau, L. Dance, P. Pardalos, and W. Chaovalitwongse, Optimization of multi-dimensional time series processing for seizure warning and prediction, International Patent Application led August 2003, Attorney Docket No. 028724-142, 2003. [29] J. Sackellares, L. Iasemidis, V. Yatsenko, D.-S. Shiau, P. Pardalos, and W. Chaovalitwongse, Multi-dimensional multi-parameter time series processing for seizure warning and prediction, International Patent Application led Septempber 2003, Attorney Docket No. 028724-143, 2003. [30] L. Iasemidis and J. Sackellares, The evolution with time of the spatial distribution of the largest Lyapunov exponent on the human epileptic cortex, in Measuring Chaos in the Human Brain, D. Duke and W. Pritchard, Eds. Singapore: World Scientic, 1991, pp. 4982. [31] A. Babloyantz and A. Destexhe, Low dimensional chaos in an instance of epilepsy, Proc. Nat. Acad. Sci. USA, vol. 83, no. 10, pp. 35133517, May 1986. [32] L. Iasemidis, H. Zaveri, J. Sackellares, and W. Williams, Phase space analysis of EEG in temporal lobe epilepsy, in Proc. 10th Annu. Int. Conf. IEEE Eng. Med. and Biol. Soc., 1988, pp. 12011203. [33] N. Packard, J. Crutcheld, and J. Farmer, Geometry from time series, Phys. Rev. Lett., vol. 45, no. 9, pp. 712716, Sep. 1980. [34] P. Rapp, I. Zimmerman, and A. M. Albano, Experimental studies of chaotic neural behavior: Cellular activity and electroencephalographic signals, in Nonlinear Oscillations in Biology and Chemistry, H. Othmer, Ed. New York: Springer-Verlag, 1986, pp. 175205. [35] F. Takens, Detecting strange attractors in turbulence, in Dynamical Systems and Turbulence. Lecture Notes in Mathematics, D. Rand and L. Young, Eds. New York: Springer-Verlag, 1981. [36] A. Wolf, J. Swift, H. Swinney, and J. Vastano, Determining Lyapunov exponents from a time series, Phys. D, vol. 16, no. 3, pp. 285317, 1985. [37] E. Keogh and M. Pazzani, Scaling up dynamic time warping for datamining applications, in Proc. 6th ACM SIGKDD Int. Conf. Knowl. Discov. and Data Mining, 2000, pp. 285289. [38] D. Berndt and J. Clifford, Using dynamic time warping to nd patterns in time series, in Proc. AAAI Workshop KDD, 1994, pp. 359370. [39] D. Gavrila and L. Davis, Towards 3-D model-based tracking and recognition of human movement: A multi-view approach, in Proc. Int. Workshop Autom. Face- and Gesture-Recog., 1995, pp. 272277. [40] M. Schmill, T. Oates, and P. Cohen, Learned models for continuous planning, in Proc. 7th Int. Workshop Artif. Intell. and Statist., 1999, pp. 278282. [41] L. Rabiner and B. Juang, Fundamentals of Speech Recognition. Englewood Cliffs, NJ: Prentice-Hall, 1993. [42] H. Shimodaira, K. I. Noma, M. Naka, and S. Sagayama, Support vector machine with dynamic time-alignment kernel for speech recognition, in Proc. Eurospeech, 2001, pp. 18411844. [43] V. Wan and J. Carmichael, Polynomial dynamic time warping kernel support vector machines for dysarthric speech recognition with sparse training data, in Proc. Interspeech, 2005, pp. 33213324. [44] E. Caiani, A. Porta, G. Baselli, M. Turiel, S. Muzzupappa, F. Pieruzzi, C. Crema, A. Malliani, and S. Cerutti, Warped-average template technique to track on a cycle-by-cycle basis the cardiac lling phases on left ventricular volume, IEEE Comput. Cardiol., vol. 25, no. 98, CH36292, 1998. [45] Y. Lee, Handwritten digit recognition using k nearest neighbour, radial basis function, and back-propagation neural networks, Neural Comput., vol. 3, no. 3, pp. 440449, 1991. [46] T. Golub, D. Slonim, P. Tamayo, C. Huard, M. Gaasenbeek, J. P. Mesirov, H. Coller, M. L. Loh, J. Downing, M. Caligiuri, C. Bloomeld, and E. Lander, Molecular classication of cancer: Class discovery and class prediction by gene expression monitoring, Science, vol. 286, no. 5439, pp. 531537, Oct. 1999. [47] S. M. Weiss, C. Apte, F. J. Damerau, D. E. Johnson, F. J. Oles, T. Goetz, and T. Hampp, Maximizing text-mining performance, IEEE Intell. Syst., vol. 14, no. 4, pp. 6369, Jul./Aug. 1999.
1016
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 37, NO. 6, NOVEMBER 2007
[48] B. Efron, Estimating the error rate of a prediction rule: Improvement on cross-validation, J. Amer. Stat. Assoc., vol. 78, no. 382, pp. 316331, Jun. 1983. [49] A. Molinaro, R. Simon, and R. Pfeiffer, Prediction error estimation: A comparison of resampling methods, Bioinformatics, vol. 21, no. 15, pp. 33013307, 2005.
Wanpracha Art Chaovalitwongse (M05) received the B.S. degree in telecommunication engineering from the King Mongkut Institute of Technology Ladkrabang, Bangkok, Thailand, in 1999 and the M.S. and Ph.D. degrees in industrial and systems engineering from the University of Florida, Gainesville, in 2000 and 2003, respectively. He is currently an Assistant Professor of industrial and systems engineering in the Department of Industrial and Systems Engineering, Rutgers University, Piscataway, NJ, where he does research on optimization, data mining, and statistical analysis in medical problems, multicast routing, and supply chain management. In 2003, he was with the NIH-funded Brain Dynamics Laboratory, Brain Institute, and the Departments of Neuroscience and Industrial and Systems Engineering, University of Florida. Before joining Rutgers University, he was with the Corporate Strategic Research, ExxonMobil Research & Engineering in 2004, where he managed research on developing efcient mathematical models and novel statistical data analyses for upstream and downstream business operations.
Rajesh C. Sachdeo received the B.S. degree in premedicine from Ewing Christian College, Allahabad, India, in 1969 and the M.D. degree from Christian Medical College, Vellore, India in 1973. He had his residency training at V.A. Hines and Loyola University Medical Center, Maywood, IL, in 1979 and his fellowship training in neurophysiology and epilepsy from Rush-Presbyterian St. Lukes Medical Center, Chicago, IL, in 1980. He is currently a Clinical Professor in the Department of Pediatrics, Jersey Shore University Medical Center, Neptune, NJ. He was a Clinical Professor of neurology at St. Peters University Hospital, New Brunswick, NJ, and the Robert Wood Johnson Medical School, University of Medicine and Dentistry of New Jersey, Piscataway, NJ. He is also the Director of the New Jersey Comprehensive Epilepsy Center. Before joining St. Peters, he was an Assistant Professor in medicine at the University of South Florida, Tampa, an Assistant Chief of neurology at V.A. Medical Center, Lyons, NJ, and an Adjunct Faculty of pharmacy and administration at Rutgers University, Piscataway. Dr. Sachdeo is a Fellow of the American Academy of Neurology, American EEG Society, American Epilepsy Society, and Academy of Medicine; an honorary member of the New Jersey Academy of Developmental Medicine; and a member of the American Association of Electromyography and Electrodiagnosis, American Academy of Clinical Neurophysiology, and Southern Neurological Society. He has received numerous awards.
Ya-Ju Fan (S07) received the B.B.A. degree in production and operations management from Fu Jen Catholic University, Taipei, Taiwan, R.O.C., and the M.S. degree in industrial engineering from the University of WisconsinMadison, Madison. She is currently working toward the Ph.D. degree at Rutgers University, Piscataway, NJ. Working with Prof. W. Art Chaovalitwongse, she is engaged in research on optimization and data mining.