DOI: 10.1055/s00391693072
Detection of StartleType Epileptic Seizures using Machine Learning Technique
Address for correspondence
Publication History
Received: 26 February 2019
Accepted after revision: 21 May 2019
Publication Date:
31 July 2019 (online)
Abstract
Background Epilepsy is a common neurological disorder characterized by seizures and can lead to lifethreatening consequences. The electroencephalogram (EEG) is a diagnostic test used to analyze brain activity in various neurological conditions including epilepsy and interpreted by the clinician for appropriate diagnosis. However, the process of EEG analysis for diagnosis can be automated using machine learning algorithms (MLAs) to aid the clinician. The objective of the study was to test different algorithms that could be used for the detection of seizures.
Materials and Methods Video EEG (vEEG) was collected from subjects diagnosed to have episodes of seizures. The epilepsy dataset thus obtained was subjected to empirical mode decomposition (EMD) and the signal was decomposed into intrinsic mode functions (IMFs). The first five levels of decomposition were considered for analysis as per the established protocol. Statistical features such as interquartile range (IQR), entropy, and mean absolute deviation (MAD) were extracted from these IMFs.
Results In this study, different MLAs such as nearest neighbor (NN), naïve Bayes (NB), and support vector machines (SVMs) were used to distinguish between normal (interictal) and abnormal (ictal) states. The demonstrated accuracy rates were 97.32% for NN, 99.02% for NB, and 93.75% for SVM.
Conclusion Based on this accuracy and sensitivity, it may be posited that the NB classifier provides significantly better results for the detection of abnormal signals indicating that MLA can detect the seizure with better accuracy.
#
Introduction
Epilepsy is a common neurological disorder characterized by seizure activity in the brain. Globally, over 65 million people suffer from epilepsy. The most commonly occurring seizures (60%) are convulsive in nature; the remaining 40% are nonconvulsive.[1] [2] [3] Startletype epilepsy as a form of reflex seizure in which startle activates the epileptic form discharge when a patient hears loud noises or sudden surprises. This is seen as an unconscious defensive response to sudden and threatening stimulus and is associated with a negative effect. It is a startle reflex reaction which serves as a mechanism to save vital parts. The subjects are specifically sensitive to one sensory modality such as temperature, taste, sound, and pressure. Sudden noise is one of the important triggers for provoked seizure.[4] [5]
Scalp electroencephalogram (EEG) is a primary test to diagnosis epilepsy. Gibbs and Gibbs (1951) described the original EEG patterns. Scalp EEG signals are nonlinear and nonstationary in nature. To analyze the signal, there are descriptive and heuristics methods. Most of the methods are based on domains such as time, frequency, time–frequency, and nonlinear methods.[6] [7] The EEG is a time series data that has nonlinear and nonstationary characteristics. Empirical mode decomposition (EMD) algorithm is applied to the EEG signals to extract intrinsic mode functions (IMFs). The mathematical rationale is developed by Huang.[8] IMFs provide instantaneous frequency as functions of time that furnish good identification of embedded structures. The IMF is based on local properties of the scalp EEG signal which makes the instantaneous frequency useful for analyzing the neuronal activity. This removes the need for spurious harmonics in the nonlinear and nonstationary signals. In classical method, the Fourier transform is used. There is a need for restrictive assumptions and conditions.[9] Some of the variations by the way of smoothing and transforming are based on a priori assumptions. However, in the case of empirical mode decomposition, the output is amenable to statistical methods over a time domain without any restrictive assumptions. These data are suitable for a composite data response involving many variables.
The scalp EEG signal is determined based on the classification and the performance using machine learning algorithms (MLAs).[7] [10] Nonlinear parameters such as interquartile range (IQR) are used to characterize the scalp EEG signal into normal and abnormal.[11] Feature extraction is performed based on IMF output. These features are mean absolute deviation (MAD), IQR, and entropy, and three different algorithms are used for classification.[12] [13]
The objective of this study was to classify the scalp EEG signal using MLA, with better accuracy for startletype epilepsy. In this study, the time–frequency domain is adapted. EMD algorithm is applied to the scalp EEG signals to extract IMFs.
#
Materials and Methods
Study Site
The research was conducted at Global Healthy City, Chennai, under the supervision of an epileptologist. The Neuro laboratory is equipped with digital monitoring videoEEG system (Nicolet One Neuro Diagnostic System; NATUS Neuro, Department of Neurology) commonly known as “epilepsy monitoring unit.” The vEEG data were finally digitized, and the sampling rate of 256 Hz was recorded and stored. The methodology of the study is shown in [Fig. 1].
#
Study Subjects
The subjects were individuals with epilepsy and recruited into the study through their parents as they are adolescents. The scalp vEEG data were acquired from 10 subjects with startletype epilepsy (4 males and 6 females; age range 13–16 years). International standard 10–20 system electrode placement was used, and a bipolar electrode montage was used for the diagnosis of the patient. Each electrode output is bandpass filtered by 0.5 to 100 Hz during recording by setting the low cut and high cut at 0.3 Hz and 70 Hz, respectively.
#
Preprocessing
To remove artifacts in the signal and power line disturbance, the signal is preprocessed with a notch filter with a frequency of 50 Hz.[14]
#
Empirical Mode Decomposition
EMD is applied to the signals and it partitions the signal to elemental IMFs. These IMFs satisfy the following conditions:

Given an input signal [y (t)], the number of zero crossings and extremes must be same or must vary by one.

At any instance, the mean value of the envelope illustrates the local minima. The local maxima must be zero.[15] [16]
The algorithm used for the study is depicted in the accompanying flow chart ([Fig. 2]).
#
Feature Extraction
The IMF coefficients are extracted to provide information that shows the distribution of the signal in time and frequency. The coefficients corresponding to the frequency bands IMF1–IMF5 were analyzed, thus reducing the number of features that represent the signal. Statistical parameters such as entropy, IQR, and MAD of the signal were calculated for the frequency bands IMF1–IMF 5.[17]
Entropy
Entropy is a measure for quantifying the degree of complexity, irregularity of time series, and is used for extracting the information from the signal. It characterizes quantification of the complexity of brain dynamics. The signal uncertainty is measured in terms of the repeatability of the signal amplitude.[18] [19]
#
Interquartile Range
The IQR is a measure of changeability and depends on dividing data into quartiles. Quartiles separate the ordered data into four equal parts. The values that divide each part are termed as the first, second, and third quartiles, denoted by,, and, respectively.[20] [21]
#
Mean Absolute Deviation
The mean absolute deviation (MAD) measures the robustness of the collected quantitative data. The MAD of a dataset is the average distance between each data point and the mean.[21]
#
#
Classification
The research envisages finding the effectiveness of popularly used parametric and nonparametric algorithms to discriminate between normal and abnormal states within a patient and these algorithms are used to classify scalp vEEG signals. A parametric method, that is, support vector machine (SVM), and nonparametric methods, that is, naïve Bayes (NB) and nearest neighbor (NN), are used.[22] The classification is performed using MATLAB functions with linear kernel, while all other parameters are set to default for increased accuracy. Features such as IQR, MAD, and entropy are applied to the classifier SVM, NB, and NN.[23] The SVM is a wellknown supervised learning algorithm for analyzing the data. It is based on the assumption of decision planes that separate a set of objects having different class memberships. For this type the training involves the minimization of error function.
This is subject to the following constraints:
In above equations, w is the vector of coefficients, b is a
constant, and represents parameters for handling nonseparable data inputs. The NB classifier works on the probability theory, where model parameters are estimated according to the distribution of the training data. It computes the probability of the new object, with respect to each class, and then chooses the most probable class. The configuration of the NB function with Gaussian distribution is used.[24] NN is a nonparametric method that searches the k value for optimality; k = 3 increases the accuracy of the classification.[24]
#
Ethics Approval
Institutional Ethical Committee (IEC) approval was obtained before the start of the study. All subjects were explained the study details in their native language and recruited into the study after informed written consent. In the case of subjects younger than 18 years of age, informed written consent was obtained from the parents/guardian.
#
#
Results
From the scalp vEEG signals, it is possible to differentiate between a normal to that of an abnormal signal. The subjects experienced normal as well as abnormal states. The corresponding IMF levels (up to 10) are depicted in [Figs. 3] [4]. The features extracted for the IMF are presented in [Table 1]. From the data it may be posited that the features are clearly different for normal and abnormal EEG signals. Furthermore, these features were provided as inputs to classifiers during the training process. These extracted features were given to classifiers. From these data, we considered 70% for the training set, 20% for testing set, and 10% for validation. The performances of the algorithm are tabulated and shown in [Table 2]. The discriminative potentials of the algorithms were quantified by diagnostic measures of accuracy given by sensitivity, specificity, positive prediction value, and negative prediction value.
#
Discussion
The analysis was undertaken to find the ability of different classifiers to classify data under supervised learning. Two parametric and one parametric algorithm were chosen. The data were subjected to analysis and the output of different algorithms are given in [Table 2]. The comparison indicates that NB provides better indices compared with other methods and thus may be considered for further analysis. Higher values of sensitivity and specificity indicate that there is good discrimination between normal and abnormal signals of the subjects. The higher indices of NB algorithm may be attributed to the nonparametric nature of the algorithm. However, the results may not indicate NB as the test of choice for classification as the sample size was small. Furthermore, falsepositive (FP) value is not possible as all the patients were diagnosed to have seizures, and epileptologist's diagnosis revealed that the subjects had startletype epilepsy, and it is one of the limitations of the study. Some subjects may be classified as not having a disease which is called falsenegative. Sensitivity is defined as a ratio of (TP/[TP + FN]). Since the sensitivity is 99.9%, indicating falsenegative is of very low value. Positive predictive value is defined as the probability of having a state of interest in a subject with a positive result. The high value of positive predictive value indicates the discrimination between patients having the disease and not having the disease is very high. Among those where subject's data were analyzed, the positive predictive value was 94.25. The high value indicates that the algorithm is able to discriminate the abnormal and normal signals. The present study is an explanatory study using data of 10 adolescent subjects. The subjects had episodes of seizures and were on antiepileptic drugs. The small dataset may not be able to correlate the clinical profile with the results. Entropy measured in the present study is likely to be affected with mode misalignments and mode mixing problems for complex data such as brain signals. Diagnostic measures such as likelihood ratio, ROC curve, and Youden index are not considered for this study.
The present study is an explanatory study using data of 10 adolescent subjects. The subjects had episodes of seizures and are on antiepileptic drugs. The small dataset may not be able to correlate clinical profile with the results thus obtained. Entropy measured in the present study is likely to be affected with mode misalignments and mode mixing problems for a complex data such as brain signals. Diagnostic measures such as likelihood ratio, ROC curve, and Youden index are not included and are planned for future studies.
#
Conclusion
Using EMD, we have attempted to extract EEG features to help us distinguish between normal and seizure signals, by calculating MAD, IQR, and entropy. Accordingly, the EEG data are classified into normal and abnormal, using a machine learning technique. The performances of the algorithms are then compared. From this, it is observed that NB has slightly higher accuracy than both SVM and NN. Further studies seem necessary to enhance the detection accuracy and resolution; this could be made possible by including more features and by applying the Random Forest classifier for optimal results.
#
#
Conflict of Interest
None declared.
References
 1 Capovilla G, Kaufman KR, Perucca E, Moshé SL, Arida RM. Epilepsy, seizures, physical exercise, and sports: a report from the ILAE Task Force on Sports and Epilepsy. Epilepsia 2016; 57 (01) 612
 2 Fisher RS, van Emde BoasW, Blume W. et al. Epileptic seizures and epilepsy: definitions proposed by the International League Against Epilepsy (ILAE) and the International Bureau for Epilepsy (IBE). Epilepsia 2005; 46 (04) 470472
 3 Radhakrishnan A. Bridging the treatment gap in epilepsy—is there an emerging trend in the use of newer antiepileptic drugs?. Neurol India 2016; 64 (06) 11401142
 4 Andermann F, Keene DL, Andermann E, Quesney LF. Startle disease or hyperekplexia: further delineation of the syndrome. Brain 1980; 103 (04) 985997
 5 Startle epilepsy (startleinduced epilepsy). Epilepsy Foundation Website. Available at: https://www.epilepsy.com/learn/professionals/aboutepilepsyseizures/reflexseizuresandrelatedepilepticsyndromes/startle
 6 Sanariya K, Garg A, More A, Bansal AR. 6 and 14 Hz positive spikes on scalp electroencephalogram. Int J Epilepsy 2018; 5 (01) 912
 7 Acharya UR, Sree SV, Swapna G, Martis RJ, Suri JS. Automated EEG analysis of epilepsy: a review. Knowledge Base Systems 2013; 45: 147165
 8 Huang NE, Shen Z, Long SR. et al. The empirical mode decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis. Proc R Soc Lond A1998 454 (1971) 903995
 9 Titchmarsh EC. Introduction to the theory of Fourier integrals. Oxford, UK: Oxford University Press; 1948
 10 Pushpa B, Najumnissa J, Keshav DN. Correlation analysis and modeling of EEG–EMG signal for startleinduced seizures. OnLine Journal of Biological Sciences 2017; 2018 18 (01) 1723
 11 Kannathal N, Acharya UR, Lim CM, Sadasivan PK. Characterization of EEG—a comparative study. Comput Methods Programs Biomed 2005; 80 (01) 1723
 12 Das AB, Mohammed IHB. Discrimination and classification of focal and nonfocal EEG signals using entropybased features in the EMDDWT domain. Biomed Signal Process Control 2016; 29: 1121
 13 Martis RJ, Acharya UR, Tan JH. et al. Application of empirical mode decomposition (EMD) for automated detection of epilepsy using EEG signals. Int J Neural Syst 2012; 22 (06) 1250027
 14 Flandrin P, Rilling G, Goncalves P. Empirical mode decomposition as a filter bank. IEEE Signal Process Lett 2004; 11 (02) 112114
 15 Zhang Z, Parhi KK. Lowcomplexity seizure prediction from iEEG/sEEG using spectral power and ratios of spectral power. IEEE Trans Biomed Circuits Syst 2016; 10 (03) 693706
 16 Lewandowski A, Williams DF, Hale P, Wang CMJ, Dienstfrey A. Covariancebased vectornetworkanalyzer uncertainty analysis for timeand frequencydomain measurements. IEEE Trans Microw Theory Tech 2010; 58 (07) 18771886
 17 Tafreshi AK, Ali MN, Omidvarnia AH. Epileptic seizure detection using empirical mode decomposition. Presented at: IEEE international symposium on signal processing and information technology Sarajevo. Serbia: December 16–18 2008
 18 Pincus SM. Approximate entropy as a measure of system complexity. Proc Natl Acad Sci U S A 1991; 88 (06) 22972301
 19 Richman JS, Moorman JR. Physiological timeseries analysis using approximate entropy and sample entropy. Am J Physiol Heart Circ Physiol 2000; 278 (06) H2039H2049
 20 Zhou W, Liu Y, Yuan Q, Li X. Epileptic seizure detection using lacunarity and Bayesian linear discriminant analysis in intracranial EEG. IEEE Trans Biomed Eng 2013; 60 (12) 33753381
 21 Ahammad N, Fathima T, Joseph P. Detection of epileptic seizure event and onset using EEG. BioMed Res Int 2014. doi: http://dx.doi.org/10.1155/2014/450573
 22 Guarnizo C, Delgado E. EEG singlechannel seizure recognition using empirical mode decomposition and normalized mutual information. In: proceedings of IEEE 10th International Conference on Signal Processing (ICSP). Beijing: October 24–28, 20102010
 23 Bernardo JM, Smith AFM. Bayesian Theory. 1st ed.. Wiley; 2000
 24 Parvinnia E. et al. Classification of EEG signals using adaptive weighted distance nearest neighbor algorithm. Journal of King Saud University—Computer and Information Sciences 2014; 26 (01) 16
Address for correspondence
References
 1 Capovilla G, Kaufman KR, Perucca E, Moshé SL, Arida RM. Epilepsy, seizures, physical exercise, and sports: a report from the ILAE Task Force on Sports and Epilepsy. Epilepsia 2016; 57 (01) 612
 2 Fisher RS, van Emde BoasW, Blume W. et al. Epileptic seizures and epilepsy: definitions proposed by the International League Against Epilepsy (ILAE) and the International Bureau for Epilepsy (IBE). Epilepsia 2005; 46 (04) 470472
 3 Radhakrishnan A. Bridging the treatment gap in epilepsy—is there an emerging trend in the use of newer antiepileptic drugs?. Neurol India 2016; 64 (06) 11401142
 4 Andermann F, Keene DL, Andermann E, Quesney LF. Startle disease or hyperekplexia: further delineation of the syndrome. Brain 1980; 103 (04) 985997
 5 Startle epilepsy (startleinduced epilepsy). Epilepsy Foundation Website. Available at: https://www.epilepsy.com/learn/professionals/aboutepilepsyseizures/reflexseizuresandrelatedepilepticsyndromes/startle
 6 Sanariya K, Garg A, More A, Bansal AR. 6 and 14 Hz positive spikes on scalp electroencephalogram. Int J Epilepsy 2018; 5 (01) 912
 7 Acharya UR, Sree SV, Swapna G, Martis RJ, Suri JS. Automated EEG analysis of epilepsy: a review. Knowledge Base Systems 2013; 45: 147165
 8 Huang NE, Shen Z, Long SR. et al. The empirical mode decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis. Proc R Soc Lond A1998 454 (1971) 903995
 9 Titchmarsh EC. Introduction to the theory of Fourier integrals. Oxford, UK: Oxford University Press; 1948
 10 Pushpa B, Najumnissa J, Keshav DN. Correlation analysis and modeling of EEG–EMG signal for startleinduced seizures. OnLine Journal of Biological Sciences 2017; 2018 18 (01) 1723
 11 Kannathal N, Acharya UR, Lim CM, Sadasivan PK. Characterization of EEG—a comparative study. Comput Methods Programs Biomed 2005; 80 (01) 1723
 12 Das AB, Mohammed IHB. Discrimination and classification of focal and nonfocal EEG signals using entropybased features in the EMDDWT domain. Biomed Signal Process Control 2016; 29: 1121
 13 Martis RJ, Acharya UR, Tan JH. et al. Application of empirical mode decomposition (EMD) for automated detection of epilepsy using EEG signals. Int J Neural Syst 2012; 22 (06) 1250027
 14 Flandrin P, Rilling G, Goncalves P. Empirical mode decomposition as a filter bank. IEEE Signal Process Lett 2004; 11 (02) 112114
 15 Zhang Z, Parhi KK. Lowcomplexity seizure prediction from iEEG/sEEG using spectral power and ratios of spectral power. IEEE Trans Biomed Circuits Syst 2016; 10 (03) 693706
 16 Lewandowski A, Williams DF, Hale P, Wang CMJ, Dienstfrey A. Covariancebased vectornetworkanalyzer uncertainty analysis for timeand frequencydomain measurements. IEEE Trans Microw Theory Tech 2010; 58 (07) 18771886
 17 Tafreshi AK, Ali MN, Omidvarnia AH. Epileptic seizure detection using empirical mode decomposition. Presented at: IEEE international symposium on signal processing and information technology Sarajevo. Serbia: December 16–18 2008
 18 Pincus SM. Approximate entropy as a measure of system complexity. Proc Natl Acad Sci U S A 1991; 88 (06) 22972301
 19 Richman JS, Moorman JR. Physiological timeseries analysis using approximate entropy and sample entropy. Am J Physiol Heart Circ Physiol 2000; 278 (06) H2039H2049
 20 Zhou W, Liu Y, Yuan Q, Li X. Epileptic seizure detection using lacunarity and Bayesian linear discriminant analysis in intracranial EEG. IEEE Trans Biomed Eng 2013; 60 (12) 33753381
 21 Ahammad N, Fathima T, Joseph P. Detection of epileptic seizure event and onset using EEG. BioMed Res Int 2014. doi: http://dx.doi.org/10.1155/2014/450573
 22 Guarnizo C, Delgado E. EEG singlechannel seizure recognition using empirical mode decomposition and normalized mutual information. In: proceedings of IEEE 10th International Conference on Signal Processing (ICSP). Beijing: October 24–28, 20102010
 23 Bernardo JM, Smith AFM. Bayesian Theory. 1st ed.. Wiley; 2000
 24 Parvinnia E. et al. Classification of EEG signals using adaptive weighted distance nearest neighbor algorithm. Journal of King Saud University—Computer and Information Sciences 2014; 26 (01) 16