Table A1.

List of articles used in this review.

Authors/Dataset Used	Year	Decomposition/Features Used	Classification	Result	Inclusion Criteria
Bhattacharyya A, Pachori R. [1] Dataset: CHB_MIT	2017	Empirical wavelet transform (EWT) Then Hadamard transform removes bias/overfitting to classifiers. Output will be the joint feature vectors.	SMOTE technique used to correct imbalance bias. RF classifier	Max average sensitivity = 97.91% Max average specificity = 99.57% Max average accuracy = 99.41%	Included due to full result specification.
Jacobs D., Hilton T., Del Campo M. et al. [5] Dataset: Toronto Western Hospital Epilepsy Monitoring Unit	2018	IcFc: cross-frequency coupling (CFC) index with a Morlet continuous wave transform	Multi-stage state classifier (MSC) based on three random forest classifiers.	Sensitivity = 87.9% Specificity and accuracy = 82.4%, Area-under-the-ROC (AUC) curve = 93.4%.	Included due to full result specification.
Shivnarayan Patidar. and Trilochan Panigrahi [6] Dataset: University of Bonn Germany	2017	Multi-stage TQWT based decomposition (TQWD) The Kraskov entropy measures and characterizes non-linearities	LS-SVM with RBF kernel functions.	Average accuracy = 97.75% Sensitivity = 97.00% Specificity = 99.00% Matthew’s correlation coefficient = 96.00%.	Included due to full result specification.
Wang D, Ren D, Li K, et al. [8] Dataset: Xi’an Jiaotong University	2018	Wavelet decomposition used with level 5 Daubechies order 4 Directed transfer function	RBF_SVM	Average accuracy = 99.4%, Average selectivity = 91.1%, Average sensitivity = 92.1% Average specificity = 99.5% Average detection rate of 95.8%.	Included due to full result specification.
Hashem Kalbkhani and Mahrokh G. Shayesteh [9] Public Bonn Epilepsy EEG Dataset	2017	Stockwell transform Kernel principal component analysis (KPCA)	Nearest neighbor classifier (kNN)	Ictal (Set E) Sensitivity = 99.42% Specificity = 99.89% Accuracy = 99.73 %	Included due to full result specification.
MuhdKaleem, Aziz Guergachi and Sridhar Krishnan. [10] Dataset: CHB MIT	2017	Level 5 Daubechies db6 wavelet is used as the mother wavelet with six vanishing moments.	Adaptive synthetic sampling (ADASYN) for imbalance problem.	Uses kNN and SVM Sensitivity = 99.8% Specificity = 99.6% Accuracy = 99.6%	Included due to full result specification.
Mingyang Li, Wanzhong Chen and TaoZhang, [13] Public Bonn Epilepsy EEG Dataset	2017	Dual-tree complex wavelet trans-form (DT-CWT)	Wilcoxon test for significance. Support vector machine (SVM)	Accuracy = 98% Sensitivity = 98% Specificity = 100%	Included due to full result specification.
JianJia, Balaji Goparaju, JiangLing Song, et al. [11] Public Bonn Epilepsy EEG Dataset	2017	Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)	Random forest classifier Kruskal–Wallis ANOVA	Sets S and (F, N) Accuracy = 99% Sensitivity = 99.5% Specificity = 100% Sets S and (F) Accuracy = 98% Sensitivity =100% Specificity = 99% Cohen’s Kappa statistics = 0.977	Included due to full result specification.
Tao Zhang, Wanzhong Chen and Mingyang Li. [16] Public Bonn Epilepsy EEG Dataset	2017	Variational mode decomposition (VMD) outputs some band-limited intrinsic mode functions (BLIMFs).	Random forest classifier	Highest accuracy is 97.352	Included due to very few papers using EMD-based extraction method, even though no full results.
Ali Yener Mutlu [17] Public Bonn Epilepsy EEG Dataset	2018	Hilbert vibration decomposition (HVD)	(LS-SVM) tested with linear, polynomial and RBF kernel with 10-fold cross-validation	Kernel function/statistical parameters/classification performance (min–max) SPC 95.00–96.50 RBF (= 0.4) ACC 97.33–97.66 SEN 96.00–98.00 SPC 97.5–98.00	Included due to full result specification.
Parvez M, Paul M [18] Dataset: Epilepsy Centre of the University Hospital of Freiburg, Germany	2017	Undulated global feature (UGF) and undulated local feature (ULF) Energy function of CFD (ECFD) and minimum mean energy concentration ratio (MECR) used as feature vector for classification	Least square-SVM classifier with RBF kernel	High prediction accuracy (i.e., 95.4%) FPR = 0.36 Average early prediction time is 22.16 s	Excluded due to no parameter on sensitivity and specificity
Sutrisno Ibrahim, Ridha Djemal and Abdullah Alsuwailem. [12] Dataset: Public Bonn Epilepsy EEG And CHB MIT	2018	Level 6 DWT Daubechies 4 (Db4) Shannon entropy and largest Lyapunov exponent (Rosenstein’s algorithm). Another two conventional methods, which are standard deviation and band power, were also used.	DWT, Shannon entropy, and k-nearest neighbor (KNN) techniques is used. K-NN used with majority vote, K = 3 Linear SVM and LDA	Accuracy = 100%	Excluded due to no parameter on sensitivity and specificity
Khan H, Marcuse L, Fields M, et al. [7] Dataset: The Mount Sinai Epilepsy Center And CHB MIT	2018	Continuous wavelet transform with Mexican mother wavelet. Pre-ictal period and prediction horizon as feature vector. Convolutional neural network used	Deep convolutional neural network used. (stochastic gradient descent with adaptive learning rate). Cross entropy loss function over three classes.	MSSM result Pre-ictal Length = 8 min FPr = 0.128/h CHB MIT result Pre-ictal Length = 6 min FPr = 0.147/h	Excluded due to no parameter on accuracy, sensitivity and specificity
Shiao H, Cherkassky V, Lee J, et al. [19] Dataset: Mayo Clinic	2017	Three feature encodings for iEEG data: Butterworth bandpass filter bank, FFT, and cross-channel correlation of two channels.	Binary SVM classification	Sensitivity = ~ 90–100%, false-positive rate =~ 0–0.3 times per day.	Excluded due to no parameter on accuracy and specificity
Nisrine Jrad, Kachenoura A, Merlet I et al., [15] Dataset: University Hospital of Rennes in France	2016	Convolution of Gabor atom function Gabor root mean square and temporal features Event of interest signals obtained from Gabor RMS	Used RBF- SVM, Receiver operating Characteristic (ROC) curves.	Sensitivity was 0.917 (0.008) for ripples and 0.728 (0.111) for fast ripples while Specificity was 0.738 (0.159) for ripples and 0.933 (0.094) for fast ripples.	Excluded due to no parameter on accuracy
MingyangLi, Wanzhong Chen and TaoZhang. [14] Dataset: Department of Epileptology, University of Bonn	2017	Level 5 Daubechies 4th order discrete wavelet transform. Envelope analysis demodulated with Hilbert transform (HT) for the following extractions: o For the envelope spectrum in each sub-band: mean, energy, standard deviation, and max value. o The mean, energy, standard deviation, and max value of the raw EEG signals.	Neural network ensemble composed of three groups of networks: five sub-nets in each group	Recognition accuracy (RA) = 98.78%	Excluded due to no parameter on sensitivity and specificity
Abeg Kumar Jaiswal, Haider Banka. [20] Dataset:Public Bonn Epilepsy EEG	2017	Local neighbor descriptive Pattern (LNDP) One-dimensional Local Gradient Pattern (1D-LGP)	Artificial neural network classifiers	Average classification accuracy of 99.82% and 99.80%, respectively	Excluded due to no parameter on sensitivity and specificity
Kostas M. Tsiouris, Sofia Markoula, Spiros Konitsiotis et al. [21] CHB MIT	2018	Four novel seizure detection conditions are proposed to isolate EEG segments called Condition I to Condition IV. The short-time Fourier Transform extract EEG energy distribution.	Signal segment used for classifications. All metrics reported here for the case of 3%, 5% and 7% of total visual inspection values respectively.	SSM4 Sensitivity = 84%, 88% and 92% FPr = 4.9 FP/h, 8.1 FP/h and 12.9 FP/h	Excluded due to no parameter on accuracy and specificity
HüseyinGöksu. [22] Public Bonn Epilepsy EEG Dataset	2018	Wavelet packet decomposition. Log energy entropy, norm entropy and energy	Multi-layer perceptron with back propagation	Accuracy = 100%	Excluded due to no parameter on sensitivity and specificity