An Epileptic Seizure Detection Technique Using EEG Signals with Mobile Application Development

Abstract

Epileptic seizure detection classification distinguishes between epileptic and non-epileptic signals and is an important step that can aid doctors in diagnosing and treating epileptic seizures. In this paper, we studied the existing epileptic seizure detection methods in terms of challenges and processes developed based on electroencephalograph (EEG) signals. To identify the research deficiencies and provide a feasible solution, we surveyed the existing techniques at each phase, including signal acquisition, pre-processing, feature extraction, and classification. Most previous and current research efforts have used traditional features and decomposing techniques. Therefore, in this paper, we introduced an enhanced and efficient epileptic seizure technique using EEG signals, for which we also developed a mobile application for monitoring the classification of EEG signals. The application triggers notifications to all associated users and sends a visual notification should an EEG signal be classified as epileptic. In this research, we have used publicly available EEG data from the University of Bonn. Our proposed method achieved an average accuracy of 98% by utilizing different machine-learning algorithms for classification, and it has outperformed recently published studies. Though there have been other mobile applications for epileptic seizure detection, they have been based on motion and falling detection, as opposed to ours, which was developed based on EEG classification. Our proposed method will have an impact in the medical field, particularly for epilepsy seizure monitoring as well as in the Human–Computer Interaction fields, majorly in the Brain–Computer Interaction (BCI) applications.

1. Introduction

1.1. Background

Epilepsy is a brain disorder that leads to seizures, which refers to abnormal or excessive neural activities in the brain [1,2]. Epilepsy is a serious neurological disease that is referred to as a disorder of the central nervous system and is characterized by the loss of consciousness and convulsions [3,4]. Epilepsy is one of the most known neurological disorders that affect young children and older adults between the ages of 65 and 70 years old. More than 1% of the world’s population has been diagnosed with epilepsy [5,6]. Most epileptic cases can be controlled by anti-epileptic medication or surgery, but a minority of epilepsy patients do not respond to typical therapeutic treatments and surgical interventions [7,8,9]. Major causes of epilepsy include head injuries, genetic and infectious illnesses, maldevelopment of the brain, brain tumors, stroke, metabolic problems, and lead poisoning, though, for some patients, the cause of their epilepsy is unknown [7]. According to a survey, at least 65 million people worldwide live with epilepsy, and each year, approximately 2.4 million people are diagnosed with epilepsy, mainly in developing countries; 150,000 of these new epileptic patients are in the United States [3]. According to studies, one out of twenty-six people worldwide are affected by epilepsy [10]. Therefore, developing an automated technique that can be implemented on a smart device would be of great importance in detecting and analyzing epileptic seizures as well as other epilepsy-related conditions.

There are several anti-epileptic medications on the market that help control the incidence of seizures but cannot cure the disease. In addition, there are side effects associated with these medications, such as tiredness, headaches, reoccurring thoughts, memory loss, and daydreaming. Surgical intervention is another approach for epilepsy treatment; however, there are significant risks involved as well as critical factors to consider, such as being able to identify the focus of the seizure so that the area of the brain involved can be safely removed without resulting in a significant deficit [10].

There are many ways to measure the activity of the human brain, and the methods can be intrusive or non-intrusive. While some techniques study the behavior of the patient, others measure the signals of the eyes, eyelid movements, the face, and even polysomnographic (PSG) recordings. However, epileptic seizure detection has typically been conducted based on physiological recordings, which include electroencephalograms (EEG), electrooculograms (EOG), electromyograms (EMG), and electrocardiograms (ECG) [11,12].

The recording of seizures and spikes is of primary importance in the evaluation of epileptic patients. Since epileptic seizure detection by PSG requires the patient to attach electrodes to their scalp and remain in the hospital for several hours to record the signals, the process is complex as well as expensive [11].

Considering these challenges, an automatic detection method for epileptic seizures would be of great assistance during the long-term monitoring of epileptic patients. In addition, it could reduce the complexity and improve classification accuracy in epileptic seizure detection. Furthermore, it could also improve the diagnosis and disease management of brain disorders. According to a survey, EEGs are a critical tool for research and diagnosis, especially in detecting epileptic seizures and brain disorder analysis. EEGs are non-invasive and do not involve X-rays, radiation, or injections. EEGs have been used in detecting and analyzing brain signals for many years and are considered safe [6]. The electrodes record EEG signals from the brain without producing any sensation. Minimal side effects include slight redness/skin irritation at the contact site of the electrode, but this typically wears off in a few hours [12]. As a result, EEGs have been widely used as a clinical tool to measure brain activity [13]. In this paper, we have not only devised methods that outperform many existing techniques in epileptic seizure detection from EEG signal analysis but also implemented the ideas as a mobile development application and a functional framework.

The computer-based monitoring systems can be effectively enhanced by utilizing the application of Internet of Things (IoT) technology [14]. Jozsef Katona et al. developed a technique in 2019 in the field of BCI. They were able to develop and control a wireless robot utilizing an EEG signal that was recorded directly from the brain via a MindWave EEG headset in relation to attention. The EEG went through pre-processing, feature extraction and classification, which was then communicated to the App using TCP/IP. Feedback was provided from the robot to the BCI App [15]. In 2020, Attila Kovari developed a BCI method proving that a relationship can be found between problem-solving requiring algorithmic thinking and executive function. The Hanoi-tower test and Flowchart-based debugging test were used, respectively. Based on the method results, a positive correlation was indicated between the level of Executive Functions and Algorithmic Problem-solving [16]. Christina Costescu et al. utilized a GP3 eye tracker to conduct research in the field of BCI in 2019 on school-age children to assess visual attention using the OpenGazeAndMouseAnalyzer (OGAMA) application. They proposed a technology-based paradigm to accurately measure psychological outcomes [17]. Tibor Ujbanyi et al. studied and tested eye-hand coordination in relation to the movement of a computer mouse cursor on the screen. The research was tested using an eye tracker and a leap motion controller in 2021 [18]. Their results are useful for developing future HCI research. In 2020, Jozsef Katona et al. were able to test an HCI method based on eye-tracking that involved the determination of the outcome of two different queries where various programming syntaxes were compared by eye parameters [19].

Due to the large amount of data recorded by EEGs, visual analysis of the data has become challenging. Moreover, epilepsy diagnosis by EEG signals requires a human evaluator and is laborious and expensive, as it requires long hours of expert analysis, which means it can be prone to human error. Therefore, many studies have focused on developing an automated, computerized model for the analysis of EEG signals [1,13].

Previously proposed methods typically perform feature extraction of the EEG signal for analysis. These features include time, frequency, and the combination of both time and frequency. Classification can then be performed utilizing a variety of widely used machine-learning algorithms such as artificial neural networks (ANNs), K-nearest neighbor, and Support-vector machines (SVMs) to distinguish between epileptic and non-epileptic EEG signals. Despite the many epileptic seizure detection techniques, achieving high accuracy, specificity, and sensitivity has always been a challenge. The complexity of most theoretical techniques has resulted in poor practical, real-time implementation. In addition, developing a method of epileptic seizure detection that can be used daily has also been a challenge. The development and implementation of a wearable device for detecting epileptic seizures based on EEG signals would be a valuable contribution to the research community [20].

1.2. Contribution and Paper Organization

In this article, we have summarized epileptic seizure detection techniques in the literature that have relied on EEG signals. We have provided an overview of the required processing factors involved in epileptic seizure detection, which includes data acquisition, pre-processing, decomposing, feature extraction, feature selection, and classification. The purpose of this overview was to provide the epilepsy research community with guidelines for the different methods that have been used for epileptic seizure detection, which may also aid in selecting the most appropriate technique for their research. Furthermore, this study focused on developing an efficient, automated epileptic seizure detection technique using EEG signals that could be easily implemented via hardware or a mobile application. In this approach, we used filter design to accommodate our technique, applied an enhanced decomposition method, and developed a new combined feature for classification and analysis. The finite impulse response (FIR) filter was designed for pre-processing; the Butterworth filter was applied and then followed by a fast Fourier transform (FFT) as the final step of filtering. Then, we applied an enhanced wavelet analysis decomposition technique using an appropriate filter for each decomposed level based on the study of each decomposed signal. The next step was feature extraction, and several features were utilized for further EEG signal processing, one of which was a new combined feature, crest range. After that, a seizure detection classification was performed using the most used machine-learning algorithms, such as neural networks (NNs), K-nearest neighbors (KNN), and multiclass Support-vector machines (SVMs). The average accuracy of 98.61% was achieved by the proposed method when KNN was applied. Finally, we developed and tested a mobile application that monitored the epilepsy classification process and provided notifications based on the results. Our research would add value to existing and future research in the medical field in regard to epilepsy seizure monitoring. Furthermore, it will have an impact on the Human–Computer Interaction field, especially in the field of Brain–Computer Interaction (BCI).

This paper is structured as follows: General structure and material summary of epilepsy and EEG signals are provided in Section 2. Section 3 describes the literature review. The proposed method is explained in Section 4. Section 5 provides a variety of techniques utilized in epileptic seizure detection. Section 6 details the results and experiments. Mobile application development is presented in Section 6. Lastly, the conclusion of this work is provided in Section 7.

2. Structure and Materials

2.1. General Structure of Epilepsy

Normally, the process of epileptic seizure detection begins by recording the patient’s EEG signals. The recording is pre-processed using different techniques such as sampling, filtering, and artifact removal. Once the signal has been pre-processed, it undergoes feature extraction to obtain important information for classification. The final stage is classification, where the signal will be analyzed by machine-learning methods to determine whether the signal is epileptic. The process is shown in Figure 1 [20].

2.2. Electroencephalograph (EEG)

An electroencephalograph (EEG) is a clinical tool that is used to image the brain while it is performing a cognitive task [21,22,23]. EEGs record the electrical activity of the brain via electrodes placed on the scalp [24]. EEGs measure the voltage fluctuations that result from the ionic flow in brain neurons over a short period of time, which could be between 20–40 min. The EEG is one of the main methods for detecting epilepsy when the abnormalities of epileptic activity are clear in the EEG recording [25,26,27], and it is an important tool for research and diagnosis, especially when the use of millisecond-range temporal resolution is required. The range of EEG signal frequencies is very wide. However, we reduced the range to 0.5–30 Hz for physiological and clinical interests [13]. This simplified the detection of the location and the magnitude of the brain activity.

3. State of the Art

Table 1. Previous Methods.

Authors	Signal	Dataset	Features	Classifiers	Results
Sinha et al., 2004 [28]	EEG	Collected from 5 patients	Spectral Power	FMQAS	Satisfactory
Arabi et al., 2009 [29]	EEG	Collected from 21 patients	Entropy, Dominant frequency, Average amplitude, and Rhythmicity	PNN	98.7% ACC
Bao et al., 2008 [30]	EEG	Bonn University	Power spectral features (PSF), Petrosian and Higuchi fractal dimensions (PFD, HFD), and Hjorth parameters	PNN	96.7% ACC
Bezobrazova and Golovko, 2007 [31]	EEG	Collected from 6 patients		ANN	96.7% ACC
Fani and Azmi, 2011 [32]	EEG	Bonn University	Mean of IF, Mean of Kaiser energy, and Energy	ANN	94% ACC
Sivasankari and Thanushkodi, 2008 [33]	EEG	Bonn University	Mean, Absolute value, and Variance	ANN	93.23% ACC
Juarez-Guerra et al., 2015 [34]	EEG	Bonn University	Mean, Absolute median, and Variance	FF-ANN	93.23% ACC
Kumar et al., 2008 [35]	EEG	Bonn University	Wavelet entropy and Spectral entropy	Recurrent NN	96.3% ACC
Kiranmaji and Udayashankara, 2013 [36]	EEG	JSS medical hospital, Mysuru, India	Maximum, Minimum, Mean, and Standard deviation	FFNN	81.67 ACC
Ghosh-Dastidar et al., 2008 [37]	EEG	Bonn University	Standard deviation, Correlation dimension, and Largest Lyapunov exponent	K-Means Clustering, Discrimination analysis, LMBPPN, and RBFNN	96.7% ACC
Dawood Dilber et al., 2016 [38]	EEG	PhysioNet	Mean, Standard deviation, Variance FFT, and Wavelet transform	K-Means and Discrimination analysis	70% ACC, 93% ACC
Mihandoos et al., 2011 [39]	EEG	Bonn University	Fourth moment of wavelet coefficient divided by the second moment, Max–Min, Zero-crossing of wavelet coefficient, Variance of wavelet coefficient, Mean of wavelet coefficient	KNN and Bayesian learning machine	96.8% ACC, 98% ACC
Kumari and Jose, 2011 [40]	EEG	Bonn University	Variance, Energy, and Power spectral density	SVM	98% ACC
Panda et al., 2010 [41]	EEG	Bonn University	Energy, Entropy, and Standard deviation	SVM	91.2% ACC
Liu et al., 2012 [42]	EEG	University Hospital of Freiburg, Germany	Relative entropy, Relative amplitude, Coefficient of variation, and Fluctuation index	RBF-SVM	95.33% ACC
Murugavel et al., 2011 [43]	EEG	Bonn University	Maximum, Minimum, Mean, and Standard deviation	PNN, RBFNN, MSVM	94% ACC, 93% ACC, 96% ACC
Schneider et al., 2009 [44]	EEG	Bonn University	Higuchi algorithm, Katz algorithm, and Sevcik algorithm	SVM	Higuchi algorithm has the higher accuracy
Shen et al., 2011 [45]	EEG	National Taiwan University Hospital	Total variation, Standard Deviation, and Energy	SVM	98.9% ACC
Seng et al., 2012 [46]	EEG	Bonn University	Mean, Coefficient of variation, Dominant frequency, Mean of power spectrum, and Variance	SVM	98% ACC
Yuan, 2010 [47]	EEG	Collected from epileptic patients	Cao’s method	PNN and SVM	96.3% ACC
Hadj-Youcef et al., 2013 [48]	EEG	Bonn University	Maximum, Minimum, Range standard deviation, Energy, and Entropy	SVM	98% ACC
Rafiuddin et al., 2011 [49]	EEG	AMU University	Energy, Coefficient of variation, IQR, and MAD	LAD	96.5% ACC
Chua et al., 2008 [50]	EEG	Bonn University	Mean of spectral magnitude, Entropy, and Power spectrum	GMM	93.11% ACC
Kumar et al., 2014 [51]	EEG	Bonn University	Normalized bispectral (NB) entropy, NB squared entropy, and NB cubed entropy, Bispectrum phase entropy, Mean bispectrum magnitude, and Moment of bispectrum	PNN, KNN, DT, and SVM	96% ACC, 96% ACC, 95% ACC, 98% ACC
Vijith et al., 2016 [52]	EEG	Government Medical College Thiruvananthapuram, Kerala	Approximate entropy, Sample entropy, and Hurst exponent	SVM	91% ACC
Rashid et al., 2017 [53]	EEG	Bonn University	Mean, Median, Maximum, Minimum, Range, Standard deviation, Median absolute deviation, Mean absolute deviation, 12 Norm, and Max Norm	NN	80% ACC
Vandecasteele et al., 2017 [54]	ECG	Collected from patient at UZ Leuven Gasthuisberg	HR peak, HR base, and STDHR base	SVM	70% Sen
Wang et al., 2017 [55]	EEG	Bonn University	Mean, Variance, Coefficient of variation, Total variation, Maximum, Minimum	SVM, KNN, LDA, NB, and LR	99.25% ACC
Gu et al., 2018 [56]	EEG	Collected from patients	Mean powers, and Peak frequency	SVM	100% Sen, 94.5% Sen
Wu, 2020 [57]	EEG	Bonn University and CHB-MIT dataset	Time domain, Frequency domain, Time–frequency domain, and Entropy-based features	CEEMD + XGBoost	99% ACC, 95.7% ACC
Maria et al., 2020 [58]	EEG, EMG, PPG	CAP Sleep Database	DWT	ANN	91.1% ACC
Molla et al., 2020 [59]	EEG	Bonn University	SODP, Squared coefficient of variation of the absolute series, Fluctuation index, Permutation entropy, Approximate entropy, and Renyi’s entropy	FFNN	99.5% ACC
Abiyev et al., 2020 [60]	EEG	Bonn University	Features obtained from the convolutional layers of CNN	CNN	96.67% ACC
Zhang et al., 2018 [61]	EEG	American Clinical Neurophysiology Society	PFD, MFD, FInfo, HComp, HMob, DFA, HuExp, TotVar, FFac, PAPR, RMS, Peak, Kurt, Skw, Var, Trough, Crest, Mean, HFD, SampEn, PeEn, SVDEn, SEn, and PSI_RIR	SVM	99.40% ACC
Mansouri et al., 2019 [62]	EEG	Physionet CHB-MIT	Power in band of interest	Adaptive threshold, Distance network	83% Sen
Shabarinath et al., 2019 [63]	EEG	Bonn University	Power spectral density and Wavelet coefficients	SVM	90.1% ACC
Adda et al., 2020 [64]	EEG	Bonn University	Amplitude and Kolmogorov complexity	SVM	97% ACC
Abedin et al., 2019 [65]	EEG	Bonn University	Mean, Standard deviation, Median, Kurtosis, Skewness, Variance, Maximum, Minimum, and Root mean square	ANN	97.33% ACC
Gupta et al., 2020 [66]	EEG	Bonn University	Mean, Standard deviation, Root mean square, Skew, Kurtosis, Maximum, Coefficient of variation, and Shannon entropy	CNN	99.29% ACC
Prasanna et al., 2019 [67]	EEG	Karunya EEG database	Mean and Standard deviation	SVM	95.8% ACC
Karim et al., 2020 [68]	EEG	Bonn University	Mean instantaneous frequencies	LS-SVM	97.66% ACC

provides a summary of reported techniques in literature between 2004 and 2020. Newer techniques are presented in other tables.

4. Proposed Method and Procedures

We proposed an enhanced, efficient automated technique for epileptic seizure detection using EEG signals that could be implemented using existing hardware to assist physicians in monitoring epilepsy patients. We also introduced the development of a mobile application for monitoring epileptic seizure activities based on the classification of EEG signals. Our proposed method, as shown in Figure 2, begins with data acquisition, signal pre-processing and filtering, signal decomposition, feature extraction including one newly designed feature, and classification using some of the well-known machine-learning techniques to determine whether the EEG signal was epileptic or non-epileptic. Furthermore, in the post-processing step, we developed a mobile application for monitoring epileptic seizures based on the result of the classification.

The proposed method for epileptic seizure detection was a fully automated technique based on EEG signal classification.

4.1. Input EEG Signal

The EEG signals used in this research were acquired from a well-known dataset available from the University of Bonn, Germany [55]. This dataset consisted of five subsets. Subsets A and B both contained normal EEG signals from healthy patients with their eyes open and closed, respectively. Subsets C and D contained seizure-free EEG signals from epileptic patients, and E contained the EEG signals of epileptic patients during seizure occurrence. Each set consisted of 100 signals. Each EEG signal consisted of 4097 samples, which were sampled at a frequency of 173.61 Hz. Each signal had a time duration of 23.6 s. Figure 3 and Figure 4 illustrate normal and epileptic EEG signals, respectively. Figure 5 illustrates seizure-free epileptic signal.

4.2. Filtering

For pre-processing EEG signals, we first utilized a finite impulse response (FIR) filter using Kaiser windows. Then, to reduce the ripples in the signal, we passed the output signal from a Butterworth filter. We used a low-pass FIR filter that accepted the default parameters with a cutoff frequency of 60 Hz and a sampling frequency of 178.6 Hz. We declared the pass-band frequency at 0.25, the stop-band frequency at 0.3, and the stop-band attenuation at 50. We then applied a Butterworth band-pass filter to the signal to obtain a maximum flat pass band for the decomposition stage.

$$\mathrm{w}[n]=\frac{I\left(\beta \sqrt{1-{\left(n-\frac{\alpha }{\alpha }\right)}^2}\right)}{I\left(\beta \right)} 0\le n\le M,0 \mathrm{elsewhere}$$

(1)

where α = M/2, M is the length of the signal, β is the shape parameter that will smooth the window, n is the window length, and I is the zeroth-order Bessel function:

$$I={\sum }_{i=0}^x[\frac{{\left(\frac{x}{2}\right)}^k}{k!}]$$

(2)

Initially, we declared the value of the stop-band ripple, pass-band ripple, pass band, and stop-band coefficients to calculate the order of the filter and then calculated the filter coefficients using these filter parameters. In the final pre-processing stage, fast Fourier transform (FFT) was employed to perform signal overlap. Filter parameters such as FFT size and block length, as well as filter coefficients, were predefined to improve spectral efficiency [69].

4.3. Decomposition

In this stage, we utilized wavelet analysis to decompose the filtered EEG signal into N levels based on the frequency composition of the signal. We tested different transforms for decomposition. However, they failed to provide better results than wavelets. Furthermore, by using wavelet decomposition, we retrieved more details from the signal. We applied a five-level decomposition on the EEG signal using a different filter at each level of decomposition, as shown in Figure 6. Furthermore, our technique applied different filters at each stage of decomposition, rather than using the same filter, based on our analysis of the decomposed signal using a spectrum analyzer. After applying these filters to the signals, we observed that they had more identifiers and more indicators. Moreover, we achieved better classification accuracy when using different filters in the decomposition stages [69].

As shown in Figure 7, representing the decomposed EEG normal signal, and Figure 8, representing the decomposed EEG epileptic signal, as well as the power spectrums of Figure 9 and Figure 10, the frequency responses showed that the frequencies from 0 to 133 Hz were impacting the signal at a high rate [11]. This range of the signal was suitable as it could discriminate between epileptic and normal EEGs.

4.4. Feature Extraction

We presented three features in order to measure the characteristic parameters of the filtered EEG signal, one of which was our designed combined feature. In this study, feature extraction was carefully studied. Both time- and frequency-domain features were extracted and used for classification. However, we selected three non-traditional time-domain features with respect to our analysis of the EEG signal [20,69].

4.4.1. Amplitude Range

To calculate the amplitude range, first, the maximum and minimum values were found, along with their indices. Then, the amplitude range was defined as the difference between the maximum and the minimum values in the EEG signal.

R = max (signal) − min (signal)

(3)

4.4.2. Band Power

Band power is defined as the average power of the given signal within a band of frequencies:

$$P_b=1/(133-0) \times (\underset{T\to \mathrm{\infty }}{\lim }\frac{1}{2\pi }{\int }_0^{133}{\left|x\left(t\right)\right|}^2 dt)$$

(4)

where x(t) is the given signal, and the band is from 0 to 133.

4.4.3. Proposed Crest Range

The crest range is a combination of the crest factor and the amplitude range. We attempted to combine several features to reach the best possible accuracy. The crest range provided outstanding performance and obtained the best accuracy results among all other features.

The crest range feature is based on the calculation of peak to peak of the signal. Then, we calculate the root mean square value of the signal so that we can calculate the crest factor.

$$\mathrm{crest} \mathrm{factor}=\frac{\mathrm{peck} \mathrm{to} \mathrm{peck} \mathrm{of} \mathrm{the} \mathrm{signal}}{\mathrm{RMS} \mathrm{value} \mathrm{of} \mathrm{the} \mathrm{signal}}$$

$$\mathrm{RMS} \mathrm{value} \mathrm{of} \mathrm{the} \mathrm{signal}=\sqrt{\frac{1}{T}{\int }_0^{133}{x}^2\left(t\right)dt}$$

(5)

Finally, to calculate the crest range, we add the crest factor to the amplitude range of the EEG signal.

crest range = Crest Factor + Amplitude Range

(6)

4.5. Machine Learning and Classification

In the classification step, we applied three of the well-known classifiers to evaluate the classification performance of epileptic or non-epileptic signals for epileptic seizure detection using EEG signals. SVM, KNN, and ANN classifiers were used in this study. For the SVM, we used a one-versus-all multiclass SVM combined with a linear kernel function. For KNN, the Mahalanobis distance metric was applied with 1, 2, 3, 4, and 5 k neighbors, and we achieved the best results when k was equal to 5. Finally, the ANN classification using a feed-forward neural network with two layers and one hidden layer that had ten neurons was also applied. In the output layer of ANN, the number of neurons was equal to the number of classes. After the careful random division of data, experiments with different training and testing percentages of 50% and 50%, 60% and 40%, 70% and 30%, and 80% and 20%, respectively, were performed. Each percentage was trained and tested randomly ten times with each classifier, and then the average accuracy was recorded.

5. Results

5.1. Dataset

Our proposed method was developed based on EEG data that were publicly available from the University of Bonn, Germany (http://epileptologie-bonn.de/cms/front_content.php?idcat=193&lang=3) accessed on 9 January 2021.

5.2. Pre-Processing

Three different filters were utilized in the pre-processing stage to eliminate any artifacts. The signal was then decomposed using enhanced wavelet analysis, as explained previously.

5.3. Feature Extraction

Ten features were extracted from the EEG data. However, the features were reduced, and three features, including amplitude range, band power, and our proposed feature crest range, were selected for classification.

5.4. Classification

Some classifiers performed better than others, and we tested many classifiers to compare their performances. However, the researchers in [70] concluded that there was no singular “best” machine-learning classifier, as one that performed well in a specific research environment might perform poorly in another. Therefore, we established our classifiers’ performances based on their accuracy, sensitivity, and specificity.

5.5. Performance

As shown in

Table 2. Results for Max–Min Feature.

Percentage Using Maximum–Minimum Feature	Classifiers
	ANN	SVM	KNN
80% Training and 20% Testing	96.49%	94.13%	98.61%
70% Training and 30% Testing	96.18%	94.57%	97.7%
60% Training and 40% Testing	95.77%	93.27%	96.54%
50% Training and 50% Testing	94.95%	93.18%	95.06%

,

Table 3. Results for Band-Power Feature.

Percentage Using Band-Power Feature	Classifiers
	ANN	SVM	KNN
80% Training and 20% Testing	90.97%	89.89%	92.92%
70% Training and 30% Testing	88.41%	89.99%	91.60%
60% Training and 40% Testing	89.01%	88.27%	90.42%
50% Training and 50% Testing	86.21%	88.27%	89.23%

,

Table 4. Results for Crest-Range Feature.

Percentage Using Crest-Range Feature	Classifiers
	ANN	SVM	KNN
80% Training and 20% Testing	98.03%	95.34%	97.92%
70% Training and 30% Testing	96.05%	93.56%	97.93%
60% Training and 40% Testing	95.17%	94.73%	96.54%
50% Training and 50% Testing	94.48%	94.92%	95.68%

,

Table 5. Results for Amplitude-Range and Band-Power Features.

Percentage Using Amplitude-Range and Band-Power Features	Classifiers
	ANN	SVM	KNN
80% Training and 20% Testing	97.87%	96.40%	96.72%
70% Training and 30% Testing	95.91%	95.97%	95.63%
60% Training and 40% Testing	94.72%	93.53%	93.62%
50% Training and 50% Testing	94.40%	93.21%	93.23%

,

Table 6. Results for Band-Power and Crest-Range Features.

Percentage Using Band-Power and Crest-Range Features	Classifiers
	ANN	SVM	KNN
80% Training and 20% Testing	96.66%	97.06%	96%
70% Training and 30% Testing	96.02%	96.20%	95.51%
60% Training and 40% Testing	95.07%	94.64%	93.70%
50% Training and 50% Testing	94.24%	93.57%	92.71%

,

Table 7. Results for Amplitude-Range and Crest-Range Features.

Percentage Using Amplitude-Range and Crest-Range Features	Classifiers
	ANN	SVM	KNN
80% Training and 20% Testing	97.25%	96.94%	98.27%
70% Training and 30% Testing	96.44%	95.40%	97.36%
60% Training and 40% Testing	94.75%	94.25%	96.23%
50% Training and 50% Testing	94.17%	93.82%	95.76%

and

Table 8. Results for Three Combined Features.

Percentage Using Three Combined Features	Classifiers
	ANN	SVM	KNN
80% Training and 20% Testing	98.43%	97.92%	98.10%
70% Training and 30% Testing	96.38%	97.20%	96.52%
60% Training and 40% Testing	96.20%	96.55%	95.68%
50% Training and 50% Testing	96.06%	94.32%	93.83%

, as a combination of more discriminative features were used, the performance accuracy improved.

When a system or an instrument completes the measurement that it was developed to carry out with the required accuracy, the validity of the system can be confirmed, whereas reliability can be obtained when the system sustains reproducibility of values when repeated tests are applied [71]. According to the study in [71], and by adapting the recommendations for reliability to measures of validity, validity is rated as good (<5%), moderate (5–10%), or poor (>10%) based on error rate. Based on our eight tables of classification results, we can see that at least two out of the three classifiers meet good validity at each level of testing, except for Table 3, where the validity would be poor–moderate. Therefore, our system’s outcome relay on the classification of selected features, which all have good validity results.

Figure 11, Figure 12 and Figure 13 illustrate the performance of all three classifiers and the combination of different features used, demonstrating the overall accuracy that was achieved by the three classifiers at different percentage levels.

Table 9. Table of Comparison.

No	Title	Author	Features	Classifier	Accuracy Results
1	Automatic epilepsy detection using the instantaneous frequency and sub-band energies of the EEG signals (2011)	Fani and Azemi [32]	frequency and the energies of the EEG signals in different sub-bands	Artificial neural network (ANN)	94%
2	Neural network classifier for the detection of epilepsy (2013)	Kiranmayi and Udayashankara [36]	maximum value in the non-redundant region, minimum value in the non-redundant region, mean value of the non-redundant region, maximum value along the principal diagonal, minimum value along the principal diagonal, standard deviation along the principal diagonal	ANN	81.67%
3	Early detection of epilepsy using EEG signals (2014)	Kumar and Ajitha [51]	bispectral entropy, bispectral squared entropy, bispectrum phase entropy, mean bispectrum magnitude, weighted center of bispectrum	PNN, KNN, DT, SVM	96% 96% 95% 98%
4	Epileptic seizure detection in EEG signals using wavelet transforms and neural networks (2015)	Juarez-Guerra et al. [34]	mean, absolute, median, variance	Feed-forward neural network	93.23%
5	EEG based detection of Epilepsy by a mixed design approach (2016)	Dilber and Kaur [38]	mean, standard deviation, variance, FFT, wavelet transform	Support-vector machine Discriminant Analysis technique	70%/93%
6	Epileptic seizure detection using nonlinear analysis of EEG (2016)	Vijith et al. [52]	approximate entropy, sample entropy, hurst exponent	Support-vector machine	89%/91%
7	Epileptic seizure classification using statistical features of EEG signal (2017)	Ahmad et al. [53]	mean, median, maximum, minimum, range, standard deviation, median absolute deviation, mean absolute deviation, l2 norm, max norm,	Neural Network (NN)	80.0% 78.7% 80.0% 79.3%
8	Mixed-band wavelet-chaos-neural network methodology for epilepsy and epileptic seizure detection (2007)	Ghosh-Dastidar and Adeli [72]	standard deviation, correlation dimension, largest Lyapunov exponent	unsupervised K-means clustering, linear and quadratic discriminant analysis, radial basis function neural network, Levenberg–Marquardt backpropagation neural network	96.7% Using (LMBPNN)
9	Seizure detection using wavelet transform and a new statistical feature (2011)	Mihandoost et al. [39]	fourth moment divided by second moment, difference between maximum and minimum, zero-crossing of the wavelet coefficients	K-nearest neighbors (KNN), Bayesian	96.83% 98.17%
10	Seizure detection in EEG using time-frequency analysis and SVM (2011)	Kumari and Jose [40]	variance, energy, power spectral density (PSD)	Support-vector machine (SVM)	98.75%
11	Classification of EEG signal using wavelet transform and support vector machine for epileptic seizure diction (2010)	Panda et al. [41]	energy, entropy, standard deviation	Support-vector machine (SVM)	91.2%
12	Automatic seizure detection using wavelet transform and SVM in long-term intracranial EEG (2012)	Liu [42]	relative energy, relative amplitude, coefficient of variation, fluctuation index	Support-vector machine (SVM)	95.33%
13	Lyapunov features based EEG signal classification by multi-class SVM (2011)	Murugavel et al. [43]	maximum, minimum, mean, standard deviation	Multi-class Support-vector machine (MSVM), Probabilistic Neural Network (PNN), Radial Basis Function Neural Network (RBFNN)	96%/94%/93%
14	Seizure detection in EEG signals using support vector machines (2012)	Seng et al. [46]	mean, variance, dominant frequency, mean power spectrum	Support-vector machine (SVM)	98%
15	Detection of epilepsy during seizure-free periods (2013)	Hadj-Youcef et al. [48]	maximum, minimum, range, standard deviation	Support-vector machine (SVM)	98%
16	Feature extraction and classification of EEG for automatic seizure detection (2011)	Rafiuddin [49]	inter-quartile range (IQR), median absolute deviation (MAD), energy of variation, coefficient of variation	Linear Discriminate Analysis (LAD)	96.5%
17	Automatic Identification of epilepsy by HOS and power spectrum parameters using EEG signals: a comparative study (2008)	Chua et al. [50]	mean of spectral magnitude for PSD, mean of spectral magnitude for HOS, entropy	Gaussian mixture model (GMM)	- (PSD) 88.78% - (HOS) 93.11%
18	Proposed method: (2023)	Lasefr et al.	amplitude range, band power, crest range	ANN, SVM, KNN	96% 95% 98%

shows a comparison between our method and several other methods using the same EEG dataset, with similar as well as different techniques, for epileptic seizure detection. As evident in Table 9, our proposed method achieved a high accuracy rate, as compared to previous research. Furthermore, the filter design, the decomposition technique, and the new features of our proposed method provided significant improvements, as shown by our results.

6. Mobile Application Development

We developed an Android-based mobile application that could monitor the behavior of an epileptic user. Our mobile application operated based on the processing and classification of the EEG signal and provided features such as sending an immediate notification once an epileptic seizure was detected. Furthermore, the application displayed the EEG spectrum of the user’s state and provided an alarm notification sent to the user as well as to family members and caretakers during the occurrence of an epileptic seizure. The application was connected to the EEG classification process via a wireless internet connection and displayed instant results. We tested the application using real-time EEG signals, as collecting epileptic signals would require clinical trials, which was beyond the scope of this paper.

The application development consisted of three modules, namely a login/registration, EEG signal, and visual and alarm notification. As shown in Figure 14 and Figure 15, the communication between the application and the server was established using socket events, and the three components were notified through an “HTTP POST” request notification to a DigitalOcean (DO) server. Then, the HTTP POST response interface was utilized by the DO server to obtain the notification from both the client and the device, as well as to relay the HTTP GET request between the application and the server [73].

Once a connection was established between the application and the server, the client was required to register and log in to the application. Then, the application informed the server that it was ready for an update event. If an epileptic signal was detected, the server would immediately send an update event to the application. After that, the application would send an HTTP GET request to the server. Furthermore, an alarm notification would be triggered as soon as an epileptic signal had been detected. Therefore, once the classification was complete, the application would receive a notification from the server and show the results on a smart device. It would also trigger an alarm.

The application required the user to initially register and then activate their account using a verification link. This enabled immediate notification and alarms to be sent to all registered clients via email ID notification. The login/registration screen is shown in Figure 16.

During the registration process, the application would track users who had requested to be notified by an alarm when an epileptic seizure had been detected. The alarm notification is shown in Figure 17.

There has not been a significant amount of published research that has involved the use of a mobile application for epileptic seizure detection. Therefore, we compared our proposed application to the most relevant methods available. As shown in

Table 10. Table of Comparison of Related Mobile Applications.

Application	Author	Method	Features
Android application for neonatal colonic seizures detection	Cattani et al. [74]	Movement of some body parts, video and image processing	A laptop required for processing. No real users.
Android application	DeVaul et al. [75]	Fall detection	Low accuracy.
Android application	Madansingh et al. [76]	Analyzing body movement using existing embedded sensors	No notification.
Fitbit devices	Fang et al. [77]	SMS and GPS alerts	Lack the compatibility. High cost.
Android application	Yavuz et al. [78]	Fall detection	GPS locations/SMS.
Proposed android application	Lasefr et al.	EEG signal processing	High accuracy. Immediate notification. Development for real-time processing.

, our proposed method outperformed the existing methods due to its use of EEG signals as well as the features that it provided.

7. Conclusions

In this study, we developed an efficient method for epileptic seizure detection that could be easily implemented in a microcontroller device as well as a mobile application for real-time detection. We used an EEG dataset, which was the best publicly available from the University of Bonn in Germany. The noise and artifacts of the signal were eliminated by applying three different filters, and in the next step, we decomposed the signal using enhanced wavelet analysis. Furthermore, out of the many features extracted, three features were selected for the classification procedure. One of these features was our newly designed feature, the crest range. In the final stage, three of the most widely used machine-learning techniques in epilepsy classification, namely ANN, KNN, and SVM, were used for training and testing. After random classification experiments with different testing percentages, including 20%, 30%, 40%, and 50%, and after calculating the average accuracy of at least ten classifications at each percentage level, we achieved the best accuracy when using the KNN classifier.

We developed a mobile application that monitored the behavior of users based on EEG processing and classification. The application was connected to the classification procedure through a cloud server and updated the results instantly whenever a new signal was processed. An alarm would be triggered should an epileptic seizure be detected. Furthermore, we have also tested real-time seizure detection using a Neurosky Mindwave headset that utilizes Bluetooth to exchange data with other devices. The device reads the EEG signal from the forehead and sends it via Bluetooth for processing. However, at this stage, we could only test the normal EEG signal since using the device on epileptic patients is out of our scope.

In future work, we will collect data from epileptic patients so that we can use and test the Mindwave device to collect epileptic EEG signals in real time. As well as trying to detect different types of seizures.