Short communicationAn accurate system to distinguish between normal and abnormal electroencephalogram records with epileptic seizure free intervals
Introduction
Epilepsy is a chronic disorder that happens in human brain and is associated with an increased vulnerability to seizures. The latter are in fact due to a disturbance in the electrical activity of the brain. When brain electrical signals are disrupted, memory can be affected and the patient may hurt himself. The analysis of the electroencephalogram (EEG) records used to measure the electrical activity in the brain is a common approach to help the physician in the diagnosis of seizures. In this regard, in recent years, various and interesting automated systems have been proposed in the literature to assist physicians and to improve the diagnosis outcome, while reducing duration of consultation [1], [2], [3], [4], [5], [6], [7], [8].
For instance, the authors in [1] employed a genetic algorithm to automatically obtain Fourier transform based characteristics and used k-nearest neighbour (k-NN) to classify segments with seizure activity against non-seizure segments either belonging to an epileptic subject or to a healthy one. The proposed system achieved 98.53% accuracy. In a similar problem, the authors in [2] employed an approach based on bandwidths of intrinsic mode functions fed to least squares support vector machine having Morlet wavelet as kernel function. The maximum (minimum) classification accuracy among ten classification results for seizure against non-seizure EEG signals was 100% (95.50%) using second intrinsic mode function based bandwidths. In a subsequent study [3], fuzzy Sugeno classifier trained with four different entropy measures reached 98.1% accuracy when used to classify healthy, pre-ictal, and ictal EEG records. An automated system that combines smoothed Hilbert-Huang transform and root mean square feature was proposed in [4] to detect seizure in epileptic brain. The system achieved 90.72% and 8.23% in terms of sensitivity and false discovery rate respectively. In an interesting study [5], fractional linear prediction was employed to model error energy in ictal and seizure-free EEG signals, and support vector machine classifier with radial basis function kernel for classification. The system achieved 96% sensitivity, 95% specificity, and 95.33% accuracy. The authors in [6] used statistical features extracted from time-frequency image in Hilbert-Huang domain and support vector machine to detect epileptic seizures in EEG records. The proposed system yielded to an average accuracy of 99.125% (97.5–100%). In [7], the proposed system for automatic seizure detection in EEG signals used spectral entropies and energy all obtained from Hilbert marginal spectrum to train support vector machine classifier. In the case of classification of healthy against seizure segments, the system achieved 99.85% accuracy and 98.80% in the case of classification of segments with seizure activity against non-seizure signals either belonging to an epileptic subject or to a healthy one. More recently [8], Gabor filter was applied to EEG signals to obtain four filter responses from which local binary pattern based traits are extracted. Finally, the latter were fed to the nearest neighbour algorithm to decide whether the original EEG signal belongs to seizure or seizure-free category. Following 10-fold cross validation protocol, the presented system achieved an average accuracy of 98.33%.
In this study, we propose a new automated system for classification of EEG signals. In particular, we aim to design an accurate, simple, and fast automated system to effectively distinguish between healthy and seizure free segments belonging to an epileptic patient. In fact, there are three reasons that motivate our study. First, it can be really difficult to diagnosis epilepsy since there is no evident sign a person has epilepsy, unless seizure occurs. Second, the investigation of this issue is definitely interesting in clinical applications. Third, prior works [1], [2], [3], [4], [5], [6], [7], [8] have not focused on this problem.
In order to tackle this problem, we propose an automated system composed of two stages. In the first stage, multifractal of the EEG signal are computed at different time scales based on generalized Hurst exponent (GHE) [9]. Then, in the second stage, support vector machine (SVM) [10] classifier is trained with computed generalized Hurst exponent estimates to distinguish between healthy EEG signals and EEG signals with seizure free segments belonging to an epileptic patient. In this work, we rely on fractality assessment of the EEG signal at different time scales to better capture its dynamics; particularly, self-similarity. For instance, measuring the long-range dependence in EEG signal at different scales could help revealing its general power-law correlations in its short and long term variations separately. Therefore, the underlying stochastic behaviour of the EEG signal could be better characterized by the level of persistence in its short and long term variations. Obviously, such intrinsic characteristics are expected to be distinct across healthy and seizure free segments belonging to an epileptic patient. In addition, the GHE provides reliable statistics [11] and was found to be effective in characterization and classification of brain magnetic resonance images [12], [13], [14], [15], [16]. Finally, the SVM classifier is chosen thanks to its ability to execute the principle of structural risk minimization; in order to avoid local minima; and capability to generalize the results [10]. Further, it was found to be effective in classification of EEG signal in the context of epilepsy detection [2], [5], [6] and also in other biomedical engineering applications [17], [18].
The rest of our work follows. Section 2 presents methods, Section 3 provides results, and Section 4 discusses our work. Finally, our study is concluded in Section 5.
Section snippets
Methods
The goal of employing the GHE [9] in this study is to measure self-similarity at different scales in healthy EEG signals and seizure free segments from EEG signals belonging to an epileptic patient. Then, the computed GHE estimates (for instance; H(q) for q = −10,…,10) are fed to the SVM [10] to distinguish between healthy and seizure free signals. Finally, the performance of the proposed system is evaluated based on accuracy, sensitivity, and specificity. In order to obtain robust and general
Data and results
In our work, the proposed automated diagnosis system is tested on publically available EEG dataset taken from the department of epileptology, University of Bonn [19]. This dataset was also studies in [1], [2], [3], [5], [6], [7], [8] and consists of five sets: A (healthy records), B (healthy records), C and D (activity measured during seizure free intervals from epileptic patients), and E (seizure activity). Each set consists of 100 signals of EEGs with 4097 samples with duration 23.6 s recorded
Discussion
Bear in mind that in our investigation, the dataset size is sufficiently large (200 healthy records and 200 unhealthy records), and the extracted features represented by generalized Hurst exponents (GHE) were found to be statistically different across healthy and unhealthy EEG records after applying parametric and nonparametric statistical tests; namely t-test, F-test, K-W test, and K-S test. Additionally, the SVM which is a robust classifier based on empirical risk minimization is employed for
Conclusion
Fractal properties of EEG signals are estimated at different scales for better description of their dynamics by using the generalized Hurst exponent methods. Parametric and nonparametric statistical tests showed strong evidence that generalized Hurst exponent estimates are statistically different among healthy and epileptic EEG signals with seizure free intervals. In addition, support vector machine classifier trained with generalized Hurst exponent estimates yielded to perfect accuracy. As
References (21)
- et al.
Automated detection and classification of high frequency oscillations (HFOs) in human intracereberal EEG
Biomed. Signal Process. Control
(2013) - et al.
Classification of ictal and seizure-free EEG signals using fractional linear prediction
Biomed. Signal Process. Control
(2014) - et al.
Classification of seizure based on the time-frequency image of EEG signals using HHT and SVM
Biomed. Signal Process. Control
(2014) - et al.
Hilbert marginal spectrum analysis for automatic seizure detection in EEG signals
Biomed. Signal Process. Control
(2015) - et al.
Classification of seizure and seizure-free EEG signals using local binary patterns
Biomed. Signal Process. Control
(2015) - et al.
On Hurst exponent estimation under heavy-tailed distributions
Physica A
(2010) Glioma detection based on multi-fractal features of segmented brain MRI by particle swarm optimization techniques
Biomed. Signal Process. Control
(2017)Parkinson’s disease detection based on dysphonia measurements
Physica A
(2017)Image characterization by fractal descriptors in variational mode decomposition domain: application to brain magnetic resonance
Physica A
(2016)- et al.
Variational mode decomposition based approach for accurate classification of color fundus images with hemorrhages
Opt. Laser Technol.
(2017)
Cited by (30)
Domain adaptation for epileptic EEG classification using adversarial learning and Riemannian manifold
2022, Biomedical Signal Processing and ControlCitation Excerpt :Each sample is an EEG segment produced by a sliding time window. The selection of seizure for training refers to the idea of the leave-one-out cross-validation (LOOCV) approach [59,60]. Such labeling method provides feasibility for popularization and application.
Recent Trends in Computer-Aided Diagnostic Systems for Skin Diseases: Theory, Implementation, and Analysis
2021, Recent Trends in Computer-aided Diagnostic Systems for Skin Diseases: Theory, Implementation, and AnalysisTime domain implementation of pediatric epileptic seizure detection system for enhancing the performance of detection and easy monitoring of pediatric patients
2020, Biomedical Signal Processing and ControlCitation Excerpt :Apart from wavelet transform, translation invariance of the wavelet has also been performed which gives stability to scaling and has been used to extract features from EEG signals [25]. Various statistical parameters like standard deviation, mean, variance [11], entropies like approximate entropy [12,13], chaotic parameters like Lyapunov exponent [6,14], Hurst exponent [15,30] and correlation dimension [14,16] have been used as features for classification of the EEG signal. Apart from using different signal processing methods and feature extraction techniques, classifiers also play an important part in distinct classification of the EEG signals.
Does the “ice-breaking” of South and North Korea affect the South Korean financial market?
2020, Chaos, Solitons and FractalsCitation Excerpt :By using multifractality, generalized Hurst exponents are calculated for distinguishing seizure and seizure-Free intervals of intracranial electroencephalogram (EEG) signals from epileptic patients [34]. The generalized Hurst exponents are also taken to classify the normal and abnormal EEG records in [35], and estimate differentiate EEG signals of healthy and epileptic patients in [36]. Li et al. [37] used the MF-DFA to analyze the human heart rate variability during exercise, which can provide useful information for athletes to choose the best sports time and accelerate the development of sports medicine.