Multivariate linear discrimination of seizures

https://doi.org/10.1016/j.clinph.2004.08.023Get rights and content

Abstract

Objective

To discriminate seizures from interictal dynamics based on multivariate synchrony measures, and to identify dynamics of a pre-seizure state.

Methods

A linear discriminator was constructed from two different measures of synchronization: cross-correlation and phase synchronization. We applied this discriminator to a sequence of seizures recorded from the intracranial EEG of a patient monitored over 6 days.

Results

Surprisingly, we found that this bivariate measure of synchronization was not a reliable seizure discriminator for 7 of 9 seizures. Furthermore, the method did not appear to reliably detect a pre-seizure state. An association between anti-convulsant dosage, frequency of clinical seizures, and discriminator performance was noted.

Conclusions

Using a bivariate measure of synchronization failed to reliably differentiate seizures from non-seizure periods in these data, nor did such methods show reliable detection of a synchronous pre-seizure state. The non-stationary variables of decreasing antiepileptic medication (without available serum concentration measurements), and concomitant increasing seizure frequency contributed to the difficulties in validating a seizure prediction tool on such data.

Significance

The finding that these seizures were not a simple reflection of increasing synchronization in the EEG has important implications. The non-stationary characteristics of human post-implantation intracranial EEG is an inherent limitation of pre-resection data sets.

Introduction

EEG is a complex signal. As a crude measure of electrical activity within the brain, we are aware of no single measure that adequately captures the nature of EEG dynamics or the dynamics of the underlying brain processes. Although there have been extensive univariate analyses of EEG and seizure dynamics, much less exploration of the role of formal multivariate analysis and discrimination has been attempted. In principle, multiple independent measures of relevant brain and seizure dynamics should help to discriminate interictal and ictal states with higher accuracy than any single measure.

Seizures have long been postulated to be a manifestation of excessive synchronization, and the earliest reference to ‘hypersynchronization’ that we are aware of was by Penfield and Jasper (1954). Since then, the presumption that seizures are a manifestation of synchronization has become pervasive (Kandel et al., 1991). Indeed, there have been multiple recent findings consistent with the possibility that a measure of synchronization might indicate changes in EEG dynamics prior to seizure onsets (Jerger et al., 2001, Lehnertz and Elger, 1995, Lehnertz and Elger, 1998, Lerner, 1996, le Van Quyen et al., 1998, le Van Quyen et al., 1999, Lopes da Silva et al., 1989, Mormann et al., 2000, Quian Quiroga et al., 2002).

Multivariate analysis of data has become highly refined in the 20th century (Anderson, 1984, Johnson and Wichern, 1998). In multivariate analysis, independent measures of data can be shown to be more sensitive than univariate discriminators. Furthermore, for correlated variables, a formal multivariate approach can be far more accurate in classifying data than multiple independent comparisons.

We here follow the approach of Flury (1997), and construct a linear discriminator based on measures of cross-correlation and phase synchronization during seizure and non-seizure periods. If seizures were a manifestation of increased synchronization, we hypothesized that a pre-seizure state could consist of a partial increase in synchronization bridging the interictal and seizure periods.

Section snippets

Methods

Data set B from the Department of Epileptology, University of Bonn, was visually inspected in its entirety by a Board Certified Neurologist and Clinical Neurophysiologist (SLW). This data set comprised intracranial recordings from a patient monitored over 6 days. Unequivocal seizure onsets and offsets were identified from the EEG as well as the earliest visually evident electrographic changes. Groups of channels demonstrating similar activity were identified as clusters. Clusters of electrodes

Results

The probability that the bivariate synchrony discriminator classified the training sets correctly varied from 100% accuracy (seizures 1 and 7), to as low as 24% (seizure 9). Quite unexpectedly, this synchrony discriminator failed to achieve over 95% accuracy in classification in 7 of 9 seizures (Table 1).

Fig. 1 shows the number of discriminations (4 of 5 consecutive threshold crossings) for the entire time series. Before 8 seizures (seizures 3, 5, 6, 7, 8, 10, 11, and 12) the frequency of

Discussion

Using a bivariate measure of synchronization failed to reliably differentiate seizures from non-seizure periods in these data, nor did such methods show reliable detection of a synchronous pre-seizure state.

There are several possibilities to account for such discriminator behavior. In using a measure of synchronization to characterize seizures, we make the assumption that a pre-seizure state would reflect a subtle development of the same dynamics that characterize the upcoming seizure. One

Acknowledgements

Work supported by NIH grants K02MH01493 and R01MH50006.

References (22)

  • R.A. Johnson et al.

    Applied multivariate statistical analysis

    (1998)
  • Cited by (42)

    • An efficient seizure prediction method using KNN-based undersampling and linear frequency measures

      2014, Journal of Neuroscience Methods
      Citation Excerpt :

      Common methods of classification process can be either simple binary threshold or machine learning (Mirowski et al., 2009). In some studies a simple binary threshold has been used to classify extracted single variable (Jerger et al., 2005; Mormann et al., 2005) which has limitations on high dimensional features (Mirowski et al., 2009). Recent studies have demonstrated that there is no unique feature offering significant discriminatory power for seizure prediction (Mormann et al., 2007; Schelter Br Timmer and Schulze-Bonhage, 2008).

    • Classification of patterns of EEG synchronization for seizure prediction

      2009, Clinical Neurophysiology
      Citation Excerpt :

      Besides, simple statistical classification not only uses simplistic linear decision boundaries, but also requires reducing the number of variables. A typical shortcoming of an ill-designed binary classification algorithm is illustrated in (Jerger et al., 2005). Hilbert-based phase-locking synchrony is computed for all frequencies without prior band-pass filtering, and cross-correlation is computed for zero delay only.

    • Seizure state detection of temporal lobe seizures by autoregressive spectral analysis of scalp EEG

      2009, Clinical Neurophysiology
      Citation Excerpt :

      The four electrodes exhibiting the largest amplitude change were selected (Table 1). Electrode selection has previously been performed by many researchers (Jerger et al., 2005; Greene et al., 2008). Fig. 2 is an example of the substantial differences in seizure EEG morphologies that can occur in epilepsy sufferers.

    • Seizure Prediction: Its Evolution and Therapeutic Potential

      2009, Blue Books of Neurology
      Citation Excerpt :

      The first problem was addressed in multiple studies in which a number of mathematical measures, starting with the correlation dimension,31 were shown to distinguish preictal and interictal periods using expanded data sets. The second problem was addressed with bivariate and multivariate measures, such as the largest Lyapunov exponent of two channels,32–34 and other methods fusing information from multiple channels and measures over time,4,35,36 simulated neuronal cell bodies,37 and measures for phase synchronization and cross correlation.36,38,39 But none of these studies addressed the lack of large data sets, containing prolonged interictal periods and sufficient numbers of seizures for classifier training and testing.

    • Practical Guide for Biomedical Signals Analysis Using Machine Learning Techniques: A MATLAB Based Approach

      2019, Practical Guide for Biomedical Signals Analysis Using Machine Learning Techniques: A MATLAB Based Approach
    View all citing articles on Scopus
    View full text