Epilepsy & Behavior 43 (2015) 30–38

Contents lists available at ScienceDirect

Epilepsy & Behavior journal homepage: www.elsevier.com/locate/yebeh

Seizure detection method based on fractal dimension and gradient boosting Yanli Zhang a,b, Weidong Zhou a,c,⁎, Shasha Yuan a,c, Qi Yuan c a b c

School of Information Science and Engineering, Shandong University, Jinan 250100, China School of Information and Electronics Engineering, Shandong Institute of Business and Technology, Yantai 264005, China Suzhou Institute of Shandong University, Suzhou 215123, China

a r t i c l e

i n f o

Article history: Received 26 August 2014 Revised 20 November 2014 Accepted 21 November 2014 Available online xxxx Keywords: EEG Seizure detection Fractal dimension Gradient boosting

a b s t r a c t Automatic seizure detection technology is necessary and crucial for the long-term electroencephalography (EEG) monitoring of patients with epilepsy. This article presents a patient-specific method for the detection of epileptic seizures. The fractal dimensions of preprocessed multichannel EEG were firstly estimated using a k-nearest neighbor algorithm. Then, the feature vector constructed for each epoch was fed into a trained gradient boosting classifier. After a series of postprocessing, including smoothing, threshold processing, collar operation, and union of seizure detections in a short time interval, a binary decision was made to determine whether the epoch belonged to seizure status or not. Both the epoch-based and event-based assessments were used for the performance evaluation of this method on the EEG data of 21 patients from the Freiburg dataset. An average epoch-based sensitivity of 91.01% and a specificity of 95.77% were achieved. For the event-based assessment, this method obtained an average sensitivity of 94.05%, with a false detection rate of 0.27/h. © 2014 Elsevier Inc. All rights reserved.

1. Introduction Epilepsy is a common neurological disorder that affects approximately 1% of the world's population. Epileptic seizures are clinical manifestation of abnormal and excessive neuronal discharges in the brain and can result in disturbance of consciousness and sudden loss of motor control [1]. In epilepsy diagnosis, such as assessment of seizure type and frequency, localization of the epileptogenic region in the brain, and EEG investigations for epilepsy surgery, long-term EEG monitoring is useful and mandatory in epilepsy surgery centers and intensive care units [2]. Though long-term EEG recordings can be visually inspected by highly trained clinicians, it is a tedious and time-consuming process. Therefore, automatic seizure detection technology can be a valuable aid to neurologists in analyzing EEG recordings. One of the first approaches for seizure detection was introduced by Gotman [3], where individual EEG signals were broken into half waves and three different measures (average amplitude, average duration, and coefficient of variation) were calculated for each epoch. Then, seizure detection was obtained by comparing these measures with some empirically predefined thresholds. The method was later improved by Gotman [4] and Qu [5]. After that, approaches for seizure detection based on pattern recognition and machine learning theory have been proposed. Qu and Gotman designed a patient-specific algorithm to ⁎ Corresponding author at: School of Information Science and Engineering, Shandong University, 27 Shanda Road, Jinan, China. Tel.: +86 531 88361551. E-mail address: [email protected] (W. Zhou).

http://dx.doi.org/10.1016/j.yebeh.2014.11.025 1525-5050/© 2014 Elsevier Inc. All rights reserved.

detect seizure onsets, where a modified nearest neighbor classifier was used after extracting five features from the time and frequency domains [6]. In the seizure warning system for intracerebral EEG introduced by Grewal and Gotman [7], the seizure probability of a section of EEG was calculated using Bayesian formulation. In addition, artificial neural network [8–10] and support vector machine [11–13] were widely used to build automatic seizure detection algorithms. The usage of pattern recognition and machine learning theory in seizure detection is premised on the extraction of representative features from epileptic EEG recordings. Features, extracted in time domain [14,15], frequency domain [16], and time-frequency domain like wavelet transform [17–19] and Hilbert–Huang transform [20], have been used in most seizure detection algorithms. At present, a growing number of studies have investigated the nonlinear features in the EEG as well, such as the largest Lyapunov exponent [21], correlation dimension [22], and entropies [23,24]. The functions of the brain at the microscopic level, i.e., the interplay of neurons, is extremely nonlinear in nature since the dynamic behavior of individual neurons is decided by threshold and saturation phenomena [25]. Electroencephalography signals recorded from the brain also have significant nonlinearity because of the close relationship with physiological and pathophysiological functions of the brain. Hence, instead of the traditional linear methods, nonlinear dynamic analysis methods may better suit the feature extraction of EEG signals. In this study, we systematically describe an automatic patientspecific seizure detection method, which extracts the nonlinear feature, fractal dimension (FD) based on the k-nearest neighbor algorithm, and discriminates the feature vector of an epoch for seizure or nonseizure

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

31

status using a gradient boosting algorithm. Fractal dimension is a powerful tool for transient detection and also a measure of the dimensional complexity of fractal objects. Fractal dimension can be calculated from a set of points using several diverse methods that differ in accuracy, sensitivity to the number of points used, and the time required for computation [26]. In contrast to the box-counting method, in which FD estimation considers the variation of mass inside fixed-size cubes, the k-nearest neighbor algorithm belongs to fixed-mass algorithms, according to which FD estimation is based on the scaling of the sizes of cubes so that they contain the same number of points (mass) [26]. The k-nearest neighbor algorithm has been previously applied for the estimation of the fractal dimension of strange attractors [27], graylevel images [28], and scalp EEG data [29]. Compared with other wellknown estimators of the fractal dimension, such as box-counting method [30], correlation algorithm [26], Katz's algorithm [31], and Higuchi's algorithm [32], it has been proven superior in terms of accuracy, dynamic range, and computational time. Gradient boosting used as a classifier in this work can build one strong classifier from many weak classifiers [33,34]. For its conceptual simplicity, gradient boosting has been used in many research domains, such as remotely sensed imagery [35] and brain–computer interface [36]. The organization of the succeeding sections of this paper is as follows. The proposed seizure detection method and the EEG data analyzed are described in Section 2. Results are presented in Section 3. Section 4 is devoted to discuss the method in terms of the analysis of the method's characteristics and missed and false detections and the comparison of related works. Finally, the conclusion is given in Section 5.

Table 1 Training dataset of 21 patients.

2. Material and methods

2.3. Estimation of fractal dimension

2.1. EEG dataset

According to the k-nearest neighbor algorithm for estimating fractal dimension [28,29], the average distance, 〈rγk 〉, of a point to its kth nearest neighbors is a function of k:

The EEG recordings used in this work come from the Epilepsy Center of the University Hospital of Freiburg, Germany. This database contains intracranial EEG signals of 21 patients. These EEG data were recorded by a Neurofile NT digital video-EEG system with 128 channels, a 256-Hz sampling rate, and a 16-bit analog-to-digital converter. Six channels of all implanted grid, strip, and depth electrodes were given for analyses, which were selected based on visual inspection of the raw data by certified epilepsy experts [37]. Three are focal channels, located near the epileptic focus, and three nonfocal channels. Only the three focal channels were chosen for seizure detection in this work. For each patient, there are 2–5 seizure events in the EEG recordings. The onset and end times of each seizure had been previously determined by experienced experts based on the identification of epileptic patterns preceding clinical manifestation of seizures in EEG recordings [38]. Additionally, each patient has approximate 24-hour interictal EEG recordings without seizure activity, i.e., nonseizure data. A detailed description of the database can be found in Ref. [39]. The task of this work was to propose a patient-specific seizure detection method. Hence, nonoverlapping training and testing datasets were created for each patient in the following way. For each of the patients, one or two seizures were chosen randomly. The EEG in these seizures and an equivalent amount of nonseizure EEG selected randomly from interictal recordings were used to construct the training dataset. Table 1 lists the detailed training dataset of 21 patients. All other available EEG data, which were not included in the training, could be used as testing data. In this work, we used 20-hour EEG data containing 1–4 seizure events to test the trained classifier for each patient. In total, 420 h of EEG data containing 59 seizure events were used for the performance assessment of the proposed seizure detection method. 2.2. Preprocessing In the preprocessing stage, a 4th-order Chebyshev band-pass digital filter was first used in order to reduce the effect of artifacts. The cutoff

Patient

Number of seizure epochs

Number of nonseizure epochs

Total EEG epochs for training

Duration of training dataset (min)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

6 16 15 12 3 7 10 20 40 189 7 8 26 68 5 22 11 7 4 20 9

6 16 15 12 3 7 10 20 40 189 7 8 26 68 5 22 11 7 4 20 9

12 32 30 24 6 14 20 40 80 378 14 16 52 136 10 44 22 14 8 40 18

2.0 5.3 5.0 4.0 1.0 2.3 3.3 6.7 13.3 63.0 2.3 2.7 8.7 22.7 1.7 7.3 3.7 2.3 1.3 6.7 3.0

frequency was empirically set to be 0.5 and 30 Hz. Then, the filtered multichannel EEG was segmented into 10-second epochs by a sliding window, without overlap between the adjacent epochs.


γ γ=Dðγ Þ rk ¼ Gðk; γ Þk

ð1Þ

where γ = (1 - q)Dq, D(γ) = Dq, and G(k, γ) is a function of k and γ, which is near unity for large k. Here Dq is the q-order generalized dimensions. For q = 0, γ = D0 and the fractal dimension is obtained. This means that the fractal dimension is the fixed point of the function D(γ), and the fractal dimension of a segment of intracranial EEG signal can be estimated iteratively as follows. Step 1. An initial value of γ, i.e., γ0, is chosen arbitrarily. In this work, we chose γ0 = 1.5. Step 2. For each point of the EEG signal, the Euclidian distance rki from its kth-nearest neighbor is calculated. Here, k = kmin, …, kmax, i = 1, 2, …, N, and N is the length of the EEG signal. In this study, kmin and kmax were selected as 1 and 100, respectively. Step 3. For j = 1, 2, …, the following recursive relations are applied:   γ j−1 D γj ¼ s j−1   γj ¼ D γj

ð2Þ ð3Þ

where sj −D 1 is Ethe   slope of the best-fitting line at the points γ and ln ðkÞ; ln r k j−1 N D γ E 1X γ r k j−1 ¼ r j−1 ; k ¼ kmin ; ⋯kmax : N i¼1 ki

ð4Þ

The calculation is repeated until the maximum number of iterations  .h  i   1 is reached or the quantity  γ j −γ j−1  drops below a 2 γ j þ γ j−1

32

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

certain value, and the γj, calculated for the last j, is the fractal dimension of the EEG signal. In the study of [28], two or three iterations were sufficient in giving an accurate estimation of the fractal dimension for an image of 1000 points. In order to obtain an accurate estimation of the fractal dimension of an EEG signal in this work, a value of maximum iterations J=5 was chosen. The fractal dimensions extracted from EEG epochs of three focal channels were composed into a three-dimensional feature vector. 2.4. Gradient boosting classifier Once features were extracted from EEG recordings, the classification between the seizure and nonseizure EEG will be carried out. The classifier used in this work was gradient boosting [34,36]. We denote the ensemble of classifiers after step m by Fm, the training data by X = {xi ∈ RK, i = 1, ⋯ N}, and the corresponding class labels by Y = {yi ∈ {0, 1}, i = 1, ⋯ N}. In this work, ordinary least square regression was used as the weak learner, and the logistic regression model was as follows: pm ðyi ¼ 1jxi Þ ¼

e e

F m ðxi Þ

F m ðxi Þ

þ e− F m ðxi Þ

; i ¼ 1; ⋯N:

ð5Þ

Gradient boosting was used to stepwise maximize the Bernoulli log-likelihood of the logistic regression model, where Bernoulli log-likelihood of Fm could be expressed as the following: N  y 1−y Lð F m Þ ¼ log ∏ pm ðyi ¼ 1jxi Þ i pm ðyi ¼ 0jxi Þ i :

ð6Þ

i¼1

The implementation of the gradient boosting algorithm is shown in the following steps: (1) Start with F0(xi) = 0, i = 1, ⋯ N and p0(yi = 1|xi) = 0.5. (2) Repeat for m = 1, 2, …, M: (a) Compute the gradient of the likelihood function with respect to F = Fm − 1: e yi ¼ 2ðyi −pm−1 ðyi ¼ 1jxi ÞÞ; i ¼ 1; ⋯N:

ð7Þ

(b) A weak classifier fm that best fits the gradient in a least squares sense is selected: f m ¼ arg min f

XN i¼1

2

ðe yi − f ðxi ÞÞ :

ð8Þ

(c) Determine the coefficient γm: γ m ¼ arg maxγ Lð F m−1 þ γ f m Þ:

ð9Þ

(d) Update the classifier: F m ¼ F m−1 þ εγ m f m :

ð10Þ

Here, γm shrinks to a small value through multiplication with a small ε at each step to improve the generalization performance of the boosting algorithm. In this paper, the value of ε is given as 0.05. (e) Compute the probability using Eq. (5). Classification consists of two steps: training and testing. In the training stage of each patient, the feature vectors generated for the training dataset were fed to train a gradient boosting classifier. All the training samples of the patient were then input into the trained gradient boosting classifier, and a threshold was adjusted in the range of 0 to 1. The threshold value minimizing the number of falsely classified training samples was selected, which would be used in the following postprocessing. In

the testing stage, the trained classifier was applied to the patient's testing data, and the probability outputs of the classifier were processed as described below. 2.5. Postprocessing Postprocessing scheme was applied to the classifier outputs and consisted of smoothing, threshold processing, collar operation, and union of some seizure detections. Primarily, a Gaussian low-pass filter was applied to the outputs of the gradient boosting classifier to reduce random noise and to remove the possible sporadic and isolated false detection. The probabilities after Gaussian smoothing were then compared with the threshold, which was set in the training process of each patient. After comparison, binary decisions were taken for continuous epochs: 1 — seizure; 0 — nonseizure. Since the smoothing procedure may make both sides of a seizure obscure and difficult to detect correctly, a collar operation was applied to make up for the possible missed seizure decisions [40]. Each detected seizure event was extended n epochs on both sides. In this work, the value of n, i.e., the length of collar, varied in different patients. To further reduce the number of false detection, we finally unified seizure detections if there was a certain distance between them. That is, for any two consecutive seizure detections obtained by the previous procedures, if the time interval between the end time of the first detection and the start time of the second detection was less than a minimum time interval, they were unified and considered as a single seizure event [38]. In this study, the minimum time interval value was set at 50 s. 2.6. Performance evaluation Epoch-based and event-based metrics were used to quantify the performance of the proposed seizure detection method [40]. For each patient, epoch-based metrics such as sensitivity, specificity, and recognition rate were defined as: sensitivity ¼

TP ; TP þ FN

specificity ¼

TN ; TN þ FP

recognition rate ¼

TP þ TN TP þ FN þ TN þ FP

where TP (true positive) was the number of epochs labeled as seizure both by the algorithm and by the experts; FN (false negative) was the number of seizure epochs incorrectly labeled as nonseizure epochs by the algorithm; TN (true negative) was the number of the correctly identified nonseizure epochs; and FP (false positive) was the number of nonseizure epochs incorrectly identified as seizure epochs by the algorithm. Unlike the epoch-based metrics, event-based metrics were thought to reflect the performance of a system for clinical application. In the event-based evaluation, the subsequent decisions of the same class were joined to create an event. If a seizure was detected by the algorithm at any time between the start and the end of an expert-labeled seizure, it was considered a true detection. False detections were those that do not overlap with expert-labeled seizure events. For each patient, the event-based sensitivity was given by the number of true detections divided by the total number of seizures marked by the experts, and false detection rate was defined as the average number of false detections per hour. 3. Results The proposed epileptic seizure detection method was implemented on MATLAB R2012b. As described in Section 2, the fractal dimensions of

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

preprocessed EEG signals were estimated using the k-nearest neighbor algorithm, and the feature vector extracted from each EEG epoch was fed into the trained gradient boosting classifier. The probability output of the classifier was processed by a series of operations, such as smoothing, threshold processing, collar operation, and union of seizure detections, to determine whether the epoch belonged to a seizure or not. Running on a desktop PC with an Intel dual-core 2.8-GHz CPU and 2 GB of RAM, the described seizure detection algorithm processed a one-hour three-channel EEG recording in about 20 min, among which the time for feature extraction accounted for more than 99%. The typical feature extraction time over one-hour signal-channel EEG was around 6.65 min. Fig. 1 shows one-hour signal-channel intracranial EEG with a seizure event of patient 6 and the corresponding curve of k-nearest neighbor fractal dimension. It can be noticed in Fig. 1 that fractal dimension values drop significantly during the seizure period compared with other periods. Statistically significant difference was also found in the fractal dimension between nonseizure EEG epochs and seizure EEG epochs by one-way analysis of variance (P b 0.05). In addition, Fig. 2 presents the mean values and standard deviations of the fractal dimensions of two types of EEG data, seizure and nonseizure. In seizure detection systems, high sensitivity and low false detection rate are desired. However, they are always in opposition to each other. The high sensitivity may trigger a rise in the false detection rate, while decreasing the false detection rate may miss some seizure events. The value of threshold, which is applied to the filtered probabilities of seizure for a binary decision, is crucial in the trade-off between the sensitivity and the false detection rate. In this work, the thresholds were selected in the classifier training process for each patient. The details of the threshold values of 21 patients have been summarized in Table 2. The mean value and the standard deviation of the 21 thresholds are 0.5158 and 0.0242, respectively. To compensate for the potential impacts due to using the smoothing procedure in postprocessing and to determine the duration of seizures more accurately, the collar operation was applied in this work. The length of collar of each patient was between 3 and 18, which was determined by means of receiver operating characteristic (ROC) curves. More specifically, the best value of the collar length was selected as the one which minimizes the distance between the corresponding (1 — specificity, sensitivity) point and the optimum point (0, 1) on the ROC curve. The collar operation could detect the onset of seizures more accurately and

33

Fig. 2. Mean values and standard deviations of the fractal dimensions of seizure and nonseizure EEG signals. Here, half the height of error bars denotes standard deviations, and KNNFD means the fractal dimensions estimated by the k-nearest neighbor algorithm.

reduce the detection delay. The average delay time of 21 patients was 2.46 s. However, the average delay time was 20.87 s if the collar operation was removed from the proposed seizure detection algorithm. Fig. 3 illustrates an example of a detected seizure. The performance of the proposed seizure detection method was assessed on all the testing data of 21 patients in the Freiburg database. The results of the epoch-based and event-based performance assessments for the presented method are listed in Table 3. It can be noticed in Table 3 that there were 12 patients with a sensitivity of 100% in the epoch-based assessment, and, among them, 9 patients had specificities and recognition accuracies greater than 95.00%. On the whole, in the epoch-based assessment, the average values of sensitivity, specificity, and recognition rate over all 21 patients were 91.01%, 95.77%, and 95.75%, respectively. In the event-based assessment, except for four patients (patients 10, 16, 19, and 20), all the others had a sensitivity of 100%. Four out of 59 seizures were missed by the proposed seizure detection method. About two-thirds of patients had a false detection rate below 0.2/h. The average sensitivity of 94.05% and false detection rate of 0.27/h over all 21 patients were obtained in the event-based assessment. 4. Discussion 4.1. Characteristics of the seizure detection system Our detection system used a nonlinear feature, fractal dimension, to detect seizures. Fractal dimension is a powerful tool for Table 2 The threshold values of 21 patients.

Fig. 1. Top panel: one-hour signal-channel intracranial EEG with a seizure event of patient 6. The seizure event marked by experts is between the two red vertical lines, and the onset of the seizure is labeled as time zero. Middle panel: greater magnification of the seizure EEG. Bottom panel: k-nearest neighbor fractal dimension (KNNFD) of the EEG recordings shown in the top panel. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Patient

Threshold

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

0.5159 0.5939 0.5241 0.5031 0.5038 0.5086 0.5198 0.5106 0.5101 0.5008 0.5295 0.4999 0.5018 0.5006 0.5088 0.5241 0.5708 0.5026 0.5011 0.5002 0.5008

34

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

Fig. 3. Example of a detected seizure from patient 3. The first three panels show one-hour three-channel raw EEG recordings. The seizure event marked by the experts is between the two vertical lines, and the onset of the seizure is labeled as time zero. The bottom panel gives the detection based on the proposed algorithm. The blue and green curves respectively represent the probability outputs of the gradient boosting classifier and the Gaussian-filtered outputs. The pink horizontal line indicates the threshold, and the seizure duration detected by the proposed algorithm is displayed with a gray rectangle. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

quantifying the complexity of EEG and for detecting epileptic seizures [41]. Compared with other methods in which phase space reconstruction is needed, estimating the fractal dimension based on the k-nearest neighbor algorithm is directly in the time domain and allows significant savings in the program running time [29]. However, the time for feature extraction in the proposed seizure detection algorithm still accounts for more than 99% of the total running time of the algorithm. The time for estimating the fractal dimension of EEG should be further reduced by optimizing the

program or by employing high-level programming languages such as C/C++ to better meet the need of online real-time detection. The gradient boosting used in this work constructs a powerful classifier for each patient through stepwise maximizing the Bernoulli log-likelihood of the logistic regression model and has the advantage of high robustness, conceptual simplicity, small number of tuning parameters, and better generalization performance. The application in the seizure detection in this work also demonstrated its high classification accuracy. In the postprocessing stage, the schemes of smoothing, threshold processing, and collar operation had been applied in previous detection study [40,42]. These operations were employed in this work in order to reduce false detections from sporadic fluctuations and to locate the seizure duration more accurately. The union of the seizure detections in a short time interval, which was applied in the last step of the proposed method, can further reduce the false detection rate significantly. Compared with using the six available EEG channels, three focal channels provided a better performance of seizure detection because the EEG obtained from the focal channels showed clear seizure patterns. Table 4 shows the performance comparison of the two cases using different numbers of channels in the proposed seizure detection algorithm. The performances were the same in the two cases for only seven patients. As described above, the proposed seizure detection method was patient-specific, that is, the seizure detectors were trained on each patient's dataset and the parameters including the threshold and length of collar were tuned for each patient. To demonstrate that patient-specific training and parameters really contribute to the performance, a patient-independent seizure detection system was also generated, with a threshold of 0.5015 and a collar length of 12. The training of the patient-independent detection system was based on 105 epochs of seizure data and 105 epochs of nonseizure data selected from 21 patients. More specifically, each patient provided 5 epochs of seizure and 5 epochs of nonseizure training data. Five epochs of seizure data came from the patients' first seizure marked by the experts or the first two seizures if the duration of the first seizure was not long enough. Five epochs of nonseizure data from each patient were taken from the beginning of interictal recordings. All

Table 3 Results of the epoch- and event-based performance assessments. Patient

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Average

Number of seizures in the testing set

2 2 3 4 4 2 2 1 3 3 3 3 1 2 3 3 4 3 3 4 4

Epoch-based performance

Event-based performance

Sensitivity (%)

Specificity (%)

Recognition rate (%)

Sensitivity (%)

False detection rate (/h)

100.00 100.00 100.00 100.00 100.00 84.21 92.86 100.00 100.00 69.23 93.85 85.71 100.00 100.00 91.67 64.58 100.00 100.00 66.67 62.50 100.00 91.01

88.42 95.60 98.70 97.57 97.35 86.20 98.68 99.93 94.91 99.65 95.60 89.91 98.68 98.00 96.92 97.25 97.49 84.77 97.21 99.03 99.37 95.77

88.43 95.61 98.71 97.58 97.36 86.19 98.65 99.93 94.93 99.53 95.58 89.90 98.68 98.01 96.87 97.03 97.51 84.79 97.17 98.87 99.37 95.75

100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 66.67 100.00 100.00 100.00 100.00 100.00 66.67 100.00 100.00 66.67 75.00 100.00 94.05

0.70 0.40 0.00 0.15 0.45 0.85 0.05 0.00 0.37 0.05 0.20 0.32 0.05 0.11 0.10 0.20 0.16 0.96 0.38 0.00 0.11 0.27

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

35

Table 4 Performance comparison of two cases where different numbers of channels were used. EEG channels used

Six channels including three focal and three nonfocal channels Only three focal channels

Epoch-based performance

Event-based performance

Average sensitivity (%)

Average specificity (%)

Average recognition rate (%)

Average sensitivity (%)

Average false detection rate (/h)

87.08 91.01

94.95 95.77

94.90 95.75

91.27 94.05

0.34 0.27

other available EEG data, which were not included in the training, were used as testing data to evaluate the performance of the patient-independent systems, and an average epoch-based sensitivity of 82.65% was achieved. Meanwhile, the average event-based sensitivity was 83.73%, and the average false detection rate was 1.99/h. Compared with the patient-specific seizure detection method proposed in this work, the patient-independent system exhibited lower sensitivity and higher false detection rate.

4.2. Missed detections and false detections In this work, there were 4 out of 59 seizures that were missed by the proposed method. Missed seizures were mainly characterized by short duration or unobvious epileptic activity as illustrated in Figs. 4 and 5. In addition, an analysis of false detection was performed in this work. Three patients (patients 1, 6, and 18) had a false detection rate above 0.5/h as shown in Table 3. The majority of the false detections were

Fig. 4. Example of missed seizures due to the short duration. (A) Six-channel intracranial EEG recordings from patient 19. The seizure event marked by experts is between the two red vertical lines, and the onset of the seizure is labeled as time zero. (B) Greater magnification of the EEG around the seizure, including 3 s of preictal data, approximately 11 s of seizure EEG and 1 s of postictal data. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

36

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

Fig. 5. Example of missed seizures due to unobvious epileptic activity. (A) Six-channel intracranial EEG recordings from patient 10. The seizure event marked by experts is between the two red vertical lines, and the onset of the seizure is labeled as time zero. (B) Greater magnification of the beginning section of the seizure, including 3 s of preictal data and 10 s of seizure EEG. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

caused by high amplitude rhythmic activities and artifacts contained in the EEG. 4.3. Comparison with other systems To date, various algorithms have been proposed for epileptic seizure detection with different degrees of success. Esteller et al. used line length, derived from Katz's fractal dimension, for epileptic seizure detection and provided promising results in all 10 patients they analyzed [43]. Because of the difference in the type and quantity of EEG data analyzed and the variations in the methodology of performance assessment, it is difficult to make a detailed comparison between different seizure detection algorithms. Here, we briefly summarize the results from recent works that used the same EEG dataset for comparison. In the work of Aarabi et al., temporal, spectral, and complexity features were extracted from IEEG segments and spatio-temporally integrated using a fuzzy rule-based system for seizure detection [38]. A

total of 302.7 h of intracranial EEG recordings from 21 patients having 78 seizures were used for the evaluation of the system. At the segment-based level, the system yielded a sensitivity of 68.9% and a specificity of 97.8% on average. At the interval-based level, the system gave an average sensitivity of 98.7% with an average false detection rate of 0.27/h. Compared with this detection method, our proposed algorithm had a much higher segment-based sensitivity. In a recent study of Chua et al., a subject-independent quadratic discriminant classifier incorporating modified features (average half-wave amplitude and duration, coefficient of variation of half-wave duration, line length, and rectified zero crossings) was first built [44]. It was then used to derive subject-specific classifiers by determining subjectspecific posterior probability thresholds via user interaction. The two schemes were tested on the intracranial EEG from 15 subjects of the EEG dataset and achieved a sensitivity of 78% and a false positive rate of 0.18/h. Compared with Chua's algorithm, our proposed algorithm provided comparable sensitivity, albeit with a little higher false positive rate.

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

37

Table 5 Related works and the results of performance assessment. Authors

Number of patients analyzed

Sensitivity (epoch-based) (%)

Specificity (%)

Sensitivity (event-based) (%)

False detection rate (/h)

Aarabi et al. [38] Chua et al. [44] Raghunathan et al. [45] Yuan et al. [42] This work

21 15 5 21 21

68.9 – 87.5 91.72 91.01

97.8 – 99.82 94.89 95.77

98.7 78 – 93.85 94.05

0.27 0.18 – 0.35 0.27

Raghunathan et al. proposed a cascaded two-stage seizure detection algorithm that was computationally efficient and resulted in a lowpower hardware implementation [45]. After separating the EEG data into desired bands of interest using the discrete wavelet transform, two linear time-based features (coastline and variance energy) were extracted to generate vector patterns and were evaluated by a pattern detector. The algorithm performed only on 5 patients with a specificity and a sensitivity of 99.82% and 87.5%, respectively. In comparison with their system, our proposed algorithm's sensitivity was better. In a more recent work, Yuan et al. presented a seizure detection system for multichannel long-term EEG [42]. In the system, the fractal intercept and the relative fluctuation index of EEG signals were extracted as a nonlinear feature and a linear feature, respectively, and the extreme learning machine was adopted as the classifier. Assessing on the data of 21 patients, a segment-based sensitivity of 91.72%, a specificity of 94.89%, an event-based sensitivity of 93.85%, and a false detection rate of 0.35/h were achieved. Compared with the system, the presented method improved the values of these performance metrics and obtained a better performance. Table 5 lists the related works and their results of performance assessment in detail. Although the proposed seizure detection method yielded comparable results, we are aware that a small number of patients were investigated in this study, and the dataset consisted of EEG recordings from 21 patients. Therefore, in order to better evaluate the proposed method and confirm the presented results, we will apply the algorithm to a larger set of epileptic EEG recorded in more realistic clinical settings in the near future. 5. Conclusion In this paper, we have presented a patient-specific epileptic seizure detection method in long-term EEG data. The detection method combined k-nearest neighbor fractal dimension, gradient boosting classifier, and a series of postprocessing. When the proposed method was tested on the Freiburg EEG dataset, a high sensitivity of 94.05% and a low false detection rate of 0.27/h were achieved. These results suggest that the proposed seizure detection method can be a candidate for clinical use. Further work includes methodological refinements for enhancing the real-time ability and algorithm validation on very large continuous, multichannel datasets. Acknowledgments The support of the Key Program of the Natural Science Foundation of Shandong Province (No. ZR2013FZ002), the Program of Science and Technology of Suzhou (No. ZXY2013030), the Development Program of Science and Technology of Shandong (No. 2014GSF118171), and the Fundamental Research Funds of Shandong University (No. 11170074611102) is gratefully acknowledged. Conflict of interest The authors declare that they have no conflicts of interest in connection with this work.

References [1] Shoeb A, Edwards H, Connolly J, Bourgeois B, Treves ST, Guttag J. Patient-specific seizure onset detection. Epilepsy Behav 2004;5(4):483–98. [2] Lagerlund TD, Cascino GD, Cicora KM, Sharbrough FW. Long-term electroencephalographic monitoring for diagnosis and management of seizures. Mayo Clin Proc 1996; 71(10):1000–6. [3] Gotman J. Automatic recognition of epileptic seizures in the EEG. Electroencephalogr Clin Neurophysiol 1982;54(5):530–40. [4] Gotman J. Automatic seizure detection: improvements and evaluation. Electroencephalogr Clin Neurophysiol 1990;76(4):317–24. [5] Qu H, Gotman J. Improvement in seizure detection performance by automatic adaptation to the EEG of each patient. Electroencephalogr Clin Neurophysiol 1993;86(2): 79–87. [6] Qu H, Gotman J. A patient-specific algorithm for the detection of seizure onset in long-term EEG monitoring: possible use as a warning device. IEEE Trans Biomed Eng 1997;44(2):115–22. [7] Grewal S, Gotman J. An automatic warning system for epileptic seizures recorded on intracerebral EEGs. Clin Neurophysiol 2005;116(10):2460–72. [8] Gabor AJ, Leach RR, Dowla FU. Automated seizure detection using a self-organizing neural network. Electroencephalogr Clin Neurophysiol 1996;99(3):257–66. [9] Wilson SB. A neural network method for automatic and incremental learning applied to patient-dependent seizure detection. Clin Neurophysiol 2005;116(8): 1785–95. [10] Srinivasan V, Eswaran C, Sriraam N. Approximate entropy-based epileptic EEG detection using artificial neural networks. IEEE Trans Inf Technol Biomed 2007; 11(3):288–95. [11] Gardner AB, Krieger AM, Vachtsevanos GJ, Litt B. One-class novelty detection for seizure analysis from intracranial EEG. J Mach Learn Res 2006;7:1025–44. [12] Meier R, Dittrich H, Schulze-Bonhage A, Aertsen A. Detecting epileptic seizures in long-term human EEG: a new approach to automatic online and real-time detection and classification of polymorphic seizure patterns. J Clin Neurophysiol 2008;25(3): 119–31. [13] Temko A, Thomas E, Marnane W, Lightbody G, Boylan G. EEG-based neonatal seizure detection with support vector machines. Clin Neurophysiol 2011;122(3):464–73. [14] Altunaya S, Telatarb Z, Erogulc O. Epileptic EEG detection using the linear prediction error energy. Expert Syst Appl 2010;37(8):5661–5. [15] Mohamed BV, Farooq O, Khan YU. Automatic seizure detection using higher order moments. International conference on recent trends in information, telecommunication and computing; 2010. p. 159–63. [16] Faust O, Acharya UR, Min LC, Sputh BH. Automatic identification of epileptic and background EEG signals using frequency domain parameters. Int J Neural Syst 2010;20(2):159–76. [17] Ayoubian L, Lacoma H, Gotman J. Automatic seizure detection in SEEG using high frequency activities in wavelet domain. Med Eng Phys 2013;35(3):319–28. [18] Conradsen I, Beniczky S, Wolf P, Kjaer TW, Sams T, Sorensen HB. Automatic multimodal intelligent seizure acquisition (MISA) system for detection of motor seizures from electromyographic data and motion data. Comput Methods Programs Biomed 2012;107(2):97–110. [19] Yuan S, Zhou W, Yuan Q, Zhang Y, Meng Q. Automatic seizure detection using diffusion distance and BLDA in intracranial EEG. Epilepsy Behav 2014;31:339–45. [20] Martis RJ, Acharya UR, Tan JH, Petznick A, Yanti R, Chua CK, et al. Application of empirical mode decomposition (emd) for automated detection of epilepsy using EEG signals. Int J Neural Syst 2012;22(6):1250027. [21] Guler NF, Ubey ED, Guler I. Recurrent neural network employing Lyapunov exponents for EEG signals classification. Expert Syst Appl 2005;29(3):506–14. [22] Ghosh-Dastidar S, Adeli H, Dadmehr N. Mixed-band wavelet-chaos-neural network methodology for epilepsy and epileptic seizure detection. IEEE Trans Biomed Eng 2007;54(9):1545–51. [23] Acharya UR, Molinari F, Sree SV, Chattopadhyay S, Ng Kwan-Hoong, Suri JS. Automated diagnosis of epileptic EEG using entropies. Biomed Signal Process 2012; 7(4):401–8. [24] Kannathal N, Choo ML, Acharya UR, Sadasivan PK. Entropies for detection of epilepsy in EEG. Comput Methods Programs Biomed 2005;80(3):187–94. [25] Acharya UR, Sree SV, Swapna G, Martis RJ, Suri JS. Automated EEG analysis of epilepsy: a review. Knowl-Based Syst 2013;45:147–65. [26] Theiler J. Estimating fractal dimension. J Opt Soc Am A 1990;7(6):1055–73. [27] Termonia Y, Alexandrowicz Z. Fractal dimension of strange attractors from radius versus size of arbitrary clusters. Phys Rev Lett 1983;51:1265–8. [28] Asvestas P, Matsopoulos GK, Nikita KS. Estimation of fractal dimension of images using a fixed mass approach. Pattern Recogn Lett 1999;20(3):347–54.

38

Y. Zhang et al. / Epilepsy & Behavior 43 (2015) 30–38

[29] Polychronaki GE, Ktonas PY, Gatzonis S, Siatouni A, Asvestas PA, Tsekou H, et al. Comparison of fractal dimension estimation algorithms for epileptic seizure onset detection. J Neural Eng 2010;7(4):046007. [30] Jin XC, Ong SH, Jayasooriah. A practical method for estimating fractal dimension. Pattern Recogn Lett 1995;16(5):457–64. [31] Katz MJ. Fractals and the analysis of waveforms. Comput Biol Med 1988;18(3): 145–56. [32] Higuchi T. Approach to an irregular time series on the basis of the fractal theory. Physica D 1988;31(2):277–83. [33] Freund Y, Schapire R. A decision-theoretic generalization of on-line learning and an application to boosting. J Comput Syst Sci 1997;55(1):119–39. [34] Friedman JH. Greedy function approximation: a gradient boosting machine. Ann Stat 2001;29:1189–232. [35] Lawrence R, Bunn A, Powell S, Zambon M. Classification of remotely sensed imagery using stochastic gradient boosting as a refinement of classification tree analysis. Remote Sens Environ 2004;90(3):331–6. [36] Hoffmann U, Garcia G, Vesin JM, Diserens K, Ebrahimi T. A boosting approach to P300 detection with application to brain–computer interfaces. Conference Proceedings of the 2nd International IEEE EMBS Conference on Neural Engineering; 2005. p. 97–100. [37] Schelter B, Winterhalder M, Maiwald T, Brandt A, Schad A, Timmer J, et al. Do false predictions of seizures depend on the state of vigilance? A report from two

[38] [39]

[40] [41] [42] [43]

[44] [45]

seizure-prediction methods and proposed remedies. Epilepsia 2006;47(12): 2058–70. Aarabi A, Fazel-Rezai R, Aghakhani Y. A fuzzy rule-based system for epileptic seizure detection in intracranial EEG. Clin Neurophysiol 2009;120(9):1648–57. Maiwald T, Winterhalder M, Aschenbrenner-Scheibe R, Voss HU, Schulze-Bonhage A, Timmer J. Comparison of three nonlinear seizure prediction methods by means of the seizure prediction characteristic. Physica D 2004;194:357–68. Temko A, Thomas E, Marnane W, Lightbody G, Boylan GB. Performance assessment for EEG-based neonatal seizure detectors. Clin Neurophysiol 2011;122(3):474–82. Zhang Y, Zhou W, Yuan Q, Wu Q. A low computation cost method for seizure prediction. Epilepsy Res 2014;108(8):1357–66. Yuan Q, Zhou W, Liu Y, Wang J. Epileptic seizure detection with linear and nonlinear features. Epilepsy Behav 2012;24(4):415–21. Esteller R, Echauz J, Tcheng T, Litt B, Pless B. Line length: an efficient feature for seizure onset detection. Proceedings of the 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2. 2001. p. 1707–10. Chua EC, Patel K, Fitzsimons M, Bleakley CJ. Improved patient specific seizure detection during pre-surgical evaluation. Clin Neurophysiol 2011;122(4):672–9. Raghunathan S, Jaitli A, Irazoqui PP. Multistage seizure detection techniques optimized for low-power hardware platforms. Epilepsy Behav 2011;22:S61–8.