Description

Animal Models are used extensively in basic epilepsy research. In many studies, there is a need to accurately score and quantify all epileptic spike and wave discharges (SWDs) as captured by electroencephalographic (EEG) recordings. Manual scoring of

Categories

Published

All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.

Share

Transcript

Abstract
—Animal Models are used extensively in basicepilepsy research. In many studies, there is a need to accuratelyscore and quantify all epileptic spike and wave discharges(SWDs) as captured by electroencephalographic (EEG)recordings. Manual scoring of long term EEG recordings is atime-consuming and tedious task that requires inordinateamount of time of laboratory personnel and an experiencedelectroencephalographer. In this paper, we adapt a SWDdetection algorithm, srcinally proposed by the authors forabsence (petit mal) seizure detection in humans, to detectSWDs appearing in EEG recordings of Fischer 334 rats. Thealgorithm is robust with respect to the threshold parameters.Results are compared to manual scoring and the effect of different threshold parameters is discussed.
I.
I
NTRODUCTION
NIMAL
modeling of absence (petit mal) seizures has been conducted in numerous studies using mice andrats. The defining EEG event in these models is a 7-12Hz generalized spike and wave discharge (SWD; Fig. 1A).In general, SWDs are episodes of abrupt onset, variableduration (seconds to minutes), and abrupt termination thatusually occur during passive wakefulness and light sleep [3][18]. The episodes are characterized by behavioral arrest anddecreased responsiveness, with or without rhythmic whisker twitching [2] [18]. Additional behavioral features that mayaccompany SWDs include accelerated breathing, head tilt,eye twitching [3], and head and neck twitching associatedwith gradual head lowering [8].Studies of generalized SWDs are divided into twocategories. Acquired SWD models are based upon the use of chemical agents to stimulate the expression of SWDs, or todecrease the threshold for their expression. SpontaneousSWD models, by contrast, are based upon inherited factorsthat lead to the development of unprovoked generalized
Manuscript received April 23, 2009. This work was partially supported by the NSF, the U.S. Air Force, the North Florida/South Georgia VeteransHealth System, an Epilepsy Foundation Research Initiative for SeniorsAward (SPN), and NINDS R01 NS046015 (KMK).P. Xanthopoulos, J-C, Zhang and P. M. Pardalos are with Industrial andSystems Engineering Department at University of Florida, 303 Weil Hall,Gainesville, FL 32611, USA (phone: 352-870-9176; fax: 352-392-3537; e-mail: petros.xanthopoulos@gmail.com).C-C, Liu, was with Department of Neurosurgery, Medical School of Johns Hopkins University, Baltimore, MD 21287, USA.Eric R. Miller, S. Nair and K. Kelly are with Allegheny-Singer ResearchInstitute, Allegheny General Hospital, Pittsburgh, PA 15212, USA.B. M. Uthman was with Weill Cornell Medical College in Qatar, Doha,Qatar.
SWD activity. Included in this category are WAG/Rij rats(Wistar Albino Glaxo strain, bred in Rijswijk, Netherlands)and genetic absence epilepsy rats from Strasbourg(GAERS), two Wistar-derived rat strains that have beenselectively inbred to increase their propensity to expressgeneralized SWDs [5]. A majority of generalized SWDstudies have been performed using these rat strains;however, SWDs can be seen in many common laboratory ratstrains, both inbred and outbred. Wistar-unrelated inbredstrains include Fischer 344 (F344), Brown Norway, and dark agouti [3] [14] [18]. Common outbred strains includeSprague-Dawley, Wistar, and Long-Evans [2] [12] [16] [18].Recently data mining, signal processing, and optimizationhave been used to provide increasing numbers of decision-assisting tools to clinical and basic researchers [10] [11].Animal models of absence seizures frequently requireaccurate scoring and quantification of the absence events(SWDs) [6] [7] [8] [9]. This task is usually performedmanually by laboratory personnel and an experiencedelectroencephalographer; manual scoring is a time-consuming and tedious task and is always subject todetection errors due to examiner experience and fatigue.Therefore, an improved automated SWD algorithm is verydesirable.There are very few studies in the literature related toautomated SWD detection in animal models. In [17] anautomatic SWD detector was introduced based on the firstderivative of EEG signals, called the “steepness of thesignals.” The SWDs are detected when the value of steepness exceeds the threshold value in certain consecutiveEEG epochs. Despite the reported high accuracy of thismethod, it sometimes misclassifies eye movement artifactsas absence seizures. In [4] Fanselow et al. described amethod based on the maximum absolute value of the EEGamplitude in the rat model; the SWDs in the EEG recordingswere
labeled when the amplitude was greater than thethreshold for a manually-defined time horizon [4]. Again,this method could not distinguish between SWDs and highamplitude artifacts. A so-called “spectral comb-based”analytical method was proposed by [15] and used for detecting SWDs in EEG recordings using the GAERS strain.The authors used the time frequency spectrum produced byShort Time Fourier Transform (STFT) to extract featuresthat enabled seizure detection.In this paper, we adapt the algorithm previously proposed by the authors for automatic detection and quantification of SWD epochs in human absence epilepsy [19] to a rat model
A robust spike and wave algorithm for detecting seizures in a geneticabsence seizure model
Petros Xanthopoulos, Chang-Chia Liu, Jicong Zhang, Eric R. Miller, S. P. Nair, Basim M.Uthman, Kevin Kelly, and Panos M. Pardalos
A
2184
31st Annual International Conference of the IEEE EMBSMinneapolis, Minnesota, USA, September 2-6, 2009
978-1-4244-3296-7/09/$25.00 ©2009 IEEE
of absence epilepsy. The paper is organized as follows: 1) insection II, we describe the methodology followed by adescription of the dataset and the algorithm; 2) in section III,we present the results together with the analysis related tothe robustness of the algorithm; and 3) in section IV, wediscuss the results and the choice of the threshold parameters.II.
M
ETHODS
A.
Dataset description
Eight-hour recordings were acquired from total four 4-month old F344 rats. In total, 6 screw electrodes wereimplanted in the skull of each animal: 2 frontal (F3, F4), 2central (C3, C4), and 2 parietal (P3, P4). The electrodeconfiguration is shown in Fig. 1B. In total, 8 differentialchannels were computed for the purpose of the study: F3-C3, C3-P3, F3-P3, F4-C4, C4-P4, F4-P4, C3-C4, P3-P4(Fig. 1A). Entire long term video-EEG recordings werevisually scanned by laboratory research assistants to detectand score SWD occurrence; identified SWDs wereconfirmed by an electroencephalographer. The exact number of epochs and their cumulative time during the 8 hours of recordings can be seen in the Table I.
TABLE I NUMBER OF SWD EPOCHS SCORED FOR EACH RAT AND THECUMMULATIVE TIME OF SWDS DURING THE 8-HOUR RECORDINGS
Number of SWDepochsCummulativeictal activity(epochs)Totalrecordingtime(hours)Rat A
53 99.33 8.27
Rat B
43 116.35 8.09
Rat C
81 368.53 8.00
Rat D
45 133.50 8.10
Total
202 717.71 32.46
B.
Algorithm
For detecting SWD discharges in a rat model, we proposea modification of our algorithm srcinally proposed for automatic SWD detection in human data [19]. The proposeddetection scheme is based in time frequency decompositionof the EEG employing the wavelet transform. Wavelettransform has profound advantages over the classical STFT because one can increase the scale (or frequency) resolutionwhile keeping the same time resolution. Subsequently, thevariance profile of the EEG is computed and seizures aredetected by a double thresholding process. The algorithmwas found to have high sensitivity and a minimal false positive detection rate for SWDs localized in the frequency band of ~ 3Hz. Here we describe our detection algorithm inthree simple steps.
1) Wavelet decomposition:
First, we decompose everydifferential channel of the raw EEG recordings, which can be represented as a time series X(t), into a time-scale domainusing the wavelet transform:, where. The function is the mother wavelet function. For this study, we used the Morlet mother wavelet function, which has analytic expression given by:. Morlet mother wavelet is used extensivelyin EEG analysis due to its minimum time-bandwidth product, its infinite differentiation, and its explicitexpression [1]. A time scale plot of a recorded absenceseizure is shown in Fig. 2.We can convert scales into frequencies using ,where is the central frequency of the mother wavelet, inour case, 0.81Hz, and is the sampling period.Among all the scales that we can decompose the EEG signal,we are interested in those that correspond to the frequency band in which SWD activity appears (~7 Hz). For this wekeep only the scales 19-25 and sum them for every channel(Fig. 3).
Fig. 1. (A) Sample of a generalized SWD recorded from a Fischer 344 rat. The F3, C3, and P3 abbreviations refer to skull screwelectrodes overlying left frontal, central, and parietal regions of theanimal’s brain, respectively; F4, C4 and P4 refer to the brain areas onthe right. An “F3-C3” label corresponds to an EEG channel produced by the output of one differential amplifier with inputs from the F3 andC3 electrodes. (B) Electrode placement used in the study.Fig. 2. Typical absence seizure and the corresponding scalogram prior tothe seizure (onset after sample 6000). Two electrode artifacts (onset after samples 2000 and 4000) were rejected by keeping only the scales (a=19-25) that correspond to the frequency band of interest
2185
2) Variance profile computation:
Based on the observationthat the high SWD activity produces wavelet profiles of highvariance, we compute the variance profile for each channelusing a sliding window of width k=200 samples (i.e.,corresponding to 1 sec of recording). All variance profilesfor all channels are summed to reject noise and artifacts thatare not generalized (i.e., not appearing in all channels). Avariance profile of the seizure can be seen in Fig. 4 (blueline).
3) Thresholding
: For accurate localization of the onset andoffset times of a seizure based on the variance profile of theEEG recordings, we used a double thresholding technique(Fig. 4). We applied a high threshold to the variance todetect the number of epochs; the high threshold was chosen,in part, to avoid detection of artifacts (false positives).
For the sample points of the variance profile curve that“hit” the high threshold, we perform a local search to specifythe exact onset and offset sample points of the seizure. More precisely, with reference to Fig 4, for the first point thatcorresponds to an offset (first point that high line hits thevariance profile), we search backwards to determine the firsttime that the variance curve falls below the low threshold(first point that low threshold intersects with variance profile). For the second point, we search forward todetermine the first point that the variance curve drops belowthe low threshold. For the third and fourth points, we repeatthe same search process (third point will correspond again toan onset and the fourth point to an offset). It is worth notingthat with the low thresholding search we are able to mergeepochs that were detected as two distinct events from thehigh thresholding (i.e., in this example the onset and theoffset of both epochs will be the same). Finally, thealgorithm returns only the unique events (it rejectsduplicates).III.
R
ESULTS
The detection sensitivity and false positive rate are highlydependent on the parameters of the algorithm (thresholds).We experimented with different thresholds to: 1) investigatethe robustness of the algorithm, i.e., the change of the outputwith small changes of the input parameters; and 2) describehow the sensitivity versus specificity changes with respect tothe input parameters. For these aims the algorithm wasapplied using 10 different thresholds and sensitivity andspecificity were computed. Sensitivity was defined as thenumber of epochs detected over the total number of scoredepochs, and false positive rate was defined as the number of false positives over the corresponding recording time.The plot of sensitivity versus specificity is the so-calledreceiver operating characteristic (ROC) curve. The ROC isvery useful in visualizing how the sensitivity percentagechanges as a function of missed seizure rate and help the enduser to decide the optimal point (corresponding threshold)that fits a specific application. In Fig. 5 (A&B), one can seethe ROC curves for the four rats separately and the meancurve for all four rats.One drawback of this definition of sensitivity and false positive rate is that all epochs are treated in the same manner without taking into consideration the epoch length. In thisway, one SWD epoch of 10 sec length will contribute thesame as an epoch of 1 sec. Given this consideration,sensitivity and specificity can be defined based on thecumulative SWD time or “cumulative epileptiform burden.”In this case, sensitivity is defined as the cumulative detectedtime (in sec) over the total time, whereas the false positiverate is defined as the cumulative missed seizure time over the corresponding time of the recordings.The ROC curves for the four rats and the mean curve for all four rats can be seen in Fig. 5(C&D). The fact that theROC curve using the cumulative time (Fig. 5B) has higher sensitivity values compared to the that with the detectednumber of epochs (Fig. 5D) means that the missed epochsare shorter compared to the detected epochs. This isconsistent with previous results in absence epilepsydetection [19]. It is worth mentioning that in the currentstudy we considered SWD epochs of all lengths. In clinical practice, a frequently encountered issue is whether SWDsare sufficiently long to result in an absence seizure (e.g., a0.5-sec SWD epoch would not be clinically significant).Under such assumptions, the detection sensitivity and false positive rate would improve drastically. Therefore, the error analysis presented here can be viewed as an upper limit of the error range. With regard to the second parameter of thealgorithm (low threshold), it determines the accuracy of theseizure onset and offset detection. It is easy to recognizefrom Fig. 5 that changes of the low thresholdwould modify the onset and offset detection point by somenumber of sample points, which correspond to 1/200 seceach.
Fig. 3. Sum of scales of interest (19 to 25)Fig. 4. High threshold illustrated in red detects two distinct epochs of SWDs. For each epoch detected, we try to find the nearest point atwhich the variance value drops below the low threshold. Firstthresholding detects two distinct epochs that are merged by the lowthresholdinsearch.
2186
IV.
D
ISCUSSION AND CONCLUSIONS
In this paper, we presented an adaptation of the algorithm proposed for human absence epilepsy [19] for the detectionof SWDs in rat model of absence epilepsy. The proposedalgorithm is robust with respect to the input parameters,which means that the output (detected epochs) cannot differ significantly when small changes are made to the threshold parameters. In addition, the ROC analysis shows that highsensitivity rates (~ >90%) can be achieved along with a lowfalse positive rate (~ 2-4 false positive epochs per hour).When considering the cumulative time of the SWD epochsinstead of the absolute number of the epochs, the algorithmdemonstrates, on average, over 90% accuracy while the totalmissed SWD time doesn’t exceed 8.5 sec per hour.Importantly, the sensitivity percentages and the false positive rates can increase dramatically when epochs longer than some predefined duration are considered. However, anoptimal threshold parameter cannot be proposed because thedesired sensitivity and false positive rate are highlyapplication-dependent. These results encourage us togeneralize the detection methodology into other generalizedEEG patterns that appear in specific frequency bands whilefurther investigation and experiments are ongoing to reveal potential imperfections of the algorithm.A
CKNOWLEDGMENT
We would like to acknowledge NSF and Air Force for partial financial support. This work was partially supported by the North Florida Foundation for Research and EducationInc., North Florida/South Georgia Veterans Health System, by an Epilepsy Foundation Research Initiative for SeniorsAward (SPN), and NINDS R01 NS046015 (KMK). Wewould also like to acknowledge Pennsylvania Department of Health Research for the Formula Fund RFA 04-07-09 SAP4100031268 (KMK).R
EFERENCES
[1]
N. Ahuja, S. Lertrattanapanich, and N. K. Bose. Propertiesdetermining choice of mother wavelet.
Vision,Image and Signal Processing, IEE Proceedings
, 2005.
[2]
G. Buzsáki, A. Smith, S. Berger, L. J. Fisher, and F. H. Gage. Petitmal epilepsy and parkinsonian tremor:
hypothesis of a common pacemaker.
Neuroscience
36(1):1-14 1990.
[3]
A. M. Coenen, W. H. Drinkenburg, M. Inoue, and E. L. vanLuijtelaar. Genetic models of absence epilepsy, with
emphasis on theWAG/Rij strain of rats.
Epilepsy Res
earch 12(2):75-86 1992.
[4]
E. Fanselow, A. P. Reid, and M.A.L. Nicolelis. Reduction of pentylenetetrazole-induced seizure
activity in awake rats by seizue-triggered trigeminal nerve stimulation.
The Journal of Neuroscience
,20:8160–8, 2000.
[5]
K. M. Kelly. Spike-wave discharges: absence or not, a commonfinding in common laboratory rats.
Epilepsy Currents
4(5): 176-1772004.
[6]
K. M. Kelly, A. Kharlamov, T. M. Hentosz, E. A. Kharlamov, J. M.Williamson, E. H. Bertram, J. Kapur,
and
D.M. Armstrong.Photothrombotic brain infarction results in seizure activity in agingFischer
344 and Sprague Dawley rats.
Epilepsy Research
47: 189-2032001.
[7]
K. M. Kelly, P. I. Jukkola, E. A. Kharlamov, K. L. Downey, J. W.McBride, R. Strong, and J. Aronowski.
Long-term video-EEGrecordings following transient unilateral middle cerebral and common
carotid artery occlusion in Long-Evans rats.
Experimental Neurology
201: 495-506 2006.
[8]
E. A. Kharlamov, P. I. Jukkola, K. L. Schmitt, and K. M. Kelly.Electrobehavioral characteristics of epileptic
rats following photothrombotic brain infarction.
Epilepsy Research
56: 185-2032003.
[9]
S. P. Nair, P. I. Jukkola, M. Quigley, A. Wilberger, D. S. Shiau, J. C.Sackellares, P. M. Pardalos, and K. M. Kelly.Absence seizures asresetting mechanisms of brain dynamics.
Cybernetics and Systems Analysis
44(5):664-672 2008.
[10]
P.M. Pardalos, V. Boginski, and A. Vazacopoulos, editors. DataMining in Biomedicine, volume 7 of
Springer Optimization and ItsApplications. Springer Verlag, 2007.
[11]
P.M. Pardalos, J.C. Sackellares, P.R. Carney, and L.D. Iasemidis,editors. Quantitative Neuroscience:
Models, Algorithms, Diagnostics,and Therapeutic Applications. Springer Verlag, 2004
[12]
K. Semba, H. Szechtman, and B. R. Komisaruk. Synchrony amongrhythmical facial tremor, neocortical 'alpha'
waves, and thalamic non-sensory neuronal bursts in intact awake rats.
Brain Research
Aug
18;195(2):281-98 1980.
[13]
O. Seref, E.O. Kundakcioglu, and P.M. Pardalos, editors. DataMining, Systems Analysis, and
Optimization in Biomedicine, volume953 of AIP Conference Proceedings. American Instintute of
Physics,2007.
[14]
E. L. Van Luijtelaar, Coenen AM. Two types of electrocortical paroxysms in an inbred strain of rats.
Neuroscience Letters
Oct20;70(3):393-7 1986.
[15]
P. Van Hese, J.P. Martens, P. Boon, S. Dedeurwaerdere, I. Lemahieu,and R. Van de Walle. Detection of
spike and wave discharges in thecortical EEG of genetic absence epilepsy rats from Strasbourg.
Physics in Medicine and Biology
, 48:1685–700, 2003.
[16]
M. Vergnes, C. Marescaux, G. Micheletti, J. Reis, A. Depaulis, L.Rumbach, and J. M. Warter. Spontaneous
paroxysmal electroclinical patterns in rat: a model of generalized non-convulsive epilepsy.
Neuroscience Letters
Nov 16;33(1):97-101 1982.
[17]
F. Westerhuis, W. Van Schaijk, and G. Van Luijtelaar. Automaticdetection of spike-wave discharges in
the cortical EEG of rats.Measuring Behavior ’96,
Int. Workshop on Methods and Techniquesin Behavioral Research
, Utrecht, The Netherlands 16-18 Oct. 1996.
[18]
J. O. Willoughby and L. Mackenzie. Nonconvulsiveelectrocorticographic paroxysms (absence epilepsy) in rat
strains.
Lab Anim Sci
. Dec;42(6):551-4 1992.
[19]
P. Xanthopoulos, S. Rebennack, C. C. Liu, P.M. Pardalos, G. L.Holmes and B. M. Uthman. A novel wavelet based algorithm for spikeand wave detection in absence epilepsy. Submitted 2008
Fig. 5. (A) ROC for four individual rats using number of epochs (B)the average curve for all rats using number of epochs (C) ROC for four individual rats using length of epochs (D) the average curve for all rats using length of epochs.
2187

We Need Your Support

Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks

SAVE OUR EARTH

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...Sign Now!

We are very appreciated for your Prompt Action!

x