Quantitative electroencephalography as a tool for objective neurophysiological seizure diagnosis: A retrospective study

INTRODUCTION

Epilepsy is a common neurological disorder worldwide. As per the WHO 60–70 Million people are affected with epilepsy globally (Fergus et al., 2016; Meng et al., 2017).^[1,2] Epileptic seizures are dysfunction of brain characterized clinically by localized or generalized convulsion and electrophysiologically by synchronous discharge of a cluster of neurons placed in the cerebral cortex as evidenced by abnormal patterns in the electroencephalography. The recent classification of epileptic seizures based on the International League Against Epilepsy (Berg et al., 2010, Chang et al., 2017)^[3,4] classifies seizure disorders into three categories: Focal, generalized, and epileptic spasms. Focal seizures usually originate in one cerebral hemisphere of the brain. Generalized seizures affect both the cerebral hemispheres with seizures originating in the cortex or subcortical area. Generalized seizures could be generalized tonic-clonic, absence, atonic, or myoclonic varieties. A third type of seizure is epileptic spasms which manifest by rapid expansion or flexion of extremity and is usually a feature of childhood epileptic syndromes.

Continuous EEG recording has become an important clinical tool for the assessment and treatment of epilepsy (Gotman, 1989).^[5] The quantitative EEG (QEEG) analysis is one of the proven tools for gauging the effects of substances on the central nervous system in the clinical, experimental, and concerned disciplines (Holler et al., 2018).^[6] Conventionally, frequency parameters are solely used for diagnosis and prediction of seizure activities. After advent of computer based EEG recording and analytical techniques, a lot of advancements have been made in the field of quantitative EEG analytics. Some of the commonly applied newer techniques of quantitative analytics include representation of frequency through wavelet transforms, independent component analysis (ICA), principal component analysis, and non-linear analytics including coherence analysis source localization (Buhimschi et al., 1998; Leman et al., 1999; Maner et al., 2006, and Diab et al., 2010).^[7-10] The energy distribution across various frequency bands has been used recently as one of the sensitive tools for the purpose of seizure detection (Omerhodzic et al., 2010).^[11] EEG offers a real-time assessment of temporal dynamics of brain signals during various activation procedures such as photo-stimulation (flashing of light) and hyperventilation. EEG signals consist of five frequency bands, that is, Gamma (30–60 Hz), Beta (13–30 Hz), Alpha (8–13 Hz), Theta (4–8 Hz), and Delta (0.3–4 Hz).

The present work aims to address the possibility of using QEEG as an objective diagnostic tool for possibility of prospective seizure attacks among the susceptible individuals. A study of energy and entropy levels among seizure and control subjects both during interictal/normal periods as well as during the actual electrophysiological seizure activities (as evidenced by the typical EEG discharges diagnostic of seizure activities without clinical presentation) could possibly yield an EEG based marker that could be used for prospective seizure prediction among the susceptible individuals.

MATERIAL AND METHODS

Work plan

Figure 1 shows the flow chart of the work plan for the study.

Close

Figure 1:

Flow chart of work plan.

Export to PPT

EEG signal analysis

EEG raw signal pre-processing in MATLAB included artifact removal, frequency division, and segmentation. A 50 Hz notch filter has been used to remove interference on the power line from the raw signals. In this study, we have used ICA for the removal of the artifacts and we reconstructed brain waves of alpha, beta, theta, and delta. Usually, two classical wavelet functions are often used to analyze the EEG. One is Daubechies (Db) wavelet and another biorthogonal wavelet; here, we have used Db4 for the waves decomposition (Kalayci and Özdamar, 1995; Petrosian et al., 2000).^[12,13]

ICA uses for the rejection of artifacts

ICA has been used to remove artifacts from electroencephalography data. The signals recorded from the scalp were obtained by a mixture of various other signals such as an eye blink and muscle artifacts. ICA is used to filter all the artifacts that may be originating from those bio-signals originating from eye blink and neuromuscular sources. The raw signal, which was allowed to go through ICA, was made to pass through a notch filter to remove the artifacts caused due to interference of 50 Hz line frequency in the EEG dataset. This was followed by a Bandpass filter (band of 0.3–45 Hz) to get the desired signal in the band of 0.3–45 Hz.

Using ICA the signal has been decomposed into various independent components. Every independent component represented a source of activity. Depending on the source which may be signal or noise, the component has been accepted or rejected.

Let the signal recorded at the scalp is X (x1, x2, x3….). The number of sources lets be S (s1, s2, s3, s4, s5…) and each source is mixed by factors let A (a, b, c…). Hence, this can be represented mathematically: X=A*S.

ICA Implementation in EEGLAB

EEGLAB allows various file formats to be imported into the lab, we have used edf format. Data were first filtered and then the direct current component was removed from the signal. ICA algorithm was implemented. Noisy component was removed from the signal based on the scalp map. Figure 2 describes the process of artifact removal by ICA in MATLAB through the scalp map. The signal is now ready for feature extraction.

Close

Figure 2:

Scalp map with (a) and without (b) eye blink artifact.

Export to PPT

Feature extraction

Feature extraction is the process of extracting useful information from the signal.

RESULTS

A total of 115 subject’s EEG data have been recorded during the period of reference, out of which 19 subjects with generalized seizure have been taken as control and 11 subjects have been taken as study subject following inclusion and exclusion criteria. The mean age of experimental group was 10.1 years where as that of control group subjects were 11 years [Figure 3]. Rest of the data have been discarded due to presence of one or more biases.

Close

Figure 3:

Average age of experimental and control.

Export to PPT

Further, data have been analyzed for energy of alpha, beta, theta, and delta waves. Mean energy level for alpha waves among control subjects was 9.78 ± 7.65 and that of study subject was 3.65 ± 2.38. Similarly, mean beta energy was 8.77 ± 6.10 for control and 4.83 ± 1.66 for test subjects. A significant reduction of alpha (P = 0.04) and beta waves (P = 0.004) was observed in the seizure group. The mean theta activity has been found slightly increased in in the experimental group in comparison to the control group, whereas the mean delta energy level showed slight decreased level, but both are statistically not significant [Table 1 and Figure 4].

Close

Figure 4:

Mean energy of α showing in figure (a), β showing in figure (b), θ showing in figure (c), and δ showing in figure (d), between normal and seizure with SD.

Export to PPT

DISCUSSION

Electroencephalography is one of the most important tools for seizure detection used across the globe. Despite of lots of technological advancements, the EEG interpretation is largely subjective and diagnosis of seizure is mostly based on the subjective observations and clinical acumen of the clinician. The present study aimed to investigate the possibility of objective diagnosis of seizure possibly from the background electroencephalographic characteristics of the prospective seizure patients even before actual seizure events.

The pre-recorded EEG data and subject’s clinical information have been analyzed following due protocol to make a comparison between the electroencephalographic patterns of non-seizure subjects and EEG of subjects with history of recent seizure. From the results of the study, the following observations could be figured out.

• The mean energy level of control subjects was higher than that of seizure patient even during non-electrophysiological seizure, but it was statistically insignificant

• There was no statistically significant difference between the energy levels during pre-seizure and post-seizure electrical activity in the study subjects

• Energy level in theta and delta band of control and study subject did not had statistically significant difference

• The energy level for alpha and beta frequency band in seizure group subjects was much lower in comparison to the control group subjects when the EEG segments taken for analysis was from the period of electrical seizure activities

• Entropy, which signifies randomness of the brain electrical activity, did not show any significant difference between the control subject and study subjects.

From the above observations, it is apparent that out of the two important parameters of the electroencephalographic signals studied, that is, energy and entropy, energy has come out as most important discriminator between the seizure and non-seizure categories. The energy of the filtered EEG signals has the optimal discriminative capability under the extreme energy difference (EED) criterion, and thus, EED can be considered as a feature extractor (Li and Sun., 2008).^[15] The previous study reported that seizure EEG changes are indicative of increased excitability, distorted synchrony, and decreased signal complexity for the period of a pre-seizure (Mormann et al., 2006 and Zaveri et al., 2009).^[16,17] As the EEG signal is a kind of non-linear signal, its dispersion characteristics are complex and difficult to estimate.^[18] The seizure subset of the data has shown a decreased mean level of energy in comparison to the non-seizure subset of individuals which were contrary to the common expectation and scientific believes. This may be attributed to the complex neural behaviors related to dispersion. The other parameter, the entropy, which is a measure of complexity did not show any significant difference between seizure and non-seizure individuals. This indicates that there is no any significant change in the signal characteristics of the background electrical activities among seizure patients in comparison to normal individuals. Another interesting and rather robust methodology for objective tool for seizure detection could be the multichannel surface EEG-based patient non-specific seizure detection approach. This approach includes DMD power, power spectral density, variance, and KFD extracted from WPD coefficients measurement for detecting seizure. In any continuous EEG signal seizure, event may be observed for seconds as compared to many minutes of non-seizure EEG signals (Dash et al., 2019).^[19] Hence, cluster-based seizure classification is preferred. Based on the observation, a future study with bigger database could be attempted with combined approach.

CONCLUSION

This a small retrospective study on the pre-recorded data to find if the normal background characteristics of an EEG signal be used for seizure detection and prediction. The study indicated changes in the energy level between the EEG characteristics of non-seizure subjects and that of seizure individuals. Further, a study will be required for better understanding the underlying mechanism.