Analysis of variations of correlation dimension and nonlinear interdependence for the prediction of pediatric myoclonic seizures – A preliminary study

Epilepsy Research 135 (2017) 102–114

Contents lists available at ScienceDirect

Epilepsy Research journal homepage: www.elsevier.com/locate/epilepsyres

Analysis of variations of correlation dimension and nonlinear interdependence for the prediction of pediatric myoclonic seizures – A preliminary study Mohamad Amin Sharifi Kolarijania,b, Susan Amirsalaria,c,d, Mohsen Reza Haidarie,

MARK

⁎

a

New Hearing Technologies Research Center, Baqiyatallah University of Medical Sciences, Molla Sadra Street, Vanak Square, Tehran, POB 14155-3, Iran Department of Bioelectrics, Faculty of Biomedical Engineering, Amirkabir University of Technology, Hafez Avenue, Tehran, POB 15875-4413, Iran c Department of Paediatrics, Faculty of Medicine, Baqiyatallah University of Medical Sciences, Molla Sadra Street, Vanak Square, Tehran, POB 14155-3, Iran d Shefa Neuroscience Research Center, Khatam al Anbia Specialty and Subspecialty Hospital, Rashid Yasemi Street, Tehran, Iran e Section of Neuroscience, Department of Neurology, Faculty of Medicine, Baqiyatallah University of Medical Sciences, Molla Sadra Street, Vanak Square, Tehran, POB 14155-3, Iran b

A R T I C L E I N F O

A B S T R A C T

Keywords: Pediatric myoclonic epilepsy Seizure prediction Scalp EEG Nonlinear analysis Correlation dimension Nonlinear interdependence

In this preliminary study, we evaluated the predictive ability of Correlation Dimension (CD) and Nonlinear Interdependence (NI) for seizures in pediatric myoclonic epilepsy patients. Scalp EEG recordings of eight diagnosed cases of myoclonic epilepsy were analyzed using Receiver Operating Curve (ROC) for discriminating the preictal period from interictal period. Furthermore, based on clinical seizure characteristics and EEG data, the spatiotemporal patterns of measures in clinically relevant areas of the brain were compared with other areas for each patient. CD showed a dominant increasing behavior in both all of the individual channels and channels of clinical interest for 75% of patients. For NI, the dominant direction was also increasing in 62.5% of patients for all of the individual channels and in 75% of patients for channels of clinical interest. However, there was no consistent general behavior in the timing of the preictal change amongst patients and within individual patient. Nonlinear measures of CD and NI can differentiate the preictal phase from the corresponding interictal phase. However, due to high variability, patient-wise tuning of possible automated systems for seizure prediction is suggested. This is the first study to employ nonlinear analysis for seizure prediction in pediatric myoclonic epilepsy.

1. Introduction Myoclonic epilepsy in children comprises up to 10% of all epilepsies and is characterized by simple seizures with nonrhythmic fast contractions of muscles in short periods of time occurring at variable intervals. These seizures are often combined with other types of generalized seizures (Camfield et al., 2013; Genton et al., 2013; Noachtar and Peters, 2009). Due to its specific etiopathology, neurophysiological and clinical features and associated disorders, appropriate treatment and management of myoclonic epilepsies in children pose special problems for the physician and the family of patients (Crespel et al., 2013; Koepp et al., 2014; Serafini et al., 2013; Wolf et al., 2015). Various techniques for automated analysis of EEG signal based on time domain (Lin et al., 2016), frequency domain (Bandarabadi et al., 2015) and time-frequency (Tzallas et al., 2007) analysis have been used in seizure detection and prediction studies. These techniques offer rapid

detection of ictal and interictal events and play important role in developing new treatment methods and improving the quality of life of patients (Kannathal et al., 2013; Litt and Echauz, 2002; Mormann and Andrzejak, 2016; Paternoster et al., 2013; Shorvon et al., 2011). EEG signals of normal brain have several characteristics of nonlinear systems, including limit cycles, instances of bursting behavior, hysteresis and amplitude-dependent frequency behavior (Carney et al., 2011). In addition, EEG signal recorded from patients with epilepsy is nonlinear and shows chaotic behavior (Mormann et al., 2005; Osorio et al., 1998; Rogowski et al., 1981). These observations suggest the possibility of detection of preictal state by dynamically monitoring minute variations of specific nonlinear measures that can signal the impending seizure before its onset. Nonlinear dynamics for characterizing epileptic EEG signals employ various univariate and bivariate nonlinear measures for the detection of preictal state such as correlation dimension (Lehnertz and Elger, 1998),

⁎ Corresponding author at: Section of Neuroscience, Department of Neurology, Faculty of Medicine, Baqiyatallah University of Medical Sciences, Molla Sadra Street, Vanak Square, Tehran, POB 14155-3, Iran. E-mail addresses: [email protected] (M.A.S. Kolarijani), [email protected] (S. Amirsalari), [email protected] (M.R. Haidari).

http://dx.doi.org/10.1016/j.eplepsyres.2017.06.011 Received 6 October 2016; Received in revised form 20 May 2017; Accepted 16 June 2017 Available online 17 June 2017 0920-1211/ © 2017 Elsevier B.V. All rights reserved.

103

Left side frontal, central, parietal and temporal regions Left side frontal, central regions

All regions Left side frontal and temporal regions

Right side of skull

LGS: Lennox-Gastaut Syndrome; CPS: Complex Partial Seizure; GTC: Generalized Tonic Clonic.

Lamotrigine, Sertraline NA 1/2 2/Frequent 19/F 10/M 7 8

15 NA

Carbamazepine, Levetiracetam 5/125 11/F 6

1

Vigabatrin, Topiramate 1/21 12/M 5

4

1/Frequent 4/2 9/M 17/M 3 4

9 17

1/5 9/M 2

9

Carbamazepine, Primidone Clonazepam

Tonic, head drop, left side clonic movements, dialeptic, staring Myoclonic, LGS Myoclonic jerk, generalized atonic, psychomotor Intractable mixed type (CPS, myoclonic and GTC) Myoclonic jerks on right side of face and right hand, staring Jerky movements of upper extremity, staring Myoclonic, partial, clonic, GTCS

Left side parietal and temporal regions Myoclonic jerk, head drop, tonic, tonic clonic

Acetazolamide, Depakine, Phenytoin Compound (Phenytoin and Phenobarbital) Clobazam, Phenobarbital, Acetazolamide, Primidone 3/5 4/M

3

Duration of observation period (day)/ No. of Seizures

1

We carried out a retrospective study using EEG data of patients who underwent long term Video-EEG monitoring due to unknown type of epilepsy or intractable seizure focus at the Shefa Neuroscience Research Center, Khatam al Anbia Hospital, Tehran, I.R. Iran. Subjects’ demographic data, including seizure characteristics and corresponding clinically relevant regions/EEG channels were obtained from their medical records (Table 1). The study sample comprised of eight patients from 2010 to 2015. This study was approved by the medical research ethics committee of the Baqiyatallah University of Medical Sciences. The patients’ medical records and EEG data were accessed with permission of the director of Shefa Neuroscience Research Center. According to the declaration of Helsinki, the confidentiality of patients was maintained throughout the data acquisition and analysis and each

Epilepsy (years)

2.1. Subjects

Age (years)/ Gender

2. Methods

Patient #

Table 1 Patients’ demographic and seizure data and EEG recording characteristics.

Drugs

Seizure Characteristics

Relevant EEG Channels/Regions

correlation density (Martinerie et al., 1998), largest Lyapunov exponent (Iasemidis et al., 1990), Kolmogorov entropy (van Drongelen et al., 2003), nonlinear interdependence (Arnhold et al., 1999) and phase synchrony (Mormann et al., 2000). These nonlinear measures provide different information about the underlying process leading to a seizure (Lehnertz et al., 2001). Therefore, automated algorithms designed to detect preictal state usually employ both univariate and bivariate measures (Brinkmann et al., 2016). Application of nonlinear measures for seizure prediction shows different spatiotemporal characteristics including the direction and timing of variations in different areas of the brain during assumed preictal period. Studies have reported either marked decrease in the dimension of EEG signals using correlation dimension (Elger and Lehnertz, 1998), or both increase and decrease in amplitude of correlation dimension during preictal period (Aarabi and He, 2012). Similarly nonlinear interdependence shows either a decrease in epileptogenic zones seconds to minutes prior to seizure (Arnhold et al., 1999), or various patterns of synchrony in different areas (Aarabi and He, 2012; Mormann et al., 2005). These variations have also been observed in the durations of preictal state ranging from seconds to hours (Bandarabadi et al., 2015; Mormann et al., 2003b). Furthermore, depending upon the type of algorithm used for the detection of preictal state, best possible results can be obtained by the placement of electrodes either over seizure onset zones (Gadhoumi et al., 2012), or over remote areas (Kuhlmann et al., 2010; Mormann et al., 2003b). Automated seizure prediction systems are designed for specific predictive measures (Gadhoumi et al., 2016) and sometimes lead to contradicting results. This calls for further investigation of spatiotemporal characteristics of nonlinear measures extracted from epileptic EEG signals. For this purpose, our study focuses on two nonlinear measures: Correlation dimension as the univariate measure, and nonlinear interdependence as the bivariate measure. The application of these measures in predicting pediatric myoclonic seizures is analyzed based on their performance in discriminating the preictal period from the interictal period. Although some studies have used accelerometry for the detection of myoclonic seizures (Ramgopal et al., 2014), to our knowledge, this is the first study that employs nonlinear EEG analysis for the prediction of myoclonic seizures. We narrowed our focus on pediatric patients as they are one of the most challenging to manage (Jerger et al., 2001; Meyer-Lindenberg, 1996; Paternoster et al., 2013; van Drongelen et al., 2003). Application of nonlinear techniques for pediatric patients needs a specific and discrete treatment (van Dijkman et al., 2016b) (van Dijkman et al., 2016a). In this preliminary study, EEG recordings of a single seizure from eight diagnosed cases of myoclonic epilepsy were analyzed for the amplitude distribution of each of these nonlinear measures during interictal and preictal periods. Also, the spatiotemporal patterns of nonlinear measures in clinically relevant areas of the brain were compared with the other areas.

Right frontal region All regions

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

1985). In order to reduce the effect of temporal correlation, a Theiler window of 0.2 s was employed in accordance with the reciprocal of the mean frequency of the power spectrum (Rosenstein et al., 1993). CD, as the univariate measure, was calculated for all the available channels. For NI2 the bivariate measure, only local combinations of channel pairs were included as shown by gray dashed lines in Fig. 1. For this purpose, the channels were grouped in five rows in the anterior-posterior plane: (Fp1-Fp2), (F9-F7-F3-Fz-F4-F8-F10), (T9-T7-C3-Cz-C4-T8-T10), (P9-P7P3-Pz-P4-P8-P10) and (O1-O2) (Klem et al., 1999). For each channel, four possible local connections were considered: one to upper row, one to lower row, one to right (in the same row) and one to left (in the same row). For Fp1, four choices were available as the connection to lower row, namely F9, F7, F3, or Fz (the channels Fp2, O1 and O2 have the similar situation). Since the standard procedure of positioning of standard 10–20 scheme includes the circumferential measurement consisting of Fp1-F7-T7-P7-O1-O2-P8-T8-F8-Fp2 (Klem et al., 1999), the local pairs were selected as (Fp1-F7), (P7-O1), (O2-P8) and (F8-Fp2) for these channels as shown in Fig. 1. Furthermore, for each channel pair i and j, the average of NIs calculated for i to j and j to i combinations was used as a general symmetrized measure of nonlinear interdependence (Mormann et al., 2005). The details of calculation of these measures are provided in Appendix. While calculating the nonlinear measures for consecutive windows of EEG data, the time profile of the measure was derived for each channel (pair). Finally, this time profile was smoothed using a backward moving average filter of 5 min length. Fig. 2 shows the process described above.

patient was assigned a code number (from 1 to 8). 2.2. EEG data All subjects showed at least one seizure during Video-EEG monitoring period. The recordings were performed using the extended 10–20 scheme including 29 channels with a sampling frequency of 500 Hz and 64ch Ref montage. Fig. 1 shows the positioning of the recorded EEG channels and the corresponding 38 local channel pairs for which nonlinear interdependence were calculated. For Patient #1, due to high level of noise and artifact, one of the channels (O1) and corresponding two channel pairs were discarded in preprocessing (see Section 2.3.1). For each subject, one myoclonic seizure was analyzed. The obtained video-EEG data was replayed and the exact onset of seizures was determined by a neurologist. Using the onset time, a 65 min window of EEG data (60 min before seizure onset and 5 min after seizure onset) was extracted. In order to avoid postictal effects, seizures for analysis were chosen so that there was at least a 90 min period between the onset of the previous seizure and that of the analyzed seizure (a minimum 30 min period was considered for the postictal effects to vanish) (Mormann et al., 2005). For all subjects the needed 65 min EEG data were available, except for patient #5. For this patient, the available EEG data included the time window between 5 min to 31.5 min before seizure onset. 2.3. Feature extraction

2.4. Statistical discrimination of preictal and interictal periods The mathematical analysis of this study, including calculation of nonlinear measures and statistical analysis of their amplitude distributions (Section 2.4) was performed using MATLAB software version R2009b.

consecutive identical values. Any window with more than 40 consecutive identical values was regarded as artifact and any channel with more than 5% of its window segments identified as artifact was discarded from analysis (Mormann et al., 2005). Based on this criterion, only channel O1 of patient #1 was discarded The line noise was canceled out using a 50 Hz notch filter. The external reference noise was canceled out using the average of all channels as the common reference. To do so, in each time step, the average of all channels was subtracted from the recorded value for each channel. The mean of signal was set to zero. To do so, the average of each channel in each window was subtracted from the recorded values of that channel.

2.4.1. Receiver operating curve analysis In order to investigate possible variations of nonlinear measures in the preictal period compared to the interictal period, the amplitude distribution of each measure in preictal period was compared with that of interictal period using Receiver Operating Curve (ROC). To illustrate it briefly, the 60 min EEG data before the seizure onset was segmented into two parts of preictal and interictal; i.e. the s min immediately before seizure onset and the remainder 60-s min were labeled as preictal and interictal, respectively. Using the values of nonlinear measure in each of these periods, the amplitude distribution of the measure in corresponding period was obtained. For this purpose, using the ProbDistUnivKernel function of MatLab, a probability distribution was derived for the amplitude of the measure. This function produces a nonparametric probability distribution for the signal based on a normal Kernel smoothing function. Finally, by varying the threshold of amplitude continuously, the sensitivity of discrimination versus one minus its specificity was plotted (ROC plot). The Area Under Curve (AUC) obtained from ROC was used to determine the magnitude and direction of variation in preictal period compared to interictal period. Based on the initial hypothesis, AUCs greater (less) than 0.5 showed a decrease (an increase) in the amplitude of nonlinear measure in preictal period compared to the interictal period. Also, the absolute value of difference between AUC and 0.5 quantifies the magnitude of discrimination.

2.3.2. Features calculation For each patient, two nonlinear measures of correlation dimension and nonlinear interdependence were calculated. The time lag for reconstruction was estimated using method described earlier (Moon et al., 1995) and resulted in values ranging from 3 to 6 for the preprocessed EEG data. For embedding dimension, the Cao’s method (Cao, 1997) was employed and yielded values between 4 and 7. Based on these results, time lag of five and embedding dimension of seven were considered to be optimal choices for comparable reconstruction of windows in all patients. CD,1 was calculated using the Takens algorithm (Takens,

2.4.2. Analysis schemes Similar to the method used by (Mormann et al., 2005), in order to investigate both global and regional effects in the preictal variation, the ROC analysis was performed for all channels together and individual channels separately. In the first scheme, the values of nonlinear measure from all channels in preictal and interictal periods were grouped together to form the corresponding amplitude distributions (Fig. 3B). In the second scheme, the amplitude distributions of nonlinear measure in preictal and interictal periods were derived and analyzed separately for each individual channel (Fig. 3C). Furthermore, in order to test if the

2.3.1. Preprocessing The raw EEG data were segmented in 20 s non-overlapping windows. Each window was then preprocessed. The preprocessing included four following steps:

• The data were scanned for dropouts or clippings in the form of

• • •

1

2

Correlation Dimension.

104

Nonlinear Interdependence.

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Fig. 1. Configuration of recorded EEG channels (black boxes) and local pairs of channels for nonlinear interdependence (dashed lines).

Fig. 2. Example of extraction of Correlation Dimension (CD). For non-overlapping 20 s windows, the EEG data is preprocessed and reconstructed in the state space. The calculated CD in consecutive windows forms the time profile of the nonlinear measure.

important role in the performance of the nonlinear measure in discriminating the preictal period from the interictal period. Therefore, in this study, each ROC analysis was performed for five different lengths of time viz. 5, 10, 15, 20 and 25 min for preictal period. The preictal length (s) and direction of change (increase/decrease) corresponding to the largest magnitude of discrimination were used to describe the timing and direction of preictal variations. Fig. 4 illustrates this process.

specific combinations of channels perform better than others in discriminating the preictal period from the interictal period, the results of ROC analysis for five best individual channels (the channels with largest magnitude of discrimination, i.e. largest absolute value of difference between their AUCs and 0.5) were reported separately. 2.4.3. Duration of preictal period The main parameter of above analysis was the duration of the preictal period (s) that determines the timing of preictal variations. Assuming that variations in a nonlinear measure that point to an impending seizure, occur with a specific timing before the seizure onset, the appropriate selection of the preictal period length plays an

2.4.4. Spatial patterns of preictal vs. interictal variations Based on the AUCs derived from individual channel analysis, the spatial pattern of preictal variation was plotted using a color map as shown in Fig. 3D. As depicted, the AUC for each channel was rescaled to 105

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Fig. 3. Example of Receiver Operating Curve (ROC) Analysis for 25 min preictal period. (A) Smoothed time profile of CD for each channel. (B) All channel ROC analysis using the amplitude distributions derived from all channels together. (C) Individual channel ROC analysis using the amplitude distributions derived for each channel separately. (D) The spatiotemporal pattern based on the individual channels’ AUCs as a measure of discrimination between preictal and interictal periods. (E) Spatial positioning of EEG channels in the color map for depicting spatial patterns.

As listed in Table 1, the medical reports of patients #3 and #8 did not mention any specific area of clinical interest; therefore, these subjects were not included in this analysis.

the [-1,1] interval and plotted using a specific color code on a map representing the surface of scalp and corresponding positions of recording EEG leads. In this color map, the direction of change, increase or decrease, is shown in red and blue, respectively, with the intensity of color corresponding to the magnitude of discrimination. The detailed positioning of different EEG channels is shown in Fig. 3E. Using the spatial patterns of preictal variations for different lengths of preictal phase (s), the spatiotemporal patterns of these variations were derived.

3. Results 3.1. Variations in preictal period versus interictal period The results of the ROC analysis comparing the amplitude distributions of nonlinear measures in preictal period vs. interictal period for three analysis schemes of all channels, individual channels, best five channels (or channel pairs) are provided in this section.

2.4.5. Preictal variations in clinical areas of interest Finally, the possible differences in timing, direction or magnitude of preictal variations in areas of the brain that were of clinical relevance based on seizure characteristics were investigated. To do so, the results of individual channel analysis were compared for the channels corresponding to these clinically relevant areas with the remaining channels.

3.1.1. Magnitude of variations Fig. 5 shows the magnitude of difference between preictal and Fig. 4. Example of ROC analysis for different durations of preictal period (s). (A) Each ROC analysis was performed for five durations of 5, 10, 15, 20 and 25 min for preictal period. Corresponding amplitude distributions and ROC curves were derived for each s. (B) The case shown in shaded row had the largest magnitude of discrimination (largest absolute difference of AUC with 0.5). This magnitude and its corresponding direction and timing (preictal duration) yielded the best performance of the measure in differentiating preictal period from interictal period.

106

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Fig. 5. Magnitude of preictal vs. interictal variations. For the individual channel analysis and best five channels, the reported magnitudes are channel-wise means. The vertical dotted lines are the patient-wise averages of the corresponding analysis schemes.

interictal periods for the considered nonlinear measures. The reported magnitudes correspond to the largest absolute difference of AUC with 0.5–which represents the level of difference between the preictal and interictal periods in ROC analysis − amongst the five durations of preictal period considered in this study (see Fig. 4). For individual channel analysis and best five channels, the magnitude of variation was calculated separately for each channel and the reported magnitudes for each patient are channel-wise averages; i.e. the gray and white bars in Fig. 5 show the average of magnitude for individual channels and the best five channels of each patient, respectively. As shown in Fig. 5A, for CD, the patient-wise average of magnitude of change is 0.15 with standard deviation of 0.11–0.15(011) – for all channels analysis, and its value increases to 0.27(0.11) for individual channel analysis. Furthermore, for the best five channels, a better average magnitude of 0.38(0.10) was derived amongst patients. The same increase in magnitude is seen in NI as depicted in Fig. 5B. The average magnitude of all channel pairs analysis, individual channel pair analysis and best five channel pair were 0.09(0.04), 0.29(0.07) and 0.44(0.09), respectively.

Table 2 Direction of preictal vs. interictal variations for ROC analysis for individual channels analysis. For each patient, mode and corresponding frequencies amongst all individual channels and five best channels are reported. Nonlinear measure

Patient

ROC analysis Direction of preictal vs. interictal variations Individual channel Analysis

Correlation Dimension

Nonlinear Interdependence

3.1.2. Direction of variations Using the results of ROC analysis for best magnitude of performance for different durations of preictal period, the corresponding direction of variations (decrease or increase) in the amplitude of nonlinear measures in preictal period compared to the interictal period were derived. Table 2 shows these directions for the individual channels analysis. The channel-wise mode of direction of change and its frequency were determined for individual channels and best five channels for each patient. For instance, for patient #1, in 57% (16 out of 28) of channels, a decrease was observed in CD, while, 4 out of 5 best channels (80%) showed a decrease in CD in preictal period. The same direction of change was derived in the all channel analysis scheme for each patient (data not shown in Table 2). The common behavior observed in the direction of change in the preictal period amongst the patients for CD was increasing with the frequency of 75%. This mode was also increasing for NI amongst patients, however, with a lower frequency of 62.5%. Furthermore, considering the direction of change amongst individual channels in each patient, for CD, in 6 out of 8 patients the mode of directions had a high frequency (greater than 70%) of occurrence in both groups of individual channels and the best five channels. The same behavior is seen in NI.

1 2 3 4 5 6 7 8 All 1 2 3 4 5 6 7 8 All

Mode

Freq. (%)a in all channels (pairs)

Freq. (%)a in best 5 channels (pairs)

Decrease Increase Increase Increase Decrease Increase Increase Increase 75% Inc. Increase Increase Increase Increase Increase Decrease Decrease Decrease 62.5% Inc.

57 59 90b 97b 93b 72b 90b 83b

80 60 100b 100b 100b 80b 100b 100b

81b 92b 100b 82b 92b 82b 66 68

80b 100b 100b 80b 100b 100b 60 80

a Percentage of channels (channel pairs) showing the same direction of change as the mode. b The case where the frequency of the mode is greater than 70% amongst all the channels and the best five channels.

various timings for preictal variations in different areas of brain in each patient. 3.2. Spatiotemporal patterns of preictal variations Based on the magnitude and direction of preictal duration in the individual channels for different preictal period durations considered for analysis, the spatiotemporal patterns of variations in the preictal period compared to the interictal period were derived. Figs. 7 and 8 show the spatiotemporal patterns for CD and NI, respectively. The obtained results are based on the corresponding montage (see Fig. 1); especially, for NI, the lines in Fig. 8 (presenting the direction and intensity of variations) match the dashed lines of Fig. 2 connecting the EEG leads (presenting the local channel pairs used for calculation of NI). Note that for better illustration, the results for preictal period of 2 min duration are also shown; however, due to the small number of data points in the preictal period for this duration (6 points), corresponding results should be analyzed with caution (for this reason the results for preictal period of 2 min were not used in other parts of the study). Also provided is the result of ROC analysis for comparison of the

3.1.3. Timing of variations Fig. 6 shows the duration of preictal period corresponding to the largest magnitude of change in the preictal period compared to the interictal period, derived from ROC analysis for each patient for different analysis schemes. For the individual channels and the best five channels, the statistics (mean and standard deviation) of preictal periods amongst channels are reported. As depicted, there is no specific behavior in timing of preictal variations amongst patients. Furthermore, large standard deviations of preictal durations in individual channels of each patient (even amongst best five channels) suggest 107

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Fig. 6. Timing of preictal vs. interictal variations for ROC analysis. For the individual channels and the best five channels, the reported durations are channel-wise means and standard deviations for each patient.

between most of the frontal, temporal and parietal areas (dark red lines between corresponding EEG channels). This general increase in NI between different areas is also the dominant behavior after seizure. Fig. 7 shows various spatiotemporal patterns of CD prior to seizure for different patients. For patients #1 and 2, various decreasing and increasing dynamics are detectable in different areas of the brain with different timings before seizure onset; while for patients #3, 4, and 7, a uniform decreasing or increasing behavior is seen prior to seizure. For patient #6, the dominant behavior in most of the brain areas is increasing for 25–15 min before seizure and decreasing for 5–2 min before seizure. Similar pattern is seen for patient #8 as the dominant behavior changes from decreasing to increasing as we get closer to the seizure onset. Patient #5 shows a weak decrease in CD, 15–10 min before seizure. Fig. 8 shows same variety in spatiotemporal patterns of NI for different patients. Except for patients #2, 3, and 5 for which a uniform

period after seizure onset (ictal and postictal periods), for two different lengths of 2 and 5 min, with the 60 min period before seizure onset. To illustrate how to interpret these patterns, Fig. 7 shows that for patient #1, CD decreases in prefrontal areas in 25–15 min before seizure (dark blue areas corresponding to Fp1 and Fp2 channels). As we get closer to the seizure onset, i.e. 5–2 min before seizure, there is an increase in CD in frontal and temporal areas (red areas corresponding to F7, F3, Fz, F4, F8, T7 and T8 channels) and a decrease in parietal and occipital areas (blue areas corresponding to P3, Pz and O2 channels). After seizure, however, a general decrease in different areas of brain is seen, especially for 5 min time window after seizure. For NI, Fig. 8 shows that for patient #1, both increasing (red) and decreasing (blue) dynamics of local interdependence is seen in different areas of the brain in the 25–15 min time interval before seizure onset. However, as we get closer to the seizure, especially in the 5–2 min time window, there is a significant increase in local interdependence

Fig. 7. Spatiotemporal pattern of CD for each patient using the area under ROC for comparison between preictal and interictal phases (the first five figures of each row) and that of after seizure and before seizure periods (the last two figures of each row). The increase and decrease in the amplitude of nonlinear measure in different areas are shown in red and blue colors, respectively. The color intensity corresponds to the magnitude of change. For patient #5, due to lack of needed EEG data, the patterns for some preictal durations were not available. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

108

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Fig. 8. Spatiotemporal pattern of NI for each patient using the area under ROC for comparison between preictal and interictal phases (the first five figures of each row) and that of after seizure and before seizure periods (the last two figures of each row). The increase and decrease in the amplitude of nonlinear measure in different areas are shown in red and blue colors, respectively. The color intensity corresponds to the magnitude of change. For patient #5, due to lack of needed EEG data, the patterns for some preictal durations were not available. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

differences in the preictal dynamics between clinical and non-clinical areas. For CD, the direction of change in patient #1 is increasing in 75% of channels in clinical areas while decreasing in 63% of the rest of the channels. For NI, three major differences are reported including timing of preictal changes in patient #1 (5 min for clinical areas and 12.50 min for non-clinical areas), direction of preictal changes in patient #6 (67% increase in clinical areas and 74% decrease in non-clinical areas), and magnitude of preictal changes in patient #7 (0.45 in clinical areas and 0.27 in non-clinical areas). Except for these major differences, the timing, direction and magnitude of preictal variations in clinical areas are generally the same as that of non-clinical areas. The last column of Table 3 lists the number of channels in areas of clinical interest that are amongst five best channels with best performance in discriminating preictal period from interictal period. As can be seen, the channels in areas of clinical interest do not necessarily provide the best discriminating power. In order to investigate the possibility of consistent behavior in timing and direction of preictal variations in clinical areas, these characteristics were specifically analyzed for channels in areas of clinical interest. Table 4 summarizes the timing and direction of preictal variations in clinical areas for all the patients. As listed, the dominant direction of variation amongst patients for both nonlinear measures is increase in channels located in the areas of clinical interest, as was also derived for all of the individual channels. For CD and NI, the dominant behavior is increase in preictal period in clinical areas for 75% of

increasing behavior is seen in local interdependence in most of the brain areas prior to seizure, other patients exhibit various increasing and decreasing behavior in different areas of the brains and in different time periods before seizure onset. 3.3. Analysis of areas of clinical importance In order to investigate the possible differences of preictal variations in areas of clinical interest compared to the other areas of the brain, the results of ROC analysis for individual channels were compared between these two areas. To do so, magnitude, direction, and timing of preictal variations were calculated for EEG channels (channel pairs) in each of these areas separately. The detailed statistics of preictal variations in clinical areas of interest and other brain areas are presented in Table 3. Note that for patients #3 and #8 no specific area was reported to be of clinical interest. Table 3 shows some differences in timing, magnitude and even direction of the preictal variations in clinical areas compared to the other areas in some patients. For instance, for patient #1, in preictal period, CD shows mostly an increase with average magnitude of 0.15 (0.12) in channels of clinical areas, while in the rest of the channels, the most frequent behavior is decrease with a higher average magnitude of 0.21 (0.09). Additionally, for NI, the timing of preictal variations is slower in clinically relevant areas compared to the other areas. As shown in shaded cells of Table 3, there are a few major 109

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Table 3 Comparison of preictal vs. interictal variations in areas of clinical interest with other areas using results of individual channels ROC analysis. Shaded cells show the major differences in preictal dynamics between clinical areas and other areas of the brain. Patient

a

1

Nonlinear measure

b

Area

CD

C NC C NC C NC C NC C NC C NC C NC C NC C NC C NC C NC C NC

NI 2

CD NI

4

CD NI

5

CD NI

6

CD NI

7

CD NI

c

Preictal vs. interictal ROC analysis

d

Nc

Preictal duration (min)

Direction

Magnitude

Mean (STD)

Mode (Freq.%)

Mean (STD)

13.75 (4.79) 12.50 (8.08) 5.00 (0.00) 12.50 (8.97) 13.46 (4.74) 15.63 (6.80) 10.36 (5.71) 11.25 (6.30) 9.29 (7.87) 6.14 (4.35) 13.00 (10.95) 16.82 (8.91) 11.82 (2.52) 12.22 (2.56) 14.58 (1.44) 14.04 (2.01) 20.00 (10.00) 16.20 (8.33) 18.33 (7.46) 16.00 (8.46) 13.33 (10.41) 13.27 (8.83) 15.00 (14.14) 13.75 (7.59)

Inc. (75) Dec. (63) Inc. (100) Inc. (71) Inc. (85) NA Inc. (86) Inc. (96) Inc. (86) Inc. (100) Inc. (60) Inc. (85) Dec. (91) Dec. (94) Inc. (100) Inc. (88) Inc. (75) Inc. (72) Inc. (67) Dec. (74) Inc. (100) Inc. (88) NA Dec. (67)

0.15 0.21 0.25 0.26 0.25 0.24 0.40 0.40 0.40 0.47 0.38 0.33 0.11 0.11 0.15 0.15 0.26 0.24 0.22 0.25 0.20 0.24 0.45 0.27

(0.12) (0.09) (0.06) (0.10) (0.08) (0.11) (0.12) (0.11) (0.09) (0.05) (0.11) (0.10) (0.04) (0.04) (0.05) (0.15) (0.05) (0.08) (0.11) (0.11) (0.09) (0.10) (0.06) (0.11)

e

1 1 1 2 0 2 2 2 1 0 0 1

a

Patients #3 and #8 are not included since no specific area was reported to be of clinical interest in their medical reports. Nonlinear measures are Correlation Dimension (CD) and Nonlinear Interdependence (NI). The results are given for two areas (group of channels): channels corresponding to areas of Clinical Interest (C) and channels corresponding to areas of No Clinical Interest (NC). These areas are derived using the medical reports of patients (see Table 1). d Shaded cells show the cases where there is a major difference in the preictal dynamics between clinical and non-clinical areas. e Number of channels of clinical interest amongst the five best channels (channels with best magnitude of performance). b c

not seen. Furthermore, the timing of these changes is highly variable between different areas of the brain in each patient. This study included two nonlinear measures of correlation dimension and nonlinear interdependence. CD, as a univariate measure, describes the general state of an area of brain, while NI, as a bivariate measure, describes the local interactions between different areas of the brain (Mormann et al., 2005). Therefore, NI is expected to be more affected by the process of seizure spread and therefore the specific mechanism of channel selection. Additionally, since univariate and bivariate measures provide different information about the underlying dynamics of seizures (Lehnertz et al., 2001), joined application of these measures is expected to exhibit higher performance as shown by recent studies (Aarabi and He, 2012; Brinkmann et al., 2016).

Table 4 Timing and direction of preictal variations in clinically relevant areas. The timings are based on the average preictal durations and the directions are based on mode of directions in channels from the areas of clinical interest. Patient

Receiver Operating Curve (ROC) Analysis Preictal vs. interictal period

1 2 3a 4 5 6 7 8a All patients

a

a

CD

NI

Increase in 14 min NA Increase in 15 min Increase in 9 min Decrease in 12 min Increase in 20 min Increase in 13 min Increase in 8 min 75% Increase in 8–20 min 12.5% Decrease in 9 min

Increase in 5 min Increase in 10 min Increase in 16 min Increase in 13 min Increase in 15 min Increase in 18 min NA Decrease in 12 min 75% Increase in 5–18 min 12.5% Decrease in 12 min

4.1. Magnitude of variations Comparison of the patient-wise averages of magnitude of variations (dotted lines in Fig. 5) shows an increase in the ability of both nonlinear measures of CD and NI in discriminating preictal period from interictal period in the individual channel analysis scheme versus the all channels analysis scheme. Furthermore, the patient-wise average of magnitude in the best five channels was higher than that of individual channels for both nonlinear measures. These observations point out not only the nonhomogeneous variety of dynamics in different areas of the brain in preictal period, but also the importance of the proper channel selection for a better seizure prediction as some areas show better performance in discriminating the preictal period from the interictal period. This process of best channels selection is actually a common technique in studies on seizure prediction (Mormann et al., 2005). Additionally, comparing the amount of increase in patient-wise magnitude of variations in different schemes for two measures shows a

For these patients, the provided results are for all the channels (see Section 3.3).

patients. 4. Discussion In this study, the spatiotemporal variations of correlation dimension and nonlinear interdependence prior to seizure in pediatric myoclonic epilepsy patients were statistically analyzed to see if these variations can signal the impending seizure. Results confirm the existence of a preictal period exhibiting specific characteristics compared to the interictal period; however, a consistent and uniform behavior in timing or direction (increase/decrease) of these variations among patients was 110

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

4.3. Timing of variations

higher level of increase for NI versus CD (Fig. 5). Specifically, for CD, the patient-wise average increased from 0.15 in the all channels analysis to 0.38 in best five channels; while, for NI, it increased from 0.09 to 0.44. This observation implies that separate analysis of channels and proper selection of channels for seizure prediction have a greater influence on performance of NI compared to CD. This may be due to the fact that NI is a bivariate measure quantifying the level of synchronization between two areas of brain (local interactions), while CD is a univariate measure quantifying the state of an area of brain (general state) (Mormann et al., 2005). Therefore, NI is expected to be more affected by the process of best channels selection.

Comparison of the preictal duration in all channels with the average preictal duration in individual channels (Fig. 6) shows that for each patient, the timings of the two schemes, as expected, matched. However, comparison of the preictal durations between patients shows no common behavior regarding the timing of the preictal variations in any of the analysis schemes. This was also expected considering the different durations reported for preictal period in previous studies that vary from minutes (D'Alessandro et al., 2003; Jerger et al., 2001; Le Van Quyen et al., 2000; Niederhauser et al., 2003; Sato et al., 2015) to hours (Esteller et al., 2005; Hively et al., 2000; Iasemidis et al., 2005; Litt et al., 2001; Mormann et al., 2003a). Furthermore, the large standard deviation of preictal duration seen in individual channels implies that the timing of preictal variations is highly variable in different areas of the brain within each patient. The same variability was also seen in the best five channels. Accounting for these timing variations in a seizure prediction system is costly; therefore, its effect on improving the performance of such a system needs further investigation.

4.2. Direction of variations Analyses of the mode of direction of variations in preictal period shows a general increasing behavior in 6 out of 8 patients for CD (see Table 2). The same behavior is seen in the best five channels. Some earlier studies, however, described a preictal decrease in temporal lobe epilepsy using intracranial recordings (Elger and Lehnertz, 1998; Lehnertz and Elger, 1998). The authors associated this decrease to lowdimensional states of brain activity due to increased synchronicity (Elger and Lehnertz, 1998; Lehnertz and Elger, 1998). A later study by Mormann et al. reported an increase in CD in preictal period compared to the interictal period (Mormann et al., 2005). Another study reported lack of a specific direction in preictal variations for CD (Aarabi and He, 2012). In this regard, some studies showed inability of correlation integral and correlation dimension in seizure prediction due to the high sensitivity of these measures to time, frequency and energy characteristics of signal (Harrison et al., 2005). Other studies also reported insufficient performance of CD for clinical applications (AschenbrennerScheibe et al., 2003; Maiwald et al., 2004). These observations further elaborate that due to the noisy and nonstationary nature of EEG signals, application of accurate definition of correlation dimension for describing the brain dynamics seems highly unrealistic, however, operational application of this measure for seizure prediction is conceivable (Aarabi and He, 2012; Elger and Lehnertz, 1998). For NI, an increasing behavior was seen in only 5 out of 8 patients in our study. An early study by Arnold et al. reported an initial decrease in NI allowing neurons to establish synchronization that led to later increase in NI (Arnhold et al., 1999). Higher levels of NI are also reported in the ictogenic areas in seizure-free periods. However, these results were based on EEG data of only two patients suffering from mesial temporal lobe epilepsy and neocortical lesional epilepsy (Arnhold et al., 1999). Our results are more in line with a later study where both preictal increase and decrease were reported (Mormann et al., 2005). The authors specifically observed a preictal increase in some areas and a preictal decrease in adjacent areas and associated it with loss of synchronization in one direction and enforced synchronization in another direction that possibly led to seizure. This is in consistence with the spatiotemporal patterns derived in our study as shown in Fig. 8. Other studies also reported various synchronization patterns between ictogenic and remote areas (Aarabi and He, 2012; Le Van Quyen et al., 2005). Finally, evaluation of the frequency of mode of direction between the individual channels and the best five channels for each patient reveals a relatively consistent behavior amongst channels of each patient. As shown in Table 2, for CD, the dominant direction (decrease or increase) is seen in more than 70% of individual channels and in more than 80% of best five channels for 6 out of 8 patients. For NI, in more than 80% of individual channels and best five channels, the same direction (decrease or increase) is reported. This illustrates that although a common behavior amongst patients regarding the direction of preictal variations may be missing, in each patient, the individual channels, especially the best channels, show the same increasing or decreasing behavior.

4.4. Spatiotemporal patterns of preictal variations The spatiotemporal patterns of preictal variations in Figs. 7 and 8 show two main classes of behaviors. For some patients, a general decrease or increase is seen in most areas of the brain, while for the other patients variable increasing and decreasing dynamics are seen in different areas as we get closer to the seizure onset. For instance, for CD in patient #4, a dominant gradual increasing behavior is seen in almost all brain areas. However, in the same patient, we saw different increasing and decreasing dynamics for NI. Our analysis shows no specific relationship between patients’ clinical characteristics and these variations in spatiotemporal patterns. This may be due to low temporal resolution considered for different preictal durations (5 min) in this study which might not match the specific dynamics of the preictal period of myoclonic seizures. Furthermore, inherent mechanisms of seizure spread in human brain are highly variable and complicated (Bromfield et al., 2006) and show a lot of interindividual variability (Jansen Holleboom, 2016). Myoclonic seizures in children have additional variability in their type and spread in different brain areas (Koepp et al., 2013; Serafini et al., 2013). Therefore, expecting uniform patterns extracted from spatiotemporal patterns of only two measures is largely unrealistic. 4.5. Analysis of areas of clinical relevance Detailed comparison of characteristics of preictal variations in areas of clinical interest with other areas in Table 3 shows some differences in magnitude, direction and timing of these variations in some patients, for both nonlinear measures. However, these differences are inconsistent amongst these patients for both measures. Nevertheless, we saw more differences in NI compared to CD, which is due to the fact that the process of channel selection and localization in analysis has a greater influence on NI’s performance in discriminating the preictal period from the interictal period as mentioned before. More importantly, the last column of Table 3 shows that the clinical channels do not necessarily show the best performance as on average, about one channel of the best five channels is amongst channels of clinical interest. This is in accordance with previous studies that reported lack of any relationship between best performing channels and seizure focal regions, as the best channels can be both ipsilateral or contralateral to these regions (Mormann et al., 2005). Actually, some studies employed a channel selection method that included channels from both near and remote regions based on seizure focus (Aarabi and He, 2012). Furthermore, as listed in Table 4, specific analysis of direction and timing of preictal variations restricted to channels of clinical interest, in 111

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

as opposed to intracranial EEG. Using intracranial EEG recordings for seizure prediction impose a significant limitation on applicability of these studies as these signals are recorded in controlled environments that are not compatible to daily lives of patients. In addition, intracranial EEG recording is not possible in developing countries as it is expensive and sophisticated technique. Furthermore, EEG recordings are under influence of artifacts caused by the patient such as eye movement, muscle artifacts, etc. These artifacts are usually non-systematic, thus, in order to remove them, some studies use specific epochs of EEG signals that are selected by human expert and free of artifacts. This usually leads to overestimation of predictive system’s performance. However, in this study, in order to minimize the human interaction and achieve a fully automated process, we avoided non-systematic filtering of unknown artifacts to derive more practical and realistic results.

search for a consistent behavior between patients in these areas, yields the same results as for all of the individual channels analysis, i.e. the channels residing in the areas of clinical interest mainly follow the dominant behavior in all brain areas, expect for NI, for which the frequency of dominant increasing behavior among patients is higher (75%). 4.6. Final remarks As described in methodology, a 5-min moving average filter was applied to calculated nonlinear measure to reduce the effects of fast dynamic artifacts such as eye blinks and even statistical fluctuations (Lai et al., 2003) in calculations of nonlinear measures. The positive effect of smoothing on performance of nonlinear measures in detecting preictal period has been reported in previous studies (Kuhlmann et al., 2010; Mormann et al., 2005). However, some of the fast variations, filtered out in this process, may characterize epileptic state of the brain. Additionally, application of other smoothing techniques such as Kalman filter has been shown to outperform the moving average filter (Zhang et al., 2014). Therefore, the possible effects of the smoothing strategy, e.g. moving average, Kalman filter, etc., and the specific length of the smoothing filter on the performance need further investigation. In this regard, it is noteworthy that in this study, preictal variations were analyzed for 5- to 25-min preictal period durations with 5-min time resolution for practical implementation. However, this specific arrangement has its own limitations as the actual preictal duration may differ among individual patients from these predefined values. In this preliminary study, we included only one seizure for each subject. Several other studies have based their nonlinear analysis on a single seizure per patient. Lehnertz and Elger, in one of their first reports on application of correlation dimension for seizure prediction, reported a marked loss of complexity based on their analysis of one seizure for 16 cases with pharmaco-resistant epilepsy (Lehnertz and Elger, 1998). Le Van Quyen and colleagues also included one seizure for 21 out of 23 patients in their study evaluating the application of nonlinear similarity for seizure prediction (Le Van Quyen et al., 2001). Other studies, also analyzed one seizure for some or all of patients (De Clercq et al., 2002; Martinerie et al., 1998; Mormann et al., 2003a). Nevertheless, inclusion of one seizure for each patient is a major limitation of our study. As EEG signal is nonstationary and affected by physiological variations in brain state (such as circadian rhythms) (Aschenbrenner-Scheibe et al., 2003; Mormann et al., 2005; Stacey et al., 2011), it leads to differences in the characteristics of individual seizures. Thus, analysis of long-term EEG data of myoclonic epilepsy patients including multiple seizures with variable ictal and interictal periods is suggested. Furthermore, considering the limited number of subjects that was available for this study, it is suggested to include more pediatric myoclonic patients for a reliable evaluation of performance of NI and CD in the prediction of incumbent seizures. Finally, in this study, in contrast to most studies, we used scalp EEG

5. Conclusion Statistical analysis of the two nonlinear measures of correlation dimension and nonlinear interdependence showed that these tools can discriminate the preictal state from interictal state and therefore signal an impending seizure. However, lack of consistency in direction (increase or decrease) and timing of these variations among patients, demonstrates the importance of patient-specific tuning of parameters of any seizure prediction system. Interestingly, the timing of preictal variations is also diverse in different areas of the brain, even among the best channels in each patient. Another pivotal factor is the proper selection of channels for better performance of the predictive system in detecting preictal state. The results showed that these channels do not necessarily reside in areas of clinical relevance and thus their selection has to be based on both the clinical semiology and EEG findings of individual patients. Finally, further assessment of the ability of these nonlinear measures by inclusion of more patients and also analysis of multiple seizures for each patient is of great importance. Conflicts of interest None. Acknowledgement The authors thank director of the Shefa Neuroscience Research Center Dr. T. Taheri for permission to access patient data, attending physicians Drs. J. Mehvari, N. Zangiabadi, and M. Rezvani for their cooperation in providing patients’ medical history and Dr. M. Nikchehreh for her help with EEG data. The authors also thank Dr S. Hashemi Fesharaki for his cooperation in data access. Cooperation of staff members of New Hearing Technologies Research Center is also appreciated. This research did not receive any specific grant from the funding agencies in the public, commercial, or not-for-profit sectors.

Appendix A For this research the Open TSTOOL software package (http://www.physik3.gwdg.de/tstool) were used for the calculation of nonlinear measures. For Correlation Dimension (CD), the following Takens estimator was employed: −1

CD = −

⎛ 2 ⎜ n (n − 1) ⎝

∑ 1≤i
log

xi − xj ⎞ forx i − x j < r r ⎟ ⎠

(A.1)

where . denotes Euclidean norm and xi is i-th point in the reconstructed state space trajectory for i = 1, …, n. r denotes the search radius of scaling region which was set to 0.05 relative to attractor diameter. Nonlinear Interdependence (NI) for the two reconstructed signals of xi and yi for i = 1, …, n, was calculated using following average Euclidean distances for each point

112

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al. k x

R i(k ) =

1 k

→ xi − → x αij

∑

2

j=1 k xy

R i(k ) =

1 k

→ xi − → x βij

∑

2

(A.2)

j=1

where αij and βij for j = 1, …, k denote the time indices of the k nearest neighbors of xi and yi, respectively. The symmetric NI was then calculated as

S=

Sx y + S y x (A.3)

2

Where

Sx y =

1 M

M

x

∑

xy

i=1

R i(k ) R i(k )

(A.4)

of the international federation: the international federation of clinical neurophysiology. Electroencephalogr. Clin. Neurophysiol. Suppl. 52, 3–6. Koepp, M.J., Woermann, F., Savic, I., Wandschneider, B., 2013. Juvenile myoclonic epilepsy–neuroimaging findings. Epilepsy Behav. 28 (Suppl 1), S40–S44. Koepp, M.J., Thomas, R.H., Wandschneider, B., Berkovic, S.F., Schmidt, D., 2014. Concepts and controversies of juvenile myoclonic epilepsy: still an enigmatic epilepsy. Expert Rev. Neurother. 14, 819–831. Kuhlmann, L., Freestone, D., Lai, A., Burkitt, A.N., Fuller, K., Grayden, D.B., Seiderer, L., Vogrin, S., Mareels, I.M., Cook, M.J., 2010. Patient-specific bivariate-synchronybased seizure prediction for short prediction horizons. Epilepsy Res. 91, 214–231. Lai, Y.C., Harrison, M.A., Frei, M.G., Osorio, I., 2003. Inability of Lyapunov exponents to predict epileptic seizures. Phys. Rev. Lett. 91, 068102. Le Van Quyen, M., Adam, C., Martinerie, J., Baulac, M., Clemenceau, S., Varela, F., 2000. Spatio-temporal characterizations of non-linear changes in intracranial activities prior to human temporal lobe seizures. Eur. J. Neurosci. 12, 2124–2134. Le Van Quyen, M., Martinerie, J., Navarro, V., Boon, P., D'Have, M., Adam, C., Renault, B., Varela, F., Baulac, M., 2001. Anticipation of epileptic seizures from standard EEG recordings. Lancet 357, 183–188. Le Van Quyen, M., Soss, J., Navarro, V., Robertson, R., Chavez, M., Baulac, M., Martinerie, J., 2005. Preictal state identification by synchronization changes in longterm intracranial EEG recordings. Clin. Neurophysiol. 116, 559–568. Lehnertz, K., Elger, C.E., 1998. Can epileptic seizures be predicted?: evidence from nonlinear time series analysis of brain electrical activity. Phys. Rev. Lett. 80, 5019–5022. Lehnertz, K., Andrzejak, R.G., Arnhold, J., Kreuz, T., Mormann, F., Rieke, C., Widman Elger, C.E., 2001. Nonlinear EEG analysis in epilepsy: its possible use for interictal focus localization, seizure anticipation, and prevention. J. Clin. Neurophysiol. 18, 209–222. Lin, L.C., Chen, S.C., Chiang, C.T., Wu, H.C., Yang, R.C., Ouyang, C.S., 2016. Classification preictal and interictal stages via integrating interchannel and timeDomain analysis of EEG features. Clin. EEG Neurosci. 48 (2), 139–145. Litt, B., Echauz, J., 2002. Prediction of epileptic seizures. Lancet Neurol. 1, 22–30. Litt, B., Esteller, R., Echauz, J., D'Alessandro, M., Shor, R., Henry, T., Pennell, P., Epstein, C., Bakay, R., Dichter, M., Vachtsevanos, G., 2001. Epileptic seizures may begin hours in advance of clinical onset: a report of five patients. Neuron 30, 51–64. Maiwald, T., Winterhalder, M., Aschenbrenner-Scheibe, R., Voss, H.U., Schulze-Bonhage, A., Timmer, J., 2004. Comparison of three nonlinear seizure prediction methods by means of the seizure prediction characteristic. Physica D 194, 357–368. Martinerie, J., Adam, C., Le Van Quyen, M., Baulac, M., Clemenceau, S., Renault, B., Varela, F.J., 1998. Epileptic seizures can be anticipated by non-linear analysis. Nat. Med. 4, 1173–1176. Meyer-Lindenberg, A., 1996. The evolution of complexity in human brain development: an EEG study. Electroencephalogr. Clin. Neurophysiol. 99, 405–411. Moon, Y.I., Rajagopalan, B., Lall, U., 1995. Estimation of mutual information using kernel density estimators. Phys. Rev. E Stat. Phys. Plasmas Fluids Relat. Interdiscip. Topics 52, 2318–2321. Mormann, F., Andrzejak, R.G., 2016. Seizure prediction: making mileage on the long and winding road. Brain 139, 1625–1627. Mormann, F., Lehnertz, K., David, P., Elger, E., 2000. Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy patients. Physica D 144, 358–369. Mormann, F., Andrzejak, R.G., Kreuz, T., Rieke, C., David, P., Elger, C.E., Lehnertz, K., 2003a. Automated detection of a preseizure state based on a decrease in synchronization in intracranial electroencephalogram recordings from epilepsy patients. Phys. Rev. E Stat. Nonlin. Soft Matter Phys. 67, 021912. Mormann, F., Kreuz, T., Andrzejak, R.G., David, P., Lehnertz, K., Elger, C.E., 2003b. Epileptic seizures are preceded by a decrease in synchronization. Epilepsy Res. 53, 173–185. Mormann, F., Kreuz, T., Rieke, C., Andrzejak, R.G., Kraskov, A., David, P., Elger, C.E., Lehnertz, K., 2005. On the predictability of epileptic seizures. Clin. Neurophysiol. 116, 569–587. Niederhauser, J.J., Esteller, R., Echauz, J., Vachtsevanos, G., Litt, B., 2003. Detection of seizure precursors from depth-EEG using a sign periodogram transform. IEEE Trans. Biomed. Eng. 50, 449–458. Noachtar, S., Peters, A.S., 2009. Semiology of epileptic seizures: a critical review.

References Aarabi, A., He, B., 2012. A rule-based seizure prediction method for focal neocortical epilepsy. Clin. Neurophysiol. 123, 1111–1122. Arnhold, J., Grassberger, P., Lehnertz, K., Elger, C.E., 1999. A robust method for detecting interdependencies: application to intracranially recorded EEG. Physica D 134, 419–430. Aschenbrenner-Scheibe, R., Maiwald, T., Winterhalder, M., Voss, H.U., Timmer, J., Schulze-Bonhage, A., 2003. How well can epileptic seizures be predicted? An evaluation of a nonlinear method. Brain 126, 2616–2626. Bandarabadi, M., Teixeira, C.A., Rasekhi, J., Dourado, A., 2015. Epileptic seizure prediction using relative spectral power features. Clin. Neurophysiol. 126, 237–248. Brinkmann, B.H., Wagenaar, J., Abbot, D., Adkins, P., Bosshard, S.C., Chen, M., Tieng, Q.M., He, J., Munoz-Almaraz, F.J., Botella-Rocamora, P., Pardo, J., Zamora-Martinez, F., Hills, M., Wu, W., Korshunova, I., Cukierski, W., Vite, C., Patterson, E.E., Litt, B., Worrell, G.A., 2016. Crowdsourcing reproducible seizure forecasting in human and canine epilepsy. Brain 139, 1713–1722. Bromfield, E., Cavazos, J., Sirven, J., 2006. Introduction to Epilepsy [Internet]. American Epilepsy Society, West Hartford (CT). Camfield, C.S., Striano, P., Camfield, P.R., 2013. Epidemiology of juvenile myoclonic epilepsy. Epilepsy Behav. 28 (Suppl 1), S15–17. Cao, L., 1997. Practical method for determining the minimum embedding dimension of a scalar time series. Physica D 110, 43–50. Carney, P.R., Myers, S., Geyer, J.D., 2011. Seizure prediction: methods. Epilepsy Behav. 22 (Suppl 1), S94–101. Crespel, A., Gelisse, P., Reed, R.C., Ferlazzo, E., Jerney, J., Schmitz, B., Genton, P., 2013. Management of juvenile myoclonic epilepsy. Epilepsy Behav. 28 (Suppl 1), S81–86. D'Alessandro, M., Esteller, R., Vachtsevanos, G., Hinson, A., Echauz, J., Litt, B., 2003. Epileptic seizure prediction using hybrid feature selection over multiple intracranial EEG electrode contacts: a report of four patients. IEEE Trans. Biomed. Eng. 50, 603–615. De Clercq, W., Lemmerling, P., Van Huffel, S., Van Paesschen, W., 2002. Evaluation of methods for determining a pre-ictal state in scalp-EEG measurements. Proceedings of the 2nd European Medical and Biological Engineering Conference 446–447. Elger, C.E., Lehnertz, K., 1998. Seizure prediction by non-linear time series analysis of brain electrical activity. Eur. J. Neurosci. 10, 786–789. Esteller, R., Echauz, J., D’Alessandro, M., Worrell, G., Cranstoun, S., Vachtsevanos, G., Litt, B., 2005. Continuous energy variation during the seizure cycle: towards an online accumulated energy. Clin. Neurophysiol. 116, 517–526. Gadhoumi, K., Lina, J.M., Gotman, J., 2012. Discriminating preictal and interictal states in patients with temporal lobe epilepsy using wavelet analysis of intracerebral EEG. Clin. Neurophysiol. 123, 1906–1916. Gadhoumi, K., Lina, J.M., Mormann, F., Gotman, J., 2016. Seizure prediction for therapeutic devices: a review. J. Neurosci. Methods 260, 270–282. Genton, P., Thomas, P., Kasteleijn-Nolst Trenite, D.G., Medina, M.T., Salas-Puig, J., 2013. Clinical aspects of juvenile myoclonic epilepsy. Epilepsy Behav. 28 (Suppl 1), S8–14. Harrison, M.A., Osorio, I., Frei, M.G., Asuri, S., Lai, Y.C., 2005. Correlation dimension and integral do not predict epileptic seizures. Chaos 15, 33106. Hively, L.M., Protopopescu, V.A., Gailey, P.C., 2000. Timely detection of dynamical change in scalp EEG signals. Chaos 10, 864–875. Iasemidis, L.D., Sackellares, J.C., Zaveri, H.P., Williams, W.J., 1990. Phase space topography and the Lyapunov exponent of electrocorticograms in partial seizures. Brain Topogr. 2, 187–201. Iasemidis, L.D., Shiau, D.S., Pardalos, P.M., Chaovalitwongse, W., Narayanan, K., Prasad, A., Tsakalis, K., Carney, P.R., Sackellares, J.C., 2005. Long-term prospective on-line real-time seizure prediction. Clin. Neurophysiol. 116, 532–544. Jansen Holleboom, L.I.M., 2016. Improving the Nurse Response to Seizures in the Epilepsy Monitoring Unit with Help of EEG-based Automatic Seizure Detection. Jerger, K.K., Netoff, T.I., Francis, J.T., Sauer, T., Pecora, L., Weinstein, S.L., Schiff, S.J., 2001. Early seizure detection. J. Clin. Neurophysiol. 18, 259–268. Kannathal, N., Chee, J., Er, K., Lim, K., Tat, O., 2013. Chaotic analysis of epileptic EEG signals. Proceedings of The 15th International Conference on Biomedical Engineering 652–654. Klem, G.H., Luders, H.O., Jasper, H.H., Elger, C., 1999. The ten-twenty electrode system

113

Epilepsy Research 135 (2017) 102–114

M.A.S. Kolarijani et al.

Stacey, W., Le Van Quyen, M., Mormann, F., Schulze-Bonhage, A., 2011. What is the present-day EEG evidence for a preictal state? Epilepsy Res. 97, 243–251. Takens, F., 1985. On the numerical determination of the dimension of an attractor. In: Braaksma, B.L.J., Broer, H.W., Takens, F. (Eds.), Dynamical Systems and Bifurcations: Proceedings of a Workshop Held in Groningen. The Netherlands, April 16–20, 1984. pp. 99–106 Springer, Berlin, Heidelberg. Tzallas, A.T., Tsipouras, M.G., Fotiadis, D.I., 2007. Automatic seizure detection based on time-Frequency analysis and artificial neural networks. Comput. Intell. Neurosci. 2007, 80510. van Dijkman, S.C., Alvarez-Jimenez, R., Danhof, M., Della Pasqua, O., 2016a. Pharmacotherapy in pediatric epilepsy: from trial and error to rational drug and dose selection − a long way to go. Expert Opin. Drug Metab. Toxicol. 1–14. van Dijkman, S.C., Alvarez-Jimenez, R., Danhof, M., Della Pasqua, O., 2016b. Pharmacotherapy in pediatric epilepsy: from trial and error to rational drug and dose selection −a long way to go. Expert Opin. Drug Metab. Toxicol. 12, 1143–1156. van Drongelen, W., Nayak, S., Frim, D.M., Kohrman, M.H., Towle, V.L., Lee, H.C., McGee, A.B., Chico, M.S., Hecox, K.E., 2003. Seizure anticipation in pediatric epilepsy: use of Kolmogorov entropy. Pediatr. Neurol. 29, 207–213. Wolf, P., Yacubian, E.M., Avanzini, G., Sander, T., Schmitz, B., Wandschneider, B., Koepp, M., 2015. Juvenile myoclonic epilepsy: a system disorder of the brain. Epilepsy Res. 114, 2–12. Zhang, Y., Zhou, W., Yuan, Q., Wu, Q., 2014. A low computation cost method for seizure prediction. Epilepsy Res. 108, 1357–1366.

Epilepsy Behav. 15, 2–9. Osorio, I., Frei, M.G., Wilkinson, S.B., 1998. Real-time automated detection and quantitative analysis of seizures and short-term prediction of clinical onset. Epilepsia 39, 615–627. Paternoster, L., Vallverdu, M., Melia, U., Claria, F., Voss, A., Caminal, P., 2013. Analysis of epileptic EEG signals in children by symbolic dynamics. Proceedings of the 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 4362–4365. Ramgopal, S., Thome-Souza, S., Jackson, M., Kadish, N.E., Sanchez Fernandez, I., Klehm, J., Bosl, W., Reinsberger, C., Schachter, S., Loddenkemper, T., 2014. Seizure detection, seizure prediction, and closed-loop warning systems in epilepsy. Epilepsy Behav. 37, 291–307. Rogowski, Z., Gath, I., Bental, E., 1981. On the prediction of epileptic seizures. Biol. Cybern. 42, 9–15. Rosenstein, M.T., Collins, J.J., De Luca, C.J., 1993. A practical method for calculating largest Lyapunov exponents from small data sets. Physica D 65, 117–134. Sato, Y., Doesburg, S.M., Wong, S.M., Ochi, A., Otsubo, H., 2015. Dynamic preictal relations in FCD type II: potential for early seizure detection in focal epilepsy. Epilepsy Res. 110, 26–31. Serafini, A., Rubboli, G., Gigli, G.L., Koutroumanidis, M., Gelisse, P., 2013. Neurophysiology of juvenile myoclonic epilepsy. Epilepsy Behav. 28 (Suppl 1), S30–39. Shorvon, S.D., Andermann, F., Guerrini, R., 2011. The Causes of Epilepsy: Common and Uncommon Causes in Adults and Children. University Press, Cambridge.

114