International Journal of Innovative Computing, Information and Control ICIC International c 2011 ISSN 1349-4198 Volume 7, Number 12, December 2011 pp. 6905 6918 CLASSIFICATION AND MEDICAL DIAGNOSIS OF SCALP EEG USING ARTIFICIAL NEURAL NETWORKS Mercedes Cabrerizo 1,2, Melvin Ayala 1, Prasanna Jayakar 2 and Malek Adjouadi 1 1 Center for Advance Technology and Education (CATE) College of Engineering and Computing Florida International University 10555 West Flagler Street, Miami, FL 33174, USA 2 Department of Brain Institute Miami Children s Hospital 3100 South West 62nd Avenue, Miami, FL 33155, USA cabreriz@fiu.edu Received August 2010; revised December 2010 Abstract. An automatic Artificial Neural Network-Aided Diagnosis (ANNAD) system is designed in this study for initial scalp EEG screening to establish whether a given subject is epileptic or not. A unique ANNAD-based decision-making process is devised to make this distinction by 1) computing all standard EEG parameters in both time and frequency domain and 2) determining which of these parameters will yield the optimal classifier. The temporal parameters include activity, mobility and complexity, and the frequency parameters include the spectral power in delta, theta, alpha, beta I and II, and gamma. A single layer perceptron was used to conduct this analysis without initially setting any conditions on the weight vector, but rather allowed for the random generation of these conditions, with as many trials as necessary. This is an important first step that confines the search space to only those EEG data that have a very high likelihood of being recorded from epileptic patients, significantly minimizing the time for accurate diagnosis. We have evaluated our system using 125 EEG files selected randomly from a database consisting of 10 subjects (5 non-epileptic and 5 epileptic). The proposed ANNAD system was capable of diagnosing subjects with epilepsy with an accuracy of 92.04% and a calculated F-measure of 93.39%. Keywords: Feature extraction, Artificial neural networks, Epileptic vs. non-epileptic EEG classification 1. Introduction. In epilepsy research, the development of algorithms to automate the detection of epileptiform activity from EEG has become a common line of investigation for many scientists due to its potential for improving disease diagnosis and treatment. An example of such efforts is presented in [1,2], where interictal spikes from EEG recordings are detected by using the Walsh Transform; the study by [3-5] compares different back propagation-based artificial neural networks (ANNs) to evaluate the best features for delineating electrodes that initiate a seizure; [6] proposes an implementation of nonlinear decision functions and identification of multidimensional classification domains to detect seizures; [7] uses EEG recordings during hand motor imagery in order to move a cursor to a target on a computer screen and [8] performs classification of EEG signals extracted during mental tasks for design of a Brain Machine Interface. Furthermore, recent advances in artificial intelligence have benefited from the problemsolving capabilities of ANNs, as exemplified by the classification of sleep apnea syndrome [9] based on wavelet transforms and ANNs; classification of seizures [10] by determining 6905
6906 M. CABRERIZO, M. AYALA, P. JAYAKAR AND M. ADJOUADI the features of temporal epilepsy and the functional anatomy of involved brain networks; epileptic seizure detection [11] by using time-frequency analysis and ANNs; classification of primary generalized epilepsy by EEG signals [12] by using multilayer perceptron neural networks; analysis of EEG signals by implementing eigenvector methods [13] for classification of EEG signals; and classification of CT images for detection of lung cancer by [14] using a modified version of the Hopfield neural network. By means of visual inspection, EEG experts are often able to qualitatively distinguish normal EEG activity from abnormalities contained within long EEG records with the purpose of diagnosing patients with epilepsy. However, abnormalities in the EEG in some patients are at times too subtle to be detected using conventional techniques. It is this dilemma that the proposed classification method is trying to address, by initially confining the search space to only those EEG data that have a very high likelihood of being recorded from epileptic patients. With this development, clinicians can now focus on those patients for further diagnosis based on the presence of spikes and seizure events. The intent then was to establish a system for the classification of key characteristics of EEG signals to differentiate non-epileptic EEG from epileptic EEG. This type of classification analysis can be conducted by using any type of supervised learning algorithm that aims at minimizing the error or risk in the classification, including ANNs, support vector machines, Bayesian classifiers or logistic regression. In this context, our ultimate goal was to find out if there are specific temporal or frequency characteristics in EEG that behave differently in epileptic and non-epileptic patients, placing the focus on an automated search for that one parameter that will best capture these differences. As a consequence, we opted for the single layer perceptron to conduct this analysis. Instead of initially setting any conditions on the weight vector, a random generation of the initial weight was performed, with as many trials as necessary. A flowchart of the steps of the proposed methodology is illustrated in Figure 1. This is based on establishing a unique descriptor matrix for any given EEG parameter to be simple in its implementation and yet robust in the anticipated outcome. The EEG parameters were extracted from EEG segments, where the seizure events were initially removed in order not to have any bias. The algorithm for computing the matrix involves statistical operations on parameters across electrodes and time. The matrix used was limited to the average (AVG), standard deviation (STD) and signal-to-noise ratio (SNR). The advantage of this matrix is that it can be used to represent EEG files of different durations and with a different number of electrodes used. The parameters investigated in this study were the Hjorth s parameters activity, mobility and complexity [15,16], as well as the spectral power in all frequency bands. The matrices were computed for all EEG files, and were then used as inputs to the ANN that was trained to classify these EEG recordings as emanating either from non-epileptic or from epileptic patients. The neural network configuration is illustrated in Figure 2. The analysis was performed on each of the aforementioned parameters across time in search of those EEG characteristics that can best serve this classification process. The main contribution of this study thus relies on the classification of off-line scalp EEG recordings in order to facilitate the screening process of potential epileptic patients. The classification of EEG signals related to epilepsy is difficult in the absence of seizures because one can only rely on the existence of abnormal behaviors in the EEG, such as interictal spikes, which represent an area of cortical irritability that may or may not be predisposed to produce epileptic seizures at some point. These interictal spikes could be seen in normal subjects who never develop epilepsy, and could be related to other neurological disorders that may or may not be associated with epilepsy at all. In addition,
CLASSIFICATION AND MEDICAL DIAGNOSIS OF SCALP EEG USING ANNS 6907 Figure 1. Flowchart of the working steps of the proposed classification strategy around 10% of epileptic patients never show evident epileptiform discharges. Other behaviors, such as abnormalities of background activity and slow activity in some regions, are much less evident than epileptiform activity and cannot be easily appreciated by looking at the EEG [17]. That is why long EEG recordings require the use of reliable and accurate computer programs that are able to extract the hidden information, so that a better diagnosis can be provided. In addition, automatic signal processing is almost indispensable when these long recordings have to be evaluated. Visual screening only relies on the EEG experts that can draw subjective conclusions at times. In this context, it is useful to automatically differentiate between the two main categories of EEG signals: epileptic EEG and non-epileptic EEG, just by using short segments of scalp interictal EEG without having to record the patient for a long time until an event associated with epilepsy happens. This last assertion constitutes the main aim of this study. 2. Materials and Methods. 2.1. Data acquisition. Scalp EEG recordings from 5 non-epileptic patients and 5 epileptic patients were used. The data were collected from routine EEG recordings from patients who are epileptic or non-epileptic, without the imposition of long recording sessions. For an unbiased comparison of files, the segments from the epileptic patients were extracted from sections without seizures which may or may not contain abnormal
6908 M. CABRERIZO, M. AYALA, P. JAYAKAR AND M. ADJOUADI Figure 2. ANN configuration discharges. Recordings were performed at Miami Children s Hospital, using XLTEK Neuroworks Ver. 3.0.5 equipment and sampled at 512 Hz and 500 Hz for non-epileptic and epileptic patients, indistinctly. Electrodes were setup following the 10-20 system using a referential montage. In total, 5 to 30 random segments were extracted from all EEG recordings; to prevent biasing, all segments were free of artifacts. The file segments from epileptic patients were only interictal (i.e., without seizure activity). All files were 4 to 45 seconds long. An overview of the patients information is given in Table 1. In retrospect, the method focused on investigating 9 well-known EEG parameters, namely activity, mobility, complexity and spectral power in the frequency bands delta (< 4 Hz), theta (4 to 8 Hz), alpha (8 to 13 Hz), beta I (13 to 20 Hz), beta II (20 to 36 Hz) and gamma (36 to 44 Hz). Frequency parameters were added to the investigation, as they were proven in earlier studies to be of high relevance for pre-surgical evaluation [18-21]. The 44 Hz was set as the limit in this study since spectral power in all file segments collected was found to be very low for frequencies above it. 2.2. Development and structure of the descriptor matrix. It should be noted that the proposed method focuses on the extraction of parameters from EEG that has been a common practice in epilepsy research with applications ranging from interictal spike detection, seizure detection, 3-D source localization, and prediction, among others. What this paper proposes is to combine some of these parameters in a unique way not found previously in the literature to classify EEG files into seizure and non-seizure files. The proposed combination of statistical measurements, called descriptors, is a matrix that consolidates common parameters (such as Horth s and frequency power parameters) into simple but effective neural networks topologies. The uniqueness of these matrices is that they can be extracted from EEG recordings of different durations and with a different number and location of electrodes, and still be used as inputs to a neural network for
CLASSIFICATION AND MEDICAL DIAGNOSIS OF SCALP EEG USING ANNS 6909 Table 1. Patients information and number of files used in the study Patient Status Age Gender Diagnosis Number Sampling of files rate (Hz) 1 Non-epileptic 15 Female 10 512 2 Non-epileptic 10 Male 15 512 3 Non-epileptic 17 Male 20 512 4 Non-epileptic 18 Female 5 512 5 Non-epileptic 8 Male 5 200 6 Epileptic 7 Male Focal area of epileptogenesis within the left frontal region involving the frontal pole and posterior frontal lobe 20 512 near the midline region primarily. 7 Epileptic 4 Female Intractable seizures over the right frontocentro-temporal 30 500 epileptogenicity. 8 Epileptic 15 Female Medically intractable seizures with epileptogenicity involving 10 500 both frontal regions. 9 Epileptic 8 Female Intractable epilepsy since age 15 months. Electrographically, onset is within the left anterior temporal lobe prior to spread anteriorly to the bilateral frontal regions. 5 512 Diffuse cerebral dysfunction 10 Epileptic 11 and a region of epileptic ac- Male Months tivity over the left posterior temporal area. 5 512 assessing the likelihood that the subject is epileptic or not. This constitutes the main contribution of the method proposed. As a consequence of this work, the processes for interictal spike detection, 3-D source localization, and ultimately diagnosis are much more enhanced to mitigate the cause and effect of seizures by focusing more appropriately on the delineated seizure-prone files for more timely and accurate diagnosis. The method for parameter extraction in the time domain when different electrodes are present is illustrated in Figure 3. For each time window, the entire window recording is replaced by a particular parameter of interest. In the example, Hjorth s parameter activity is computed for each electrode, using time windows of 1 second. As can be seen from Figure 3(a), extracting parameters electrode by electrode has the disadvantage of generating a set of values which depend on the number of electrodes. To cope with this issue, this study used statistical features, such as average, standard deviation, and signal-to-noise ratio, and computed them across all electrode parameters for each time window, as depicted in Figure 3(b), where the grand average across all electrodes is used as a representative feature. However, since the purpose was to analyze and compare EEG recordings regardless of their duration, further statistical feature extraction was performed along the time axis to provide a single feature descriptor for the whole EEG file. Figure 3(c) illustrates this last step using the grand average of the inter-electrode average activity. The statistical features chosen in this investigation were the AVG, STD and SNR. All steps of the feature extraction procedure are given in Figure 4, which depicts how an entire EEG segment can be described by a matrix obtained from each of the 9 parameters. The
6910 M. CABRERIZO, M. AYALA, P. JAYAKAR AND M. ADJOUADI Figure 3. (a) Extraction of EEG activity A(Ek) from 1-second windows; (b) extraction of inter-electrode average; (c) extraction of grand average over time of the inter-electrode average convention that will be followed in explaining this method is to use uppercase for the inter-electrode features, and lowercase for time series features. Using three features across electrodes and another three for each time series resulted in a total of 9 features, represented for convenience in matrix form in Table 2, starting with the avg(avg), i.e., the average over time (avg) of the inter-electrode average (AVG), and ending with snr(snr) with similar reasoning. This descriptor has 2 dimensions because it contains information consolidated across electrodes and time. Table 2. Descriptor matrix of an EEG segment for a particular parameter avg std snr AVG avg(avg) std(avg) snr(avg) STD avg(std) std(std) snr(std) SNR avg(snr) std(snr) snr(snr) Following the approach with the steps that were described in Figures 3 and 4, a descriptor matrix for each of the 9 parameters (as reflected in Table 2) was computed for each EEG segment. The intent was to represent each data segment by 9 descriptor matrices, as one matrix per parameter. It is worthy to note that the descriptor matrices for all 125 files and for each parameter were obtained using a general window size of 1 second for each file, and consequently, a classifier for each parameter was created using the different descriptor matrices. 2.3. Analysis on linear separability. Before deciding on the training strategy for the classifier, we investigated if all the data collected was linearly separable in the input space. The overlap index between the two classes that we used was applied to each descriptor dimension individually, and was computed as: β = C P N + C N P C t (1) where C t is the total number of cases in the study, C P N is the number of positives cases (epileptic) whose values overlap with the negative cases (non-epileptic) and C N P is the number of negative cases whose values overlap with the positive cases. The overlap index
CLASSIFICATION AND MEDICAL DIAGNOSIS OF SCALP EEG USING ANNS 6911 Figure 4. Proposed approach to feature extraction from EEG data segments, reduced to a 3 3 matrix obtained from the common statistical features avg, std and snr was defined so as to yield values between 0 and 1. The overlap indexes obtained for each descriptor component are shown in Table 3. In this table, the lowest index found is 0.432 (for descriptor components avg(avg) of complexity, which is an indication of overlap. Within this context, the overlap index values provided in Table 3 clearly show that all 9 dimensions present overlap of the two populations when used individually. This analysis was extended to using these dimensions 2 at a time and 3 at a time, assuming all the possible combinations. Each time, the overlap varied in extent with no optimal solution to be found. This is the reason we opted for the neural network to contend with the 9 dimensions, as expressed in Equation (2). For this purpose, we used a linear classifier consisting of a perceptron without hidden layers, and identity transfer function in order to simulate a linear output for the 9 features of the form: y = w 0 + w 1 x 1 + w 2 x 2 +... + w 9 x 9 (2)
6912 M. CABRERIZO, M. AYALA, P. JAYAKAR AND M. ADJOUADI Table 3. Overlap index of descriptor components between all epileptic and non-epileptic cases per parameter Activity Mobility Complexity Delta Theta Alpha Beta I Beta II Gamma avg std snr AVG 0.776 0.944 0.688 STD 0.768 0.752 0.824 SNR 0.760 0.776 0.552 AVG 0.592 0.720 0.760 STD 0.960 0.960 0.728 SNR 0.664 0.680 0.640 AVG 0.432 0.712 0.760 STD 0.984 0.984 0.736 SNR 0.712 0.784 0.760 AVG 0.672 0.664 0.832 STD 0.832 0.960 0.712 SNR 0.728 0.656 0.728 AVG 0.520 0.552 0.808 STD 616 0.624 0.976 SNR 0.944 0.904 0.872 AVG 0.784 0.792 0.968 STD 752 0.704 0.784 SNR 0.952 0.864 0.800 AVG 0.728 0.712 0.936 STD 744 0.744 0.896 SNR 0.912 0.864 0.912 AVG 0.944 0.816 0.616 STD 752 0.656 0.816 SNR 0.856 0.904 0.960 AVG 0.840 0.688 0.872 STD 752 0.784 0.968 SNR 0.920 0.864 0.984 where w 0 to w 9 are the weights of the classifier, and x 1 to x 9 correspond to avg(avg) to snr(snr), as given in Table 2. Likewise, we used all available data to fully train that classifier, using the delta rule of learning. Training was performed in repeated trials for a long period of time (maximum 1 hour), but the network never converged to 100% accuracy. This fact allowed us to conclude that the data was not linearly separable in the input space. The analysis of separability was further extended by using different perceptron configurations (adding 10 to 20 hidden units in one and two hidden layers and logistic and radial basis transfer functions); however, the classification performance was not significantly improved on average. Therefore, we decided to simplify the analysis and use a linear classifier instead without hidden units. 2.4. Classification algorithm. The architecture of the proposed method is formed by four steps: 1. Collection of descriptor data; 2. Generation of the training and testing tables; 3. Training and testing of the classifiers; 4. Evaluation of the classifiers.
CLASSIFICATION AND MEDICAL DIAGNOSIS OF SCALP EEG USING ANNS 6913 In the first step, we extracted from each file 9 descriptor matrixes (one for each parameter). For a total of 125 file segments, we generated 1125 3 3-matrixes. However, we conducted the analysis using one parameter at a time, i.e., we used only 125 3 3-matrixes in each classifier. In the second step, we generated the training and testing tables using the extracted descriptor matrixes. To that end, we partitioned the new set of matrixes into approximately 40% (52 files) for pure training, 10% (13 files) for cross-validation and 50% (60 files) for testing. All tables were structured using the 9 matrix elements as inputs, and one additional column for the target ( 1 for non-epileptic file and 1 for epileptic file). In the third step, we created 9 two-layer perceptrons, each with 9 inputs and 1 output. As part of the training, 15 trials were performed on each classifier in order to prevent the network from being trapped in a local minimum, or leading to overflow. After each trial, the set of weights that yielded the best solution was stored. The intention of repeating the training was to rule out the possibility that any differences in the results were obtained just by chance. All training repetitions were stopped by either cross-validation stop criterion or by reaching a time limit. This procedure was applied consistently across all 9 classifiers. In the last step, all classifiers were compared based on their performances. The computational time it took to train the neural network through the use of the descriptor matrix was close to 5 minutes, while the testing phase when an EEG file is classified as coming from of an epileptic or a non-epileptic control subject took 40 to 50 seconds. From all available measures, such as accuracy, sensitivity, specificity, precision and F- measure, we decided to use the F-measure as the selection criteria. The best classifier was defined as the one with the highest F-measure. This measure represents a compromise which combines precision and sensitivity, as follows: ( ) precision sensitivity F = 2 (3) precision + sensitivity Recall that precision itself embeds within its definition specificity, making Equation (3) an appropriate measure that takes into consideration both the undesirable false negatives, as well as the more tolerable false positives. 3. Results. There are five key steps that were considered in the implementation of the proposed method: 1. Determine the overlap index of the 9 descriptor components between all epileptic and non-epileptic cases per each parameter in order to assess linear separability. 2. Use the F-measure to evaluate the performance of the 9 different ANNs. The choice of the F-measure is intentional as it includes the critical measures of both false positives and false negatives. 3. Test the results using the classifier obtained from each parameter to evaluate the performance of the 9 classifiers based on ROC parameters. 4. Apply the two-sample t-test assuming unequal variances to compare the performance metrics of the activity networks with the performance metrics of the remaining parameters. 5. Show the 3 best performing parameters based on the five performance metrics used (accuracy, sensitivity, specificity, precision and F-measure). These five steps are taken to ensure the soundness of the proposed method. The discussion about the suitability of these descriptors will focus on a statistical analysis of the classification results obtained across all sets. After training the 9 networks 15 times, all 125 results were sorted in descending order of the F-measure, which considers
6914 M. CABRERIZO, M. AYALA, P. JAYAKAR AND M. ADJOUADI Table 4. Performance of the 9 different ANNs based on the F-measure F-measure Activity 0.93388850 Alpha 0.86578947 Beta1 0.80933984 Gamma 0.80067610 Beta2 0.77039872 Theta 0.74367829 Mobility 0.74201213 Delta 0.66126537 Complexity 0.54736815 both precision and sensitivity. As can be seen from Table 4, the best parameter that resulted from testing the networks based on F-measure was activity, defined in Equation (4) in the way it was implemented for convenience: activity = σ 2 (input(t)) (4) where input(t) is the EEG signal for a window of 1 second. In general, ranking showed predominance of networks associated with activity. A summary of the best performance of the networks associated with each of the 9 parameters is given in Table 5. From this table, it can be observed that the alpha frequency power outperformed activity in sensitivity; however, it performed poorly in accuracy, and even worst in specificity. As an average, with the exception of the 85.33% obtained for specificity, activity proved to have acceptable values for accuracy and F-measure, which makes it the most suitable parameter for the type of classification proposed in this study. Table 5. Performance evaluation of the 9 classifiers based on the testing set Accuracy Sensitivity Specificity Precision F-measure Activity 0.92037037 0.96825397 0.85333334 0.90355970 0.93388850 Mobility 0.74444445 0.85396825 0.59111111 0.66559010 0.74201213 Complexity 0.57777778 0.61587302 0.52444444 0.55759566 0.54736815 Delta 0.59074074 0.84761905 0.23111111 0.59310606 0.66126537 Theta 0.61481481 0.95873016 0.13333333 0.60750892 0.74367829 Alpha 0.81666666 1.00000000 0.56000000 0.76481482 0.86578947 Beta1 0.73518519 0.95238095 0.43111111 0.70626780 0.80933984 Beta2 0.65555555 0.98412698 0.19555555 0.63478702 0.77039872 Gamma 0.72222222 0.95238095 0.40000000 0.69171380 0.80067610 In order to rule out the possibility that the higher performance of the activity-based network was obtained just by chance, a t-test was performed. The test was conducted for the F-measure values. For this purpose, two data sets were created, one including the F-measure of the 15 training/testing trials conducted with the activity network (recall that it was trained 15 times), and another including the performance metric of the remaining 120 trials (8 parameters 15 training/testing trials). The results of these tests are provided in Table 6. The two sample mean values (variance) are 0.9338 (0.000552) and 0.742566 (0.038313). The p-value was p = 1.5 10 18. Since the p-value is less than the 0.05 confidence range, and the obtained t-value exceeds the critical value, this provides sufficient substantiation to reject the null hypothesis of equal means. It was concluded
CLASSIFICATION AND MEDICAL DIAGNOSIS OF SCALP EEG USING ANNS 6915 that the performance of activity was not due to chance, and that activity can indeed be regarded as a more reliable parameter for EEG classification. Table 6. Results of the two-sample t-test assuming unequal variances: comparing performance metrics of activity with performance metrics of remaining parameters T-Test: Two-Sample Assuming Unequal Variances Variable 1 Variable 2 (Activity) (remaining parameters) Mean 0.933888 0.742566 Variance 0.000552 0.038313 Observations 15 120 Hypothesized Mean Difference 0 df 133 t Stat 10.13912 P(T<=t) one-tail 1.5E-18 t Critical one-tail 1.656391 After applying these descriptors to both the time and frequency domains, as provided in Table 7, the method proved that activity was indeed by far the best parameter in our dataset to differentiate EEG of non-epileptic controls from epileptic patients. Table 7. Three best performing parameters based on the five performance metrics used (accuracy, sensitivity, specificity, precision and F-measure) Best Performing Classifiers Accuracy activity (92.04%), alpha (81.67%), mobility (74.44) Sensitivity alpha (100%), beta II (98.41%), activity (96.83%) Specificity activity (85.33%), mobility (59.11%), alpha (56.00%) Precision activity (90.36%), alpha (76.48%), beta I (70.63%) F-measure activity (93.39%), alpha (86.58%), beta I (80.93%) For a clinical interpretation of the results, we finally turned our attention to the original dataset and computed an average activity descriptor for all epileptic and non-epileptic cases. Since the SNR dimension was much higher than the other dimensions (Figure 5(a)), it was removed simply for visual appreciation of the other dimensions. Figure 5(b) thus shows that all components in the activity descriptor are on average much higher in the epileptic cases, which is the clinical expectation in epileptic patients who are prone to having chaotic and disorganized EEG (abnormal discharges) behavior. 4. Conclusion. In this study, an automated classification process of scalp EEG for epilepsy diagnosis is implemented and validated. The uniqueness of the method is that it proved that a new descriptor matrix could be extracted from a scalp EEG segment in order to associate it with either a non-epileptic or an epileptic patient. Furthermore, such a descriptor can be used to detect signs of epilepsy in EEG, even in the absence of abnormal discharges. The performance of the proposed neural network configuration was evaluated in terms of the F-measure, and the results confirmed that the proposed network structure has great potential in classifying epileptic from non-epileptic EEG signals. When training
6916 M. CABRERIZO, M. AYALA, P. JAYAKAR AND M. ADJOUADI (a) (b) Figure 5. Averaged activity descriptor for epileptic and non-epileptic subjects
CLASSIFICATION AND MEDICAL DIAGNOSIS OF SCALP EEG USING ANNS 6917 each classifier 15 times with cross-validation, the network corresponding to activity performed best on the testing set, yielding 92.04% accuracy, 96.83% sensitivity and 93.39% F-measure. The alpha classifier ranked second, with an average F-measure of 86.58%. The t-test proved further that the superiority of activity was statistically significant. This method constitutes a first attempt toward a solution to the EEG classification problem that would delineate EEG segments of epileptic patients from EEG segment of non-epileptic controls. Another interesting aspect of the proposed method is that is easily extendable to include other features of interest such as coherence or entropy. And since duration of the EEG files is not an issue in the way the algorithm is designed, it would be worthwhile to also investigate window sizes greater than 1 second to see how they could affect the results in terms of both processing time and accuracy in the classification results. From a clinical point of view, as more patients are added to the analysis, the inputs to the neural network could be fine-tuned to seek higher accuracy and more stable inputs across large datasets. Likewise, sampling frequencies higher than 512 Hz are also of interest to see if the spectrum reveals high frequency powers which could be of value for this type of classification. Based on this line of thought, these initial empirical results that were obtained lead us to believe that some course of action could be considered to address the complex problem of seizure prediction from a different research perspective, by initially focusing on what delineates EEG files that may lead to seizure (as recorded from epileptic patients) from EEG files that are recorded from non-epileptic patients. The obtained results look very promising considering the number of patients that were used in this study. However, we expect to augment the performance of the classifier by adding more patients. Further improvements of the study could be achieved by adding all parameters together as the input to the neural network for more accurate results. Acknowledgments. The authors appreciate the support provided by the National Science Foundation under grants CNS-0959985, CNS-1042341 and HRD-0833093. The authors are also grateful for the clinical support provided through the Ware Foundation and the joint Neuro-Engineering Program with Miami Children s Hospital. REFERENCES [1] M. Adjouadi, M. Cabrerizo, M. Ayala, D. Sanchez, P. Jayakar, I. Yaylali and A. Barreto, A detection of interictal spikes and artifactual data through orthogonal transformations, J. Clin. Neurophysiol, vol.22, no.1, pp.53-64, 2005. [2] M. Adjouadi, M. Cabrerizo, M. Ayala and N. Mirkovic, Seizing lesions in 3-D, IEEE Potentials, vol.24, no.5, pp.11-17, 2005. [3] M. Cabrerizo, M. Adjouadi, M. Ayala and P. Jayakar, Subdural interictal EEG analysis for extracting discriminating features towards electrode classification using artificial neural networks, in Brain Mapping Research Progress, I. C. Girard and J. S. Andre (eds.), Hauppauge, New York, NOVA Science Publishers, 2009. [4] M. Cabrerizo, M. Adjouadi and M. Ayala, An application of eigensystem and frequency analysis in brain functional mapping, in Progress in Brain Mapping Research, F. J. Chen (ed.), Hauppauge, New York, NOVA Science Publishers, 2006. [5] M. Cabrerizo, M. Adjouadi, M. Ayala, K. Nunez, P. Jayakar and I. Yaylali, Integrated study of topographical functional based on an auditory-comprehension paradigm using an eigensystem study and spectrum analysis, Brain Topogr, vol.17, no.3, pp.151-163, 2005. [6] M. Tito, M. Cabrerizo, M. Ayala, A. Barreto, I. Miller, P. Jayakar and M. Adjouadi, Classification of electroencephalographic seizure recordings into ictal and interictal files using correlation sum, Comp. Biol. Med., vol.39, no.7, pp.604-614, 2009.
6918 M. CABRERIZO, M. AYALA, P. JAYAKAR AND M. ADJOUADI [7] C. R. Hema, M. P. Paulraj, R. Nagarajan, S. Yaacob and A. H. Adom, Pattern recognition of EEG signal during motor imagery by using SOM, International Journal of Innovative Computing, Information and Control, vol.4, no.10, pp.2617-2630, 2008. [8] C. R. Hema, M. P. Paulraj, R. Nagarajan, S. Yaacob and A. H. Adom, Brain machine interface: A comparison between fuzzy and neural classifiers, International Journal of Innovative Computing, Information and Control, vol.5, no.7, pp.1819-1827, 2009. [9] N. Sezgin and M. E. Tagluk, Energy based feature extraction for classification of sleep apnea syndrome, Comp. Biol. Med., vol.9, no.11, pp.1043-1050, 2009. [10] L. Vaugier, S. Aubert, A. McGonigal, A. Trebuchon, M. Guye, M. Gavaret, J. Regis, P. Chauvel, F. Wendling and F. Bartolomei, Neural networks underlying hyperkinetic seizures of temporal lobe origin, Epilepsy Res., vol.86, no.2-3, pp.200-208, 2009. [11] A. T. Tzallas, M. G. Tsipouras and D. I. Fotiadis, Epileptic seizure detection in EEGs using timefrequency analysis, IEEE Trans. Information Technology in Biomed., vol.13, no.5, pp.703-710, 2009. [12] S. N. Ogulata, C. Sahin and R. Erol, Neural network-based computer-aided diagnosis in classification of primary generalized epilepsy by EEG signals, J. Med. Syst., vol.33, no.2, pp.107-112, 2009. [13] E. D. Ubeyli, Analysis of EEG signals by implementing eigenvector methods/recurrent neural networks, Digital Signal Processing, vol.19, no.1, pp.134-143, 2009. [14] S. Rachid, J. Hassan and M. Sammouda, CT image analysis for early detection of lung cancer, International Journal of Innovative Computing, Information and Control, vol.4, no.11, pp.2847-2860, 2008. [15] B. Hjorth, EEG analysis based on time domain properties, Electroenceph. Clin. Neurophysiol., vol.29, no.3, pp.306-310, 1970. [16] B. Hjorth, The physical significance of time domain descriptors in EEG analysis, Electroenceph. Clin. Neurophysiol., vol.34, no.3, pp.321-325, 1973. [17] S. J. M. Smith, EEG in the diagnosis, classification, and management of patients with epilepsy, J. Neurol Neurosurg Psychiatry, vol.76, pp.ii2-ii7, 2005. [18] R. S. Fisher, W. R. Webber, R. P. Lesser, S. Arroyo and S. Uematsu, High-frequency EEG activity at the start of seizures, J. Clin. Neurophysiol., vol.9, no.3, pp.441-448, 1992. [19] E. Urrestarazu, R. Chander, F. Dubeau and J. Gotman, Interictal high-frequency oscillations (100-500 Hz) in the intracerebral EEG of epileptic patients, Brain, vol.130, no.9, pp.2354-2366, 2007. [20] G. A. Worrell, T. D. Lagerlund, F. W. Sharbrough, B. H. Brinkmann, N. E. Busacker, K. M. Cicora and T. J. O Brien, Localization of the epileptic focus by low-resolution electromagnetic tomography in patients with a lesion demonstrated by MRI, Brain Topogr, vol.12, no.4, pp.273-282, 2000.