2nd Reading May 14, 2015 15:53 1550015

International Journal of Neural Systems, Vol. 25, No. 5 (2015) 1550015 (18 pages) c World Scientific Publishing Company  DOI: 10.1142/S012906571550015X

Automated Seizure Onset Zone Approximation Based on Nonharmonic High-Frequency Oscillations in Human Interictal Intracranial EEGs

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Evelien E. Geertsema MIRA-Institute for Biomedical Technology and Technical Medicine University of Twente, Enschede, The Netherlands Stichting Epilepsie Instellingen Nederland (SEIN) Heemstede, The Netherlands Gerhard H. Visser† Stichting Epilepsie Instellingen Nederland (SEIN) Heemstede, The Netherlands Demetrios N. Velis Stichting Epilepsie Instellingen Nederland (SEIN) Heemstede, The Netherlands Department of Neurosurgery, Academic Center for Neurosurgery VUmc, Free University Medical Center Amsterdam, The Netherlands Steven P. Claus Stichting Epilepsie Instellingen Nederland (SEIN) Heemstede, The Netherlands Department of Clinical Neurophysiology VUmc, Free University Medical Center Amsterdam, The Netherlands Maeike Zijlmans† Stichting Epilepsie Instellingen Nederland (SEIN) Heemstede, The Netherlands Department of Neurology and Neurosurgery, Brain Center Rudolf Magnus University Medical Center Utrecht, The Netherlands Stiliyan N. Kalitzin∗ Stichting Epilepsie Instellingen Nederland (SEIN), Achterweg 5, 2130SW, Heemstede, The Netherlands Image Sciences Institute, University Medical Center Utrecht, Utrecht, The Netherlands [email protected] Accepted 5 March 2015 Published Online 18 May 2015 A novel automated algorithm is proposed to approximate the seizure onset zone (SOZ), while providing reproducible output. The SOZ, a surrogate marker for the epileptogenic zone (EZ), was approximated from intracranial electroencephalograms (iEEG) of nine people with temporal lobe epilepsy (TLE), ∗ †

Current address: Postbus 540, 2130SW, Hoofddorp, The Netherlands. Both authors contributed equally. 1550015-1

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

using three methods: (1) Total ripple length (T RL): Manually segmented high-frequency oscillations, (2) Rippleness (R): Area under the curve (AUC) of the autocorrelation functions envelope, and (3) Autoregressive model residual variation (ARR, novel algorithm): Time-variation of residuals from autoregressive models of iEEG windows. TRL, R, and ARR results were compared in terms of separability, using Kolmogorov–Smirnov tests, and performance, using receiver operating characteristic (ROC) curves, to the gold standard for SOZ delineation: visual observation of ictal video-iEEGs. TRL, R, and ARR can distinguish signals from iEEG channels located within the SOZ from those outside it (p < 0.01). The ROC AUC was 0.82 for ARR, while it was 0.79 for TRL, and 0.64 for R. ARR outperforms TRL and R, and may be applied to identify channels in the SOZ automatically in interictal iEEGs of people with TLE. ARR, interpreted as evidence for nonharmonicity of high-frequency EEG components, could provide a new way to delineate the EZ, thus contributing to presurgical workup.

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Keywords: Seizure onset zone; high-frequency oscillations; epilepsy surgery.

1.

Introduction

Epilepsy is a chronic brain disorder that is characterized by recurrent epileptic seizures.1 When a patient has an epileptic syndrome that is possibly remediable by surgery, and seizure control could not be attained with multiple trials of antiepileptic drugs, the patient qualifies to enter presurgical evaluation. The objective of epilepsy surgery is to resect completely or to disconnect the epileptogenic zone (EZ), which is defined as the minimum amount of cortical tissue that must be resected to produce seizure-freedom.2 The EZ is a theoretical concept. In practice, surrogate markers for the EZ are used, such as the seizure onset zone (SOZ), defined as the area of the cortex from which clinical seizures are generated.3 The gold standard method of determining the SOZ is visual observation of ictal recordings from intracranial electroencephalograms (iEEG) and synchronized video.3 During such observations habitual seizures need to be recorded. Appropriate positioning of electrodes with respect to the seizure onset zone is assumed when (a) the unequivocal electroencephalographic onset (UEO) of each seizure is identified as a transition from the interictal to the ictal recording of the individual undergoing iEEG recordings and (b) the UEO precedes unequivocal clinical onset (UCO) of that seizure.4,5 UEOs and UCOs are both patientspecific. Reliable identification of what constitutes ictal and what constitutes interictal EEG in depth recordings is a matter of familiarity with that particular intracranial record, and thus the record should provide a representative sample of multiple seizures to prove reproducibility of the findings. Therefore, several days of video-EEG recording may be required (at least one week). As a consequence, review of the

attained long-term video-EEG is both laborious and time consuming. Interictal high-frequency oscillations (HFOs) have been associated with epileptic properties of neuronal tissue in the brain.6–10 HFOs have been shown to be reliable markers of the SOZ,8,11–16 and to provide better SOZ localization than interictal spikes.11,13,14 The use of interictal HFOs to delineate the SOZ reliably may result in a reduced need for seizure recordings, thus shortening EEG registration time.17,18 Nevertheless, such a technique may only be used clinically after extensive validation and when there is proof of appropriate positioning of the intracranial electrodes. Visual detection of HFOs is, like visual review of the EEG record, very laborious and time consuming. This is due to the small time frame needed to observe the oscillations because of their short duration and small amplitude.19 Moreover, subjectivity in visual detection is hard to avoid: studies in which expert human reviewers have marked HFO-like activity on the iEEG have shown poor inter-rater reliability and reproducibility.20 Automatic detection of HFOs could possibly solve these problems, and several possibilities to implement such a detector have been proposed in the literature.20–29 HFOs, and especially ripples (80–250 Hz oscillations), are not necessarily a sign of epileptogenicity alone; they can also have a physiological origin. Physiological ripples (∼200 Hz) have been measured in the CA1 region of the hippocampus, the entorrhinal cortex,30 the mesiotemporal lobe, and the extratemporal neocortex.31 Therefore, when using HFOs to identify the SOZ, it is important to realize that HFOs may also be found in areas outside the SOZ; e.g. in the contralateral hippocampus. Furthermore, fast

1550015-2

2nd Reading May 14, 2015 15:53 1550015

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Automated SOZ Approximation

oscillations can result from filtering sharp transients in the EEG.32,33 B´enar et al. showed that high-pass filtering of epileptic transients or nonsinusoidal oscillations may lead to spurious oscillations, or false ripples. Consequently, separation of epileptic HFOs and other HFO-like signals remains a challenge. In this study a novel methodology is proposed. The main assumption underlying this methodology is that epileptic HFOs are of the nonharmonic type. In other words, they cannot be exactly described as solutions of a linear differential equation of a certain order. There can be a variety of reasons for deviation from harmonic behavior, such as noise input, nonstationary parameters, and nonlinear dynamics. Various quantifiers that characterize for example nonlinearity related to a single recording site have proved effective in aiding the lateralization and localization of the EZ.34–44 Most of these studies focus only on lateralization of the SOZ, while the capacity for SOZ localization of the corresponding algorithms is only reported in a sub-selection of those studies.37,42,43 This study aims to design an automated algorithm that can independently, reliably, and reproducibly approximate the SOZ, as a surrogate marker for the EZ. By quantifying certain nonharmonic features of high-frequency EEG components we address the issue of less arbitrary selection of regions that may be closest to the SOZ. As a secondary goal, it is investigated if ‘rippleness’, obtained with an algorithm previously proposed by our group,43 could possibly be used for fast automated estimation of the amount of HFOs. We investigated whether the results from this previous study, showing correspondence between rippleness and the total length of manually scored HFOs in the ripple range (80– 250 Hz), can be reproduced in our data set. Subsequently, the newly designed autoregressive model residual variation (ARR) algorithm, which quantifies nonharmonic aspects of the signal, is compared with manually scored HFOs and rippleness in terms of their ability to approximate the SOZ. 2. Methods 2.1. Subjects Nine individuals with intractable temporal lobe epilepsy were included. The selected individuals were candidates for epilepsy surgery between 2008 and 2012, and the registered iEEGs were part of the presurgical work up. The iEEG registrations were

performed at Stichting Epilepsie Instellingen Nederland (SEIN), in Heemstede, The Netherlands. No further exclusion criteria were applied beyond the sampling frequency of the stored data (>1000 Hz). Data were analyzed retrospectively; individuals had epilepsy surgery at least one year prior to conducting this study. Relevant clinical and ancillary information on the subjects in this study is shown in Table 1. 2.2. Data 2.2.1. Intracranial EEGs All patients underwent iEEG registration for at least one week. Depth electrodes were implanted at the Academic Center for Neurosurgery of the Free University Medical Center (VUmc) in Amsterdam, The Netherlands. Ad-Tech (Racine, WI, USA) electrodes with a 2.4 mm2 contact size were used. Electrode placement was guided by instructions from clinical neurophysiologists S. Claus and D. Velis, according to clinical hypotheses regarding the SOZ and ictal spreading patterns. These hypotheses were formed on the basis of scalp video EEG and MRI, supplemented by MEG, PET, and SPECT results.46 iEEG data was recorded using a 65-channel Schwarzer amplifier (Schwarzer GmBH, Germany) implying 0.016 Hz hardware high-pass filtering. The amplifier was coupled to a workstation with Harmonie 6.2 acquisition software (Stellate Systems, Montreal, PQ Canada). During registration of the iEEGs, a silent electrode situated in the white matter was chosen as a reference. An external ground electrode was placed on the forehead of the subject. A sampling frequency of 1000 Hz was used. Antiepileptic drugs were tapered according to the guidelines in Ref. 47 to increase the chance of a seizure occurring during registration; this was part of the standard operating procedure. Based on visual inspection, EEG channels with severe artefact contamination were excluded from further processing. All recordings were performed using referential montages. Bipolar montages were secondarily constructed by subtracting the signals from the neighboring contacts on each depth electrode. 2.2.2. Clinical data The SOZ was determined by visual observation of ictal recordings from the video-iEEG by trained

1550015-3

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Table 1. Subject information. Age is defined at time of registration; registration day is defined as the day, during registration, after medication lowering. Surgical outcome is defined in terms of the UCLA classification, and follow up period is given in years. For electrode locations and corresponding abbreviations, see Table A.1 in Appendix A. UCLA 1A: Free of disabling seizures and auras, UCLA 1B: Free of disabling seizures, UCLA 2: Rare disabling seizures, UCLA 3: Worthwile improvement, UCLA 4: No worthwile improvement.45 Subject

Age

Sex

Registration day

Seizure onset zone

Resected area (year)

Surgical outcome (follow up period)

1

22

F

3

ATHCL1, MTHCL1, IAIL2-4, AML1-2, IAPL2-4, TPL1-3

L temporal + insula (2011)

UCLA IA (1.5)

2

27

M

8

LesAntR1-2, LesSupR2-3, LesAnIR1-3, InsAntR1-5

R temporo-insular frontal (2012)

UCLA IA (1)

3

63

M

2

(1) HAR1-3, HPR1-3, AR1-4 (2) HAR1-3 IAIR1

R temporal + amygdalohippocampectomy (2011)

UCLA III (2)

4

28

M

7

(1) HCR1-5 (2) HCL1-5 (1 > 2)

R temporal + amygdalohippocampectomy (2008)

UCLA II (4.5)

5

13

F

3

FODL4-7a

Superior to earlier resection area (L frontal, 2009) + L SMA (2010)

UCLA IV (3.5)

6

35

F

6

HCL1-2, AML1-2

L fronto-temporal + amygdalohippocampectomy + frontoopercular + frontobasal (2008)

UCLA IB (5)

7

26

F

4

(1) ATR1-2, AHR1-2, MHR1-2, IAR2-4, PHR1-2 (2) MHL1-2 (1 > 2)

R hippocampectomy (2011)

UCLA II (2)

8

26

F

3

HCR1-3

R temporal + leasionectomy + amygdalohippocampectomy (2010)

UCLA IB (3)

9

39

M

10

TBL1-3, HC1L1-3, HC2L1-3, AML1-2

L temporal (2011)

UCLA IA (2)

a

There is possibly a second, less frequent seizure type (involving PSMAL, PSMAR, TOL, and PFL bundles in its SOZ), but due to many artefacts and unclear seizure semiology, this hypothesis could not be substantiated.

clinical neurophysiologists (D. Velis and S. Claus), based on ictal changes in time-related context of the semiology. These findings were used as the gold standard. The outcome of surgery was assessed according to the UCLA classification proposed by Engel et al., simplified to discriminate classes IA, IB, II, III, and IV.45 For part of the analysis in this study, subjects are subdivided into two outcome groups: subjects with good postoperative seizure control (UCLA class IA, IB, II; i.e. subjects 1, 2, 4, 6–9) and those with poor postoperative seizure control (UCLA class III, IV; i.e. subjects 3, 5). The SOZ and surgical outcome

are shown in Table 1. A total of 77 channels of EEG data were considered to be inside the SOZ, with 364 channels outside the SOZ. 2.2.3. EEG data selection Five consecutive minutes of slow-wave-sleep (SWS) were analyzed from the iEEG data of each subject. HFOs and spikes occur more frequently during SWS,11,48 and the signals of the most superficial iEEG channels (in closest proximity to the skull) are less influenced by muscle activity in this sleep

1550015-4

2nd Reading May 14, 2015 15:53 1550015

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Automated SOZ Approximation

stage. During the iEEG registrations no surface EEG, electro-oculograms, or electro-myograms were recorded, so we relied on information on sleep staging that could be gathered from the iEEG. SWS epochs were found with a method similar to the one used by D¨ umpelmann et al.29 The relative delta power was tracked automatically in the iEEG, with relative delta power being the energy in the delta band (0.5–4 Hz) divided by the energy in a large frequency band (0.5–100 Hz). The module to calculate spectral features in Stellate Harmony (Stellate Systems, Montreal, PQ Canada) was used to obtain the relative delta power trend. The relative delta power trend was obtained throughout a nightly recording in 30-s windows in three superficial iEEG channels from different bundles. A 5-min epoch, in which the relative delta power exceeded 70%, was selected for each subject and used in the analysis for this study.

iEEGs: all channels, bipolar montage

5 minutes of SWS

80 Hz highpass filtering

80 Hz Huangfiltering

Windowing

Visual scoring of HFOs

ACF

Residuals from AR-modeling

Total ripple length

AUC (Rippleness)

ARR

SOZ

2.3. Data analysis Data were analyzed retrospectively using three methods independently: Visual detection of HFOs, the rippleness algorithm, and a novel algorithm called the ARR algorithm. The rippleness algorithm uses the autocorrelation function (ACF) and its area under the curve (AUC) to quantify the extent of oscillation in a signal.43 The ARR algorithm uses the presence of high residual signal variation after autoregressive (AR) model fit as a biomarker to identify contacts in or close to the SOZ. High ARR values are interpreted as evidence of the presence of nonharmonicity in iEEG signals. The rippleness and ARR algorithms are further explained below. Results from all three methods were compared with the gold standard (visual SOZ determination). An overview of the analysis steps is shown in Fig. 1. Data analysis was performed using Matlab (Mathworks, Natick, MA, USA) version 7.5.0. 2.3.1. Total ripple length (TRL) Each 5-min iEEG epoch was assessed visually for HFOs using the method currently standard for manual HFO segmentation. The method used is similar to the one described by Jacobs et al.13 After HFO segmentation the TRL (sum of the lengths of all ripples) per channel could be calculated. An 80 Hz high-pass filter (Finite Impulse Response filter, order 63) was applied to allow the HFOs to be discerned

Fig. 1. Informational flow diagram of how the data in this study was used. Dashed arrows indicate comparison of data. (iEEG: Intracranial electroencephalogram, SWS: slow-wave sleep, HFOs: high-frequency oscillations, ACF: autocorrelation function, AUC: area under the curve, SOZ: seizure onset zone, AR: autoregressive, ARR: autoregressive model residual variation.)

from the relatively high-voltage background activity. The HFOs were visualized in eight channels simultaneously, while displaying the traces at maximum time-resolution. Events were regarded as ripples if their amplitude was clearly higher than the remaining high-frequency baseline activity of the reviewed channel, and the oscillation consisted of at least four periods. HFOs were distinguished from presumed artefacts based on their frequency content and sharpness and co-occurrence over several channels.18,49 For the entire 5-min epoch, HFOs were marked by one of the authors (E. Geertsema) and revised by an experienced observer (M. Zijlmans). By filtering above 80 Hz, manually segmented HFOs lay in the ripple (80–250 Hz) range. With the sampling frequency (1000 Hz) used, it was unfeasible to take fast ripples (250–500 Hz) into account. Digital sampling at five times the oscillation frequency of interest is advisable to sample adequately the temporal dynamics of an HFO.50

1550015-5

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

2.3.2. Rippleness (R) Channel rippleness was obtained by applying an algorithm that uses the ACF and its AUC. This algorithm is described in detail by Kalitzin et al.43 A brief description is provided here. First, the selected 5-min epoch of iEEG data was 80 Hz high-pass filtered, using a morphological decomposition of the signal, the Huang–Hilbert transform. This filtering method is applied because unlike commonly used FIR filters, it does not cause spurious oscillations which can be mistaken for ripples.32,43 Subsequently, the ACF is calculated using ACF(τ ) = S(τ )S(t + τ )τ ,

(1)

where S(t) is the filtered signal through time, S(t+τ ) the shifted signal with lag τ (t and τ in samples). The lowest frequency of interest was selected as 80 Hz and a maximal number of 10 cycles to constitute an HFO was postulated. Therefore, a maximum lag of 10/80 of a second, or 125 samples (with a 1000 Hz sampling frequency), was chosen. The envelope of the ACF is obtained by the Hilbert transform, after which the area under the envelope, or curve (AUC), can be calculated. This quantity represents the total power of oscillations above 80 Hz, or rippleness (R).

2.3.3. Autoregressive model residual variation (ARR) We developed and evaluated a novel algorithm designed to select automatically the channels from which signals had been obtained closest to the SOZ. AR modeling was selected because the frequencies and damping coefficients resulting from AR modeling are able to quantify oscillatory properties of an EEG.51,52 A preliminary study in part of the data (subjects 1, 2, 3, 4, 5, and 7, results not shown here) showed that windows of iEEG containing ripples measured in areas close to the SOZ exhibited higher residual signal variation after AR model fit than windows with ripples from areas outside the SOZ (for example, from the contralateral hippocampus). Moreover, when the residual signal variation from all iEEG windows in a 5-min epoch was obtained, channels in the SOZ showed frequent windows with an excessively high residual. By

quantifying these frequently occurring high residuals using the coefficient of variation (CV) (i.e. a normalized standard deviation), a classifier was found for the approximation of the SOZ, named the ARR. Because the AR model describes harmonic oscillators of given order that are stationary and linear, the ARR measures the nonharmonic component in the system dynamics. Typically the residual variance of the signal after AR fit is interpreted as noise, but in our application it can be reminiscent of the nonlinear or nonstationary dynamic patterns of the underlying system generating the signal.53,54 It was not the purpose of this study to reconstruct which is the primary source of the ARR; in fact both the nonlinear and the nonstationary effects may contribute to the classification power of the method. Our approach is therefore an attempt to detect the “boundaries of harmonicity” of the neuronal system that may indirectly quantify closeness to epileptic transition. The processing steps executed by the ARR algorithm are shown in Fig. 2. For each channel, the 5-min SWS iEEG epochs were divided into windows of 40 samples, with 50% overlap. Three-pole AR models were estimated, thereby obtaining the residual signal variance (r) for each window. Preliminary investigation in a part of the data (subjects 1, 2, 3, 4, 5, and 7) showed that a 40 sample window length and model order 3 provided the best distinction between channels inside and channels outside the SOZ. Highfrequency components of the EEG are presumably quantified with these parameters. Because of the 40sample window length used for each AR model, the lowest oscillatory frequency that can be modeled in this window, using a sampling frequency of 1000 Hz, is (1/40) ∗ 1000 = 25 Hz. Nonharmonicities in low frequencies therefore do not influence the obtained residual signal variation. Since we hypothesized that the important information for SOZ approximation is in the intermittent occurrence of excessive r values (high-valued outliers) in signal windows, the variation of r in time was quantified by calculating the CV, which represents the extent of variability in relation to the mean: σ(r) . (2) ARR = CV(r) = µ(r) Here σ is the standard deviation and µ is the mean of the residuals after AR modeling (r-values) for

1550015-6

2nd Reading May 14, 2015 15:53 1550015

Automated SOZ Approximation

(a)

Amplitude ( µV)

w1

w2

w3

w4

w5

80

100

w6

w7

0

−50

−100

0

20

40

60

120

140

160

100 iEEG

(b)

Amplitude (µV)

Modelled iEEG

50

0

−50

−100 40

50

60

70

80

Time (samples)

100 Residuals

(c)

Amplitude (µV)

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Time (samples)

50

r3 = Var (Residuals)

0

−50

−100 40

50

60

70

80

Time (samples)

(d)

ARR =

σ (r) µ (r)

with r = [r1, r2, r3, r4, r5, r6, r7] Fig. 2. The processing steps executed by the ARR algorithm. (a) The iEEG epoch, which has a length of 160 samples in this example, is divided into windows of 40 samples, with 50% overlap. (b) Of each window, an order 3 AR model is estimated. (c) The residual signal variation (r) in the window is obtained. (d) The CV is determined for the iEEG epoch, defined as the standard deviation of the r values from the epoch, divided by the mean of those r values. This constitutes the ARR value for the iEEG epoch.

1550015-7

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

the series of time windows. Using CV instead of the standard deviation provides normalization of values, thus allowing comparison of results between subjects. Thus CV provides the variation of r-values after AR modeling, thereby obtaining the ARR for each channel.

2.4. Statistical analysis

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

2.4.1. Comparison between R and TRL results The TRL method is based on manual identification and segmentation of ripples while the R technique intends to provide an automated estimation of the amount of ripples. Therefore these two quantities are directly comparable and the association between them is a measure of the quality of the automated ripple quantification given the manual segmentation of ripples as gold standard. In a previous study we found a correspondence between TRL and R values per channel.43 In the current study, we compared TRL and R in a larger dataset in order to test reproducibility. The R results from the EEG of all channels were compared with TRL results, using the h2 index; a measure for the association between pairs of signals in general (linear or nonlinear).55–57 The h2 method belongs to the class of unidirectional association measures and as such it gives extra information about the functional relations between two sets of data.57 Namely, if the values of one set are functionally determined by the other one, this does not imply automatically that the second data set is functionally determined by the first, as it would be in case of nonmonotonic, noninvertible functions. We obtained the h2 index in two directions, to test the fraction of variation of R that can be explained by TRL, and vice versa; to test the fraction of variation of TRL that can be explained by R. This results in a value between 0 and 1, where h2 (TRL, R) = 1 implies that all variation of TRL is explained by R. The h2 findings are significant if h2 exceeds the critical value, or significance margin, in a bootstrapped test with α = 0.05.57 A bi-directional association reflects the invertibility of the function representing the mapping between R and TRL or TRL and R, while unidirectional associations indicate noninvertible mapping.

2.4.2. Comparison of TRL, R, and ARR results with the SOZ TRL, R, and ARR results were compared with the SOZ, for each individual subject, and for the whole group, using two methods for analysis. Firstly, the ability of each of the three methods to distinguish between signals from channels in the SOZ as opposed to those outside it (termed separability for the purpose of this investigation), was assessed using a Kolmogorov–Smirnov test (K–S test). The K–S test was used in two ways. The cumulative distribution functions (CDF) of two datasets (values for signals from channels in versus outside the SOZ) were compared to determine whether they had distinctly different distributions (with α = 0.01). Moreover, the K–S test was used to test the hypothesis that the values obtained for channels in the seizure onset zone are higher than the values for those outside it (with α = 0.01). This is done by testing whether CDF(SOZ) < CDF(non-SOZ). The CDF of values for channels in the SOZ saturates to 1 at higher values. In other words, smaller CDF indicates a distribution with higher values. Secondly, the performance of TRL, R, and ARR at approximating the SOZ is determined. This was done by obtaining the sensitivity and specificity for various thresholds. Sensitivity is defined as the proportion of actual positives which are correctly classified as such. Specificity is defined as the proportion of negatives which are correctly classified as such. A channel is classified as positive (i.e. being potentially in the SOZ) by one of the three methods, when its TRL, R, or ARR value exceeds a varying threshold. Otherwise, the channel is classified as negative according to that method. With the sensitivity and specificity at various thresholds, a receiver operating characteristic (ROC) curve can be drawn, where the AUC represents algorithm performance. An AUC of 0.5 suggests a random classifier, whereas an AUC of 1 suggests a perfect classifier. Performances are determined from the TRL, R, and ARR results per subject and at group level. In the latter case, all channels from all subjects are taken together. The same method to obtain algorithm performance was applied to only the cases with good postoperative seizure control. Assuming that in those cases the SOZ as delineated by the gold standard

1550015-8

2nd Reading May 14, 2015 15:53 1550015

Automated SOZ Approximation

contains the whole epileptogenic area, this analysis shows possible causal relationships between HFO generation and nonharmonic dynamics, and the processes of transitions to seizures.

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

2.4.3. ARR algorithm performance and surgical outcome We also investigated whether surgical outcome can possibly explain the performance results of the ARR algorithm. For this purpose, ARR performance results are viewed in respect to surgical outcome of all subjects. Poorer performance of the ARR algorithm to approximate the SOZ can be caused by either false positives (FPs) or false negatives (FNs). ARR FPs occurring more in data from subjects with poor postoperative seizure control, could suggest that some of the electrodes corresponding to those FPs were actually situated in an epileptogenic area. This area could have been missed in the (gold standard) delineation of the SOZ, or the area could be outside the SOZ, while situated in the EZ. ARR FNs occurring in data from subjects with poor postoperative seizure control suggest that electrodes corresponding to those FNs were actually situated outside the epileptogenic area. It is therefore also informative to investigate whether diminished performance of the ARR algorithm is caused by FPs rather than FNs, or vice versa. This was done by comparing distributions of ARR values from subjects with good postoperative seizure control with ARR values from subjects with poor postoperative seizure control. This comparison is made separately for channels recorded from inside and outside the SOZ.

2.4.4. ARR performance and antiepileptic drugs Antiepileptic drug (AED) tapering is standard operating procedure to increase the chance of seizures during iEEG registration. The day of registration after drug tapering differs between subjects (see Table 1), because the choice of stored interictal data was limited. We investigated whether the registration day of the used iEEG recording, reflecting the level of AED tapering, could explain performance results of the ARR algorithm.

3.

Results

3.1. TRL and R results TRL and R results are shown in scatterplots per subject in Fig. 3. Results from h2 signal association analysis are shown in the boxes in each boxplot. The rising trends that can be observed in the scatterplots in this figure suggest that R and TRL values are correlated. However, this correspondence is not strictly monotonic, and some scatter is observed. The h2 index findings show significant signal association in both directions (R to TRL and TRL to R) in the iEEG data from subjects 1, 2, 4, 5, 7, and 9. In data from subjects 3 and 6, signal association was only significant in one direction (TRL to R and R to TRL, respectively). Signal association is thus not favored in one particular direction. In subject 8, signal association was not significant in either direction. Although scatter is shown, especially at TRL = 0, a rising trend can be observed in the data points where TRL > 0. When combining all R and TRL data points from the total dataset, h2 (TRL, R) ≈ 0.989, and h2 (R, TRL) ≈ 0.805 were found. Both h2 findings were significant with p < 0.01. These results suggest correspondence between TRL and R values per channel. 3.2. TRL, R, and ARR results in relation to the SOZ The distributions of TRL, R, and ARR values inside and outside the SOZ are shown in Figs. 4(a)–4(c), respectively. Only iEEG channels of individuals with a good postoperative seizure control were used to obtain the distributions. By applying a K–S test it was found that TRL, R, and ARR values from inside and outside the SOZ have independent and distinctly different distributions (p < 0.01). Furthermore, the values for signals from channels in the SOZ were higher than values for the signals obtained from channels outside the SOZ for all three methods (p < 0.01). The performance of TRL, R, and ARR at approximating the SOZ is provided in an ROC AUC per subject in Table 2. For all subjects except subject 3, ARR provides the highest ROC AUC value. For subject 3, TRL has the highest ROC AUC value. Performance of TRL, R, and ARR on group level

1550015-9

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

Subject 1 2 0 Log (R)

Subject 2 2

h2(TRL,R) ≈ 0.941 * h2(R,TRL) ≈ 0.854 *

0 −2

−2

−4

−4

−4

0

5

−6

10

0

5

Subject 4 2 Log (R)

−6

2

2

h2(TRL,R) ≈ 0.804 * 2 h (R,TRL) ≈ 0.833 *

0

−2

−4

−4

−4

5

−6

10

0

5

Subject 7 2

2

10

−6

2

−2

−2

−4

−4

−4

0

5

−6

5 10 Log (TRL+1)

0

10

h2(TRL,R) ≈ 0.898 *

2 0 h (R,TRL) ≈ 0.877 *

−2

−6

0

Subject 9

h2(TRL,R) ≈ 0.601 2 h (R,TRL) ≈ 0.153

0

10

h2(TRL,R) ≈ 0.605

Subject 8

h2(TRL,R) ≈ 0.927 * 2 h (R,TRL) ≈ 0.969 *

0

5

2 0 h (R,TRL) ≈ 0.532 *

−2

0

0

Subject 6

−2

−6

Log (R)

10

Subject 5

h2(TRL,R) ≈ 0.974 * 2 h (R,TRL) ≈ 0.993 *

0

h2(TRL,R) ≈ 0.857 * h2(R,TRL) ≈ 0.286

0

−2

−6

5 10 Log (TRL+1)

−6

0

5 10 Log (TRL+1)

Fig. 3. Scatterplots showing the results from the visual assessment of ripples, the TRL, versus the results from the rippleness algorithm (R). Each data point represents the TRL and R result for an individual channel. Both axes are scaled logarithmically. For each subject, the h2 indexes for signal association are shown in the top-left corner of the plot, followed by an asterisk if the found association was significant (p < 0.05). Note that log(TRL + 1) is on the x-axis. This was done so that for TRL = 0 when no ripples were found visually, log(TRL + 1) = 0, thus providing a datapoint at zero on the x-axis.

12000

0.25

2.5

0.2

2

10000

6000

ARR values

8000

R values

TRL values

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Subject 3 2

h2(TRL,R) ≈ 0.726 * h2(R,TRL) ≈ 0.873 *

0.15

0.1

1.5

1

4000 0.05

2000

0.5 0 outside SOZ

(a) TRL

in SOZ

0

outside SOZ

(b) R

in SOZ

outside SOZ

in SOZ

(c) ARR

Fig. 4. Distributions of (a) TRL, (b) R, and (c) ARR values per channel inside versus outside the seizure onset zone. Only iEEG channels of the subjects with a good postoperative seizure control (UCLA IA, IB and II, i.e. subjects 1, 2, 4, 6–9) were used to obtain the distributions. Outliers are not shown.

1550015-10

2nd Reading May 14, 2015 15:53 1550015

Automated SOZ Approximation

Subject

Surgical outcome (good/poor)

TRL

R

ARR

1 2 3 4 5 6 7 8 9

UCLA IA (good) UCLA IA (good) UCLA III (poor) UCLA II (good) UCLA IV (poor) UCLA IB (good) UCLA II (good) UCLA IB (good) UCLA IA (good)

0.88 0.78 0.87 0.92 0.71 0.79 0.75 0.83 0.82

0.60 0.66 0.48 0.81 0.53 0.46 0.77 0.67 0.63

0.92 0.81 0.70 0.99 0.84 0.94 0.78 0.92 0.87

All

0.79

0.64

0.82

Only UCLA IA, IB, II

0.79

0.65

0.83

In Fig. 7, distributions of ARR values for channels outside the SOZ are shown for good postoperative seizure control and poor postoperative seizure control. Thus the shown distributions hold only the true negative and FP findings. Distributions of ARR values for channels inside the SOZ would hold only the true positive and FN findings (results not shown here). Applying K–S tests showed (1) separate distributions for good versus poor postoperative seizure control in channels from outside the SOZ (p < 0.05), (2) nonsignificant separation of distributions for good versus poor postoperative seizure control in channels from inside the SOZ, (3) significantly higher ARR values outside the SOZ from subjects with poor postoperative seizure control than from subjects with good postoperative seizure control (p < 0.01), and (4) ARR values measured inside the SOZ were not significantly higher from subjects with good postoperative seizure control compared to ARR values from subjects with poor postoperative seizure control. Thus, there are significantly more FPs, rather than FNs, in the poor postoperative seizure control group than in the good postoperative seizure control group.

with only data from subjects with good postoperative seizure control is shown in the ROC curves in Figs. 5(a)–5(c), respectively. The ROC AUC is 0.79 for TRL, 0.64 for R, and 0.82 for ARR. 3.3. ARR performance and surgical outcome

3.4. ARR performance and antiepileptic drugs

In Fig. 6(a), ARR performances are plotted against surgical outcome per subject. Surgical outcome showing UCLA class II or higher implies that a subject is not free of disabling seizures. A declining trend can be observed for ARR performance with lower surgical outcome. Linear regression performed on this data was however not significant.

In Fig. 6(b), the performance of the ARR algorithm is plotted against the registration day after medication lowering for each subject. A rising trend can be observed in the ARR results with increasing registration day. Linear regression performed on this data was however not significant.

1

1

0.8

0.8

0.8

0.6 0.4 AUC = 0.79

0.2 0

0

0.5

(a)

sensitivity

1

sensitivity

sensitivity

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Table 2. Performance expressed in ROC AUC values per subject for the three classifiers: TRL, R, and ARR. Group performance for all subjects and for the good postoperative seizure control group (UCLA IA, IB, II) are shown separately.

0.6 0.4 AUC = 0.65

0.2

1

0

0

0.5

(b)

0.6 0.4 AUC = 0.83

0.2

1

0

0

0.5

1

(c)

Fig. 5. ROC curves for the performance of (a) TRL, (b) R, and (c) ARR for predicting the seizure onset zone. Algorithm performance is provided by the AUC, indicated in the textboxes in each figure. Only results from the subjects with good postoperative seizure control (UCLA IA, IB and II; i.e. subjects 1, 2, 4, 6–9) were used to obtain the ROC curves. 1550015-11

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

1 Algorithm performance

Algorithm performance

1 0.9 0.8 0.7

0.9 0.8 0.7

2

IA IB II III IV Surgical outcome (UCLA class)

10

(b)

Fig. 6. Scatter plots showing the influence of (a) surgical outcome and (b) iEEG registration day after medication lowering on algorithm performance. In both figures, performance (in ROC AUC) of the ARR algorithm is shown on the y-axis. Each data point represents the results found for one of the subjects in this study. Note that in panel (b), subjects 1 and 8 share the same data point [3, 0.92]. 0.8

Fig. 7. Distributions of ARR values for channels outside the SOZ, according to the gold standard for SOZ delineation. These distributions therefore hold only true negative and FP observations. Separate distributions are shown for good postoperative seizure control (UCLA class IA, IB, and II) on the left and poor postoperative seizure control (UCLA class III and IV) on the right. Outliers are not shown.

Three methods for reliable approximation of the SOZ — TRL, R, and ARR — were compared to the current gold standard for determining the SOZ. All three methods possess the ability to distinguish between signals obtained from channels in the SOZ and those from outside it. Moreover, all three methods show higher values for signals obtained from channels inside than for channels outside the SOZ. The ARR algorithm outperforms TRL and R in the extent to which the algorithm achieves correct classification of channels in the SOZ. Moreover, the ARR and R algorithms process a 5-min 1000 Hz sampled epoch of iEEG data rapidly (within a minute) compared to the current gold standard of determination of the SOZ (visual observation of ictal recordings from invasive video-EEG) and manual segmentation of HFOs, which are both very time consuming and can take several hours.

4.

4.1. Surgical outcome and ARR results

0.7 0.6 ARR values

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

(a)

4 6 8 Registration day

0.5 0.4 0.3 0.2 0.1 0

Good

Poor

Discussion

The results from this study suggest that ARR, interpreted as evidence for nonharmonicity measured in short interictal epochs of iEEGs, may be used to automatically identify channels in the SOZ, a surrogate marker for the EZ, in people with temporal lobe epilepsy. Regarding reliable identification of the SOZ, the ARR algorithm performs better than the R algorithm, and better than the TRL, although the difference in performance is smaller in the latter case. Moreover, the findings in this study suggest correspondence between TRL and R values per channel, thus confirming results from our previous study.43 Therefore, the R measure can possibly be used for fast automated estimation of the TRL.

ARR performance shows a declining trend with poorer surgical outcome. Linear regression performed on the ARR performance versus surgical outcome data was however not significant, possibly due to the small number of subjects. Moreover, there are significantly more FPs, rather than FNs, in the poor postoperative seizure control group than in the good postoperative seizure control group. This result implies that diminished ARR algorithm performance with poorer postoperative seizure control is due to FPs rather than FNs. Some of the channels corresponding to these FPs could actually be located in the EZ and could thus potentially be used for improvement of EZ localization.

1550015-12

2nd Reading May 14, 2015 15:53 1550015

Automated SOZ Approximation

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

4.2. AED levels and ARR results In the current study it was not possible to control for AED tapering, which could influence TRL, R, ARR, and even gold standard SOZ results.58,59 A rising trend can be observed in the ARR performance at later recording days, but it was not found to be significant (possibly because of the limited number of subjects in this study). A similar rising trend was observed in the TRL results (results not shown). These findings could imply better agreement between the TRL and ARR algorithms and the SOZ when fewer AEDs or a lower concentration of them is used. HFOs have been found to be attenuated by AEDs,58 and evidence for the attenuation of the values measured with certain nonlinear quantifiers has also been found.59 It could be that due to this attenuation, the difference between normal and epileptic brain areas, regarding HFOs and nonharmonicity, is smaller. If this hypothesis is correct, it would be advisable to quantify ripples and nonharmonicity in an interictal iEEG epoch from a later registration day, when AED levels have decreased. The effect of AEDs on ARR results should therefore be further investigated. 4.3. Examination of study results Traditionally, the residual signal variance after an AR fit is regarded as an “unexplained noisy component”. It could, however, also be due to nonharmonic processes in the underlying neuronal dynamics. Assuming that the noisy component is always present, the intermittent generation of nonharmonic events such as ripples will cause variations in the r quantity, or the residual signal variation after AR modeling. Moreover, because to obtain an ARR value the CV of r rather than the mean value per channel is calculated, we can assume that the noisy source of ARR has been largely suppressed. This motivates the introduction of the CV to obtain ARR in Eq. (2). Certain nonlinear EEG features have been analyzed in several studies for spike detection, seizure detection, and seizure prediction.60 Various univariate quantifiers, which characterize nonlinearity related to a single recording site, have been used for lateralization and localization of the EZ. Used quantifiers include neuronal complexity loss,34–36 correlation dimension,37,38 Lyapunov exponents,39 ξ,40 mean phase coherence,41 and relative phase

clustering index (rPCI).42,43 The ARR algorithm provides global evidence of nonharmonicity of system dynamics, meaning that the specific type of nonlinearity or nonstationarity causing the high ARR values, is disregarded. One might argue that by focusing on the frequently appearing high residual values in EEG data, which is essentially what the ARR algorithm does, interictal spikes are actually detected, as previous research showed is possible.61,62 Because spikes were not quantified in this study, it cannot be confirmed whether this is the case or not. However, spikes are also often observed in channels outside the SOZ.13 The zone with HFOs often has a smaller spatial extent than the irritative zone,18 defined as the area of the cortex generating spikes.2 Results from this study suggest similar or better performance of ARR compared with manually segmented HFOs. It therefore seems unlikely that ARR locates the irritative zone rather than the SOZ. Furthermore, it is noted by Lehnertz that certain nonlinear quantifiers of the EEG do not necessarily coincide with interictal discharges, but seem to coincide quite well with the EZ as determined by presurgical evaluation techniques.60 Closer examination of residual signal variation values (r) and HFO results, led to a hypothesis. It was observed in the study data that in the signals obtained from channels located in the SOZ, the iEEG windows with high residual signal variation values (r) often coincide with HFOs. In signals from channels contralateral to a SOZ that showed evidence of HFOs, r values were not high in iEEG windows during HFOs. This is consistent with findings from preliminary investigation, where we found that HFOs from the SOZ had higher nonharmonicity (results not shown). We hypothesize therefore, that at least two properties are needed in order for a neuronal network to be prone to producing seizures, i.e. recurrent connections and evidence of nonharmonicity in its behavior. A combination of recurrent connections and nonharmonicity promotes a pathological oscillatory state, and thus increases the chance of seizure occurrence.

4.4. Value of the ARR method A great advantage of the ARR algorithm is that it yields results very rapidly, as opposed to visual

1550015-13

2nd Reading May 14, 2015 15:53 1550015

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

E. E. Geertsema et al.

observation of video-EEG, manual segmentation of HFOs, and even to some nonlinear time series analysis algorithms63 that are characterized by high computational complexity. Another advantage of the ARR algorithm (which also applies to the R algorithm) is that the algorithm’s findings are not sensitive to spurious, filter-induced oscillations, which can be mistaken for HFOs when manually segmenting those. These spurious oscillations reflect the inherent impulse response of the filter or the harmonic nature of the signal.32 ARR could potentially be used in combination with the current gold standard for delineation of the SOZ (visual observation of ictal recordings from the invasive video-EEG) as an extra control during presurgical workup. Sub-optimal postoperative seizure control could, among other reasons, possibly be caused by sub-optimal electrode placement or even a SOZ which is unidentifiable with the implantation method used (e.g. a “hidden” or a “missed” SOZ in which no electrodes were located), which were consequently not resected.64 When the ARR findings and the SOZ as found by the current gold standard are not in agreement, this could suggest that poorer postoperative seizure control might be expected. Combining ARR with the current gold standard for delineation of the SOZ could therefore be beneficial for presurgical workup.

A question now arises; what is causing FP detections by ARR in data from subjects where the gold standard SOZ was in all probability correct, with good postoperative seizure control? When visually inspecting the EEG data from channels presenting FPs, two reasons for the FPs in this study were observed. Firstly, in some channels activity resembling continuous rippling can be observed. This continuous high-frequency activity has recently been found to be a physiological rather than pathological finding.65,66 Secondly, FPs could be caused by recurrent artefacts, the morphology of which could give rise to spuriously high residual values. Small numbers of artefacts are, however, unlikely to influence the CV used to obtain ARR; the artefacts would have to be recurrent to give rise to high ARR values. Additionally, channels with many artefacts are often excluded from EEG storage, if the channels are unlikely to be located in the SOZ. The choice of 5-min slow-wave sleep epochs of iEEG for analysis was based on findings regarding the stability of the HFO rate.19 Often only one short nightly interictal iEEG recording was stored, thus investigating ARR stationarity during the total registration, which often lasted a week, was not possible.

4.6. Future work 4.5. Study limitations As in many studies about novel EZ biomarkers, ARR results were compared with the current gold standard for delineating the SOZ. This is in turn only a surrogate marker for the EZ. If ARR findings and gold standard SOZ are not in agreement, reassessment of the EEG data may provide extra insights in retrospect in people with difficult-to-localize SOZ. Sub-optimal agreement of a new algorithm with an imperfect gold standard is possible, even if the algorithm performs its task well. FPs produced by any algorithm in light of the SOZ could in fact be true positives in light of the EZ. Validation of the ARR algorithm for its added value in the identification of the SOZ as well as in the delineation of the EZ is required. In such a validation study ARR findings should be compared with long term surgical outcome, as this is currently the only way to derive the EZ.

Further investigations, required to validate the ARR algorithm are justified by the promising results from the pilot test performed in this study. The added value of the ARR algorithm could be further investigated by comparing ARR findings within the resected EZ with those outside it in individuals with poor versus good surgical outcomes. The EZ could be approximated by reviewing resection and long term surgical outcome data. Before performing such a validation study, a number of properties should be further investigated. Firstly, optimization of the chosen length of the AR modeling window should be investigated. As mentioned in the methods section, the ARR algorithm possesses filtering capacities in accordance with the chosen window length. One could decrease the window length to focus on even higher frequencies (e.g. in the ripple, or even fast ripple range). In doing this, however, one is limited by the sampling frequency; the modeling window (40 samples in this

1550015-14

2nd Reading May 14, 2015 15:53 1550015

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Automated SOZ Approximation

study) must have sufficient samples to model the signal adequately. Secondly, general stationarity of the measured EEG characteristic should be further investigated by comparing results from various registration nights and from various epoch-lengths. That way the appropriate recording moment and length of an iEEG sample epoch can be determined. It is possible for example, that ARR values in SOZ channels increase when there is an increased chance of a seizure. AED concentration fluctuations during tapering and reintroduction in the course of long term iEEG monitoring may influence ARR results as well. Furthermore, the choice of a sample of slow-wave sleep as opposed to other epoch sampling methodologies in wakefulness as well as in sleep for ARR determination should be further studied.

Table A.1. Subj

Subj

Abbreviation

Electrode location

1

AML ATHCL

IPAR PT1R

Left amygdala Left anterior-temporal hippocampus Left anterior insula Left mid-temporal hippocampus Left insula anterior posterior Left parieto-temporal Left temporal pole Right fronto-basal Right amygdala Right anterior-temporal hippocampus Right insula posterior anterior Right parieto-temporal

LesAntR LesSupR LesAnIR TPamR HCR

Right Right Right Right Right

IAIL MTHCL IAPL PT1L TPL FBR AMR ATHCR

2

lesion anterior lesion superior lesion anterior insula temporal pole amygdala hippocampus

Electrode location

FOrbR FPoIR DLSMAR

Right fronto-orbital Right fronto-polar insula Right dorso-lateral supplementary motor area Right lateral gyrus cinguli anterior Right lateral gyrus cinguli medial Right insula anterior

LGCMdR InsAntR 3

FBL HAL FBR AR HAR HPR TBR TPOR IAIR IPIR TOPAR

Left fronto-basal Left hippocampus anterior Right fronto-basal Right amygdala Right hippocampus anterior Left hippocampus posterior Right temporo-basal Right temporo-parietal-occipital Right insula anterior Right insula posterior inferior Right temporal-occipital posterior amygdala

4

LFL INL HCL TOR FOR LFR INR MFR HCR

Left lateral frontal Left insula Left hippocampus Right temporo-occipital Right frontal operculum Right lateral frontal Right insula Right medial frontal Right hippocampus

5

PSMAL

Left posterior supplementary motor area Left frontal operculum Left parietal frontal Left fronto-orbital Left temporal operculum Left hippocampus Right hippocampus Right posterior supplementary motor area

Appendix A Table A.1. Implanted intracranial electrodes per subject, and used abbreviations.

Abbreviation

LGCAnR

Acknowledgments The authors would like to thank J. C. Baayen, for the implantation of the depth electrodes and for the subsequent epilepsy surgery on the patients included in this study. Gerhard H. Visser and Maeike Zijlmans contributed equally.

Continued

FODL PFL FORBL TOL HCL HCR PSMAR 6

FOL OFL GCL MFL OTL HCL AML OFR MFR HCR

Left fronto-orbital Left frontal operculum Left gyrus cinguli Left medial frontal Left temporal operculum Left hippocampus Left amygdala Right frontal operculum Left medial frontal Right hippocampus

7

LFL MHL

Left lateral frontal Left medial hippocampus

1550015-15

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

Table A.1.

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

Subj

Continued

Abbreviation

Electrode location

IAR IPR IIR FCR ATR AHR MHR PHR FBR TPOR

Right Right Right Right Right Right Right Right Right Right

8

HippL FrInsR Peril FrOrbR Antgr Postgr HippR

Left hippocampus Right frontal insula Peri lesional Right fronto-orbital Anterior grid Posterior grid Right hippocampus

9

FBL TPL HC1L HC2L AML IAL TBL TPOL S2PIL MIL FBR TPR HC1R TBR TPOR

Left fronto-basal Left temporal pole Left hippocampus Left hippocampus Left amygdala Left insula anterior Left temporo-basal Left temporo-parieto-occipital Left posterior insula left mid-insula Right fronto-basal Right temporal pole Right hippocampus Right temporo-basal Right temporo-parietal-occipital

insula anterior insula posterior insula inferior frontal cingular anterior temporal anterior hippocampus medial hippocampus posterior hippocampus fronto-basal temporo-parietal-occipital

References 1. W. T. Blume, H. O. L¨ uders, E. Mizrahi et al., Glossary of descriptive terminology for ictal semiology: Report of the ILAE task force on classification and terminology, Epilepsia 42(9) (2001) 1212–1218. 2. H. O. L¨ uders, I. Najm, D. Nair et al., The epileptogenic zone: General principles, Epileptic Disord. 8(Suppl 2) (2006) S1–S9. 3. F. Rosenow and H. L¨ uders, Presurgical evaluation of epilepsy, Brain 124 (2001) 1683–1700. 4. B. Litt, R. Esteller, J. Echauz et al., Epileptic seizures may begin hours in advance of clinical onset: A report of five patients, Neuron 30(1) (2001) 51–64. 5. B. Litt and K. Lehnertz, Seizure prediction and the preseizure period, Curr. Opin. Neurol. 15(2) (2002) 173–177. 6. A. Bragin, J. Engel, C. L. Wilson et al., Highfrequency oscillations in human brain, Hippocampus 9(2) (1999) 137–142.

7. A. Bragin, J. Engel, C. L. Wilson et al., Hippocampal and entorhinal cortex high-frequency oscillations (100–500 Hz) in human epileptic brain and in kainic acid–treated rats with chronic seizures, Epilepsia 40(2) (1999) 127–137. 8. G. A. Worrell, L. Parish, S. D. Cranstoun et al., High-frequency oscillations and seizure generation in neocortical epilepsy, Brain 127(7) (2004) 1496– 1506. 9. E. Urrestarazu, J. D. Jirsch, P. LeVan et al., Highfrequency intracerebral EEG activity (100–500 Hz) following interictal spikes, Epilepsia 47(9) (2006) 1465–1476. 10. A. Bragin, J. Engel and R. J. Staba, High-frequency oscillations in epileptic brain, Curr. Opin. Neurol. 23(2) (2010) 151–156. 11. R. J. Staba, C. L. Wilson, A. Bragin et al., Highfrequency oscillations recorded in human medial temporal lobe during sleep, Ann. Neurol. 56(1) (2004) 108–115. 12. J. D. Jirsch, E. Urrestarazu, P. LeVan et al., Highfrequency oscillations during human focal seizures, Brain 129(Pt 6) (2006) 1593–1608. 13. J. Jacobs, P. LeVan, R. Chander et al., Interictal high-frequency oscillations (80–500 Hz) are an indicator of seizure onset areas independent of spikes in the human epileptic brain, Epilepsia 49(11) (2008) 1893–1907. 14. J. Jacobs, M. Zijlmans, R. Zelmann et al., Highfrequency electroencephalographic oscillations correlate with outcome of epilepsy surgery, Ann. Neurol. 67(2) (2010) 209–220. 15. J. Y. Wu, R. Sankar, J. T. Lerner et al., Removing interictal fast ripples on electrocorticography linked with seizure freedom in children, Neurology 75(19) (2010) 1686–1694. 16. P. N. Modur, S. Zhang and T. W. Vitaz, Ictal highfrequency oscillations in neocortical epilepsy: Implications for seizure localization and surgical resection, Epilepsia 52(10) (2011) 1792–1801. 17. M. A. Klooster van’t, M. Zijlmans, F. S. S. Leijten et al., Time-frequency analysis of single pulse electrical stimulation to assist delineation of epileptogenic cortex, Brain 134 (2011) 2855–2866. 18. M. Zijlmans, P. Jiruska, R. Zelmann et al., Highfrequency oscillations as a new biomarker in epilepsy, Ann. Neurol. 71(2) (2012) 169–178. 19. R. Zelmann, M. Zijlmans, J. Jacobs et al., Clinical neurophysiology improving the identification of high frequency oscillations, Clin. Neurophysiol. 120(8) (2009) 1457–1464. 20. A. B. Gardner, G. A. Worrell, E. Marsh et al., Human and automated detection of high-frequency oscillations in clinical intracranial EEG recordings, Clin. Neurophysiol. 118(5) (2007) 1134–1143. 21. R. J. Staba, C. L. Wilson, A. Bragin et al., Quantitative analysis of high-frequency oscillations (80– 500 Hz) recorded in human epileptic hippocampus

1550015-16

2nd Reading May 14, 2015 15:53 1550015

Automated SOZ Approximation

22.

23.

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

24.

25.

26.

27.

28.

29.

30.

31.

32.

33.

34.

35.

and entorhinal cortex, J. Neurophysiol. 88 (2002) 1743–1752. R. Nelson, S. M. Myers, J. D. Simonotto et al., Detection of high frequency oscillations with Teager energy in an animal model of limbic epilepsy, in Conf. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Vol. 1 (2006) 2578–2580. H. Firpi, O. Smart, G. Worrell et al., High-frequency oscillations detected in epileptic networks using swarmed neural-network features, Ann. Biomed. Eng. 35(9) (2007) 1573–1584. G. A. Worrell, A. B. Gardner, S. M. Stead et al., High-frequency oscillations in human temporal lobe: Simultaneous microwire and clinical macroelectrode recordings, Brain 131(Pt 4) (2008) 928–937. J. A. Blanco, M. Stead, A. Krieger et al., Unsupervised classification of high-frequency oscillations in human neocortical epilepsy and control patients, J. Neurophysiol. 104(5) (2010) 2900–2912. R. Zelmann, F. Mari, J. Jacobs et al., Automatic detector of high frequency oscillations for human recordings with macroelectrodes, 32nd Annu. Int. Conf. IEEE EMBS, Vol. 32 (2010), pp. 2329–2333. B. Cr´epon, V. Navarro, D. Hasboun et al., Mapping interictal oscillations greater than 200 Hz recorded with intracranial macroelectrodes in human epilepsy, Brain 133(Pt 1) (2010) 33–45. G. Birot, A. Kachenoura, L. Albera et al., Automatic detection of fast ripples, J. Neurosci. Methods 213 (2013) 1–14. M. D¨ umpelmann, J. Jacobs, K. Kerber and A. Schulze-Bonhage, Automatic 80–250 Hz “ripple” high frequency oscillation detection in invasive subdural grid and strip recordings in epilepsy by a radial basis function neural network, Clin. Neurophysiol. 123(9) (2012) 1721–1731. J. Csicsvari, H. Hirase, A. Czurk´ o et al., Fast network oscillations in the hippocampal CA1 region of the behaving rat, J. Neurosci. 19(16) (1999) RC20. G. Curio, B.-M. M. Mackert, M. Burghoff et al., Localization of evoked neuromagnetic 600 Hz activity in the cerebral somatosensory system, Electroencephalogr. Clin. Neurophysiol. 91(6) (1994) 483–487. C. G. B´enar, L. Chauvi`ere, F. Bartolomei and F. Wendling, Pitfalls of high-pass filtering for detecting epileptic oscillations: A technical note on “false” ripples, Clin. Neurophysiol. 121(3) (2010) 301–310. N. Jmail, M. Gavaret, F. Wendling et al., A comparison of methods for separation of transient and oscillatory signals in EEG, J. Neurosci. Methods 199(2) (2011) 273–289. K. Lehnertz and C. E. Elger, Spatio-temporal dynamics of the primary epileptogenic area in temporal lobe epilepsy characterized by neuronal complexity loss, Electroencephalogr. Clin. Neurophysiol. 95(2) (1995) 108–117. B. Weber, K. Lehnertz, C. E. Elger and H. G. Wieser, Neuronal complexity loss in interictal EEG

36.

37.

38.

39.

40.

41.

42.

43.

44.

45.

46.

47.

48.

1550015-17

recorded with foramen ovale electrodes predicts side of primary epileptogenic area in temporal lobe epilepsy: A replication study, Epilepsia 39(9) (1998) 922–927. G. Widman, K. Lehnertz, H. Urbach and C. E. Elger, Spatial distribution of neuronal complexity loss in neocortical lesional epilepsies, Epilepsia 41(7) (2000) 811–817. M. C. Casdagli, L. D. Iasemidis, R. S. Savit et al., Non-linearity in invasive EEG recordings from patients with temporal lobe epilepsy, Electroencephalogr. Clin. Neurophysiol. 102(2) (1997) 98–105. H. Jing, M. Takigawa and A. Benasich, Relationship of nonlinear analysis, MRI and SPECT in the lateralization of temporal lobe epilepsy, Eur. Neurol. 48(1) (2002) 11–19. Z. J. Kowalik, A. Schnitzler, H. Freund and O. W. Witte, Local Lyapunov exponents detect epileptic zones in spike-less interictal MEG recordings, Clin. Neurophysiol. 112(1) (2001) 60–67. R. G. Andrzejak, G. Widman, K. Lehnertz et al., The epileptic process as nonlinear deterministic dynamics in a stochastic environment: An evaluation on mesial temporal lobe epilepsy, Epilepsy Res. 44(2–3) (2001) 129–140. F. Mormann, K. Lehnertz, P. David and C. E. Elger, Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy patients, Physica D 144 (2000) 358–369. S. Kalitzin, D. Velis, P. Suffczynski et al., Electrical brain-stimulation paradigm for estimating the seizure onset site and the time to ictal transition in temporal lobe epilepsy, Clin. Neurophysiol. 116(3) (2005) 718–728. S. Kalitzin, M. Zijlmans, G. Petkov et al., Quantification of spontaneous and evoked HFO’s in SEEG recording and prospective for pre-surgical diagnostics. Case study, in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, Vol. 6 (2012), pp. 1024– 1027. H. Adeli and S. Ghosh-Dastidar, Automated EEGBased Diagnosis of Neurological Disorders: Inventing the Future of Neurology (CRC Press, Boca Raton, Florida, 2010). J. Engel Jr., P. van Ness, T. Rasmussen and G. Ojemann, Outcome with respect to epileptic seizures, in Surgical Treatment of the Epilepsies, 2nd edn. (Raven Press, New York, 1993), pp. 609–621. L. D. Olson and M. S. Perry, Localization of epileptic foci using multimodality neuroimaging, Int. J. Neural Syst. 23(1) (2013) 1250031. S. Claus, D. Velis and W. van Emde Boas, Antiepileptic drug withdrawal in presurgical evaluation: Advantages, disadvantages, and guidelines, in ed. H. O. Luders, Textbook Epilepsy Surgergy (CRC Press, 2008), pp. 580–587. A. P. Bagshaw, J. Jacobs, P. LeVan et al., Effect of sleep stage on interictal high-frequency oscillations

2nd Reading May 14, 2015 15:53 1550015

E. E. Geertsema et al.

49.

50.

Int. J. Neur. Syst. 2015.25. Downloaded from www.worldscientific.com by NANYANG TECHNOLOGICAL UNIVERSITY on 11/20/15. For personal use only.

51.

52.

53.

54.

55.

56.

57.

recorded from depth macroelectrodes in patients with focal epilepsy, Epilepsia 50(4) (2009) 617– 628. H. Otsubo, A. Ochi, K. Imai et al., High-frequency oscillations of ictal muscle activity and epileptogenic discharges on intracranial EEG in a temporal lobe epilepsy patient, Clin. Neurophysiol. 119(4) (2008) 862–868. E. Niedermeyer and F. Lopes da Silva, Electroencephalography: Basic Principals, Clinical Applications and Related Fields (Lipponcott and Wilkins, Philadelphia, 2011). P. Franaszczuk and K. Blinowska, Linear model of brain electrical activity — EEG as a superposition of damped oscillatory modes, Biol. Cybern. 53(1) (1985) 19–25. V. Perumalsamy, Sleep spindles detection from human sleep EEG signals using autoregressive (AR) model: A surrogate data approach, J. Biomed. Sci. Eng. 2(5) (2009) 294–303. J. Timmer, The power of surrogate data testing with respect to non-stationarity, Phys. Rev. E 58(4) (1998) 5153–5156. C. Rieke, F. Mormann, R. G. Andrzejak et al., Discerning nonstationarity from nonlinearity in seizurefree and preseizure EEG recordings from epilepsy patients, IEEE Trans. Biomed. Eng. 50(5) (2003) 634–639. F. Lopes da Silva, J. P. Pijn and P. Boeijinga, Interdependence of EEG signals: Linear vs. nonlinear associations and the significance of time delays and phase shifts, Brain Topogr. 2 (1989) 9–18. S. Kalitzin, J. Parra, D. Velis and F. Lopes da Silva, Quantification of unidirectional nonlinear associations between multidimensional physiological signals, Eur. Symp. Biomed. Eng., Patras, Greece (2006) S5.13. S. N. Kalitzin, J. Parra, D. N. Velis and F. H. Lopes da Silva, Quantification of unidirectional nonlinear

58.

59.

60. 61.

62.

63.

64.

65.

66.

1550015-18

associations between multidimensional signals, IEEE Trans. Biomed. Eng. 54(3) (2007) 454–461. M. Zijlmans, J. Jacobs, R. Zelmann et al., Highfrequency oscillations mirror disease activity in patients with epilepsy, Neurology 72(11) (2009) 979– 986. K. Lehnertz and C. E. Elger, Neuronal complexity loss in temporal lobe epilepsy: Effects of carbamazepine on the dynamics of the epileptogenic focus, Electroencephalogr. Clin. Neurophysiol. 103(3) (1997) 376–380. K. Lehnertz, Epilepsy and nonlinear dynamics, J. Biol. Phys. 34(3–4) (2008) 253–266. F. H. Lopes da Silva, A. Dijk and H. Smits, Detection of Nonstationarities in EEGs Using the Autoregressive Model an Application to EEGs of Epileptics (Gustav Fischer Verlag, Stuttgart, 1975). A. T. Tzallas, V. P. Oikonomou and D. I. Fotiadis, Epileptic spike detection using a Kalman filter based approach, in Conf. Proc. IEEE Eng. Med. Biol. Soc. Vol. 1 (2006), pp. 501–504. R. G. Andrzejak, F. Mormann, G. Widman et al., Improved spatial characterization of the epileptic brain by focusing on nonlinearity, Epilepsy Res. 69(1) (2006) 30–44. A. Harroud, A. Bouthillier, A. G. Weil and D. K. Nguyen, Temporal lobe epilepsy surgery failures: A review, Epilepsy Res. Treat. 2012 (2012) 1–10. F. Melani, R. Zelmann, F. Mari and J. Gotman, Continuous high frequency activity: A peculiar SEEG pattern related to specific brain regions, Clin. Neurophysiol. 124(8) (2013) 1507–1516. K. Kerber, M. D¨ umpelmann, B. Schelter et al., Differentiation of specific ripple patterns helps to identify epileptogenic areas for surgical procedures, Clin. Neurophysiol. 125(7) (2014) 1339–45.

Automated Seizure Onset Zone Approximation Based on Nonharmonic High-Frequency Oscillations in Human Interictal Intracranial EEGs.

A novel automated algorithm is proposed to approximate the seizure onset zone (SOZ), while providing reproducible output. The SOZ, a surrogate marker ...
507KB Sizes 0 Downloads 13 Views