Distinguishing childhood absence epilepsy patients from controls by the analysis of their background brain electrical activity (II): A combinatorial optimization approach for electrode selection

https://doi.org/10.1016/j.jneumeth.2009.04.028Get rights and content

Abstract

In this sequel to our previous work [Rosso OA, Mendes A, Rostas JA, Hunter M, Moscato P. Distinguishing childhood absence epilepsy patients from controls by the analysis of their background brain electrical activity. J. Neurosci. Methods 2009;177:461–68], we extend the analysis of background electroencephalography (EEG), recorded with scalp electrodes in a clinical setting, in children with childhood absence epilepsy (CAE) and control individuals. The same set of individuals was considered—five CAE patients, all right-handed females and aged 6–8 years. The EEG was obtained using bipolar connections from a standard 10–20 electrode placement. The functional activity between electrodes was evaluated using a wavelet decomposition in conjunction with the Wootters distance. In the previous study, a Kruskal–Wallis statistical test was used to select the pairs of electrodes with differentiated behavior between CAE and control samples (classes). In this contribution, we present the results for a combinatorial optimization approach to select the pairs of electrodes. The new method produces a better separation between the classes, and at the same time uses a smaller number of features (pairs of electrodes). It managed to almost halve the number of features and also improves the separation between the CAE and control samples. The new results strengthen the hypothesis that mostly fronto-central electrodes carry useful information and patterns that can help to discriminate CAE cases from controls. Finally, we provide a comprehensive set of tests and in-depth explanation of the method and results.

Introduction

The electroencephalogram (EEG) is the recording of the spontaneous brain electrical activity by means of electrodes located on the scalp. One can assume that the EEG is a signal containing information about the condition of the brain. We can also accept as working hypothesis that the EEG recorded under resting conditions is representative of the global state of the brain. As a consequence, we can expect that individual patterns might become prominent (Elbert et al., 1992). Then, a plausible working hypothesis is that EEG time series (background EEG) corresponding to healthy controls is different from patients with pathologies (e.g. epilepsy) or neurodegenerative diseases (e.g. schizophrenia, Alzheimer, and multiple sclerosis). Moreover, the same hypothesis can be applied to discover and specify electrocortical substrates for cognitive, emotional, and behavioural conditions. However, it is currently accepted that a human observer cannot discriminate EEG traces of healthy controls from schizophrenics or inter-ictal epileptic activity—just to mention some cases Elbert et al., 1992, Jan, 2002, Engel and Pedley, 1997. Quantitative EEG analysis using computational methods can therefore assist in the background EEG characterization.

Epilepsy is one of the most common human brain disorders and refers to a group of neurologic disorders. It is often accompanied by disturbances in behaviour, brain dysfunction, and cognitive impairment. The incidence of epilepsy is estimated in about 0.6–1% of the world’s population, according to the World Health Organization.1 Moreover, the disease can appear at any age. This generally peaks in childhood and advanced age, meaning that a large proportion of patients have this chronic disease for most of their lives. This supports the importance of identifying this population as early as possible such that the clinician can prescribe the necessary medication to stop its progression. Childhood epilepsies are different from many adult epilepsies as onset and offset are often determined by age, i.e. there is an important influence of, or link with, maturity. Archetypical age-related epilepsies would include the two best described childhood absence epilepsy syndromes: childhood absence epilepsy (CAE) and juvenile absence epilepsy (JAE). Undoubtedly, the EEG constitutes a very important clinical tool in investigating children with various neurological disorders, particularly epilepsy. Indeed, even when the diagnosis of seizures and epileptic syndromes are primarily clinical, the EEG records provide supportive evidence and helps in the seizure classification (Sundaram et al., 1999).

Absence epilepsy is characterized by generalized non-convulsive epileptic seizures expressed predominantly as disturbances of consciousness with no, or relatively little motor activity Aicardi, 1994, Holmes, 1997. Day dreaming in class and at home in children is a major source of referral to the hospital for consideration of absence seizures. Typical absence seizures take place in otherwise apparently normal children and adults, and have an EEG hallmark of brief ictal and inter-ictal 3–3.5 Hz spike-and-wave discharges with a maximum amplitude over the fronto-rolandic regions. The EEG background is otherwise normal to standard visual analysis in the vast majority of cases Blume and Kaibara, 1999, Wyllie et al., 1991, Aicardi, 1994, Holmes, 1997, Jan, 2002and as consequence is disregarded by the clinicians for diagnostic purposes (Jan, 2002). Given the strong link with development in absence epilepsy in terms of onset and sometimes remission, it may be that the quantitative background EEG differs from those of unaffected children of the same age.

In our previous work (Rosso et al., 2009) we combined: (a) the scalp EEG signal (channel) wavelet decomposition with, (b) a statistical feature selection method based on the Kruskal–Wallis test, in order to select electrodes to differentiate children with childhood absence epilepsy (CAE) and control individuals. The Relative Wavelet Energy (RWE) was used in order to evaluate the inter channel similarity activity matrices, that represent the functional connectivities (Stam, 2005) and the functional activity between electrodes was evaluated in conjunction with the Wootters distance. The goal was to find a clear differentiation of the functional networks between EEG background corresponding to a sample set of child patients presenting CAE and healthy control individuals, and also to identify which set of electrodes provide the maximum differentiation. The analysis procedure provided a clear discrimination between classes, finding patterns of functional networks (cerebral activity) which can be used to diagnose CAE and thus have an impact in future clinical practice.

In the present work, we use a combinatorial optimization approach – based on the (α,β)-Feature Set problem Berretta et al., 2007, Berretta et al., 2008, Cotta et al., 2004– to select the pairs of electrodes (i.e. features). The new approach produces the same, or better levels of separation between the two classes (control and CAE) and at the same time requires a smaller number of features. This is a clear improvement over the statistical-only approach results, as from the mathematical point of view, it is advantageous to explain some condition using less variables, if the level of differentiation is the same or better. That is especially true for classification purposes because the use of less – but still equally meaningful – variables makes the classifier less sensitive to over-fitting on the training set. From the clinical point of view, requiring the use of information from fewer electrodes will allow faster calculation and may produce more accurate predictions of diagnosis. It may help as well to narrow down the search of possible mechanisms involved. The new method was able to almost halve the number of features; improve the differentiation of the classes; and select mostly fronto-central electrodes, as expected from the clinical experience about CAE, thus giving more evidence to this hypothesis.

Section snippets

Materials

The cases selected for this study are the same of our previous paper (Rosso et al., 2009). They were selected by searching in the ‘history’ field of the John Hunter Hospital EEG Database, Newcastle, Australia (Hunter et al., 2005) for children referred for the investigation of absence epilepsy. Inclusion criteria were a history of normal cognition, with absence seizures only, and free of anticonvulsant, sedation and other medications. EEG inclusion criteria were normal EEG background to

Methods

An orthogonal decimated discrete wavelet transform (ODWT) was applied to the EEG signals, in which orthogonal cubic spline functions were used as mother wavelets Rosso and Mairal, 2002, Rosso et al., 2001, Rosso et al., 2006, Rosso, 2007. For a detailed description of the wavelet transformation; wavelet energy, entropy and resolution levels; as well as how the Wootters distance was used to calculated the distances between pairs of electrodes, we refer the reader to Appendix 1.

In the present

Computational results

As mentioned before, the class-information entropy discretization method was applied on the original Wootters distance matrix with 171 pairs of electrodes, resulting in a binary matrix with 74 pairs (see Fig. 1). Regarding these 74 features, we must point that the majority of them contains at least one fronto-central electrode. Just to illustrate that, if we consider the pairs containing fronto-central electrodes only, i.e. Fp1, Fp2, F7, F3, Fz, F4, F8, C3, Cz and C4, we have 37, or exactly

Conclusions

In this study, we extend the results of our previous paper, where a Kruskal–Wallis statistical test was used to select pairs of electrodes to differentiate between CAE patients and control individuals. We have applied a combinatorial optimization approach, based on the (α,β)-Feature Set problem, aiming to find a smaller set of features, i.e. pairs of electrodes, but with the same, or better, ability to separate the two groups of individuals. As a result, we were able to reduce the number of

Acknowledgements

The authors wish to thank Dr. W. Hyslop and Dr. R.L.L. Smith for their help in the selection and clinical evaluation of the EEG recordings used in this work. This work was partially supported by the Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Argentina. O.A. Rosso gratefully acknowledges the support from the Australian Research Council (ARC) Centre of Excellence in Bioinformatics, Australia. The comments of the two anonymous reviewers were responsible for a significant

References (28)

  • R. Berretta et al.

    Integer programming models and algorithms for molecular classification of cancer from microarray data

  • R. Berretta et al.

    Selection of discriminative genes in microarray experiments using mathematical programming

    J Res Pract Inf Technol

    (2007)
  • A. Berretta et al.

    Combinatorial optimization models for finding genetic signature from gene expression

  • W.T. Blume et al.

    Role of electroencephalogram in some pediatric neurological problems

  • Cited by (18)

    • Classification of epilepsy seizure phase using interval type-2 fuzzy support vector machines

      2016, Neurocomputing
      Citation Excerpt :

      Research into the existing literature provides evidence to suggest that the 19 channels of the EEG data vary in importance with regard to classification. It was observed that some of the channels have a lesser impact on the classification of the EEG and the exclusion of these channels has been investigated in [29,30]. Both studies have discovered that some of the channels (F3, Fz, F4, C3 and Cz) are the most significant ones for the classification between the seizure-free and seizure patients and the remaining electrodes are found to have relevant information for the classification between the different seizure phases.

    • Variable weight neural networks and their applications on material surface and epilepsy seizure phase classifications

      2015, Neurocomputing
      Citation Excerpt :

      In this study, we selected the most useful channels by considering different channel combinations. It was found that the 1st, 2nd, 3rd, 4th, 5th, 6th, 11th, 12th, 13th, 14th channels out of the 19 channels contain the most significant information for classification, which compile with the results in [88,89] that channels F3, Fz, F4, C3 and Cz contain the most important information. From each of the chosen channels, a feature vector consisting of time-domain and frequency-domain components is formed.

    • Dynamic characteristics of absence EEG recordings with multiscale permutation entropy analysis

      2013, Epilepsy Research
      Citation Excerpt :

      However, the prediction of sudden and abrupt seizures by detectable dynamic changes in the EEG is still debated in absence patients (Li et al., 2007; Stacey and Litt, 2008). It is challenging to understand the transition of brain activities towards an absence seizure and look for some precursor activities (Crunelli et al., 2011; Rosso et al., 2009a, 2009b; Gupta et al., 2011). Our previous analysis of dynamic changes in the EEG (in Genetic Absence Epilepsy Rats from Strasbourg) has demonstrated that EEG epochs prior to seizures exhibit a higher degree of regularity/predictability than seizure-free EEG epochs, but they present a lower degree than that in seizure EEG epochs (Li et al., 2007; Ouyang et al., 2008).

    View all citing articles on Scopus
    View full text