^{1, 2}

^{3}

^{1}

^{3}

^{3}

^{4}

^{5}

^{5}

^{1}

^{2}

^{3}

^{4}

^{5}

There is an important evidence of differences in the EEG frequency spectrum of control subjects as compared to epileptic subjects. In particular, the study of children presents difficulties due to the early stages of brain development and the various forms of epilepsy indications. In this study, we consider children that developed epileptic crises in the past but without any other clinical, psychological, or visible neurophysiological findings. The aim of the paper is to develop reliable techniques for testing if such controlled epilepsy induces related spectral differences in the EEG. Spectral features extracted by using nonparametric, signal representation techniques (Fourier and wavelet transform) and a parametric, signal modeling technique (ARMA) are compared and their effect on the classification of the two groups is analyzed. The subjects performed two different tasks: a control (rest) task and a relatively difficult math task. The results show that spectral features extracted by modeling the EEG signals recorded from individual channels by an ARMA model give a higher discrimination between the two subject groups for the control task, where classification scores of up to 100% were obtained with a linear discriminant classifier.

Epilepsy is one of the most common
human brain disorders. It is often accompanied by disturbances in behavior,
brain dysfunction, and cognitive impairment. According to the World Health
Organization, 0.7% to 1% of the world’s population suffer from epilepsy and
this generally peaks in childhood and advanced age, meaning that a large
proportion of patients have this chronic disease for most of their lives [

Various studies have been carried out to promote
our understanding on the development of this disease and on how epileptic
subjects differ from normal subjects. Most of the work [

Literature indicates that various parametric [

Both in adult epilepsy and child epilepsy, most of
the published work is focussed on the seizure itself or on related events such
as the ictal, preictal, interictal, postictal parts of the
seizure and spikes [

Patients suffering from epilepsy are most often
under the control of antiepileptic drugs (AEDs). The effect of these drugs can
also be a reason behind the resulting significant differences between epileptic
patients and controls. Salinsky et al. [

The discrepancy between the EEGs of epileptic
subjects as compared to controls has been studied mostly in adults. Most of the
work focuses on the alpha band which is the dominant frequency in the human
scalp EEG of adults [

In this paper, we address the possibility of identifying changes in epileptic subjects versus control subjects at an early stage, when just a few seizures occurred in the past. The epileptic population consists of children selected from the pool of paediatric neurology outpatient clinics of two hospitals in Heraklion, Crete, where they were diagnosed and followed at regular intervals. It should be noted that they were diagnosed with no psychological findings, they were not suffering from severe epileptic syndromes and the visual inspection of their EEG was normal. These children, referred to as controlled epileptic, were put under scrutiny because of their early symptoms, without any detected brain damage; they had one or more epileptic seizures in the past and some of them were under monotherapy with drugs in low doses, without clinical side-effects.

The EEG study of such children compared with matched age controls is important from both the clinical andtechnical perspectives. Thus, the purpose of this paper is twofold. First, we address the question of whether controlled-epileptic children exhibit spectral differences in their EEGs in comparison to an age-matched control group during the performance of a control task and a mental task. Second, we address the development of a sensitive and reliable measure for discrimination between the two groups. According to our knowledge, such an analysis has not been carried out so far. We compare two different approaches of localizing activity differences and retrieving relevant information for classifying the two populations. In particular, we elaborate the differences in classification results obtained when using a nonparametric signal representation approach such as Fourier transform or wavelets and a parametric signal modeling approach such as autoregressive moving average (ARMA).

The paper proceeds as follows. Section

The studied population consisted of twenty children aged 9–13 (9 boys, 11
girls) with controlled epileptic seizures, but without any clinical or
laboratory findings of brain dysfunction, and twenty (age and sex) matched
controls on a volunteer basis. Inclusion criteria for patients and controls
consisted of the following: (a) age of 9–13 years old; (b)
normal intellectual potential (assessed with WISC-III); (c) absence of
neurological damage documented by
neurological evaluation for patients and controls and by brain CT and/or MRI
scan for patients; and (d) absence of psychiatric problems (based on parent’s
interview). It should be noted that the EEG signals recorded in both groups were
visually evaluated as normal; and detailed clinical, laboratory, and
neuropsychological findings could not indicate any population differences; the
only clinical indication for the epileptic population was the medical diagnosis
of repeated epileptic seizures in the past (the last epileptic event was
diagnosed between a few weeks to 1 and half years before this study). These
children were treated using common antiepileptic medication only after they
exhibit at least two seizures or absences. The types of seizures diagnosed were the most common
ones in childhood: Rolandic epilepsy (4 children), idiopathic
generalized seizures (5 children), focal seizures (3 children), focal secondary
generalized seizures without detectable brain damage (6 children), and absence seizures (2 children). More specifically, the children with absences were
free from seizures from the beginning of the treatment with Depakine. The other forms such as
generalized tonic-clonic seizures or those with rolandic spikes had a history
of two to five episodes, which were prevented after the treatment with common therapeutic (low) dosages of Tegretol. Especially
in the case of the absences the treatment is effective from the very beginning
and these children were monitored, while treated, for one to two years. During
this period no seizures were identified. Absences and idiopathic tonic-clonic
seizures are generated from the brainstem, while rolandic seizures are generated from
the rolandic area [

Patients and controls, all right-handed, were individually evaluated in the clinical neurophysiology laboratory, at the Medical School of the University of Crete. All parents of children involved in the study signed a written consent form, after having been informed about the study’s purpose and the required procedures. The study was approved by the Local Ethics Committee.

Continuous EEGs were recorded in an electrically
shielded, sound and light attenuated room while participants sat in a reclined
chair. The EEG signals were recorded from 30 electrodes placed according to the
10/20 international system, referred to linked A1+A2 electrodes. This electrode
montage is shown in Figure

Electrode montage consisting of 30 electrodes placed according to the 10/20 international system.

In this study, two different tasks were analyzed to identify differences
in brain dysfunction under tasks with different brain operations. During the
control task (Task 1) subjects were at rest and had their eyes fixed on a point
displayed on a computer screen to reduce eye artefacts. The second task was a
mathematical task (Task 2) involving the subtraction of two digit numbers [

The goal of this analysis is to find discriminating features between
epileptic and control children that result in high classification scores. Ideally,
a preprocessing step is used to filter out irrelevant data and enhance the
discriminating features of the signal. A subsequent spectral analysis step
could then be applied to extract those suitable biomarkers. This nonparametric approach,
which is labelled as approach (1) in Figure

(1) Nonparametric and (2) parametric approaches for feature extraction and classification.

In order to compare these different approaches in classifying the two
subject groups, both methods were implemented and the results obtained were similarly
analyzed. For the nonparametric approach both a global Fourier Transform (FT)
and wavelets were used for the spectral analysis stage and the biomarkers
extracted from each method were compared. The FT gives an average spectral plot
over the time period considered. On the other hand, wavelets are mathematical functions that divide the data into different
frequency components and then analyze each component with a resolution matched
to their scale. Thus, instead of working on a single time or frequency scale,
they work on a multiscale basis [

For the second approach, a time-frequency spectrum
is generated using the estimated parameters of the ARMA model derived from each
EEG signal. Parametric models are known to enhance the time-frequency
resolution of power spectra estimation [

The FT transforms a signal in the time domain into
its frequency domain representation. By definition, a signal

Over the past decade, the WT has been developed into an important tool
for analysis of time series that contain nonstationary power at many different
frequencies (such as the EEG signal), and it has proved to be a powerful
feature extraction method [

The continuous wavelet transform (CWT) was preferred in this work, so
that the time and scale parameters can be considered as continuous variables.
In the CWT, the notion of scale

The first stage of the feature extraction method is
based on capturing the time-averaged power spectrum

The autoregressive moving average (ARMA) or Box-Jenkins model is a
parametric model where the estimate of the time series at a time instant
depends on its past values (deterministic part) and on a random disturbance
(stochastic part) [

A parametric method can provide adequate spectral estimates only when the correct model
order is chosen. Various techniques have been developed to estimate the optimal
order, the most renowned being the Akaike's information criterion (AIC) [

An ARMA

Let

Once an estimate of the ARMA parameters

Time-frequency plots of log-power distribution for ARMA model orders (5,2), (12,3), (18,5), and (24,6) are shown.

This study proposes a statistical method for mining
the most significant lobes, resembling the way many clinical neurophysiological
studies evaluate the brain activation patterns. Since the goal is to find
significant differences between two groups, the independent two-sample

The former statistical analysis technique was used to identify which channels and frequency bands give significant differences between the epileptic subject group and the control group for both the signal representation approaches and the signal modeling approach.

In this study, the epileptic and control groups were classified by
using a strictly linear discriminant analysis (LDA) classifier based on a linear
discriminant function that fits a multivariate normal density to each group,
with a pooled estimate of covariance as implemented in the MATLAB statistics
toolbox [

The capabilities of the methods described in Section

During the tasks performed, neurological examinations showed that there are no differences in achievement between children younger than 11 years old and children in the age of 11 and above. Therefore, in subsequent analyses the subjects were not divided into different age groups. Both nonparametric and parametric approaches were then applied to the real EEG data, where each signal was initially set to zero mean and unit variance. In each case, we compute the channel/band significance, as well as the corresponding classification scores with sensitivity-specificity measures.

Figure

Topographic maps showing the

Topographic maps showing the classification scores above 65% over all 27 channels.

The results for Task 2 show less discriminative differences, especially
when using the ARMA approach. WT succeeds in identifying weak spectral
differences within the alpha band (8–13 Hz) for a
number of channels in the left frontal area. Notice that in all cases the
discrimination levels achieved by either method are weak and do not support any
significant differentiation between the studied populations in this task. The
global FT method has also found the left brain area to show significant
differences between the epileptic and the control group especially in the alpha
and beta band, but once again the classification levels (Figure

The classification scores for both Tasks 1 and 2 are shown in Figure

Classification scores, sensitivity, and specificity results for Task 1.

Classification scores, sensitivity and specificity results for Task 2.

Both Figures

For Task 2, the classification scores are much lower and the differences
between the parametric and nonparametric approaches are not as clear as for
Task 1. The highest classification score of 80% was obtained by the WT approach
over the frontal channels within the alpha band, particularly over Fp1, which
was also found to be significant (

In order to identify further differences in the feature distributions of
the two subject groups, probability density estimates of the feature values of
the patients and controls over different frequency bands and channel locations
were also computed. Figure

Power spectral feature distributions for control and epileptic subjects over channel Cz. Features were extracted from the delta band when the subjects were performing the control (rest) Task 1: (a) shows the results for the FT approach, (b) shows the results for the WT approach, and (c) shows the results for the ARMA approach.

The set of
plots shown in Figure

Power spectral feature distributions for epileptics and controls comparing the differences between the different frequency bands. (a) and (b) show the results for FT, (c) and (d) show the results for wavelets, and (e) and (f) show the results for ARMA.

This work considers methods for the discrimination of two groups of age-matched children, that is, controls and children with controlled epilepsy. Initial clinical and psychological examinations, as well as visual EEG inspection, do not provide any information leading to possible differences. On the original EEG data we apply two types of methodologies, one based on direct signal representation (through nonparametric techniques, mainly the Fourier transform and the wavelet transform) and the other at modeling the signal dynamics (through a parametric ARMA model). The spectral features extracted by these methods in each channel and spectral band are examined through significance tests, classification accuracies, and statistical distributions of biomarkers. This work indicates that parametric modeling of the EEG dynamics provides better representation of the significant EEG content than nonparametric techniques for feature extraction. The features extracted by the ARMA model for the control task provide higher discrimination power than those extracted by the Fourier transform and wavelet approaches. The results for the Fourier transform have shown to be slightly superior than those of the wavelet transform in this case where the biomarker is an average of the spectral power over the whole period of data. This may be the cause of artefacts introduced by the windowing leakage effect of wavelets which is less dominant in the global Fourier transform approach where a single window was considered. In other situations where the temporal resolution is taken into consideration, it is expected that wavelets outperform the Fourier transform technique.

Comparing the control and math tasks, the methods derive significant differences during the control (rest) task, but they are unable to identify any consistent differences during the more demanding mathematical task where the discrimination of the specific brain dysfunction seems more difficult.

The potential clinical benefit of this work is the analysis of EEG data
towards the identification of children with mild epilepsy at early stages, where
classical, neurological, and clinical examinations and detailed psychological
and neuropsychological testing are unable
to identify any signs of brain dysfunction. The ARMA results show that
epileptic children during the control task have higher activity in frequency
bands up to the gamma1 band, but this activity becomes similar in both groups
for frequencies within the gamma2 band. When analyzing an adult-patient group,
Tuunainen et al. [

This work involved the study of children with mild epilepsy who had epileptic seizures in the past but who did not exhibit any clinical, physiological, or visible neurophysiological symptoms during the study. The goal of this analysis was to develop reliable techniques to test if such controlled epileptic conditions induce related spectral differences in the EEG. The results show that parametric ARMA modeling techniques extract more reliable biomarkers than the nonparametric Fourier and wavelet transform techniques implemented here. For the control task, the ARMA technique led to classification scores up to 100% across all channels for frequency bands ranging from the delta to the gamma1 band.

Diagnosis of epilepsy was here conducted by considering biomarkers on an average of the spectral power over the whole 10.24-second period of data available. Future work will investigate whether taking into account the temporal information enhances these classification scores specifically for the math task where the complexity of the task made it difficult to capture any brain dysfunction through global biomarkers. In the latter case it is expected that wavelets will outperform the Fourier transform technique and lead to results which are comparable to its parametric counterpart.

This work was supported in part by the EC-IST project Biopattern, Contract no. 508803, and by the internal research grant of the University of Malta LBA-73-967.