Download - eeg and ecg

8/12/2019 eeg and ecg

1/12

Combination of EEG and ECG for improved automaticneonatal seizure detection

Barry R. Greene a,*, Geraldine B. Boylan b, Richard B. Reilly a,c,Philip de Chazal a, Sean Connolly d

a School of Electrical, Electronic & Mechanical Engineering, University College Dublin, Irelandb Department of Paediatrics and Child Health, University College Cork, Ireland

c Cognitive Neurophysiology Laboratory, St. Vincents Hospital, Fairview, Dublin, Irelandd Department of Clinical Neurophysiology, St. Vincents University Hospital, Dublin, Ireland

Accepted 7 February 2007Available online 29 March 2007

Abstract

Objective:Neonatal seizures are the most common central nervous system disorder in newborn infants. A system that could automat-ically detect the presence of seizures in neonates would be a significant advance facilitating timely medical intervention.Methods: A novel method is proposed for the robust detection of neonatal seizures through the combination of simultaneously-recordedelectroencephalogram (EEG) and electrocardiogram (ECG). A patient-specific and a patient-independent system are considered,employing statistical classifier models.Results: Results for the signals combined are compared to results for each signal individually. For the patient-specific system, 617 of 633(97.52%) expert-labelled seizures were correctly detected with a false detection rate of 13.18%. For the patient-independent system, 516 of

633 (81.44%) expert-labelled seizures were correctly detected with a false detection rate of 28.57%.Conclusions: A novel algorithm for neonatal seizure detection is proposed. The combination of an ECG-based classifier system with anovel multi-channel EEG-based classifier system has led to improved seizure detection performance. The algorithm was evaluated using alarge data-set containing ECG and multi-channel EEG of realistic duration and quality.Significance:Analysis of simultaneously-recorded EEG and ECG represents a new approach in seizure detection research and the detec-tion performance of the proposed system is a significant improvement on previous reported results for automated neonatal seizuredetection. 2007 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.

Keywords: Neonatal seizure detection; EEG; ECG; EKG

1. Introduction

Seizures in the neonate require immediate medical atten-tion and represent a distinctive sign of central nervous sys-tem dysfunction. There is increasing evidence that neonatalseizures have an adverse effect on neurodevelopmental out-come, and predispose to cognitive, behavioural, or epilepticcomplications in later life (Levene, 2002). Neonatal seizures

occur in 6% of low birth-weight infants (Volpe, 2001) and

in approximately 2% of all newborns admitted to a neona-tal ICU (Scher et al., 1993a). Seizures in this age-group areoften subtle, difficult to diagnose and may be clinicallysilent, particularly after antiepileptic drug treatment, mak-ing diagnosis by clinical observation alone very unreliable(Boylan et al., 2002). Electroencephalography (EEG) isthe most reliable method available to detect the majorityof neonatal seizures but interpretation requires specialexpertise that is not readily available in most neonatalintensive care units least so on a 24-h basis. A systemthat could automatically detect the presence of seizures in

1388-2457/$32.00 2007 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.

doi:10.1016/j.clinph.2007.02.015

* Corresponding author. Tel.: +353 21 490 3793.E-mail address:[email protected](B.R. Greene).

www.elsevier.com/locate/clinph

Clinical Neurophysiology 118 (2007) 13481359
mailto:[email protected]:[email protected]


2/12

newborn babies would be a significant advance, facilitatingtimely medical intervention.

A number of studies have reported neonatal seizuredetection methods based on the EEG (Liu et al., 1992; Got-man et al., 1997a; Celka and Colditz, 2002; Altenburget al., 2003).Faul et al. (2005)provided a review and exper-

imental comparison of three of the most commonly citedneonatal seizure detection algorithms. None performedsufficiently to be deemed suitable for use in the neonatalintensive care unit (ICU). Karayiannis et al. (2001)reported a video-based method for distinguishing myo-clonic from focal clonic seizures and differentiating thesetypes of seizures from normal infant behaviours. However,this approach does not provide a complete solution to theproblem, as many neonatal seizures are not accompaniedby this spectrum of body movements.

The importance of autonomic changes may be underes-timated in neonatal seizure detection research. Neonatalseizures are often associated with changes in heart and res-

piration rate (Greene et al., 2006b). Significant changes inheart rate may alert the clinician to the possibility of sei-zures and instigate further investigation with EEG. Thesefindings led to the development of a neonatal seizure detec-tion system based exclusively on the electrocardiogram(ECG) (Greene et al., 2006a).

The aim of this study was to attempt to improve theneonatal seizure detection rate by combining simulta-

neously-acquired ECG and EEG data. To the best of ourknowledge this is the first method to combine the ECGwith the EEG for seizure detection.

2. Data-set

A data-set of 12 records from 10 term neonates contain-ing 633 labelled seizure events, with mean seizure durationof 4.60 min, were recorded and analysed. The records had amean duration of 12.84 h. Each record contained 712channels of EEG and one channel of simultaneously-acquired ECG. Ten records, sampled at 256 Hz, were madein the neonatal intensive care units of the Unified Mater-nity Hospitals in Cork, Ireland, using the Viasys NicOnevideo-EEG system. The remaining recording, sampled at200 Hz, was recorded at Kings College Hospital, London,on a Telefactor Beehive video-EEG system. A total of154.1 h of EEG and ECG were analyzed.

The data-set used in this research is a resource of con-

tinuously-recorded digital video-EEG data and otherphysiological parameters in newborns with seizures inthe first 3 days from birth. All newborns were full term(GA: 4042 weeks) and had hypoxic ischaemic encepha-lopathy (HIE). All the data for each recording wereincluded in the analysis regardless of record length orquality. Electrographic seizures were identified and anno-tated by an expert in neonatal EEG (GBB). Fig. 1 shows

Fig. 1. Example of a multi-channel electrographic seizure. Seizure onset and duration are marked.

B.R. Greene et al. / Clinical Neurophysiology 118 (2007) 13481359 1349


3/12

an example of an electrographic seizure from record 12,detected by both the patient-specific and patient-indepen-dent systems.

Annotations give information on the time of onset andthe duration of each electrographic seizure.Table 1detailsthe number of seizure events per record, the duration ofeach record and the mean seizure duration for each record.As the ECG and EEG signals were recorded simulta-neously these annotations can be related directly in timeto the ECG signal.

The data-set contained a wide variety of seizure dura-

tions and seizure types. While the mean seizure durationacross the data-set was 4.60 min, the mean seizure durationfor each patient ranged from 1.05 min to 11.64 min. Thedata-set contained Electrographic-only seizures as wellas Electroclinical seizures. Four records 2, 3, 10, 12 con-tained only Electrographic-only seizures. Two records 9and 11 contained only Electroclinical seizures. Theremaining recordings contained both Electrographic-onlyand Electroclinical seizures. Furthermore, the data-setcontained focal, multi-focal and generalized seizures.

3. Method

The combination of EEG and ECG for neonatal sei-zure detection was considered in the context of bothpatient-specific and patient-independent seizure detectionclassifiers. While the ideal scenario for this applicationis a patient-independent system capable of identifyingall seizures from any patient with a zero false detectionrate, a patient-specific system might also represent anadvance in neonatal ICU monitoring. The algorithmsconsidered in this study are epoch-based, so each seizureevent was rounded to the nearest epoch length whenmapping time annotations to epochs. An epoch contain-ing P50% electrographic seizure activity was labelled as

a seizure epoch.

3.1. ECG

The algorithm reported in this paper utilises the sameECG features described previously, based on the RRintervals for 60-s epochs of ECG (Greene et al., 2006a).

3.1.1. ECG pre-processingAll ECG signals were filtered with a 20th order FIRband-pass filter (corner frequencies 8 and 18 Hz) to removebaseline wander, power-line noise and out of band noise.Before filtering, the mean of the ECG was removed fromthe signal.

3.1.2. RR interval calculation

The RR interval is defined as the time in secondsbetween adjacent R-wave maximum (QRS) points. Robustdetection of the QRS point is determined using a QRSdetection algorithm as described byBenitez et al. (2001).

The Hilbert transform of the first derivative of the signal

was used to emphasize the R peaks. A moving windowpeak search was carried out with an adaptive threshold.As neonatal ECG often manifests elevated P-wave, a stepback search was performed to isolate the P peak ensuringrobust detection of the R-wave maximum. Correction formissing and extra QRS points was implemented asdescribed byde Chazal et al. (2003).

3.1.3. ECG feature extraction

The six ECG feature types considered in this study werecalculated on a 60-s (15,360 samples for a record sampledat 256 Hz, 12,000 samples for 200 Hz) non-overlapping

epoch basis. Features are based on the RR intervals asso-ciated with each 60-s epoch. The features used in this study:

Mean RR interval (lRR) Std. Dev. RR intervals (rRR) Mean RR interval spectral entropy (RR H) Mean change in the RR interval (DRR) RR interval coefficient of variation (dRR) RR interval power spectral density (RR PSD)

The mean RR interval, RR interval standard devia-tion, RR interval coefficient of variation, mean changein adjacent RR intervals per epoch (DRR) as well as RR interval spectral entropy each contributed one featureto the ECG feature vector. Relative features for lRR,rRR, DRR and dRR were obtained by subtracting themean of the feature for the four preceding epochs as wellas the mean for the four subsequent epochs (called lRR 0,rRR 0,DRR 0 anddRR 0 here) and each contributed one fea-ture. The RR PSD features were calculated on an intervalbasis (Teich et al., 2000). The mean of the RR intervalsfor each epoch was subtracted to yield a zero meansequence. The sequence was then zero-padded to length256 and the fast Fourier transform (FFT) taken. Theresulting sequence was multiplied by its complex conjugate

to yield a periodogram estimate of the RR interval power

Table 1Data characteristics for each record: number of seizure events, duration ofrecording, mean seizure duration

Record No. of seizure events

Recordduration (h)

Mean seizureduration (min)

1 90 10.01 2.772 22 10.42 7.33

3 21 24.53 5.414 60 14.25 1.565 35 14.40 10.026 29 10.01 2.157 155 24.04 5.288 56 13.17 1.999 60 5.20 1.05

10 41 5.69 1.1611 50 17.33 4.8812 14 5.05 11.64

Total Total Mean633 154.10 4.60

1350 B.R. Greene et al. / Clinical Neurophysiology 118 (2007) 13481359


4/12

spectral density. A 64-point periodogram was obtained byaveraging the values in four adjacent frequency bins. Onlythe first 33 of these constituted a valid PSD, with the first32 of these points taken as a feature in its own right foreach 60-s epoch. In total each epoch produced 41 featuresfor each ECG feature vector.

3.1.4. ECG artefact detection

There are a variety of artefacts that may be found in anECG signal. In this paper we have attempted to reject twokinds of artefact from subsequent analysis, namely: move-ment artefacts which are large signal spikes caused bymovement of the electrodes, and zero-signal artefactcaused by the amplifier being powered-off in the courseof a recording.

A new signal, Qecg, was constructed with the same samplerate as the ECG and provided a binary flag for the presenceor absence of artefact for each sample of the ECG signal.

To identify the artefact sections of the ECG a zero mean

ECG signal was first calculated by subtracting the mean ofthe ECG from each sample and then processing this signalas follows:

The standard deviation of the absolute value of signal wascalculated and any signal samples greater than six timesthe standard deviation were flagged as movement arte-fact andQecgwas assigned the value 1 at these samples.

Any 10 sample epoch whose mean was 100 times smallerthan the 5% trimmed mean of the signal was flaggedzero-signal artefact and each sample Qecgwas assignedthe value 1.

Unflagged samples ofQecg were assigned the value 0.

This artefact measure was then associated with each 60-sepoch with the mean value for each epoch assigned as theQecg value for that epoch. Fig. 2 shows the operation ofthe ECG artefact detector on section of the ECG recordingfor Record 1.

3.2. EEG

3.2.1. EEG pre-processing

The EEG for each channel was low-pass filtered using atype II Chebyshev IIR filter with a corner frequency of34 Hz to remove power-line noise along with out-of-band

noise.

3.2.2. EEG feature extraction

A set of EEG features and a novel multi-channel EEGclassifier architecture (as shown in Fig. 3) were used. Inorder to determine the optimum method for combininginformation across EEG channels the authors carried outa separate study comparing the score level or Late Integra-tion of EEG channels (i.e. each channel processed inde-pendently and the scores combined) with feature level orEarly integration of multi-channel neonatal EEG (as dis-cussed in this paper). We found that early integration pro-vided greatly superior results to late integration for this

application. Results from this study have recently beenpublished (Greene et al., 2006c). The fundamental differ-ence between our use of multiple EEG channels to otherprevious methods is that our method exploits the statisticalinter-relationships between EEG channels. ProcessingEEG channels independently assumes an equal weightingfor each EEG channel in the decision function and is lessequipped to handle redundant features from not-involvedEEG channels. This is a weak assumption when one con-siders neonatal seizure EEG which is often multi-focaland migrates across EEG channels.

Feature vectors containing m features from n channels

were concatenated into a large feature vector which wasthen fed to a pattern classifier. Six features were extractedfrom each 2048 sample non-overlapping EEG epoch foreach channel. The six features calculated per EEG epochwere:

Dominant spectral peak (F) Power ratio (P) Bandwidth of dominant spectral peak (BW) Nonlinear energy (N) Spectral entropy (H) Line length (L)

The features for each channel were sorted accordingto feature type and then each group of features sortedinto numerical order. All the grouped-sorted featureswere then concatenated into a super feature vector.The sorting function removes information about the spa-tial location of the seizure from the training set, prevent-ing the classifier from expecting seizure activity in aparticular channel. The sorting function behaves as anumerical feature selector for the patient-independentclassifier, using the numerical differences between featurevalues of channels involved in a seizure and channelsnot-involved. Features of involved and not-involved

channels will be placed at opposing ends of the sorted,

354 356 358 360 362 364 366

-1

-0.5

0

0.5

1

1.5

2

Time (s)

Amplitude(uV)

ECG Art efact Detector

ECG

Arte fact Detec torArtefact

Fig. 2. Example of the operation of the ECG artefact detector on a

movement artefact for record 1.



5/12

concatenated feature vector. A classifier can then learn todistinguish the features of involved channels from thoseof not-involved channels using their rank in the sortedconcatenated feature vector.

The dominant spectral peak, power ratio and band-width features employed are those reported by Gotmanet al. (1997b). The frequency spectrum was calculatedfor each epoch using the FFT. The dominant frequency

was defined to be the frequency in the spectrum with thelargest average power in its bandwidth. The bandwidthof the dominant spectral peak was defined as the widthin Hertz (Hz) between the two half power points ofthe dominant spectral peak. The power ratio was definedas the ratio of the power in the dominant spectral peakto the power at the same frequency in the backgroundEEG, where the background EEG is the average of thethree epochs 60 s behind the current epoch (Gotmanet al., 1997b).

Recent evidence suggests that seizure activity representsa reduction in the complexity of the signal (Celka and Col-

ditz, 2002). Spectral entropy can be interpreted as a mea-sure of signal complexity and so represents a potentialfeature for seizure detection. DAlessandro et al. (2003)employed the mean nonlinear energy of an EEG epoch inpredicting epileptic seizure in adults. Esteller et al. (2001)proposed line length, an approximate measure of the frac-tal properties of the signal, as a potential feature for epilep-tic seizure onset detection.

3.2.3. EEG artefact rejection

The stability of an EEG epoch has been used as anEEG signal quality measure Qeeg (Gotman et al., 1997b).

The larger this value is relative to unity, the more likely

it is to contain artefact. The mean Qeeg across n EEGchannels was taken as the EEG signal quality measure,Qeeg for each epoch, where Qeeg can be written as:Qeeg

1n

Pni1Qeegi .

3.3. Classifier model

Classification based on a linear discriminant (LD) clas-sifier model was employed for all signal modes and config-urations. Linear discriminant classifier models utilise classconditional mean vectors and a common covariancematrix. They provide optimal performance when featureswithin a class have a normal distribution and the same var-iance across classes. The class conditional mean vectorsand a common covariance matrix were estimated sepa-rately from the training data for the patient-specific andpatient-independent classifiers. Weighting of the class con-ditional mean vectors and common covariance matrix bythe duration of each record was implemented for thepatient-independent classifier (Greene et al., 2006a). This

ensures that records of differing lengths contribute equallyto the training of a patient-independent classifier.

3.4. Combining the ECG and EEG information

Two schemes for combining the information determinedfrom the EEG and ECG signals were considered- the earlyintegration (EI) scheme and the late integration (LI)scheme, and are discussed separately below.

3.4.1. Early integration of EEG and ECG features

The EI configuration, hereafter referred to as the EI

fusion configuration, involves concatenating the EEG

Feature

Extraction

mx

x

x

1

12

11

ns

ns

ns

s

s

s

s

s

m

m

x

x

x

x

x

x

x

x

2

1

2

1

2

1

2

2

2

1

1

mx

x

x

2

22

21

nm

n

n

x

x

x

2

1

ClassifierFeature

Extraction

Feature

Extraction

Channel 1

Channel 2

Channel n

nx

1x

2x Decision

sP

Feature

Sort

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Fig. 3. EEG classifier configuration.



6/12

and ECG feature vectors into a single feature vector andfeeding this super feature vector to a pattern classifier.Fig. 4gives a graphical description of the EI configuration.If the signal quality measures Qecg and Qeeg for an epochwere both over an empirically-derived threshold, thatepoch is considered to contain artefact and the epoch is

neglected from analysis for both signals.

3.4.2. Late integration of EEG and ECG classifications

The late integration (LI) configuration, hereafterreferred to as the LI fusion configuration, employs separateclassifiers for each signal to determine a probability of sei-zure for each signal mode. These two probabilities are thencombined to provide an overall probability of seizure.Fig. 5gives a graphical description of the LI fusion config-uration. The combination of the EEG and ECG signalmodes at using the classifiers confidence score, as is per-formed in this configuration, allows each signal to beweighted for improved classification performance.

Static and dynamic weighting of the two signals wasinvestigated. An expression for the overall probability ofseizure is given in Eq.(1). The scalarsasand bsare the sta-tic weights for the ECG and EEG signals, respectively. Forthe static weighting case,as= 1 bs. Ifa is varied over therange 01 the optimum static weights for each signal can bedetermined for both the patient-specific classifier and thepatient-independent classifier.Fig. 6shows the mean clas-sification accuracy for the patient-specific classifier for thecombined classifiers as the EEG static weight, as is variedfrom 0 to 1 in increments of 0.1.

Psz asPecg Peeg 1

Dynamic weighting takes account of a measure of qual-ity in each signal (shown as Q in Fig. 5). If an epoch isdetermined to contain artefact in either mode the corre-sponding weight for that mode is reduced appropriatelycausing the system to favour the decision for the other sig-nal. The dynamic weight for each signal is calculated by

subtracting the quality measure, scaled by dividing by themaximum value of that quality measure, from the staticweight for each signal (It should be noted that a real-timesystem would require an empirically determined value inplace of the maximum value of the quality measure usedhere, such a value should be chosen to be the largest valuethat may reasonably occur for this parameter and shouldnot significantly differ from the Qmax value used here).Expressions for the dynamic weights for both signals aregiven in Eqs.(2) and (3).

adasQecg

maxfQecgg 2

bd bsQeeg

maxfQeegg 3

Consequently, the probability of seizure Psz with dynamicweighting can be determined from Eq. (1) using ad andbd as the mode weights in place ofas and bs.

An epoch is labelled as seizure ifPszis over a given deci-sion threshold. A decision threshold equal to 0.5 is used inthis study.

3.4.3. Interpolation of mode frame rates

The ECG was considered in non-overlapping epochs of60 s (16,384 samples at 256 Hz). However the EEG was

Classifier

Sort

Nx6

42

EEG

ECG

Feature

Extraction

Q

Q

ns

ns

ns

s

s

s

s

s

x

xx

x

x

x

x

x

2

2

2

1

1

zy

y

y

2

1z

ns

s

s

s

y

y

y

x

x

x

x

2

1

2

2

1

Interpolate

zy

y

y

2

1

N

Concatenate

Feature

Extraction

sP

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Fig. 4. Early integration (EI) neonatal seizure detection configuration.



7/12

considered in 2048 sample non-overlapping epochs. Inorder to facilitate direct comparison and fusion of thetwo signals the ECG frame rate must be matched to theEEG frame rate by means of interpolation. The EEG framerate is a multiple of the ECG frame rate. The interpolationfactor is the integer closest to this multiple, the frames arethen shifted to ensure that the EEG and ECG windowsremain synchronized. In the EI configuration thisinterpolation was performed at a feature level. A superfeature vector with a frame rate matching that of theEEG was passed to the classifier. In the LI configuration,this interpolation was performed at the score level.The output probability from the ECG classifier was

interpolated after sub-dividing the output for each 60 sECG epoch into eight 8-s epochs (for the 256 Hz case) tomatch the frame rate from the EEG classifier and thetwo combined, as discussed in Section3.4.2, for an overallprobability of seizure.

3.5. Classifier performance estimation

Each classifier configuration was considered as both apatient-specific and a patient-independent classifier. Theperformance of each patient-specific classifier was esti-mated using m fold cross-validation on each record. Thisinvolves randomly splitting each record into msections orfolds: m1 of these folds are then used to train theclassifier and the remaining fold is then used to test theperformance of the classifier. By shuffling the data andrepeating this procedureq times and averaging the result-ing accuracies for the training and test sets an unbiased,low variance estimate of the classifier performance canbe obtained. In this study 10 folds and 10 shuffles ofthe data were used. The patient-specific performance mea-sures were then taken as the average of each measureacross records.

The performance of the generalized or patient-indepen-dent classifier was estimated using cross-validation acrossall records. This involved training the classifier model on(z1) of the z records and using the zth record to testthe classifier performance and then rotating through the zpossible combinations of training and test sets. The meanof the results for all iterations is taken as the patient-inde-pendent performance estimate. This test provides ameasure of the classifiers ability to generalize from the

training set and classify from unseen records.

Sort

Nx6

42

Feature

Extraction

Feature

Extraction

Q

Q

Classifier

Decision

ECG

ns

ns

ns

s

s

s

s

s

x

x

x

x

x

x

x

x

2

2

2

1

1

sP

zy

y

y

2

1

Interpolate

zy

y

y

2

1

ClassifierEEG

eegP

ecgP

N

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

Fig. 5. Late integration (LI) neonatal seizure detection configuration.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 160

65

70

75

80

85

90Fusion weights vs mean Fusion Accuracy

EEG fusion weight

Accuracy(%)

mean Fusion accuracy

mean EEG accuracy

mean ECG accuracy

Fig. 6. Patient-specific classification accuracy for LI EEG fusion classifieras the EEG static weight is varied in the range 01 (ECG weight equals

1-EEG weight).



8/12

3.6. Classifier performance measures

The classifier architectures considered in this study wereepoch-based. As a result, classification accuracy is anepoch-based measure. In contrast, the percentage of sei-zures detected by the system is an event-based measure.

There is much inconsistency in the literature regardingthe format of reported results. For this reason, we havepresented our results in both epoch format and event for-mat. The classification accuracy is defined as the percent-age of epochs correctly classified by the system. Thesensitivity is defined as the percentage of seizure epochs(as labelled by an expert in neonatal EEG) correctly iden-tified as seizure epochs by the system. The specificity isdefined as the percentage of labelled non-seizure epochscorrectly classified as non-seizure by the system. The falsedetection rate (FDR) is defined as the percentage of non-seizure epochs incorrectly identified as seizure epochs andis equivalent to 100-specificity (%). Caution must be exer-

cised in reporting false detection results. Many algorithmsreport these in terms of clusters of false detections perhour (Gotman et al., 1997b). Although this can be a usefulmeasure of clinical utility of the algorithm, it is not alwaysan accurate assessment of algorithm performance as it ispossible for an entire hour of false detections to be takenas a single false detection for that hour. False detectionsless than 30 s apart are grouped as a single false detection.The mean false detection per hour (FD/h) is included herefor completeness.

The seizure sensitivity or good detection rate (GDR) isdefined as the percentage of electrographic seizure events

as labelled by an expert in neonatal EEG (G.B.B.) correctlyidentified by the system. If a seizure was detected any timebetween the start and end of a labelled seizure this was con-sidered a good detection.

A receiver operating characteristic (ROC) curve is agraphical representation of class sensitivity against specific-ity as a threshold parameter is varied. The area under theROC curve (calculated using trapezoidal numerical inte-gration) is an effective way of comparing the performanceof different features or classifiers and is equivalent to theMannWhitney version of the Wilcoxon rank-sum statistic(Zweig and Campbell, 1993). A random discrimination willgive an area of 0.5 under the curve while perfect discrimi-nation between classes will give unity area under theROC curve.

4. Results

The presented system is an epoch-based system so itmakes intuitive sense to quantify its performance in termsof epoch-based measures such as accuracy, sensitivity andspecificity as detailed above. From a clinical viewpoint,the most important measure of the clinical utility of a sei-zure detection system is the percentage of seizure eventscorrectly detected by the system (GDR) along with the

number of false detections. For this reason we have given

our results in terms of both epoch measures and event-based measures.

The results are divided into two sections: the patient-specific classifier results appear in Section 4.1 and thepatient-independent classifier results appear in Section4.2. Within each section results are presented for each sig-

nal individually as well as the results from the combinationof the EEG and ECG signals. It should be noted thatalthough artefact rejection was employed in both Fusionconfigurations, each of the 633 seizure events in the data-set was included in our analysis.

4.1. Patient-specific results

The mean patient-specific ECG GDR was found to be99.36% with an FDR of 29.80%. On an epoch basis theECG classifier had a mean classification accuracy of69.09% with associated sensitivity and specificity of

60.06% and 69.48%, respectively. The patient-specificEEG classifier had a GDR of 93.64% and an FDR of11.47%. The EEG classification accuracy was 84.55% withsensitivity 71.02% and specificity 89.23%.

The EI fusion patient-specific classifier had a meanGDR of 95.82% with an FDR of 11.23%. The EI fusionclassification accuracy was 86.32% with sensitivity of76.37% and specificity of 88.77%.

When static weighting of modes was employed, the LIfusion patient-specific classifier had a GDR of 95.18%and an FDR of 10.77% when as= 0.7 and bs= 0.3. Theaccuracy was 85.99%, sensitivity 73.69% and specificity89.23%.Fig. 6shows the classification accuracy asasis var-

ied from 0 to 1.With dynamic weighting of modes the LI patient-specific

GDR was 97.52% with an FDR of 13.18%. The mean clas-sification accuracy was 84.66% with sensitivity 74.08% andspecificity 86.82%. Table 2 gives a breakdown for thepatient-specific results for all modes and configurations.The results given for the LI fusion configuration are thosefor dynamic weighting of modes as this method gave supe-rior performance to static weighting.

4.2. Patient-independent results

The mean patient-independent ECG GDR was 63.54%with an FDR of 38.00%. The mean classification accuracy

Table 2Patient-specific results

ECG EEG Fusion

EI LI

GDR (%) 99.19 93.64 95.82 97.52FD/h 0.68 5.52 5.63 3.96FDR (%) 30.79 11.47 11.23 13.18Accuracy (%) 68.98 84.55 86.32 84.66Sensitivity (%) 59.69 71.02 76.37 74.08Specificity (%) 69.21 88.53 88.77 86.82

Results for LI fusion with dynamic weighting for each signal.



9/12

was 63.54% with associated sensitivity and specificity of69.63% and 61.63%, respectively.

The mean patient-independent EEG GDR was 80.41%with an FDR of 28.57%. The mean classification accuracywas 71.51% with associated sensitivity and specificity of68.18% and 73.95%.

The EI fusion patient-independent GDR was 81.44%with an FDR of 28.57%. The mean classification accuracywas 71.51% with sensitivity of 71.73% and specificity of71.43%.

With static weighting of modes the LI fusion patient-independent classifier had a GDR of 80.41% and anFDR of 27.87% when the mode weights were as= 0.7and bs= 0.3. With dynamic weighting the GDR for thisclassifier was 81.27% with an FDR of 33.05%.Table 3out-lines the patient-independent results. The results given forthe LI fusion configuration are those for dynamic weight-ing of modes as this method gave superior performance.

These results were confirmed by ROC analysis. Fig. 6shows the ROC curves for the ECG, EEG and LI Fusionclassifiers. The ECG ROC area was 0.68 while the EEGROC area was 0.76. The LI fusion ROC area was 0.76while the EI fusion ROC area was 0.77 (Fig. 7).

Fig. 8a gives two examples of a good detection of a

seizure event, showing the system output in terms of theprobability of seizure Psz. Fig. 8b also shows an exampleof a false detection. In both cases an epoch was classifiedas seizure ifPsz was greater than or equal to the decisionthreshold.

5. Discussion

An approach is proposed for combining simultaneously-recorded ECG and EEG signals for more accurate androbust detection of neonatal seizures.

Recent research by the authors has suggested that the

ECG is suitable in its own right for use in seizure detection

Table 3Patient-independent results

ECG EEG Fusion

EI LI

GDR (%) 82.33 80.41 81.44 81.27FD/h 1.71 3.42 3.15 3.05

FDR (%) 37.78 26.05 28.57 33.05Accuracy (%) 63.97 72.45 71.51 68.89Sensitivity (%) 69.51 68.18 71.73 74.39Specificity (%) 62.22 73.95 71.43 66.95

Results for the LI fusion classifier employed dynamic weighting of eachsignal.

0 20 40 60 80 100

0

10

20

30

40

50

60

70

80

90

100

Sensitivity[%]

Specificity [%]

Patient Independent ROC curve

Fusion EI ROC

ECG ROC

Fusion LI ROC

EEG ROC

Fig. 7. Patient-independent LI Fusion ROC curves with ECG and EEG

ROC curves.

200 400 600 800 1000

0.45

0.5

0.55

0.6

0.65

0.7

Epoch No.

ProbabilityofSeizure

Patient Independent Classifier: Output Probability

0 100 200 300 400 500 600 700150

100

50

0

50

Time (s)

Amplitude(uV)

EEG Seizure

Seizure Onset

SeizureSeizure

Label

Probability of Seizure

Decision Threshold

EEG

50 100 150 200 250 300 3500.2

0.4

0.6

0.8

1

Epoch

Proba

bilityofseizure

370 372 374 376 378 380 382 384 386 388 390

200

100

0

100

200

Time (s)

Amplitude(uV)

Patient Independent Classifier: Output Probability

EEG Seizure

Good Detection

False Detections

Probability of Seizure

Label

Decision threshold = 0.5

a

b

Fig. 8. (a) An example of a good detection for the patient-independentclassifier, for a seizure in record 5. The top panel shows an EEG channelwith seizure onset marked with a black arrow. Bottom panel shows theprobability of seizure generated by the system for that seizure. (b) Anexample of a good detection and false detection for the patient-independent classifier, for a seizure in record 6. The top panel shows theEEG seizure event denoted by the dashed black line in the bottom panel.



10/12

algorithms, due to the fact that RR interval timing, com-plexity and variability changes appear to be associated withneonatal seizures (Greene et al., 2006b). There are a num-ber of existing methods for EEG-based neonatal seizuredetection. Many are either based on a single channel ofEEG (Gotman et al., 1997b; Celka and Colditz, 2002; Has-

sanpour et al., 2004) or use empirically-based decisionthresholds (Altenburg et al., 2003; Liu et al., 1992) asopposed to a classifier model, trained on real multi-channelEEG. The novel EEG-based classifier architecture reportedhere exploits the statistical inter-relationships and synchro-nously recorded nature of the EEG by processing all avail-able EEG channels, while employing a statistical classifiermodel.

Manifestations of seizure were observed simultaneouslyin the EEG and ECG signals. The combination of the twosignals supplies the neonatal seizure detection system witha broader seizure-specific information base, offering poten-tially superior seizure detection performance. The ECG

and multi-channel EEG data used to evaluate the algo-rithms developed in this study are of the same durationand quality as that found in the neonatal ICU and socan be said to faithfully reflect the performance of the algo-rithms under real-world conditions. Many previous studieshave selected small numbers of seizure and non-seizureEEG epochs instead of long-duration EEG recordings toevaluate their algorithms.

Scher and co-workers have reported that the neonatalsleep cycle is approximately 1 h in duration (Scher et al.,1993b). Results for algorithms trained and validated onsmall non-continuous tracts of EEG shorter than 1 h will

not reflect the performance of such algorithms on theunique stage-specific characteristics of the neonatal EEGsleep cycle. A similar observation can be made of algo-rithms that are defined and validated using a single channelof EEG. Performance figures given for such algorithms donot reflect the performance of such algorithms on realmulti-channel EEG data. A robust system must be ableto cope with all EEG records regardless of record qualityand duration.

Gotman et al. (1997b)reported a GDR of 71% and Liuet al. (1992)reported a GDR of 84% for their patient-inde-pendent neonatal seizure detection methods (Liu et al.,1992), with 1.7 false detections per hour and a false detec-tion rate of 1.7%, respectively. Celka and Colditz reporteda GDR of 93% with an FDR of 4% for their patient-spe-cific neonatal seizure detection method (Celka and Colditz,2002). An independent evaluation of these three methods,performed on the same data-set as is used here, found thatthe results reported in the source papers overestimated theperformance of these algorithms and found none were suit-able for use in a clinical environment (Faul et al., 2005).Our patient-independent results for EEG alone and forLI and EI combined ECG and EEG are an improvementon those reported by the evaluation of Faul et al. Theresults for the Gotman method were validated by a subse-

quent paper by results for a separate data-set containing

281 h of EEG data from 54 patients in three centres ( Got-man et al., 1997a). The mean seizure detection rate for thisset was 69% with a mean of 2.3 false detections per hour.The size of this data-set must lend credence to these results.Our patient-independent results for EEG and ECG com-bined were an improvement on those reported by Gotman,

and achieved using a methodology to ensure robust repro-ducible results. The data-set used by Celka and Colditzcontained 4 neonates and does not detail the number of sei-zures or the duration of the recordings used (Celka andColditz, 2002). Furthermore the results are based on a sin-gle channel of EEG. The data-set of Liu et al. used 58 30-sseizure epochs, selected for prototypicality, this may havehad a biasing effect on their results as noted by Gotmanet al.

It has been noted that different classifier models offerpotentially complementary information about the patternsto be classified, which could be harnessed to improve theperformance of the selected classifier (Kittler et al., 1998).

As a result it has been found that combining classifiersfrom different modes with generalized knowledge of thepatterns to be classified generally yields improved, morerobust, classification performance. Our results confirm thisfinding. The combined EEG and ECG classifiers out-per-formed both the ECG and EEG classifiers individual per-formances. While the GDR or FDR for an individualsignal may have been comparable to that for the early inte-gration (EI) or late integration (LI) fusion classifiers, whentaken together, results for fusion were always superior tothose for each signal individually. As a result, the combina-tion of EEG and ECG has led to a more robust system for

neonatal seizure detection than a system based exclusivelyon the EEG.

Two methods for combining ECG and EEG were con-sidered in this study. The EI fusion configuration was gen-erally found to give better performance than the LI fusionconfiguration. The one exception to this trend gave ahigher GDR to the patient-specific LI fusion classifier thanthe EI classifier. A LI configuration would possess a dis-tinct advantage over an EI configuration in a real-worldpatient-monitoring scenario. The use of dynamic weightingallows the system to deal with the presence of artefact, elec-trode drop-off or interference. Such a configuration couldalso take into account local variations in feature character-istics allowing weighting of each signal in the decisionfunction.

Patient-specific neonatal seizure detection may have util-ity in the modern neonatal ICU. When an electroencepha-lographer is alerted to the presence of electrographicseizure, relevant sections and channels of the precedingEEG could then be labelled as seizure. These annotated sei-zures could then be used to adaptively train a base patient-independent classifier towards the individual patients elec-troclinical seizure characteristics. A patient-specific systemas discussed in this paper, while an improvement on cur-rent systems falls short of the ideal neonatal seizure detec-

tion system. This is due to the fact that it would require



11/12

significant manual intervention from an electroencephalog-rapher, as outlined above, to ensure robust operation.However, such a system may prove to be more clinicallyuseful than a patient-independent system which we havefound to have potentially lower false detection rates.

Monitors such as the CFM are often used in the neona-

tal intensive care unit despite the fact that they were origi-nally designed for adult intensive care use. They arecurrently used in both term and preterm neonates forseizure detection, prognosis, and to assess the severity ofencephalopathy in trials of therapeutic hypothermia.

The cerebral function monitor (CFM) produces a onechannel amplitude integrated EEG signal. Despite attemptsto develop more sophisticated cerebral function monitorssuch as the compressed spectral array (CSA) system, it isthe CFM which remains dominant in the NICU today.There have been criticisms of the CFM because of its lim-itation to a single EEG channel (plus a simultaneous arte-fact detection channel) and the lack of detailed

information compared with the conventional multi-channelEEG, especially when used for detection of neonatal sei-zure discharges (Eaton et al., 1994; Klebermass et al.,2001; Toet et al., 2002; Rennie et al., 2004). In the studyby Rennie et al. up to 50% of seizures were missed partic-ularly those that were of short duration, focal or of lowamplitude. Therefore we would have to say that while thesedevices are in use in the NICU they have serious limitationsand therefore the need for automated seizure detectionfrom multi-channel EEG is even greater.

The signal framework introduced here raises the possi-bility of multi-channel, multi-signal intelligent neonatal

monitoring by taking account of, and combining, all avail-able physiological parameters, for monitoring the state andwellbeing of newborns in the ICU. This framework couldbe further extended to all clinical patient-monitoringsituations.

It should be noted that the clinical utility of our patient-independent system is limited by the relatively high falsedetection rates reported for the patient-independent classi-fier in this paper. In future research we hope to reduce falsedetection rates for both patient-specific and patient-inde-pendent classifiers through the use of more advanced arte-fact detection and rejection algorithms. A moresophisticated normalisation scheme may lead to improvedpatient-independent performance. Further increases in sys-tem performance might be achieved by taking account ofother recorded physiological signals such as the electrooc-ulogram and cerebral blood flow velocity. However, inorder for these signals to be included in multi-signal neona-tal seizure detection systems, the spatial and temporal rela-tion of these seizures to the electrographic seizure must firstbe quantified. The inclusion of non-signal information,such as gestational age, weight, maternal history, etc., intoan automatic neonatal monitoring system has the potentialto greatly improve the performance of neonatal seizuredetection systems due to the highly variable age-dependent

characteristics of the neonatal period.

6. Conclusion

We describe a novel algorithm for neonatal seizuredetection. Combination of an ECG-based classifier systemwith a novel multi-channel EEG-based classifier system ledto improved seizure detection performance. The algorithm

was evaluated using a large data-set containing ECG andmulti-channel EEG of realistic duration and quality.Future work is needed to develop improvements in thesealgorithms and to explore the possible added diagnosticvalue of other combinations of physiological data in theautomatic identification of seizures in this age-group.

Acknowledgements

This project was funded by an Irish Higher Educationauthority grant (HEA 9300) and an interdisciplinarygrant from the Health Research Board of Ireland. The

authors would like to acknowledge the helpful technicalassistance of Dr. Edmund Lalor and Mr. Brian OMullane.We also acknowledge the help and support of the nursingand medical staff of the Unified Maternity Services, Cork,and the parents and families of the babies involved in thisstudy.

References

Altenburg J, Vermeulen RJ, Strijers RLM, Fetter WPF, Stam CJ. Seizuredetection in the neonatal EEG with synchronization likelihood. ClinNeurophysiol 2003;114:505.

Benitez D, Gaydecki PA, Zaidi A, Fitzpatrick AP. The use of the Hilbert

transform in ECG signal analysis. Comput Biol Med 2001;31:399406.Boylan GB, Rennie JM, Pressler RM, Wilson G, Morton M, Binnie CD.

Phenobarbitone, neonatal seizures, and video-EEG. Arch Dis ChildFetal Neonatal Ed 2002;86:16570.

Celka P, Colditz P. A computer-aided detection of EEG seizures ininfants: a singular-spectrum approach and performance comparison.IEEE Trans Biomed Eng 2002;49:45562.

DAlessandro M, Esteller R, Vachtsevanos G, Hinson A, Echauz J, Litt B.Epileptic seizure prediction using hybrid feature selection over multipleintracranial EEG electrode contacts: a report of four patients. IEEETrans Biomed Eng 2003;50:60315.

de Chazal P, Heneghan C, Sheridan E, Reilly R, Nolan P, OMalley M.Automated processing of the single-lead electrocardiogram for thedetection of obstructive sleep apnoea. IEEE Trans Biomed Eng2003;50:68696.

Eaton DM, Toet M, Livingston J, Smith I, Levene M. Evaluation of theCerebro Trac 2500 for monitoring of cerebral function in the neonatalintensive care. Neuropediatrics 1994;25:1228.

Esteller R, Echauz J, Tcheng T, Litt B, Pless B. Line length: an efficientfeature for seizure onset detection. In: Proceedings of the 23rd annualinternational conference of the IEEE engineering in medicine andbiology society, 2001, 2; 2001, p. 170710, vol. 2.

Faul S, Boylan G, Connolly S, Marnane L, Lightbody G. An evaluationof automated neonatal seizure detection methods. Clin Neurophysiol2005;116:153341.

Gotman J, Flanagan D, Rosenblatt B, Bye A, Mizrahi EM. Evaluation ofan automatic seizure detection method for the newborn EEG.Electroencephalography Clin Neurophysiol 1997a;103:3639.

Gotman J, Flanagan D, Zhang J, Rosenblatt B, Bye A, Mizrahi EM.Automatic seizure detection in newborns: methods and initial evalu-

ation. Electroencephalography Clin Neurophysiol 1997b;103:35662.



12/12

Greene BR, de Chazal P, Boylan GB, Reilly RB, Connolly S. Electro-cardiogram based Neonatal Seizure detection. IEEE Trans BiomedEng, TBME-00350-2005.R2, in press.

Greene BR, deChazal P, Boylan GB, Reilly RB, OBrien C, Connolly S.Heart and respiration rate changes in the neonate during electroen-cephalographic seizure. Med Biol Eng Comput 2006b;44:2734.

Greene BR, Reilly RB, Boylan G, de Chazal, P, Connolly S. Multi-channel EEG based neonatal seizure detection. In: 28th internationalconferences of the IEEE-EMBS conference; 2006c.

Hassanpour H, Mesbah M, Boashash B. Time-frequency feature extrac-tion of newborn EEG seizure using SVD-based techniques. EURASIPJ Appl Signal Process 2004;16:254454.

Karayiannis NB, Srinivasan S, Bhattacharya R, Wise MS, Frost Jr JD,Mizrahi EM. Extraction of motion strength and motor activity signalsfrom video recordings of neonatal seizures. IEEE Trans Med Imag2001;20:96580.

Kittler J, Hatef M, Duin RPW, Matas J. On combining classifiers. IEEETrans Patt Anal Mach Intell 1998;20:22639.

Klebermass K, Kuhle S, Kohlhauser-Vollmuth C, Pollak A, Weninger M.Evaluation of the cerebral function monitor as a tool for neurophysi-ological surveillance in neonatal intensive care patients. Childs NervSyst 2001;17:54450.

Levene M. The clinical conundrum of neonatal seizures. Arch Dis Child2002;86:757.

Liu A, Hahn JS, Heldt GP, Coen RW. Detection of neonatal seizuresthrough computerized EEG analysis. Electroencephalography CliniNeurophysiol 1992;82:307.

Rennie JM, Chorley G, Boylan GB, Pressler R, Nguyen Y, Hooper R.Non-expert use of the cerebral function monitor for neonatal seizuredetection. Arch Dis Child Fetal Neonatal Ed 2004;89:F3740.

Scher MS, Aso K, Beggarly ME, Hamid MY, Steppe DA, Painter MJ.Electrographic seizures in preterm and full-term neonates: clinicalcorrelates, associated brain lesions, and risk for neurologic sequelae.Pediatrics 1993a;91:12834.

Scher MS, Hamid MY, Steppe DA, Beggarly ME, Painter MJ. Ictal andinterictal electrographic seizure durations in preterm and termneonates. Epilepsia 1993b;34:2848.

Teich MC, Lowen SB, Jost BM, Vibe-Rheymer K, Heneghan C. In: AkayM, editor. Nonlinear biomedical signal processing, vol. II. Piscataway(NJ): IEEE Press; 2000.

Toet MC, van der Meij W, de Vries LS, Uiterwaal CSPM, van HuffelenKC. Comparison between simultaneously recorded amplitude inte-grated electroencephalogram (cerebral function monitor) and standardelectroencephalogram in neonates. Pediatrics 2002;109:7729.

Volpe JJ. Neurology of the newborn. Philadelphia (PA): Saunders; 2001.Zweig M, Campbell G. Receiver-operating characteristic (ROC) plots: a

fundamental evaluation tool in clinical medicine. Clin Chem1993;39:56177.
