A new feature for the classification of non-stationary signals based on the direction of signal energy in the time–frequency domain

Highlights

• A new feature based on the direction of signal energy in time-frequency domain has been defined.

• The proposed feature outperforms all other features in terms of area under receiver operator characteristics curve.

• The proposed classification methodology outperforms state of the art.

Abstract

The detection of seizure activity in electroencephalogram (EEG) segments is very important for the classification and localization of epileptic seizures. The evolution of a seizure in an EEG usually appears as a train of non-uniformly spaced spikes and/or as piecewise linear frequency modulated signals. If a seizure is present, then the energy of the EEG is concentrated along the time axis and the frequency axis in the time–frequency plane. However, in the absence of a seizure, the energy of the EEG signal is uniformly distributed along all directions in the time–frequency plane. Based on this observation, we propose a new approach for the detection of a seizure. In this paper, we develop a new feature that exploits the direction of the energy of the signal in the time–frequency domain to distinguish between seizures and non-seizures in an EEG. Our experimental results indicate the superiority of the proposed approach over other conventional time–frequency approaches; for example, the proposed feature set achieves a classification accuracy of 98.25% by only using five features.

Introduction

Epilepsy is one of the most common disorders of the human nervous system and it is characterised by recurring seizures. Epileptic seizures are defined as temporary and brief disturbances in the routine electrical activity of the human brain. When too many nerve cells in the brain fire very quickly, they cause an electrical storm [36]. Because of their non-invasive nature, EEG recordings are frequently used to monitor the brain's activities. Consequently, brain abnormalities, such as seizures, can be diagnosed on the basis of deviations of the EEG recording from the normal pattern. However, the manual detection of seizures in EEG recordings requires a neurologist to be available at all times of the day and night. Therefore, there is considerable interest in the automatic detection of seizures in EEG recordings using a computer.

Machine learning algorithms are normally used in the automatic detection of seizures, which can be classified into the following categories:

• Model-based approaches: These approaches are based on the assumption that the signal is generated by a mathematical model. The most commonly used signal models are: the auto-regressive moving average model, the moving average model, the auto-regressive model, and the linear predictive model. Linear predictive models have been successfully employed for the detection of seizures [5,15,18].

• Feature-extraction approaches: Feature extraction is an operation that transforms an EEG recording into a feature vector, which represents the pattern in a reduced dimensional space. These approaches involve two stages of processing: the first stage extracts the set of features from the signal and the second stage uses the extracted features to train a classifier. The features can be extracted from the time-domain representation [30], the frequency-domain representation [16], or from joint time-frequency (TF) representations [8,19,34,37].

The extracted features are then employed by classifiers to detect the presence of a seizure in a recording. The most frequently used classifiers are: support vector machines [17], least squares support vector machine, artificial neural networks [22], genetic algorithms [19], random forest classifiers [1], and extreme learning machines [38].

EEG signals have non-stationary characteristics. This means that these signals are difficult to analyse in the time domain alone or in the frequency domain alone because their spectral content changes with time. Time–Frequency distributions (TFDs) are frequently employed for the analysis of non-stationary signals. Recently, there has been considerable interest in using TFDs for the automatic detection of changes or abnormalities in non-stationary signals, such as the detection of seizures in an EEG [21]. Although TFDs are rich in information, analysing a non-stationary signal in the TF domain increases the dimension of the signal and this complicates the signal classification problem due to the curse of dimensionality. The dimension of a TFD can be reduced by methods such as principal component analysis and singular value decomposition [9,25]. Another approach is to extract a small set of features from the TFD, which represents the same information required for discriminating between different classes of data. TF features can be broadly classified into the following categories:

• Instantaneous-frequency features: Amplitude-modulation frequency-modulation (AM-FM) is a widely used model for the analysis of non-stationary signals. AM-FM signals are completely described by parameters such as the instantaneous frequency, instantaneous amplitude, and the total number of components. Therefore, instantaneous frequency (IF) based features are frequently used in the detection of seizures [20,24,31,35].

• The IF can be estimated from quadratic TFDs (QTFDs) by applying sophisticated image processing algorithms [7,32] or by using signal decomposition methods, such as empirical mode decomposition or matching pursuit [27]. However, the main limitation of this approach is that estimating the IF of multi-component signals is a challenging task, especially when the IFs of the signals cross each other. Therefore, these methods are only applicable to a signal whose IFs do not overlap.

• Image-based features: A TFD is a 2D representation, which is just like an image. Consequently, image related features, such as Harlic features and local binary patterns, can be extracted from a TFD [6,23,33]. These methods have shown good performance in the detection of seizure activity and in the classification of the stages of sleep [2,12].

• TF features obtained by extending time-only or frequency only features: These approaches derive TF features by extending the conventional time domain and frequency domain features to the joint TF domain. Previous studies have shown that these features lead to better performance when compared to original time-only or frequency-only features [9].

In this study, a new approach to the extraction of features from a TF representation, which is based on the direction of signal energy in the TF domain, is developed and its efficacy is demonstrated using a newborn EEG database. The proposed approach first uses differentiation, which whitens the EEG background and enhances the spike-like characteristics of a seizure. A TF analysis of the EEG signal (after filtering) reveals that in the case of an EEG seizure, most of the signal energy is distributed along the frequency axis due to the spike-like signatures or along the time axis due to the tone-like characteristics. Meanwhile, in the case of the background, the signal energy is uniformly distributed along all directions [20,35]. Hence, the energy concentration of the signal distributed along the time axis or frequency axis, when normalised by the total signal energy, is considered as a feature for signal classification. Our experimental results indicate that the proposed feature outperforms in terms of the AUC, time-only, frequency-only, and commonly used TF features. It is also demonstrated that a classification system based on the proposed features achieves a total accuracy of 98.25%, which is 0.5% greater than the accuracy obtained by a combination of TF signal and image related algorithms.