A Novel Wavelet-Based Index to Detect Epileptic SeizuresUsing Scalp EEG Signals

Ali Shahidi Zandi∗, Student Member, IEEE, Guy A. Dumont∗, Fellow, IEEE, Manouchehr Javidan†,Reza Tafreshi∗∗, Member, IEEE, Bernard A. MacLeod‡, Craig R. Ries‡, and Ernie Puil‡

Abstract— In this paper, we propose a novel wavelet-basedalgorithm for the detection of epileptic seizures. The algorithmis based on the recognition of rhythmic activities associated withictal states in surface EEG recordings. Using a moving-windowanalysis, we first decomposed each EEG segment into a waveletpacket tree. Then, we extracted the coefficients correspondingto the frequency band of interest defined for rhythmic activities.Finally, a normalized index sensitive to both the rhythmicity andenergy of the EEG signal was derived, based on the resultingcoefficients. In our study, we evaluated this combined indexfor real-time detection of epileptic seizures using a dataset of∼11.5 hours of multichannel scalp EEG recordings from threepatients and compared it to our previously proposed wavelet-based index. In this dataset, the novel combined index detectedall epileptic seizures with a false detection rate of 0.52/hr.

I. INTRODUCTION

Epilepsy is the second most common neurological disorder(after stroke), affecting more than 50 million patients aroundthe world [1], [2]. This serious disorder is associated withrecurrent, unprovoked epileptic seizures that result froma sudden disturbance of brain function, characterized byabnormal synchronous firing in neuronal populations.

Monitoring the patient’s electroencephalogram (EEG) forseveral days is often needed for diagnosis and objectiverecord of the epileptic seizure events. This process is tedious,expensive, and time-consuming. Accordingly, a reliable on-line seizure detection system would facilitate long-term mon-itoring and treatment of epilepsy.

Automatic diagnosis of epileptic seizures using EEG hasbeen the goal of several studies. Some of these have focusedon both intracranial and scalp EEG recordings to detect ictalstates [3], and others have employed either surface or depthEEG electrodes [4], [5], [6], [7], [8]. Although proposedseizure detection algorithms provide promising results, theyare still far from being able to recognize most epilepticseizures in real time with high sensitivity and high specificity.One algorithm [3] tested on long-term EEG recordingsshowed 76% sensitivity and a false detection rate in theorder of 1/hr, whereas another study [5] revealed nearly 93%sensitivity with an average false detection rate of 1.35/hr.

Manuscript received April 10, 2008.∗Department of Electrical & Computer Engineering at The University of

British Columbia (UBC), Vancouver, BC, V6T 1Z4, Canada†Division of Neurology, Department of Medicine at The University of

British Columbia (UBC); Director, Neurophysiology Lab. at VancouverGeneral Hospital (VGH), Vancouver, BC, V5Z 1M9, Canada∗∗Department of Mechanical Engineering at Texas A&M University at

Qatar, Doha, P.O. Box 23874, Qatar‡Department of Anesthesiology, Pharmacology & Therapeutics at The

University of British Columbia (UBC), Vancouver, BC, V6T 1Z3, Canada

Also, a recently developed seizure detection algorithm [8]had an average sensitivity of 76% and a false detection rateof 0.34/hr.

In this paper, we propose a novel wavelet-based index fordetecting epileptic seizures based on recognition of rhythmicactivities in scalp EEG signals. The ultimate objective of ourresearch is to increase both sensitivity and specificity of theseizure detection algorithm, as well as to reduce the detectiondelay.

II. WAVELET TRANSFORM

As a linear time-frequency transform, wavelet transform(WT) is a suitable analytical tool in pattern recognitionand signal processing especially in the analysis of transientand non-stationary phenomena [9], [10]. Therefore, it hasbeen utilized widely in biomedical signal processing [6], [8],[11]. Our group recently developed a wavelet packet seizuredetection procedure [12], [13], after having developed anindex to estimate hypnotic depth in an anesthetized patient’sEEG signals using wavelet analysis, discriminating betweenconscious and anesthetized states [14].

In discrete wavelet analysis, a multi-resolution descriptionis used to decompose a given signal x(t) into increasinglyfiner details based on two sets of basis functions [9], thewavelets and the scaling functions, as follows:

x(t) = ∑k

2 j0/2c j0(k)ϕ(2 j0t− k)

+∞

∑j= j0

∑k

2 j/2d j(k)ψ(2 jt− k) (1)

where functions ϕ(t) and ψ(t) are the basic scaling andmother wavelet respectively. In the above expansion, the firstsummation presents an approximation of x(t) based on thescale index of j0, while the second term adds more detailsusing larger j (finer scales). The coefficients in this waveletexpansion are called the discrete wavelet transform (DWT)of the signal x(t). When the wavelets are orthogonal [9],these coefficients can be calculated by

c j(k) =∫

∞

−∞

2 j/2x(t)ϕ(2 jt− k)dt (2)

d j(k) =∫

∞

−∞

2 j/2x(t)ψ(2 jt− k)dt (3)

where c j(k) and d j(k) are, respectively, the scaling (ap-proximation) and wavelet (detail) coefficients. In the DWT,the frequency axis is divided into dyadic intervals towards

30th Annual International IEEE EMBS ConferenceVancouver, British Columbia, Canada, August 20-24, 2008

978-1-4244-1815-2/08/$25.00 ©2008 IEEE. 919

Fig. 1. Decomposition tree in a wavelet packet.

the lower frequencies while the bandwidth length decreasesexponentially [9]. The wavelet packet (WP) transform isa generalization of the DWT in which the decompositionprocedure is done in both directions (lower and higher fre-quencies). This general decomposition offers a greater rangeof possibilities for signal analysis than the discrete waveletdecomposition. In the WP tree, each node is recognized bythe decomposition level (scale) l with respect to the WPtree root and the frequency band f . In other words, node(l, f ), which corresponds to subspace Ωl, f , is the f th nodeat level l from the root of the WP tree, for l = 0,1, . . . ,Land f = 0,1, . . . ,2l − 1, where L = log2n (n is the signaldimensionality). Figure 1 shows a typical scheme for a WPtree with three levels of decomposition.

III. METHODS

A significant number of epileptic seizures are associatedwith rhythmic activities in the frequency range of 3 to29 Hz [8], [15]; therefore, the rhythmicity of EEG canbe considered as an indicator of ictal states in differentchannels. In this paper, we propose a novel index based onrecognition of regularity in EEG signals to detect epilepticseizures. However, the EEG includes different types of non-epileptic rhythmic patterns which may affect the performanceof such a detection approach. Accordingly, to increase thespecificity of our detection algorithm, we also consider theenergy (amplitude) of EEG. The proposed measure is acombined index sensitive to both the rhythmicity and energyof EEG signals and is defined as the product of regularityand amplitude indices which are computed as follows.

A. The Regularity Index as a Measure of Rhythmicity

To compute the regularity index, it is first necessaryto define a frequency band in which the rhythmicity ofEEG is considered as a probable sign for ictal activity.This frequency band is termed the regularity band. In amoving-window analysis, each EEG epoch is decomposedusing the WP transform to obtain the coefficient vector V0corresponding to the regularity band at the coarsest scale(last decomposition level) in the WP tree which resultsin the maximum frequency resolution. Finding the first Nlargest absolute values of V0 elements (here N = 5), theone which has the highest energy in its vicinity is chosenas the dominant coefficient in vector V0 (the energy isdefined as the summation of the corresponding coefficient

square values). The frequency band corresponding to thevicinity of the dominant coefficient is considered as F0 witha central frequency of fc. Suppose y(t) is the time-domainsignal corresponding to coefficient vector V0 (inverse WPtransform); then, the raw regularity index R0 is computed asthe maximum normalized cross-correlation between y(t) andthe reference signal x(t) by

R0 = maxτ

∣∣∣∣∣∣∫

∞

−∞x(t + τ)y(t)dt√∫

∞

−∞x2(t)dt×

∫∞

−∞y2(t)dt

∣∣∣∣∣∣ (4)

where x(t) is a pure sinusoidal waveform defined as x(t) =sin(2π fct). By definition, this index is confined to [0,1]. Themore similar these two signals are, the higher the obtainedvalue of R0 is. The raw regularity index is, then, averagedover consecutive epochs to obtain a smoother waveformtermed Rav. Since there is usually a minimum level ofrhythmicity in EEG signals, a nonlinear scaling is appliedto the averaged regularity index Rav to highlight the highregularity values. This scaling results in the regularity indexR and is performed by

R(k) = 1− exp(− R4

av(k)1−R4

av(k)

)(5)

B. The Amplitude Index as a Measure of Relative EnergyTo decrease the false detection rate of the detection

algorithm, it is necessary to ignore low amplitude activitiesin the regularity band. However, it should be also consideredthat the absolute value of the EEG amplitude may notcharacterize seizure severity. That is, seizures associated withhigher EEG amplitude may not be more severe than seizureswith low EEG amplitude. Therefore, a relative index isemployed to reduce the effect of the low-amplitude rhythmicactivities according to typical EEG amplitude of seizures,without highlighting the high-amplitude seizure activities.After determining a typical seizure amplitude termed Atyp,a pure sinusoidal waveform is defined as z(t) = Atypx(t) =Atyp sin(2π fct) for each EEG epoch. Then, applying theWP transform to this waveform, the WP coefficient withmaximum absolute value in the coarsest scale in the WP treeis found, and the energy stored in its vicinity is computed.This energy, named E2, is then used to compute the rawamplitude index A0, as shown by

A0(k) =

E1/E2, if E1/E2 < 11, if E1/E2 ≥ 1

(6)

where k is the epoch number, and E1 is the energy storedin the vicinity of the dominant coefficient in V0. Then, theamplitude index A is obtained by applying a moving-averagefilter to A0.

Finally, the combined index C, which is used in seizuredetection, is formed by (7) for the kth EEG epoch, aftercomputation of the regularity and amplitude indices.

C(k) = A(k)×R(k) (7)

The resultant index monitors both the rhythmicity and rela-tive energy of EEG epochs and shows a significant increase

920

Fig. 2. A 15-second multichannel EEG segment in bipolar montage. Dashedline shows the seizure onset.

in its value as seizures occur. Due to the definition of thecombined index, its value is confined to [0,1]; i.e., it is anormalized index. This provides an opportunity to comparedifferent seizures.

IV. RESULTS

A. Epilepsy Data

To evaluate the performance of the proposed seizure de-tection index (combined index), an EEG dataset provided bythe EEG department of Vancouver General Hospital (VGH)after ethics approval was used. This dataset included ∼11.5hours of 23-channel surface EEG recorded from 3 patientswith temporal lobe epilepsy (TLE), contained 7 seizures, andwas sampled at 256 Hz.

To apply a moving-window analysis, each EEG recordingwas segmented into two-second windows with one-secondoverlap. Since this segmentation results in epochs withthe length which is power of two, it is suitable for EEGanalysis using wavelets [10]. Figure 2 shows a 15-secondmultichannel EEG segment of a patient with TLE in bipolarmontage, in which the dashed line indicates the seizure onset.Channels F7− T3, T3− T5, SP1− T3, and T3−C3 representsignificant rhythmic activities during the ictal state.

B. Wavelet Packet Energy Ratio

Previously proposed by our group [12], [13], the waveletpacket energy ratio (WPER) is an index for detecting epilep-tic seizures based on the relative energy of the EEG in twodominant frequency bands. After computing the energy ofeach band by summing the square values of the correspond-ing WP coefficients, the WPER index is determined as theenergy of higher band divided by that of lower one. Changesin the frequency content of the EEG during the seizure causea significant increase in the WPER index, which is used tomark seizure occurrence. To investigate the performance ofthe proposed combined index, we also applied the WPER tothe epilepsy dataset in this paper.

TABLE ITHE FREQUENCY BANDS SELECTED TO ANALYZE EEG RECORDINGS

FROM 3 PATIENTS USING THE COMBINED INDEX AND WPER.

Patient Combined Index WPERRegularity Band (Hz) Dominant Bands (Hz)

1 2-6 3-6 & 1-2.52 4-10 5.5-9 & 1.5-43 4-10 5.5-9 & 1.5-4

Fig. 3. Seizure detection indices for an EEG segment from channel T3−T5of patient 3 with left TLE. (a) Average WPER. (b) Combined Index.

C. Seizure Detection Results

To analyze the EEG recordings of each patient, thefrequency bands required to compute the combined indexand the WPER, i.e., the regularity band in the combinedindex and the dominant frequency bands in WPER, weredetermined separately using the first EEG recording of thatpatient. Table I shows this selection for each of 3 patientsincluded in this study. To compute the combined index, thetypical seizure amplitude Atyp was selected as 80 µV forall patients. The length of the moving-average filters used incomputing the combined index was set to 5. Also, to smooththe profile of the WPER index, a moving-average filter withthe same length was applied to this index. Computationof both measures was done using Daubechies-12 wavelet.Figures 3 and 4 present examples of the combined indexand average WPER for two EEG segments each of which in-cludes both nonseizure and seizure intervals. In figure 3, thetemporal changes of these indices for an EEG segment fromchannel T3−T5 of patient 3 (with left TLE) are depicted, andfigure 4 shows theses measures for a part of EEG recordingrelated to channel T4−T6 of patient 1 suffering from rightTLE. According to the EEG technologist, the seizure onsetsin the first and second cases are, respectively, about 2740sand 2615s. Based on these results, WPER is much noisierthan the combined index. In addition, while there is a bigdifference between the maximum WPER amplitude valuesin these two cases (15 and 2.3 in figure 3 and figure 4,respectively), the combined index introduces a fixed ampli-

921

Fig. 4. Seizure detection indices for an EEG segment from channel T4−T6of patient 1 with right TLE. (a) Average WPER. (b) Combined Index.

TABLE IITHE RESULTS OF APPLYING THE COMBINED INDEX AND WPER TO THE

EPILEPSY DATA.

Measure Sensitivity (%) False Detection RateWPER 85.7 4.32/hr

Combined Index 100 0.52/hr

tude range as it is a normalized index. In the quantitativeevaluation of these two indices, due to the small number ofEEG recordings, fixed thresholds were used for all patients(0.35 for the combined index and 2 for the WPER). For eachindex, a channel detection was made when the index valuesurpassed the predefined threshold, and a seizure detectionalarm was produced when at least two channel detections(not in the same channel) occurred within four consecutiveepochs (5s). Successive detections were assumed as a singledetection, provided that their time difference was less than2 min. The results of this investigation are presented intable II. Perfect sensitivity was achieved using the combinedindex, while WPER detected 6 out of 7 seizures (85.7%).According to the results, the false detection rate was foundto be significantly lower for the combined index.

V. CONCLUSION

The purpose of this work was to evaluate the performanceof the combined index, as a wavelet-based measure, inreal-time epileptic seizure detection and compare it to theWPER index which was previously developed by our group.Results revealed high sensitivity for both indices. However,the false detection rate obtained by the combined indexwas significantly lower than the rate resulting from WPER.Moreover, not only is the combined index less noisy, but it isalso a normalized measure which can facilitate comparisonamong different ictal states.

Since the combined index looks for rhythmic activitiesduring the ictal period, if the early stages of seizure perioddo not contain significant rhythmic waveforms, this indexmay not be able to detect the seizure onset. In other words,

there may be a delay in epileptic seizure detection using thecombined index.

Therefore, in the future, we intend to improve the proposedindex so that it can detect epileptic seizures at earlier stages.Moreover, we will apply this index to data from a largernumber of patients, to confirm the preliminary results andevaluate the index more completely.

ACKNOWLEDGMENTS

We acknowledge the diligent work of Mr. Peter VanRienen, EEG technologist, and Mr. Larry Stevenson, biomed-ical engineer, at the Neurophysiology Lab. of VancouverGeneral Hospital for preparation of EEG data in this study.

REFERENCES

[1] H. Witte, L. D. Iasemidis, and B. Litt, “Special issue on epilepticseizure prediction,” IEEE Transactions on Biomedical Engineering,vol. 50, pp. 537–539, 2003.

[2] M. D. Alessandro, R. Esteller, G. Vachtsevanos, A. Hinson, J. Echauz,and B. Litt, “Epileptic seizure prediction using hybrid feature selectionover multiple intracranial eeg electrode contacts: A report of fourpatients,” IEEE Transactions on Biomedical Engineering, vol. 50,pp. 603–615, 2003.

[3] J. Gotman, “Automatic seizure detection: improvements and evalua-tion,” Electroencephalography and Clinical Neurophysiology, vol. 76,pp. 317–324, 1990.

[4] I. Osorio, M. G. Frei, and S. B. Wilkinson, “Real-time automated de-tection and quantitative analysis of seizures and short-term predictionof clinical onset,” Epilepsia, vol. 39, pp. 615–627, 1998.

[5] A. J. Gabor, “Seizure detection using a self-organizing neural network:validation and comparison with other detection strategies,” Electroen-cephalography and Clinical Neurophysiology, vol. 107, pp. 27–32,1998.

[6] Y. U. Khan and J. Gotman, “Wavelet based automatic seizure detec-tion in intracerebral electroencephalogram,” Clinical Neurophysiology,vol. 114, pp. 898–908, 2003.

[7] N. S. O’Neill, M. Javidan, and Z. J. Koles, “Identification of the tem-poral components of seizure onset in the scalp EEG,” The CanadianJournal of Neurological Sciences, vol. 28, pp. 245–253, 2001.

[8] M. E. Saab and J. Gotman, “A system to detect the onset of epilepticseizures in scalp EEG,” Clinical Neurophysiology, vol. 116, pp. 427–442, 2005.

[9] C. S. Burrus, R. A. Gopinath, and H. Guo, Introduction to waveletsand wavelet transforms: A primer. Prentice-Hall, Upper Saddle River,NJ, 1998.

[10] S. Mallat, A wavelet tour of signal processing. Academic Press, 1999.[11] A. S. Zandi and M. H. Moradi, “Quantitative evaluation of a wavelet-

based method in ventricular late potential detection,” Pattern Recog-nition, vol. 39, pp. 1369–1379, 2006.

[12] R. Tafreshi, G. Dumont, D. Gross, C. R. Ries, E. Puil, and B. A.MacLeod, “Seizure detection by a novel wavelet packet method,”pp. 6141–6144, in Proc. of the 28th Annual International Conferenceof the IEEE EMBS, New York City, USA, 2006.

[13] A. S. Zandi, R. Tafreshi, G. A. Dumont, C. R. Ries, B. A. MacLeod,and E. Puil, “Electroconvulsive therapy: A model for seizure detectionby a wavelet packet algorithm,” pp. 1916–1919, in Proc. of the 29thAnnual International Conference of the IEEE EMBS Cite Interna-tionale, Lyon, France, August 2007.

[14] T. Zikov, S. Bibian, G. A. Dumont, M. Huzmezan, and C. R. Ries,“Quantifying cortical activity during general anesthesia using waveletanalysis,” IEEE Transactions on Biomedical Engineering, vol. 53,pp. 617–632, 2006.

[15] J. Gotman, “Automatic recognition of epileptic seizures in theEEG,” Electroencephalography and Clinical Neurophysiology, vol. 54,pp. 530–540, 1982.

922