Elsevier

Pattern Recognition Letters

Volume 133, May 2020, Pages 202-209
Pattern Recognition Letters

Feature extraction from EEG spectrograms for epileptic seizure detection

https://doi.org/10.1016/j.patrec.2020.03.006Get rights and content

Highlights

  • Three different approaches to extracting features from EEG spectrograms are proposed with relevant results.

  • Type window and overlapping from short time Fourier Transform is justified based on spectrum energy.

  • A method for selecting the window length from short-time Fourier transform based on signal frequency is proposed.

  • For the two-classes problem case the here proposed method is competitive in accuracy when compared against the literature.

  • This approach requires fewer features than others.

Abstract

Identification of EEG signals is currently an open problem where performance analysis in terms of accuracy is relevant in several fields, such as biomedicine and brain computer interfaces. Nevertheless, performance depends on the feature extraction phase, where the aim is to find relevant patterns related to different mental activities. Thus, in this work, an approach to extract features from EEG signals is proposed based on spectrograms: Firstly, STFT is applied to EEG to obtain time-frequency representations, where parameters such as window length and type are experimented based on the EEG signal frequency. After that, spectral peaks are found to be used as reference in order to obtain descriptors per spectrogram. Three ways for extracting features from EEG are presented, the first based on frequency and surfaces, the second using K-means to extract features and the adaptation of local ternary pattern, and finally, a third using maximum peaks. The extracted descriptors are evaluated by means of a multilayer perceptron, support vector machines, and k-nearest neighbors. The proposed approach was evaluated using the dataset from Bonn University, identifying a healthy person and an epileptic attack classes as main task. According to the experimental results, the proposed method obtains acceptable accuracy (100%) in several cases by considering fewer features than those extracted by other related works.

Introduction

Epilepsy is one of the most common neurological disorders, affecting individuals around the world. It is caused by temporary and unanticipated electrical disturbances in the brain, resulting in altered behaviours such as a temporary loss of consciousness, respiration, or memory [27] [23].

Electroencephalography (EEG) is a measurement of electrical activities in the human brain along the scalp. This provides information about how the brain works over time and is used by physicians and scientists to analyze and study brain functions and diagnose neurological disorders such as epilepsy, Alzheimer’s disease, brain tumors, head injuries, sleep disorders, dementia, among others [13], [27].

EEG is an important tool for the diagnosis and analysis of epilepsy, and research into patient management is ongoing, mainly due to the manifestation of abnormal cortical excitability that underlies epilepsy [27][31]. An important task in diagnosis and treatment is the detection of seizures in EEGs [27]; this task is commonly performed manually because it requires highly-skilled neurophysiologists and long recordings time [23] [30]. Two categories of electrical activity can be analyzed in EEGs: ictal (during an epileptic seizure) and interictal (between seizures) [27], and tools to detect these categories can help neurophysiologists to detect seizures [23].

EEG signals contain a large amount of data in addition to complex random and non-stationary signals, and visually scanning these is time-consuming and imprecise [3]. Hence, EEG signal analysis represents a problem that can be observed upon classification; the key to its solution is to obtain useful information from the EEG through different methods [26].

Different methodologies have been proposed to identify epileptic seizures in EEG signals based on frequency, time, wavelet transforms, and Gabor filters [4], [11], [23]. However, EEG signals are non-stationary, which involves specific aspects when employing techniques based on frequency or time features extracted [4], as they do not provide enough information [3]. Time-frequency(t-f) techniques are powerful tools that decompose signals such as EEGs into both time and frequency, allowing the analysis of non-stationary signals. Implementing different distributions in time-frequency in EEG signals allow viewing them as images: therefore, several features can be obtained directly from the resulting mappings. These techniques have shown good results in terms of accuracy for different applications [22]. Short Time Fourier Transform (STFT) is a distribution commonly used in different works with distinct objectives and with acceptable classification results. STFT is a time-frequency representation commonly used in signal analysis, as well as in digital image processing, voice processing, biology, and medicine [14], Electroencephalographic (EEG) signal analysis (STFT allows a representation in time-frequency of the EEGs, which shows a different visualization for analysis).

A study carried out in 2017 on features obtained from EEG spectrograms showed the relevance of STFT for classifying EEG signals [20], where different features from several works are described as energy, power spectrum density, local binary pattern codes, and gradient histograms, among others, used to different applications as in epilepsy, drowsiness, alcoholism, mental tasks, among others. STFT has been used with different goals, and its performance (which depends on the features extracted) has been acceptable. However, some works do not show the performance modifying parameters of STFT or their justification. The results could depend on the parameters used in the transform, which modify the spectrogram resolution and the features obtained; one of the most important parameter is known as window length, which determines the spectrogram resolution. Although there are different STFT parameters that affect spectrogram resolution, the main is the window length. In [14] it is stated that the Wavelet Transform (WT) is an alternative to this weakness and Nimmy et al., proposed a solution to alleviate a smearing of energy observed in the STFT [16]; however, using defined parameters, the resolution and the results can be improved.

The approach proposed in this paper consists of the application of the STFT to EEG signals; then, the STFT parameters, such as window length, window type, and overlapping are fixed; these are justified and not only by experimentation. On one side, a window size represents a complicated parameter to propose because of this defines the spectrogram resolution, so that, a formula is presented, which is based on the minimum frequency of any EEG signal; this guarantees a good resolution in time and frequency of the spectrogram, obtaining features with a better time-frequency resolution through three different approaches. In the first, the use of maximum frequency responses per window as spectral peaks is proposed, and from these the local ternary pattern (LTP) is adapted to obtain the features in terms of time-frequency and energy. A second approach extracts features from spectral peaks using K-means in order to obtain relevant points from the centroids, peaks volume, peaks area and total energy. Finally, the LTP combination with the smallest peak also is proposed, giving the best results. In general,this paper presents a methodological alternative for the identification of epilepsy, as well as its application for other types of signals based on the parameterization of the STFT and the proposed approaches, which have not yet been analyzed in the literature [20]. For classification, multilayer perceptron (MLP), support vector machines (SVM) and k-nearest neighbours (kNN) were used. According to different experiments carried out to evaluate this proposal, in some cases 100% accuracy was obtained for the different classifiers.

This paper is organized as follows: Section 2 describes the methodology proposed; Section 3 shows the corresponding results. Finally, Section 4 presents conclusions and future work.

Section snippets

Proposed approach

The proposed approach to extract features from EEG signals, specifically from healthy people, and undergoing epileptic seizures, is shown in Fig 1.: Initially, STFT is applied to EEG signals from the Bonn Dataset to obtain time-frequency representations, and spectral peaks are obtained from these; some are selected as features to train MLP, SVM, and kNN classifiers to identify EEG signals, showing the results in terms of accuracy.

Results

The results of three experiments are shown and described in this section. Initially, the proposed FOSP approach is applied, after which each of the three methods described in Section 2.3 is used to extract features, which are evaluated using three different classifiers (MLP, SVM, and kNN), showing the results in terms of accuracy.

Conclusions

The parameters that affect the resolution of the STFT were analyzed: window type, length, and overlap. An acceptable window was found with respect to the spectral lobes features over time, since it is similar both in time and in frequency to its spectrum, being a variant of the raised cosine; the lengths were found with respect to the frequency of the waves; finally, an overlapping was used, as seen in the literature. These parameters allow more energy in the spectral peaks, and these are what

Future work

As future work, the methodology based on maximum peaks will be tested with using other datasets to compare the different classes of epilepsy as well as interictal and preictal epilepsy a multi class classification, as these are some of the most important issues in the field.

On the other hand, our approach can be improved, because when spectral peaks are found, neighbors peak are not taken into account, only the maximum is obtained from the spectrogram columns, however, based on maximum level

Acknowledgments

Authors thank the reviewers for their valuable comments and the first author would like to express his gratitude for support from CONACyT-México through scholarship number 457637.

References (33)

  • D. Bansal et al.

    Chapter 5 - cognitive analysis: frequency domain

  • B. Boashash

    Chapter 2 - heuristic formulation of time-frequency distributions

    Time-Frequency Signal Analysis and Processing (Second Edition)

    (2016)
  • M.M. Goodwin, The STFT, Sinusoidal Models, and Speech Modification, Springer Berlin Heidelberg, Berlin, Heidelberg,...
  • C.-S. Huang et al.

    Knowledge-based identification of sleep stages based on two forehead electroencephalogram channels

    Frontiers in Neuroscience

    (2014)
  • P. Kovacs et al.

    On application of rational discrete short time fourier transform in epileptic seizure classification

    2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    (2014)
  • T.S. Kumar et al.

    Classification of seizure and seizure-free EEG signals using local binary patterns

    Biomed. Signal Process. Control

    (2015)
  • Cited by (0)

    View full text