A METHOD FOR DETECTING EEG SEIZURES IN A NEWBORN OR A YOUNG CHILD

TECHNICAL FIELD

The present invention relates to a method for detecting electroencephalogram (EEG) seizures in a newborn or a young child.

BACKGROUND

The reference to any prior art in this specification is not, and should not be taken as an acknowledgement or any form of suggestion that the prior art forms part of the common general knowledge.

A high incidence of newborn EEG seizures [1] has resulted in considerable mortality and morbidity rates in neonates [2-5]. Dysfunction in the newborn central nervous system (CNS) is often first identified through seizures. Accordingly, accurate and rapid detection of neonatal seizure is highly desirable with a view to providing appropriate treatment and therapy to newborns that may be prone to CNS dysfunction.

Presently, EEG experts (e.g. highly qualified pediatric neurologists) manually review newborn EEGs for characteristics which are evidence of seizure activity. This review process can be quite arduous and is time intensive. The process is also subjective and some seizures may go undetected when reviewing EEGs. The EEG recordings are typically of a relatively short duration (i.e. between 20-40 minutes) and are not suited for detecting the ongoing presence, frequency and duration of seizures.

A cerebral function monitor (CFM) system is one time-based technique which can be used for displaying the integrated amplitude of an EEG channel continuously over a very long time basis. CFM systems have been proposed for the long term monitoring of newborn EEGs and allow for the determination

of global neurological status. However, CFM systems must be manually monitored for seizures and do not allow for automated seizure detection.

Alternatively, techniques based on spectral estimation, which involves estimating the frequency content of an EEG signal and providing a frequency domain representation, can be used to detect EEG seizures. One such newborn EEG seizure detection technique, developed by Gotman et al. [6], incorporates features from the frequency domain representation of 10 second EEG epochs. However, the newborn EEG signal has been shown to be highly nonstationary, particularly during EEG seizure periods [7]. Therefore, the features obtainable using stationary signal processing techniques are likely to be suboptimal and lead to suboptimal automatic newborn EEG seizure detection methods.

A seizure detection monitor developed by "Brainz" involves long term EEG recording by virtue of a two channel recording configuration. The Brainz monitor displays basic signal features such as CFM and spectral edge measurement. These features, while useful in identifying seizure manually, are not suitable for automatic detection.

A seizure detection system developed by "Stellate Harmonie E" can record in a full channel configuration. This system uses the frequency-based Gotman algorithm (discussed above) to detect newborn EEG seizures.

It is an object of the present invention to provide a method for detecting seizures in newborns or young children using an alternative EEG processing technique.

SUMMARY OF THE INVENTION

According to one aspect of the present invention, there is provided a method for detecting a seizure in a newborn or a young child, the method including the steps of: transforming electroencephalogram (EEG) data of the newborn or young child to form a time-frequency representation; extracting features from the time-frequency representation using a component linking technique; and detecting the seizure using the extracted features.

The inventors have found that time-frequency signal analysis/processing techniques are highly suitable for the analysis and processing of non- stationary signal data, such as newborn EEG data. Time-frequency signal processing techniques can effectively account for the non-stationary nature of the EEG signals by representing them in a joint time-frequency domain representation.

In one embodiment, the step of transforming involves forming the time- frequency representation of the EEG data using either a linear or a quadratic transformation. The step of transforming may involve the step of selecting the transformation. The step of transforming may involve the further step of selecting suitable parameters of the transformation.

The step of extracting features may involve the steps of: detecting significant local peaks of the time-frequency representation; and linking the detected significant local peaks together to form linked time- frequency components.

The linked time-frequency components may represent an estimate of the instantaneous frequencies (IFs) of the multicomponent EEG signal, and the step of. extracting features may further involve extracting a set of time- frequency features based on the estimated IFs.

The step of extracting features may further involve extracting a set of time- domain features and frequency-domain features to enhance a true positive rate (to account for different types of EEG seizures) and to reduce a false positive rate.

The step of detecting the seizure may involve the step of comparing the extracted features to a number of thresholds previously derived from the analysis of a plurality of EEG seizures.

The step of detecting the seizure may further involve combining the results of the threshold comparisons using logical operators to determine whether or not a seizure is present. The duration of the detected seizure may be determined by performing a gap closing procedure.

The step of detecting may further include the step of setting a flag responsive to the compared extracted features satisfying predetermined seizure conditions. A seizure may be detected upon the setting of a plurality of consecutive flags.

The method may further include the step of generating an alarm to signify the detection of the seizure.

Preferably, prior to the step of transforming, the method further includes the step of removing artefacts from the EEG data. The artefacts may be removed using a matching pursuit (MP) decomposition method.

Even more preferably, between the steps of removing artefacts and transforming, the method further involves the step of segmenting the EEG data into a plurality of epochs. The step of segmenting may involve using a sliding rectangular window of length 12.8 seconds with an overlap of 6.8 seconds between epochs.

Prior to the step of removing artefacts, the method may further include the step of re-sampling the EEG data to reduce the amount of data to be subsequently processed.

Prior to the step of re-sampling the EEG data, the method may further include the step of filtering the EEG data.

Prior to the step of removing artefacts, the method may further include the step of sampling EEG signals from the newborn or young child using a plurality of channels to form the EEG data. In one embodiment, two channels are used.

According to a further aspect of the present invention, there is provided a medium, such as a magnetic or optical disk or solid state memory, containing machine readable instructions for execution by one or more processors to thereby perform the method.

According to a further aspect of the present invention, there is provided a computational device which is loaded with computer readable instructions to perform the method.

According to a further aspect of the present invention, there is provided a seizure detection apparatus, the seizure detection apparatus including: one or more connectors for connecting to electroencephalogram (EEG) measurement electrodes; and one or more processors in communication with said connectors, and for receiving EEG signals from the connectors and storing related EEG data in a storage means, the processors being configured to: transform the stored EEG data of a newborn or young child to form a time-frequency representation; extract features from the time-frequency representation using a component linking technique; and detect a seizure using the extracted features.

The processors may be in communication with an electronic memory, the electronic memory storing computer readable instructions to: transform the stored EEG data of a newborn or young child to form a time-frequency representation; extract features from the time-frequency representation using a component linking technique; and detect a seizure using the extracted features

The storage means may be an electronic memory or a disk.

BRIEF DESCRIPTION OF THE DRAWINGS

Preferred features, embodiments and variations of the invention may be discerned from the following Detailed Description which provides sufficient information for those skilled in the art to perform the invention. The Detailed Description is not to be regarded as limiting the scope of the preceding Summary of the Invention in any way. The Detailed Description will make reference to a number of drawings as follows:

Figure 1 is a schematic view of a hardware system for performing a method for detecting a seizure in a newborn or young child in accordance with an embodiment of the present invention.

Figure 2 is a block diagram of a software product in accordance with the embodiment.

Figure 3 is a flowchart showing the steps performed by a preprocessing/signal conditioning submodule of the software product of Figure 2.

Figure 4 is a graph showing raw EEG data and processed EEG data with artefacts removed by the preprocessing/signal conditioning submodule of Figure 3.

Figure 5 is a data flow diagram relating to a time-frequency representation submodule of the software product of Figure 2.

Figure 6 is a flowchart showing the steps performed by a time-frequency feature extraction submodule of the software product of Figure 2.

Figure 7 is a flowchart showing the steps performed by an EEG state identification submodule of the software product of Figure 2.

Figure 8 is a control flow diagram of an output/user interface submodule of the software product of Figure 2.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

HARDWARE SYSTEM 2

According to an embodiment of the present invention, there is provided a method for detecting an electroencephalogram (EEG) seizure in a newborn or young child using time-frequency signal processing techniques. The method is performed using the hardware system 2 (i.e. seizure detection apparatus) shown in Figure 1.

The system 2 includes two or more electrodes 4 for attachment to the scalp 3 of the newborn for measuring two or more channel EEG signals. The electrodes 4 sense EEG signals 6 and are releasably attached to an amplifier 8 via electrical connectors 9. An optical protection circuit may be provided between the electrodes 4 and amplifier 8 for electrically isolating the newborn from the amplifier 8. The amplifier 8 serves to amplify the measured EEG signals 6 which are subsequently passed though an anti-aliasing (low pass) filter 10 to remove any high frequency noise. The filtered analog EEG signals are coupled to an analog to digital converter (ADC) 11 and the resulting raw digital EEG data is subsequently transferred to and recorded on a storage medium 15 via disk drive 13.

The processor 12 (e.g. personal computer) is interfaced to a monitor display 14 for displaying information and user inputs 16 including a mouse and a keyboard for receiving user input. In operation, the processor 12 loads software product 17 into memory 7, and executes software product 17 to perform the method for detecting an EEG seizure as described in detail below. The software product 17 may be provided as a set of computer readable instructions on an optical disk 15 or other like storage media.

SOFTWARE PRODUCT 17

Referring to Figure 2, the software product 17 includes instructions for processor 12 to record measured raw digital EEG data. The software product 17 further includes instructions for processor 12 to process the recorded EEG data to determine whether an EEG seizure condition exists. The software product 17 includes an input data module 18, a seizure detection module 20 and an output/user interface module 22.

Input data module 18

The input data module 18 retrieves and passes the recorded EEG data 64 from the storage medium 12 to the seizure detection module 20. The retrieved data 64 is also provided to the output/user interface module 22.

Seizure detection module 20

The seizure detection module 20 includes four submodules: a preprocessing/signal conditioning submodule 24, a time-frequency representation submodule 25, a feature extraction submodule 26 and an EEG state identification submodule 28. These submodules 24, 25, 26, 28 are described in detail below.

Preprocessing/Signal conditioning submodule 24

The preprocessing/signal conditioning submodule 24 receives the recorded EEG data 64 and performs a variety of signal alterations (i.e. conditioning) to enhance the desired signal characteristics and remove any unwanted signal components.

The flowchart of Figure 3 shows the preprocessing method 30 performed by the preprocessing/signal conditioning submodule 24.

At step 32, the digitally recorded raw EEG data 64 is received by the preprocessing/signal conditioning submodule 24.

At step 34, the signal alterations begin with a band-pass filtering procedure to isolate the frequency band of interest, based on previous characterisation of newborn EEGs. Cutoff frequencies are selected to be 0.5Hz and Fs _d /2 Hz, where FS _d refers to the desired sampling frequency at which the signal will be resampled. Typically, Fs _d = 20Hz.

At step 36, the EEG data is resampled to FS _d to reduce the size of the data for further processing. In this manner, redundant data is removed and, accordingly, further processing times are reduced.

At step 37, the software segments the resampled EEG data 41 into manageable EEG epochs. Typically, the epoch length is chosen to be N = 256 samples or 12.8 seconds in length (i.e. with Fs _d = 20Hz).

At step 38, physiological and extraphysiological artefacts are removed from the signal. Artefacts are those inherent, unintended and unwanted aberrations which occur while recording EEG data. The artefact removal involves a time- frequency based signal filtering process as described in detail below.

Newborn EEG data can be severely affected by other physiological signals, such as the ECG (electrocardiogram), EOG (electrooculogram) and EMG (electromyogram). Interference from these signals on the EEG data can have a masking effect, hiding seizure patterns that lead to missed EEG seizure detections [8]. Of particular importance are artefacts that are characterized by their short durations and high amplitude. These artefacts can severely mask the true EEGs, making them unreadable. Artefacts also cause many problems with automated seizure detection systems and can lead to false positives (false detections) or false negatives (missed detections), depending on the seizure detection method and the type of artefact. Therefore, careful removal of these artefacts allow for improved newborn EEG seizure detection.

Artefacts occur randomly in the newborn EEG data and have varying time durations and frequency supports. Therefore, time-invariant filtering procedures cannot be used in the removal of these artefacts from EEG data. In the present embodiment, the artefact removal process involves using a time-frequency signal processing technique to adaptively remove energy caused by short-time and high amplitude artefacts from the EEG data.

The artefact removal technique is based on a matching pursuit method [9]. This method can be effectively used to obtain a time-frequency representation/parameterization of non-stationary signals. The matching pursuit method is an iterative process which decomposes the signal into a weighted sum of atoms selected from a predefined dictionary [9]. At each iteration, the matching pursuit method selects the atom that best represents the signal residue in an inner product sense. Using a matching pursuit decomposition, a signal, x , is represented as:

^x = σ 7 ^a A (D

where a _γ is the coefficient corresponding to the atom φ _γ and / is an indexing vector which uniquely identifies an atom from the dictionary. A

signal, x , can also be approximated, using the matching pursuit decomposition, with m atoms, as

where the residual signal is given as

R ^m X ~ X - χ (3)

At each iteration, the atom is selected such that

where (f,g) represents the inner product of the functions /and g .

To account for the observed characteristics of the newborn EEG seizure, the time-frequency dictionary used in the artefact removal procedure is a combination of linear frequency modulated (LFM) atoms and Gabor atoms, similar to that which was proposed in [10].

For the real Gabor dictionary, the indexing vector includes the atom scale, s, translation, u, modulation, ^ , and phase, θ (i.e. γ = [s,u,ξ,θ] ). The Gabor atom is expressed as

φøABθR(0 + β) (5)

where φ(t) is a Gaussian function.

For the LFM dictionary, the indexing vector includes the initial frequency, ξ, , the rate of change in frequency, ξ _r , and phase, θ (i.e.

Y - [4 _t >ζ _r - _> θ] )-The LFM atom is expressed as

φ _γLFM it) ^ ∞s^π[ξ _{i +} ^t) + θ)

(6)

The fundamental idea behind the artefact removal technique is that short- time and high amplitude artefacts will be well represented by Gabor atoms with small scales in the first couple of iterations of the matching pursuit decomposition. Therefore, if these atoms are removed from the signal representation, an artefact reduced signal will be obtained. The inclusion of the LFM atoms is to prevent atoms representing EEG seizure from being removed by the artefact removal method as detailed below.

Artefact Removal Method

This artefact removal method uses the matching pursuit decomposition to iteratively determine whether or not an artefact atom has been selected. If the atom selected is not an artefact atom, the procedure is exited. If, however, an artefact atom is selected, the atom is removed from the signal and the procedure is repeated on the residual signal.

Table I below shows the atom scale (width) and associated percentage of signal energy which corresponds to atoms representing short-time and high amplitude artefacts. These values were selected based on the analysis of newborn EEG data corrupted by artefacts. The scale values and percentage of signal energy were based on a sampling frequency of

Fs _d = 20Hz and an epoch length of 256 samples.

Table I. Parameters for an atom representing a short-time and high amplitude artefact.

Pseudocode of the artefact removal method for EEG epoch, x is shown below:

A. Initialization: EEG epoch = x, i = l

B. Run matching pursuit iteration, / ^' , to get

γι ^•

C. If selected atom is LFM OR scale > 64 samples

Exit the artefact removal process Else if scale < 64 AND associated atom energy ratio < Listed in

Table 1.

Exit the artefact removal process Else

Form the residual signal x = x - a^φ^ D. / = / + 1 , Return to step B.

Upon completion of step 38, the preprocessed EEG data 41 is stored. An example of the proposed artefact removal technique applied to an EEG epoch which is corrupted by a short-time and high amplitude artefact is shown in

Figure 4. The dashed line shows an original EEG data 64 and the continuous line shows the preprocessed EEG data 41.

Time-Frequency Representation submodule 25

As shown in Figure 5, the time-frequency representation submodule 25 receives the preprocessed EEG 41 which, in turn, is transformed to a time- frequency representation 51. The time-frequency representation 51 can be obtained through using linear or quadratic time-frequency transformations in the time-frequency representation submodule 25. The time-frequency representation 51 is preferably a reduced interference quadratic time- frequency distribution (QTFD).

The general form of a QTFD for a real signal x(t) is given by [11]

λ ('./) = \ \h{τ)G{t-u,τ)z{u-y ₂ )z\u-y ₂ )e- ^J2φ dτ ₍₇ )

where z{t) is the analytic associate of the real signal x{t) .

The time-lag kernel G(t, τ) defines the particular characteristics of a QTFD. Therefore, decision on G(t,τ) , and hence on which QTFD to use, must be made by a user using the QTFD decision submodule 42, as indicated in the time-frequency representation submodule 25 in Figure 5.

A subclass of QTFDs, which are suitable for real life applications such as newborn EEG seizure detection, are the reduced interference distributions

(RIDs). The time-lag kernel of RIDs generally incorporates a smoothing parameter which can be adjusted to provide optimal time-frequency representation. There are also a number of other parameters, such as the length of the lag window, h{τ) , which require setting by the user using the parameter decision submodule 44 in the time-frequency representation submodule 25 in Figure 5.

Typically, time-lag kernel G(t,τ) , is chosen to be

with parameter β = 0.02

A time-frequency representation 51 of the segmented newborn EEG data is then obtained at the output of this time-frequency representation submodule 25.

Feature Extraction module 26

Newborn EEG seizure can be characterized by dominant and distinct frequency modulated components in the time-frequency domain [12, 13]. In contrast, the newborn EEG background lacks any dominant frequency modulated components. Hence, features which indicate the presence of clear, dominant and distinct frequency modulated components are extracted from the time-frequency representation 51 in the feature extraction submodule 26.

The feature extraction submodule 26 receives the time-frequency representation 51 which is subsequently processed using image processing techniques which allow for feature extraction according to the method 52 shown in Figure 6.

Firstly at step 54, the significant local peaks of the time-frequency representation 51 are detected. The significance of the peaks is determined by a predefined threshold parameter 53.

The detected significant local peaks are then linked together at step 56 using a component linking procedure. This procedure links the detected significant

peaks to form an estimate of the instantaneous frequency of a frequency modulated component. A frequency modulated component is identified if significant peaks in the time-frequency representation 51 are continuously linked for a predefined period of time in accordance with a predetermined time threshold 55.

Based on those estimated IFs, a set of time-frequency features are extracted. The following is a list of output extracted features 46 from the input time- frequency representation 51:

Feat. 1 : TF-Comp: is a feature taking the value of 0 or 1, indicating whether a DOMINANT time-frequency component exists in the newborn EEG time- frequency representation 51. This feature is extracted using the local peak detection step 54 with a threshold 53 of.

?7 = 0.3 x max {IF.D}

and the component linking step 56 with a time threshold 55 of 9.5 seconds. If a linked component is found, the feature is assigned the value 1. Otherwise the feature is assigned the value 0.

Feat. 2: Min-IF-Freq: is the minimum instantaneous frequency of the TF- Comp.

Feat. 3: Mean-Freq-Fund: is the mean frequency associated with TF-Comp.

Feat. 4: Num-Close-Peaks: is the number of significant local peaks which are within a predefined number of pixels (typically 10) of the TF-Comp.

Feat 5: Harmonic-Comp: is a feature which is used for determining whether there is a harmonic component present in the epoch. However, a new binary image of significant local peaks, different to that in "Feat. 1" is computed using a threshold 53 of

η = 0.05 x max{TFD}

A component must have a predefined minimum duration (e.g., 9.5 seconds) to be considered as a true component. This feature is a three parameter feature consisting of 1) Intersection of component time supports, 2) mean ratio between component IFs and 3) standard deviation of ratio between component IFs.

Feat. 6: Components: is the exact number of linked components in the EEG epoch when using the η = 0.05 x max {TFD} threshold 53. The number of linked components can be seen as a measure of deterministic complexity.

The following is a list of features extracted from the time domain and frequency domain introduced to account for more types of EEG seizures and to help reduce the false positive rate.

Time domain features: the following features are extracted directly from the EEG epoch.

Feat. 7: Max-Amp: is the maximum of the absolute value of the voltages in each EEG epoch.

Feat. 8: Epoch-Energy: is the total energy of the EEG epoch.

Feat. 9: TP-Ratio: is the ratio between the number of turning points in an EEG epoch and Feat 3. Mean-Freq-Fund.

Frequency domain features: the following features are extracted from the power spectrum of the EEG epoch. The power spectrum is typically estimated using Burg's method with autoregressive prediction model order of 10.

Feat. 10: Peak-Freq: is the frequency component of an EEG epoch which has the largest magnitude.

Feat. 11 : H-Spec: is the spectral entropy of the EEG epoch. It is calculated as the Shannon entropy using the normalized spectrum as a probability density function [14].

EEG state identification submodule 28

The EEG state identification submodule 28 receives the EEG features 46 extracted from the newborn EEG epoch. The extracted features 46 are then evaluated in accordance with the method 90 of Figure 7. A detailed description of the EEG state identification method 90 is provided below.

At step 92, the extracted features 46 are compared with predefined seizure threshold values 91 and conditions. These seizure thresholds 91 for the extracted EEG state identifying features 46, that indicate whether or not a seizure is occurring, are determined based on an analysis of a large database of newborn EEGs containing seizures, with the seizure events being identified by a neurologist.

If all of the extracted features 46 are in accordance with their corresponding seizure threshold values 91 and are thereby indicative of seizure, a seizure event is marked at step 94. Otherwise, a non-seizure event is marked at step 96.

Elaborating further, the first step in determining whether there is a seizure, is to determine whether the EEG epoch has sufficient energy or the maximum voltage is excessively large (i.e. typically, if threshold values 91 Epoch-Energy < 5χlO ^~8 or Max-Amp > 300). If either of these criteria is met, a nonseizure event is recorded as the data is not suitable to make the determination. Otherwise, the epoch has met the criteria to be assessed for seizure characteristics based on the five seizure subtypes defined below.

The five seizure subtypes are labeled as

1. Linked Component Seizure (LCS) 2. High Frequency Seizure (HFS)

3. Dominant Harmonic Seizure (DHS)

4. Dominant Linked Component and Multiple Components Seizure (DLCMCS)

5. Continual Multiple Components Seizure (CMCS)

The following pseudocode provides typical criteria for the five seizure subtypes to be registered as either seizure at step 94 or nonseizure at step 96 for a given EEG epoch.

LCS:

if (TF-Comp = 1) & (TP-Ratio > 0.17) & (Min-IF-Freq > 0.6Hz) & (Num-Close- Peaks < 12)

LCS = I Else

LCS = O End

HFS:

if {Peak-Freq > 3.125Hz) & {H-Spec < 6.5)

HFS = I Else

HFS = O End

DHS:

if (Components = 2) & (H-Spec < 6.5)

p1 = Harm-Comp: 1) p2 = absolute value of Harm-Comp: 2) - rounded value of Harm-Comp: 2) p3 = Harm-Comp: 3) if (p1 > δseconds) & (((p2 <= 0.03) & (p3 < 0.125)) OR ((p2 <= 0.07) &

(p3 < 0.09)))

DHS = I Else

DHS = O End

End

DLCMCS:

if (TF-Comp = 1 ) & (Components >= 3)

DLCMCS = 1 Else

DLCMCS = 0; End

CMCS:

if (Components > 2)

Get Next Epoch If (Components_(next_epoch) > 2)

Get Next Epoch If (Components_(next_epoch) > 2)

CMCS & CMCS(next epoch) & (CMCS(next-next epoch) — 1 Else

CMCS & CMCS(next epoch) & (CMCS(next-next epoch) = 0

End End

End

If any of the foregoing seizure subtypes is detected in any channel for a given epoch of time, a seizure event is flagged in a vector of binary flags 60. A seizure is indicated by assigning the flag vector 60 a value "1". Otherwise the flag vector is assigned a value "0" to indicate that no seizure event was detected for that epoch of time.

The flag vector 60 indicating the decision on the state of the EEG (e.g. seizure or nonseizure) is then sent to the Output/User interface module 22.

Output/User interface module 22

The output/user interface module 22 receives the flag vector 60 from the EEG state identification submodule 28 and the digitally recorded raw EEG data 64 from the input data module 18.

At box 68, the raw EEG data 64 and associated flag data 66 are displayed on the monitor display 14 in real-time.

At box 70, the raw EEG data 64 and flag data 66 are stored on a storage medium of personal computer 12 for future review at a later time if necessary.

At box 72, the output/user interface module 22 determines in real-time whether a seizure state exists based upon the flag data 66. An audible alarm is triggered when the flag data 66 indicates a seizure positive state at box 74 for three or more successive epochs. Otherwise, the audible alarm is off (i.e. not triggered) as the flag data 66 indicates a seizure negative state.

A person skilled in the art will appreciate that many embodiments and variations can be made without departing from the ambit of the present invention.

The preferred embodiment was described in relation to seizure detection techniques applied to the newborn, but these techniques for seizure detection are also suitable for the newborn and the young child.

In compliance with the statute, the invention has been described in language more or less specific to structural or methodical features. It is to be understood that the invention is not limited to specific features shown or described since the means herein described comprises preferred forms of putting the invention into effect. The invention is, therefore, claimed in any of its forms or modifications within the proper scope of the appended claims appropriately interpreted by those skilled in the art.

BIBLIOGRAPHY

[1] E. Mizrahi and P. Kellaway, Diagnosis and Management of Neonatal

Seizure. Philadelphia: Lippincott-Raven, 1998

[2] M. Scher, M. Painter, I. Bergman, M. Barmada, and J. Brunberg, "EEG diagnosis of neonatal seizures: Clinical correlations and outcome," Pediatric Neurology, vol. 5, pp. 17-24, 1989

[3] J. Connell, R. Oozer, L.D. Vries, L. Dubowitz, and V. Dubowitz, "Continuous EEG monitoring of neonatal seizures, diagnositic and prognostic considerations," Archives of diseases in children, vol. 64, pp. 425-458, 1989

[4] M. Andre, M. Matisse, P. Vert, and C. Debruille, "Neonatal seizures- recent aspects," Neuropediatrics, vol. 19, 1988 [5] B. Wical, "Neonatal seizures and electrographic analysis: Evaluation and outcomes," Pediatric Neurology, vol. 10, pp. 271-275, 1995

[6] J. Gotman, D. Flanagan, J. Zhang, and B. Rosenblatt, "Automatic seizure detection in the newborn: methods and initial evaluation," Electroenceph. Clin. Neurophys., vol. 103, no. 3, pp. 356-362. 1997

[7] B. Boashash and M. Mesbah, "Time-frequency methodology for newborn electroencephalographic seizure detection," in Applications in Time-Frequency Signal Processing, A. Papandreou-Suppappola, Ed., Boca Raton: CRC Press, 2003

[8] P. Celka, B. Boashash and P. Colditz, "Preprocessing and Time- Frequency Analysis of Newborn EEG Seizures," IEEE Engineering in Medicine and Biology, vol. 20, iss. 5, pp. 30-39, 2001

[9] S. Mallat and Z. Zhang, "Matching pursuits with time-frequency dictionaries," IEEE Trans, on Signal Processing, vol. 41, no. 12, pp. 3397-3415, Dec. 1993

[10] L. Rankine, M. Mesbah and B. Boashash, "A Novel Algorithm for

Newborn EEG Seizure Detection using Matching Pursuits with a Coherent Time-Frequency Dictionary," Proc. International Conference on Scientific and Engineering Computing, Singapore, July 2004, CD-ROM

[1 1] B. Boashash, "Theory of quadratic TFDs," in Time- Frequency Signal Analysis and Processing: A Comprehensive Reference, B. Boashash, Ed., London: Elsevier, 2003, pp. 59-81 [12] B. Boashash and M. Mesbah, "A time-frequency approach for newborn seizure detection," IEEE Engineering in Medicine and Biology Magazine, vol. 20, no. 5, pp. 54-64, 2001

[13] L. Rankine, N. Stevenson, M. Mesbah and B. Boashash, "A nonstationary model of newborn EEG," IEEE Transactions on

Biomedical Engineering, accepted, July 2006

[14] N. Kannathal, M. L. Choo, U. Rajendra Acharya, P. K. Sadasivan, "Entropies for detection in epilepsy in EEG," Computer Methods and Programs in Biomedicine, vol. 80, 2005, pp. 187-194