This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2015.2424074, IEEE Journal of Biomedical and Health Informatics

JBHI-00674-2014

1

Epileptic Seizure Detection Based on Partial Directed Coherence Analysis Gang Wang*, Member, IEEE, Zhongjiang Sun, Ran Tao, Kuo Li, Gang Bao, Xiangguo Yan  Abstract—Long-term video EEG epilepsy monitoring can help doctors diagnose and cure the epilepsy. The workload of doctors to read the EEG signals of epilepsy patients can be effectively reduced by automatic seizure detection. The application of partial directed coherence (PDC) analysis as mechanism for feature extraction in the scalp EEG recordings for seizure detection could reflect the physiological changes of brain activity before and after seizure onsets. In this study, a new approach on the basis of PDC was proposed to detect the seizure intervals of epilepsy patients. First of all, the multivariate autoregressive (MVAR) model was established for a moving window and the direction and intensity of information flow based on PDC analysis was calculated. Then, the outflow information related to certain EEG channel could be obtained by summing up the intensity of information flow propagated to other EEG channels in order to reduce the feature dimensionality. At last, according to the pathological features of epileptic seizures, the outflow information was regarded as the input vectors to a support vector machine (SVM) classifier for discriminating interictal periods and ictal periods of EEG signals. The proposed method had achieved a good performance with the correct rate of 98.3%, the selectivity rate of 67.88%, the sensitivity rate of 91.44%, the specificity rate of 99.34% and the average detection rate of 95.39%, which demonstrated that this method was suitable for detecting the seizure intervals of epilepsy patients. By comparing with other existing techniques, the proposed method based on PDC analysis achieved significant improvement in terms of seizure detection. Index Terms—partial directed coherence, support vector machine, seizure detection, information flow, cross validation

T

I. INTRODUCTION

HE epileptic seizure is a transitory cerebral disorder caused by the paroxysmal abnormality and the excessive electrical activities of brain neurons [1]. The electroencephalogram (EEG) is a non-invasive tool for epileptic seizure detection. The long-term video EEG epilepsy monitoring can provide the information about cerebral activities and the frequencies of epileptic seizures, which is helpful for clinical diagnosis and Manuscript received November 13, 2014. This work was supported by the National Natural Science Foundation of China under Grant no. 81201162, no. 61471291, the Natural Science Basic Research Plan in Shaanxi Province of China under Grant No. 2013JQ8007, and the Fundamental Research Funds for the Central Universities of China. Asterisk indicates corresponding author. * G. Wang, Z. Sun, R. Tao, and X. Yan are with the Key Laboratory of Biomedical Information Engineering of Ministry of Education, Institute of Biomedical Engineering, School of Life Science and Technology, Xi’an Jiaotong University, Xi’an 710049, China (e-mail: [email protected]). K. Li and G. Bao are with the Department of Neurosurgery, First Affiliated Hospital, Xi’an Jiaotong University, Xi’an 710061, China.

treatment of epilepsy patients. The visual inspection of experienced experts is required to confirm the seizure intervals, which is a weary task. The diagnosis results depend much on the experience and the subjective judgment. Hence, it is necessary for the automatic detection technique to recognize the seizure intervals of epilepsy patients. The development of methodologies to extract the effective features still remains as a primary challenge for detecting the seizure intervals. Many attempts have been made for epileptic seizure detection. In time domain, linear prediction error energy, relative amplitude, and rhythmicity were reported as efficient ways to differentiate between normal and epileptic EEG signals [2, 3]. In frequency domain, power density spectrum was extracted to perform the automatic identification of epileptic and background EEG signals [4, 5]. In time-frequency domain, short-time Fourier transform, wavelet transform and wavelet packet transform were demonstrated to have the potential for the non-invasive detection of epileptic seizures [6-8]. Additionally, the Hilbert-Huang transform as an adaptive model was applied to recognizing seizure activities from the non-seizure ones for patients with medically resistant epilepsy [9, 10]. Recently, a great number of investigations have applied to the nonlinear features of EEG signals. Lyapunov exponents, correlation dimension and recurrence quantification analysis were employed for the detection of epileptic seizures after wavelet analysis [11-13]. A patient-specific method based on fractal dimension and gradient boosting was presented to discriminate the EEG signals [14]. Kolmogorov entropy, spectral entropy, approximate entropy and fuzzy entropy were applied to EEG data for detecting whether there is an epileptic attack [15-17]. However, further improved detection accuracy would benefit the treatment and diagnosis of epilepsy patients. To the best of our knowledge, this is the first experimental study to consider the exchange of information flow between brain regions for the seizure detection by regarding brain activities as an organic whole. To this end, the partial directed coherence (PDC) analysis was used to extract the direction and intensity of information flow between brain regions. The PDC algorithm could describe the Granger causality in the frequency domain to analyze the causal relationship between EEG channels [18]. In this study, from the perspective of the application of PDC analysis as mechanism for feature extraction in the scalp EEG recordings for epileptic seizure detection, a new approach based on PDC was proposed to detect the seizure intervals of epilepsy patients.

2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2015.2424074, IEEE Journal of Biomedical and Health Informatics

JBHI-00674-2014

2

II. MATERIALS AND METHOD A. Patients and EEG Data acquisition Ten patients with medically intractable partial epilepsy were studied using a protocol approved by the Institutional Review Boards of Xi'an Jiaotong University. The patients reported are consecutive patients selected according to the following inclusion criteria: (1) ictal activities were recorded by long-term video EEG epilepsy monitoring, (2) patients underwent respective surgery after EEG monitoring, and (3) there were MRI visible putative epileptogenic lesions in the pre-operative MRIs. All patients in this study signed their consent forms for data acquisition and subsequent analysis of their EEG-recordings. The patients were admitted to the epilepsy monitoring unit at the First Affiliated Hospital of Xi'an Jiaotong University and underwent noninvasive presurgical evaluations that included a seizure protocol structural MRI, and video scalp EEG epilepsy monitoring. The location of epileptogenic foci for each patient was identified by experienced epileptologists using high resolution anatomical MRI and ictal scalp EEG. Each patient had a resection of the epileptogenic zone, and follow-up at 6 months or longer after the surgical resection. Following surgical resection, all patients were seizure free. TABLE I CLINICAL INFORMATION FOR TEN EPILEPSY PATIENTS

Patient

Sex

No. of seizures

Seizure duration (s)

Data length (min)

1

M

9

302

270

Seizure origin Right frontal lobe and central region

2

M

8

1680

206

Frontal lobe

3

F

5

229

612

Right central region

4

M

5

340

900

Right frontal lobe and temporal lobe

5

M

4

455

424

Absence seizure

6

M

3

305

38

Right frontal lobe

7

M

3

498

2632

Left front and middle temporal lobe

8

M

3

299

721

Left temporal lobe

9

M

2

83

735

10

F

2

152

771

Frontal lobe and temporal lobe Left and middle temporal lobe

The scalp EEG data were acquired with the EEG-1100 digital video-EEG system produced by NIHON KOHDEN Company from 19 scalp electrodes placed according to the standard international 10-20 system. The channel CZ at the top of the head was used as the reference electrode. Differential amplifiers with band-pass filters between 0.5 and 60 Hz were used to minimize the effects of high frequency noise and low frequency artifacts. The sampling rate of signals was 200 Hz. During EEG data recording, the patients were asked to stay at a relaxed state with eyes closed and tried to avoid unnecessary body movements. Additionally, some abnormal EEG waves such as spikes and sharp waves could be observed clearly in EEG signals and the clinical seizures of epilepsy patients could also be watched in the video surveillance. The seizure onset and

offset timepoints were determined by two experienced epileptologists based on the identification of EEG recordings and clinical manifestations. The length of the entire EEG data was 120.6 hours including 44 seizures in total. The average length of all ictal periods was 98.7 seconds. The number of seizures for different patients ranged from 2 to 9. Table I shows the detailed clinical information of ten epilepsy patients. B. Analysis paradigm The analysis flowchart of seizure detection algorithm utilized in the present study is illustrated in Fig. 1. For each patient, after the EEG data were screened to eliminate some artifacts containing much EMG interference caused by large body movements through the entire recording time, the EEG data were firstly segmented according to the length of analysis window. The multivariate autoregressive (MVAR) model was established for each moving window. Subsequently, the direction and intensity of the information flow between brain regions were calculated through the PDC algorithm. Secondly, in order to reduce the feature dimensionality, the intensity of information flow propagated to other EEG channels was summed up to obtain the outflow information related to certain EEG channel. In the end, according to the pathological features of epileptic seizures, the outflow information of each channel was used as the inputs to a SVM classifier for distinguishing interictal periods and ictal periods of EEG signals. The performance of proposed algorithm was evaluated by using 5-fold cross-validation method in order to avoid the deviation caused by multiple seizures of single patient. 1) Feature extraction: The scalp EEG signals were band-pass filtered between 0.5 and 60 Hz in this experiment. To resolve the lowest end of this frequency band range, the detection algorithm should extract the features from the EEG segments of over 2 seconds. Hence, the length of analysis window was selected as 2 s in this study. Then, the features were extracted for each EEG segment. The PDC analysis based on the estimation of coefficients of MVAR model would construct a directed network which could both reflect the association strength between EEG channels and characterize the direction of information flow. The N channels of EEG signals at time t can be defined as T a vector X (t )  [ X1 (t ), X 2 (t ),..., X N (t )] , where T denotes the

transpose of matrix and X n (t )(n  1, 2,...N ) is the nth channel of EEG signals. For each EEG segment, a MVAR model with p order is built as follows: p

X (t )   Ar X (t  r )  E (t )

(1)

r 1

where p is the order of MVAR model, matrix Ar with size 19×19 can be calculated by Levison-Wiggins-Robinson (LWR) algorithm [19] and E (t ) denotes the estimate error which is unrelated white noise sequence with zero mean. The fitting results of MVAR model will be influenced by the choice of model order p , which can be determined by the Akaike Information Criterion (AIC) [20]. Then, by taking the Fourier transform of coefficients of

2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2015.2424074, IEEE Journal of Biomedical and Health Informatics

JBHI-00674-2014

3

Fig. 1. A diagram outlining the PDC-based method utilized in the present study. p

MVAR model, A( f ) can be obtained by A( f )   Ar e 2 irf . r 1

The transfer function for the N channels of EEG signals can be represented by

A( f )  1  A( f )  [a1 ( f ), a2 ( f ),..., aN ( f )] ,

where the elements of A( f ) is defined as p  2 i rf ,i  j 1  aij  r  e .  r 1 Aij  f    p   a  r  e 2 i rf , i  j ij   r 1

(2)

Finally, the information flow can be calculated by [18]:

PDC j i ( f )  Aij ( f )

T

a j ( f )a j ( f )

(3)

where a j ( f ) is the jth column of matrix and PDC j i ( f ) represents the intensity and direction of information flow from channel j to channel i at frequency f . 2) Feature Dimensionality Reduction: Because the sampling rate of EEG signals was 200Hz, the PDC values between any two channels fell within the 0-100 Hz frequency band. After the PDC values from channel j to channel i (i  1, 2,...,19; j  1, 2,...,19) at frequency f were obtained for each EEG segment, the square of all PDC values between 0 and 100 Hz were summed up to obtain the intensity of information flow from channel j to channel i . Then, the 19×19 matrixes were extracted from the original EEG signals. It was still hard to discriminate interictal and ictal periods of EEG segments, because there were much computation complexity and bad results of seizure detection when directly dealing with these PDC matrixes. Hence, it was of great necessity for feature dimensionality reduction to improve the generalization capability of the classifier and facilitate its design. In accordance with the definitions of information inflow and outflow from directed transfer function (DTF) analysis [21], the inflow and outflow information based on PDC analysis were proposed in this study. Given the intensity of information flow from channel j to channel i is  ji , the inflow and outflow information of channel m can be calculated by the formula as

flowin   j 1  jm and flowout   i 1  mi . The value of inflow 19

19

information of channel m represents the sum of information which is received from other channels. The value of outflow information of channel m denotes the sum of information which is propagated to other channels. The inflow and outflow information can be used to characterize the exchange process of information flow between different brain regions. Because there is much information flowing out from epileptogenic foci during epileptic seizures, the outflow information can characterize the variation of information flow before and after seizure onsets. Therefore, by calculating the outflow information of each EEG channel, a high-dimensional matrix with size 19×19 was reduced to a low-dimensional vector with size 19×1, which can be used as the inputs to the classifier for discriminating interictal periods and ictal periods of EEG signals. 3) Classification of Interictal Periods and Ictal Periods: In this study, the SVM based on the radial basis function (RBF) kernel [22] was used to distinguish interictal periods and ictal periods of EEG signals. If both training and testing the classifier model were based on the same data set, the high sensitivity and specificity would be obtained because of overfitting. Such good effect could not be obtained in the practical situation [23]. In order to guarantee the accuracy of epileptic seizure detection when using the SVM, the 5-fold cross-validation method was used for in-sample optimization and out-of-sample test. For each patient, the data set of EEG segments was randomly divided into five mutually exclusive folds of equal size and each fold contained the same percentage of interictal and ictal samples. In these five folds, one fold was left aside for the test set and the remaining four folds were regarded as the training set. This process was repeated until every fold was tested. The cross-validation results were obtained by calculating the average of five individual measures. III. RESULTS The EEG signals were firstly segmented according to the length of analysis window. In this study, the length of analysis window was assigned as 2 seconds. By comparing the seizure detection results obtained by the proposed algorithm with those determined by the experienced epileptologists, the performance

2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2015.2424074, IEEE Journal of Biomedical and Health Informatics

JBHI-00674-2014

4

of the proposed seizure detection method can be assessed by the following indices [24]. The correct rate (CR) can be used to measure the capability of detecting interictal and ictal periods of EEG segments by (TP  TN ) (TN  FP  TP  FN ) . The selectivity (SEL) can be used to measure the detection accuracy of ictal periods of EEG segments by TP (TP  FP ) . The sensitivity (SEN) can be used to measure the capability of detecting ictal periods of EEG segments by TP (TP  FN ) . The specificity (SPE) can be used to measure the capability of detecting interictal periods of EEG segments by TN (TN  FP) . The average detection rate (ADR) can be denoted by the mean of sensitivity and specificity. Among the above evaluation indices, true positive (TP) is the correct number of ictal periods, false positive (FP) denotes the wrong number of ictal periods, true negative (TN) is the correct number of interictal periods and false negative (FN) represents the wrong number of interictal periods. TABLE II THE RESULTS OF EPILEPTIC SEIZURE DETECTION USING THE PROPOSED ALGORITHM BASED ON PDC ANALYSIS IN TEN EPILEPSY PATIENTS Patient

CR (%)

SEL (%)

SEN (%)

SPE (%)

ADR (%)

1

99.82

45.43

87.50

99.95

93.72

2

99.40

54.62

93.19

99.90

96.55

3

99.33

43.53

96.09

99.89

97.99

4

99.87

84.90

94.46

99.95

97.20

5

99.12

82.38

95.04

99.72

97.38

6

99.72

92.18

96.71

99.86

98.28

7

96.28

86.99

96.32

97.97

97.15

8

99.50

70.83

87.75

99.91

93.83

9

99.66

48.57

87.24

99.97

93.61

10

90.26

69.41

80.10

96.30

88.20

Average

98.30

67.88

91.44

99.34

95.39

The results of epileptic seizure detection using the proposed method for ten epilepsy patients are summarized in Table II. The total time length of EEG recordings for all patients was 120.6 hours. There were 44 seizures determined by two experienced epileptologists based on long-term video EEG epilepsy monitoring and the total time length of ictal periods was 72.38 minutes. Shown in this table, the correct rate using the proposed algorithm ranged from 90.26% to 99.87% and the correct rate reached to 99% for the majority of patients. Overall, the seizure detection results of the proposed method were the average correct rate of 98.3%, the average selectivity of 67.88%, the average sensitivity of 91.44%, the average specificity of 99.34%, and the average positive rate of 95.39%. Although the correct rate of seizure detection was very high, there were still some errors and omissions. It was very difficult for the proposed method to accurately identify all seizure intervals because the seizure attacks lasted for a short time or the EEG amplitudes were very small for some seizures [25]. Additionally, because the data lengths of ictal periods were significantly less than those of interictal periods, the average selectivity was significantly lower than the average specificity.

In addition to SVM classifier, the artificial neural network (ANN) classifier based on error back-propagation (BP) rule was also applied to epileptic seizure detection [26]. The same training and test sets were used in this comparison. The differences of seizure detection results between the proposed method and the combination approach of PDC analysis and ANN-BP classifier (PDC-BP) are shown in Fig. 2. When using the proposed method, the averaged evaluation indices ranged between 67.88% and 98.30%. In terms of the PDC-BP method, the averaged evaluation indices were between 27.88% and 94.39 %. By applying the one-way analysis of variance (ANOVA) [27], the results of seizure detection using the proposed method were significantly superior when compared to those using the PDC-BP method for all five evaluation indices. Generally speaking, the SVM classifier maximizes the margin between the most similar samples in each group and the decision boundary, while the ANN-BP classifier maximizes the classification accuracy of training samples to obtain the decision boundary. The SVM achieves better classification results for unseen samples. Moreover, the SVM successfully classify the unbalanced number of interictal and ictal EEG segments, as is the case. Thus, the SVM classifier can provide better evaluation indices than the ANN-BP classifier for seizure detection.

Fig. 2. Comparison of averaged evaluation indices across ten patients between the proposed method and the PDC-BP method. Mean values corresponding to different approaches are represented by different color bars. Black bars: standard deviations. An asterisk represents that the results of the proposed method are significantly better than those of the PDC-BP method.

IV. DISCUSSION To explore the change of cerebral functional connectivity during ictal periods, the outflow information of each EEG channel within two patients was calculated for 3-to-6 seconds EEG signals before a seizure onset, 3 seconds EEG signals before a seizure onset and 3 seconds EEG signals after a seizure onset. The brain electrical activity mappings (BEAMs) of outflow information for the different time periods in patient 1 and 2 are shown in Fig. 3. The clinical diagnosis of patient 1 indicated that the epileptogenic foci were located in right frontal lobe and central region. As shown in Fig. 3(a), the values of outflow information were small while the location of

2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2015.2424074, IEEE Journal of Biomedical and Health Informatics

JBHI-00674-2014 channels with the strongest outflow was inconspicuous before the seizure onset. According to the BEAM analysis of 3 s EEG segments after the seizure onset, it was observed that the outflow information was obviously enhanced and the areas whose outflow was strongest concentrated on the right frontal and central region. These results were accordant with the clinical diagnosis of patient 1 and the strong change of outflow information could clearly reflect the time point of seizure onset. The clinical diagnosis of the patient 2 was that the epileptogenic foci were located in frontal lobe. As can be seen from Fig. 3(b), the outflow information was also small and the brain areas where outflow was strongest were comparatively scattered before the seizure onset. During the epileptic seizures, outflow information was evidently increased and the areas with the strongest outflow shifted to the right frontal lobe. This indicated that the right frontal brain area discharged violently and the information flow directed from epileptogenic foci to other brain regions. In summary, the values and the strongest areas of outflow information have obvious changes before and after the seizure onset. Comparing with interictal periods, the outflow information was obviously increased and the brain areas with the strongest outflow focused on epileptogenic foci during ictal periods. This reflected the pathological features of seizures that severe paroxysmal brain dysfunction resulted from excitability increase and violent discharge of neurons of epileptogenic foci during the ictal period. The change of cerebral functional connectivity obtained by PDC analysis was consistent with clinical diagnosis and had clear physiological and pathological significance.

5 length of analysis window was set at 2 seconds, the averaged evaluation indices ranged between 67.88% and 98.30%. For the data length of 3 seconds, the averaged evaluation indices were between 47.72% and 97.26%. In terms of the EEG segments of 5 seconds, the averaged evaluation indices ranged between 51.25% and 97.26%. By performing the statistical analysis of ANOVA, there were no significant differences among three kinds of analysis window for all five evaluation indices. This indicated that the results of seizure detection had no evident distinctions while the length of analysis window varied. Additionally, if the length of analysis window was too large, the computational burden would also be increased remarkably. Therefore, in order to enhance the computation efficiency of epileptic seizure detection, the data length of 2 seconds served as a suitable choice.

Fig. 4. Comparison of averaged evaluation indices across ten patients for different lengths of EEG segments by using the proposed method. Mean values corresponding to different lengths are represented by different color bars. Black bars: standard deviations.

Fig. 3. The brain electrical activity mappings (BEAMs) of outflow information for the different time periods in patient 1 (a) and patient 2 (b).

In order to investigate the effect of the length of analysis window on epileptic seizure detection, three different analysis windows were used in this study. The scalp EEG signals including physiologic and pathophysiologic activity were band-pass filtered between 0.5 and 60 Hz in this experiment. To resolve the lowest end of this frequency band range, the detection algorithm should extract the features from the EEG segments of over 2 seconds [28]. The evaluation indices of seizure detection using the proposed method for different lengths of EEG segments are summarized in Fig. 4. When the

In addition to different lengths of EEG segments, different feature extraction approaches also made a great impact on the detection of epileptic seizures. To further discuss the performance of the proposed method, we compared this method with other existing algorithms. Firstly, the combination of approximate entropy and SVM (ApEn-SVM) was employed to detect the seizure intervals of epilepsy patients [17]. To facilitate the effectiveness comparison, the same training and test sets were applied to this method. The ApEn values were calculated for each channel of 2s EEG segments. A feature vector containing 19 elements was subsequently constructed. Then, the classification scheme based on SVM classifier was carried out. The numerical differences of averaged evaluation indices across ten patients between the proposed method and the ApEn-SVM method are illustrated in Fig. 5. It could be seen from this figure that the correct rate, sensitivity, specificity, and average detection rate of the proposed method were higher than those of the ApEn-SVM method. Especially, the mean selectivity of the proposed method was higher than that of the ApEn-SVM method by 43.88%. Then, the ANOVA was used to statistically compare the results of seizure detection between two methods. The proposed method resulted in significantly higher values than the ApEn-SVM method for all five evaluation indices. Secondly, the auto-regression model and

2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2015.2424074, IEEE Journal of Biomedical and Health Informatics

JBHI-00674-2014 SVM classifier (AR-SVM) was used in this comparison task [29]. The coefficients of AR model of EEG segments were extracted from each channel individually and the features of 19 channels were combined to construct new feature vectors. Since new feature vectors had a relatively high dimensionality, the principle components analysis was used to reduce feature dimensionality. Subsequently, the feature vectors were put into a SVM classifier to evaluate the results of seizure detection. The seizure detection differences between the proposed method and the AR-SVM method are shown in Fig. 6. The statistical analysis demonstrated that the superiority of the proposed method was significant when compared with the AR-SVM method for all evaluation indices. It is generally considered that the excessive activities of neuron firing firstly occur at epileptogenic foci and then are transferred to surrounding brain areas. The epileptiform propagation areas were different from the seizure onset zones [30]. Therefore, there will be much information flowing out from epileptogenic foci to epileptiform propagation areas during ictal periods. The application of PDC analysis as mechanism for feature extraction in the scalp EEG

6 recordings for seizure detection can reflect the physiological changes of information flow before and after seizure onsets. Thus, the proposed method is more effective than other algorithms when detecting epileptic seizures. V. CONCLUSION In this study, a new method based on the PDC analysis was proposed from the perspective of the direction and intensity of information flow between brain regions. By applying 5-fold cross-validation, the proposed method achieved the correct rate of 98.3% and the specificity of 99.34%. This demonstrated that the proposed method reflected the change of the causal relationship between brain areas before and after the seizure onset and was suitable for the epileptic seizure detection. In comparison with other existing techniques, the proposed method had significant superiority in all five evaluation indices. Additionally, according to the BEAM analysis of outflow information, there were obvious changes before and after the seizure onset for outflow values and the brain areas with the strongest outflow. The outflow information was remarkably enhanced and the brain regions whose outflow was strongest were approximately consistent with epileptogenic foci during ictal periods. The BEAM analysis might be used to achieve the localization of epileptogenic foci for epilepsy patients. REFERENCES

Fig. 5. Comparison of averaged evaluation indices across ten patients between the proposed method and the ApEn-SVM method. Mean values corresponding to different approaches are represented by different color bars. Black bars: standard deviations. An asterisk represents that the results of the proposed method are significantly better than those of the ApEn-SVM method.

Fig. 6. Comparison of averaged evaluation indices across ten patients between the proposed method and the AR-SVM method. Mean values corresponding to different approaches are represented by different color bars. Black bars: standard deviations. An asterisk represents that the results of proposed method are significantly better than those of AR-SVM method.

[1] H. Witte, L. D. Iasemidis, and B. Litt, “Special issue on epileptic seizure prediction,” IEEE Transactions on Biomedical Engineering, vol. 50, no. 5, pp. 537-539, 2003. [2] S. Altunay, Z. Telatar, and O. Erogul, “Epileptic EEG detection using the linear prediction error energy,” Expert Systems with Applications, vol. 37, no. 8, pp. 5661-5665, 2010. [3] A. M. Murro, D. W. King, J. R. Smith, B. B. Gallagher, H. F. Flanigin, and K. Meador, “Computerized seizure detection of complex partial seizures,” Electroencephalogr Clin Neurophysiol, vol. 79, no. 4, pp. 330-3, Oct, 1991. [4] O. Faust, U. R. Acharya, L. C. Min, and B. H. Sputh, “Automatic identification of epileptic and background EEG signals using frequency domain parameters,” Int J Neural Syst, vol. 20, no. 2, pp. 159-76, Apr, 2010. [5] S. Grewal, and J. Gotman, “An automatic warning system for epileptic seizures recorded on intracerebral EEGs,” Clin Neurophysiol, vol. 116, no. 10, pp. 2460-72, Oct, 2005. [6] I. Conradsen, S. Beniczky, P. Wolf, T. W. Kjaer, T. Sams, and H. B. Sorensen, “Automatic multi-modal intelligent seizure acquisition (MISA) system for detection of motor seizures from electromyographic data and motion data,” Comput Methods Programs Biomed, vol. 107, no. 2, pp. 97-110, Aug, 2012. [7] L. Ayoubian, H. Lacoma, and J. Gotman, “Automatic seizure detection in SEEG using high frequency activities in wavelet domain,” Medical Engineering & Physics, vol. 35, no. 3, pp. 319-328, 2013. [8] A. T. Tzallas, M. G. Tsipouras, and D. I. Fotiadis, “Epileptic seizure detection in EEGs using time-frequency analysis,” IEEE Trans Inf Technol Biomed, vol. 13, no. 5, pp. 703-10, Sep, 2009. [9] R. J. Martis, U. R. Acharya, J. H. Tan, A. Petznick, R. Yanti, C. K. Chua, E. Y. Ng, and L. Tong, “Application of empirical mode decomposition (emd) for automated detection of epilepsy using EEG signals,” Int J Neural Syst, vol. 22, no. 6, pp. 1250027, Dec, 2012. [10] S. M. Alam, and M. I. Bhuiyan, “Detection of seizure and epilepsy using higher order statistics in the EMD domain,” IEEE J Biomed Health Inform, vol. 17, no. 2, pp. 312-8, Mar, 2013. [11] N. F. Güler, E. D. Übeyli, and İ. Güler, “Recurrent neural networks employing Lyapunov exponents for EEG signals classification,” Expert Systems with Applications, vol. 29, no. 3, pp. 506-514, 2005. [12] S. Ghosh-Dastidar, H. Adeli, and N. Dadmehr, “Mixed-band wavelet-chaos-neural network methodology for epilepsy and epileptic seizure detection,” IEEE Trans Biomed Eng, vol. 54, no. 9, pp. 1545-51, Sep, 2007.

2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2015.2424074, IEEE Journal of Biomedical and Health Informatics

JBHI-00674-2014 [13] M. Niknazar, S. R. Mousavi, B. V. Vahdat, and M. Sayyah, “A new framework based on recurrence quantification analysis for epileptic seizure detection,” IEEE J Biomed Health Inform, vol. 17, no. 3, pp. 572-8, May, 2013. [14] Y. Zhang, W. Zhou, S. Yuan, and Q. Yuan, “Seizure detection method based on fractal dimension and gradient boosting,” Epilepsy Behav, vol. 43C, pp. 30-38, Dec 27, 2014. [15] N. Kannathal, M. L. Choo, U. R. Acharya, and P. K. Sadasivan, “Entropies for detection of epilepsy in EEG,” Comput Methods Programs Biomed, vol. 80, no. 3, pp. 187-94, Dec, 2005. [16] J. Xiang, C. Li, H. Li, R. Cao, B. Wang, X. Han, and J. Chen, “The detection of epileptic seizure signals based on fuzzy entropy,” J Neurosci Methods, Jan 19, 2015. [17] V. Srinivasan, C. Eswaran, and N. Sriraam, “Approximate entropy-based epileptic EEG detection using artificial neural networks,” IEEE Trans Inf Technol Biomed, vol. 11, no. 3, pp. 288-95, May, 2007. [18] L. Baccalá, and K. Sameshima, “Partial directed coherence: a new concept in neural structure determination.,” Biol Cybern, vol. 84, no. 6, pp. 463-74, Jun, 2001. [19] M. Kaminski, and H. Liang, “Causal influence: advances in neurosignal analysis,” Crit Rev Biomed Eng, vol. 33, no. 4, pp. 347-430, 2005. [20] H. Akaike, “A new look at the statistical model identification,” IEEE Transactions on Automatic Control, vol. 19, no. 6, pp. 716-723, 1974. [21] B. He, Y. Dai, L. Astolfi, F. Babiloni, H. Yuan, and L. Yang, “eConnectome: A MATLAB toolbox for mapping and imaging of brain functional connectivity,” J Neurosci Methods, vol. 195, no. 2, pp. 261-9, Feb 15, 2011. [22] N. Cristianini, and J. Shawe-Taylor, "Application of support vector machines," in An introduction to support vector machines and other kernel-based learning methods, 1st ed., United Kingdom: Cambridge university press, 2000, pp. 158-172. [23] F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehnertz, “Seizure prediction: the long and winding road,” Brain, vol. 130, no. Pt 2, pp. 314-33, Feb, 2007. [24] A. Aarabi, R. Grebe, and F. Wallois, “A multistage knowledge-based system for EEG seizure detection in newborn infants,” Clin Neurophysiol, vol. 118, no. 12, pp. 2781-97, Dec, 2007. [25] J. Mitra, J. R. Glover, P. Y. Ktonas, A. Thitai Kumar, A. Mukherjee, N. B. Karayiannis, J. D. Frost, Jr., R. A. Hrachovy, and E. M. Mizrahi, “A multistage system for the automated detection of epileptic seizures in neonatal electroencephalography,” J Clin Neurophysiol, vol. 26, no. 4, pp. 218-26, Aug, 2009. [26] G. Wang, Z. Wang, W. Chen, and J. Zhuang, “Classification of surface EMG signals using optimal wavelet packet method based on Davies-Bouldin criterion.,” Med Biol Eng Comput, vol. 44, no. 10, pp. 865-72, Oct, 2006. [27] G. Wang, and D. Ren, “Classification of surface electromyographic signals by means of multifractal singularity spectrum,” Med Biol Eng Comput, vol. 51, no. 3, pp. 277-84, Mar, 2013. [28] A. H. Shoeb, “Application of machine learning to epileptic seizure onset detection and treatment,” Ph.D. dissertation, Harvard-MIT Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, 2009. [29] A. Subasi, “Selection of optimal AR spectral estimation method for EEG signals using Cramer-Rao bound,” Comput Biol Med, vol. 37, no. 2, pp. 183-94, Feb, 2007. [30] G. Wang, G. Worrell, L. Yang, C. Wilke, and B. He, “Interictal spike analysis of high-density EEG in patients with partial epilepsy,” Clin Neurophysiol, vol. 122, no. 6, pp. 1098-105, Jun, 2011. Gang Wang (M’12) received the B.S. degree in mechanical engineering from Wuhan Polytechnic University, Wuhan, China, in 2001, the M.S. degree in mechatronics engineering from Nanchang University, Nanchang, China, in 2004, and the Ph.D. degree in biomedical engineering from Shanghai Jiao Tong University, Shanghai, China, in 2008. He was a postdoctoral associate at the Department of Biomedical Engineering at the University of Minnesota, Twin Cities. Since 2011, he has been with the Faculty of Xi’an Jiaotong University as an Assistant Professor at the School of Life Science and Technology and the Institute of Biomedical Engineering. His research interests include biomedical functional neuroimaging, biomedical signal processing, neural engineering, and bioelectromagnetism.

7 Zhongjiang Sun received his B.S. and M.S. degrees in Biomedical Engineering from Xi’an Jiaotong University, Xi’an, China, in 2013 and 2015, respectively. His research interests are the analysis of epilepsy seizure and biomedical signal processing.

Ran Tao received her B.S. degree in Biomedical Engineering from Xi’an Jiaotong University, Xi’an, China. Her research interests mainly focus on biomedical signal processing.

Kuo Li receiver his B.S. degree in Clinical Medicine from Xi’an Jiaotong University, Xi’an, China. His research interests mainly focus on EEG and Epilepsy Surgery.

Gang Bao received the B.S. degree in Clinical Medicine from Xi’an Medical University, Xi’an, China, in 1996, the M.S. degree in Clinical medicine from Xi’an Medical University, Xi’an, China, in 1998, and the M.D. degree in Clinical medicine from Xi’an Jiao Tong University, Xi’an, China, in 2008. His research interests are the analysis of epilepsy seizure and the treatment of brain-tumor.

Xiangguo Yan received the B.S. degree in industrial automation from Zhengzhou University of Light Industry, Zhengzhou, China, in 1983, the M.S. degree in automatic control and the Ph.D. degree in biomedical engineering from Xi’an Jiaotong University, Xi’an, China, in 1990 and 1995, respectively. From 1996 to 1998, he was a Visiting Scientist at Juelich research center, Germany. He is currently a Professor in the School of Life Science and Technology at Xi’an Jiaotong University. His main research interests are biomedical signal and image processing, and the development of medical instrumentation.

2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.