Introduction
Epileptic seizure focus
Cortical zones | Definition | Clinical diagnosis methods |
---|---|---|
Epileptogenic zone (EGZ) | Area of cortex responsible for generating epileptic seizures; resection yields seizure freedom | Scalp EEG and iEEG |
Irritative zone | Area of cortex that generates interictal spikes | EEG and MEG |
Seizure onset zone (SOZ) | Area of cortex from which clinical seizures originate | SPECT, scalp EEG, and iEEG |
Epileptogenic lesion | Structural lesion that is related to the epilepsy | High-resolution MRI |
Ictal symtomatogenic zone | Area of cortex that generates the seizure symptoms or signs | Ictal video recording |
Functional deficit zone | Area of cortex that is not functioning normally in the interictal period | Neurological examination, Neuropsychological testing, Interictal PET and SPECT, Non-epileptiform EEG, and MEG |
Dataset name | No. of subjects | Electrode type | yEEG Type | Sampling frequency in Hz | Goal of the datasets |
---|---|---|---|---|---|
Bonn (Andrzejak et al. 2001) | 5 | Single-channel | Scalp EEG iEEG | 173.61 | Epileptic and non-epileptic patient detection |
Flint Hill (Osorio et al. 2001) | 10 | Multi-channels | iEEG | 240 | Seizure detection |
Freiburg (Winterhalder et al. 2003) | 21 | Multi-channels | iEEG | 256 | Seizure detection |
n CHB-MIT\(^{1}\) (Shoeb and Guttag 2010) | 23 | Multi-channels | Scalp EEG | 256 | Seizure detection |
Epilepsiae (Ihle et al. 2012) | 275 | Multi-channels | Scalp EEG iEEG | 250-2500 | Seizure detection |
TUSZ\(^{2}\) (Obeid and Picone 2016) | 315 | Multi-channels | Scalp EEG | 250 | Seizure detection y |
Bern-Barcelona (Andrzejak et al. 2012) | 5 | Binary-channels | iEEG | 512 | Epileptic focus detection |
Public datasets
-
The Temple University Hospital (TUH) EEG Corpus (Harati et al. 2014), which is a large size and contains various sub-datasets, including abnormality detection, seizure detection, and artifact classification.
-
The University of Bonn EEG Dataset (Andrzejak et al. 2001) contains several different classes of data that are recorded from healthy volunteers and patients.
-
Other common datasets include IEEG.org and the European Epilepsy Dataset.
-
It includes no information regarding locations of electrodes, which is essential for focal identification.
-
Signals are provided as independent segments without patient labels.
-
The highest frequency is limited to 150 Hz, even though recent neurological findings indicate that high frequency components (>100 Hz) are crucial to identify the epileptic focus.
Bio-marker-based approach
High frequency oscillation
Automated methods for HFOs detection
Phase-amplitude coupling (PAC)
Authors | Dataset | EEG type | Significant PAC range |
---|---|---|---|
Guirgis et al. (2015) | 7 patients with ETLE | iEEG | \( MI_{30-450 Hz \& 0.5-4 Hz}\) |
Amiri et al. (2016) | 25 consecutive epileptic patients (Montreal Neurological Institute and Hospital) | Scalp EEG | \( MI_{30-260 Hz \& 0.3-13 Hz}\) |
Weiss et al. (2016) | 12 patients with MTLE (UCLA Seizure Disorder Center) | iEEG | PAC between ripple amplitude and epileptiform spike phase |
Elahian et al. (2017) | 10 patients with epilepsy (Le Bonheur Children’s Hospital) | ECoG | \( MI_{80-150 Hz \& 4-30 Hz}\) |
Motoi et al. (2018) | 123 patients with drug-resistant focal epilepsy (Children’s Hospital of Michigan and Harper University Hospital in Detroit) | ECoG | \( MI_{150-300 Hz \& 3-4 Hz}\) |
Varatharajah et al. (2018) | 82 patients with focal epilepsy (Mayo Clinic, Rochester, MN) | iEEG | \( MI_{65-115 Hz \& 0.1-30 Hz}\) |
Amiri et al. (2019) | 18 patients with mTLE (Montreal Neurological Institute and Hospital) | iEEG | \( MI_{HFOs \& 4-8 Hz}\) |
Interictal epileptiform discharges (IEDs)
Information theoretic methods | Statistical methods |
---|---|
Sample entropy (SE), permutation entropy (PE), delay permutation entropy (DPE), approximate entropy (APE), fuzzy entropy (FzE),y Reny’s entropy (REN), Shannon entropy (SE), Tsallis entropy (Ts), phase entropy (S1 and S2), wavelet entropy (WE), k-nearest neighbors- entropy (kNNE), centered correentropy (CCE), Stein’s unbiased risk- estimate entropy (SUREE), log-energy entropy (LEE), multi-variate entropy (MVE) | Mean, variance (Var), standard deviation (SD), coefficient of variation, mean absolute value,modified mean absolute value (MMAV), MMAV2, fluctuation index, log detector median frequency (MDF), mean frequency (MNF), katz fractal dimension (KFD), fractal dimension (FD), skewness, kurtosis, different types of quartile: (Q1, Q3, interquartile range), largest lyapunov exponent (LLE), root mean square (RMS), band power (BP), zero crossing (ZC), Hjorth parameter: (activity, mobility, and complexity), teager energy, 1st and 2nd derivative: (mean, SD, var), recurrence qualitative analysis (RQA): mean diagonal line length (MDLL), laminarity (LAM), trapping time (TT), longest vertical line (LVL), longest diagonal line (LDL), recurrence times (RT), Kolmogorov Complexity (KC), Lempel-Ziv complexity (LC) |
Statistical feature extraction
Authors | Features | Classifier | Evaluation |
---|---|---|---|
Zhu et al. (2013) | Information theoretic features | SVM | ACC: 84% |
Sharma et al. (2014) | EMD+Information theoretic features | LS-SVM | ACC: 85% |
Sharma et al. (2015a) | EMD+Information theoretic features | LS-SVM | ACC: 87%; Sen: 90%; Spe: 84% |
Sharma et al. (2015b) | DWT+Information theoretic features | PNN, kNN, FSC, LS-SVM | yACC: 84%; Sen: 84%; Spe: 84% |
Deivasigamani et al. (2016) | DTCWT+ Statistical methods | ANFIS | ACC: 99%; Sen: 98%; Spe: 100% |
Das and Bhuiyan (2016) | EMD-DWT+ Information theoretic features | kNN | Acc: 89.40%; Sen: 90.70%; Spe: 88.10% |
Sharma et al. (2017) | Wavelet FB+Information theoretic features | SVM | ACC: 94.25%; Sen: 91.95%; Spe: 96.56% |
Sharma et al. (2017) | TQWT+Information theoretic, statistical features | SVM | ACC: 95%; |
Gupta et al. (2017) | FAWT+Information theoretic features | kNN, LS-SVM | ACC: 94.41%; Sen: 93.25%; Spe: 95.57% |
Bhattacharyya et al. (2017) | TQWT+ Information theoretic features | LS-SVM | Acc=84.67%; Sen=83.86%; Spe=85.46% |
Sriraam and Raghu (2017) | Information theoretic, statistical features | SVM | ACC: 92.15%; Sen: 94.56%; Spe: 89.74% |
Arunkumar et al. (2017) | Information theoretic features | NB, SVM, kNN, NNge, BFDT | ACC: 98%; Sen: 100%; Spe: 96% |
Itakura and Tanaka (2017) | BEMD+Information theoretic features | RBF SVM | ACC: 86.89% |
Chen et al. (2017) | DWT+ Statistical features | RBF SVM | ACC: 88% |
Bhattacharyya et al. (2018) | EWT+ Statistical features | LS-SVM | yACC: 90%; Sen: 98%; Spe: 92% |
Acharya et al. (2019) | Statistical features | LS-SVM | yACC: 87.93%; Sen: 89.97%; Spe: 85.89% |
Dalal et al. (2019) | FA-WT+Statistical features | RELS-TSVM | ACC: 90.2% |
Subasi et al. (2019) | EMD+DWT+WPD features | RF | ACC: 99.92% |
Gupta and Pachori (2020) | WT+Information theoretic features | LS-SVM | ACC: 95.85%; Sen: 95.47%; Spe: 96.24% |
Sharma et al. (2020) | Statistical features | SVM | ACC: 99% |
Neural networks: end-to-end approach
Structure of neural network
-
The convolutional layer consists of a set of learnable filters (or kernels), each of which has a small receptive field. Dot product (inner product) is performed between the filter weights and region in the input data. The output of the convolutional layer is called the feature map; its depth can be controlled by the number of filters. The stride is set to control how much the filter convolves across the input data.
-
The recurrent layer operates within the cyclical nature of data input and output; each output builds upon the one before it. The RNN with a tanh activation function can be defined as:where \(x_{t}\) is the input data of the time t, and \(h_{t}\) and \(h_{t-1}\) are the hidden states of the time t and \(t-1\), respectively. \(W_{x}\) and \(W_{h}\) are the learnable parameter matrices used for learning input data and hidden state, respectively. b is the parameter vector of the bias.$$\begin{aligned} h_{t} = \tanh (W_{x}x_{t} + W_{h}h_{t-1} + b), \end{aligned}$$(4)
-
The pooling layer performs the downsampling operation for the input data, which can lower the calculation complexity and prevent overfitting. Some commonly-used pooling operations are max pooling and average pooling, both of which partition the input data into a subregion set. For each subregion, the output is either the maximum or the average value.
-
The fully connected layer is used to compute the class scores in the last layer. A one-dimensional feature vector immediately precedes this layer and functions as its input. In a fully-connected layer, each neuron is connected to all the numbers in the previous volume, which is identical to the traditional multi-layer perceptron neural network.
-
The activation function is a non-linear mathematical operation between the current neuron and its output to the next layer. The definitions of some of the commonly-used activation functions are as follows: \(\text{ Sigmoid } (x) = \frac{1}{1 + e^{-x}}\), \(\tanh (x) = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}\), and \(\text{ ReLU }(x) = \max (0, x)\), in each of these, x is the input variable.
-
The batch normalization layer is used for re-centering and re-scaling the input data, which stabilizes the neural networks by allowing for faster convergence. The calculation includes two steps. The first calculates the mean \(E(\cdot )\) and variance \(\text{ Var }(\cdot )\) of a batch data. In the second step, each sample is centered by subtracting the mean and dividing it by the standard deviation: \(y = \frac{x - E(x)}{\sqrt{\text{ Var }(x)}}\), where x is the input variable and y is the normalized result.
VGG Model Structure | |||||
---|---|---|---|---|---|
A | A-LRN | B | C | D | E |
Input: 224*224 RGB Image | |||||
Conv3-64 | Conv3-64 | Conv3-64 | Conv3-64 | Conv3-64 | Conv3-64 |
LRN | Conv3-64 | Conv3-64 | Conv3-64 | Conv3-64 | |
Maxpool | |||||
Conv3-128 | Conv3-128 | Conv3-128 | Conv3-128 | Conv3-128 | Conv3-128 |
Conv3-128 | Conv3-128 | Conv3-128 | Conv3-128 | ||
Maxpool | |||||
Conv3-256 | Conv3-256 | Conv3-256 | Conv3-256 | Conv3-256 | Conv3-256 |
Conv3-256 | Conv3-256 | Conv3-256 | Conv3-256 | Conv3-256 | Conv3-256 |
Conv3-256 | Conv3-256 | Conv3-256 | |||
Conv3-256 | |||||
Maxpool | |||||
Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 |
Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 |
Conv3-512 | Conv3-512 | Conv3-512 | |||
Conv3-512 | |||||
Maxpool | |||||
Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 |
Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 | Conv3-512 |
Conv3-512 | Conv3-512 | Conv3-512 | |||
Conv3-512 | |||||
Maxpool | |||||
FC-4096 | |||||
FC-4096 | |||||
FC-1000 | |||||
Softmax |
Methods based on neural networks
Authors | Feature | Classifier | Dataset | Accuracy |
---|---|---|---|---|
Sui et al. (2019) | STFT | CNN | Barcelona | 91.8 % |
Subathra et al. (2020) | FWHT | ANN | Barcelona | 92.8 % |
Siddharth et al. (2019) | SM-SSA | SAE-RBFN | Barcelona | 99.11 % |
Zhao et al. (2018) | Entropy | CNN | Barcelona | 83.0 % |
San-Segundo et al. (2019) | FT, WT & EMD | CNN | Barcelona | 98.9 % |
Gagliano et al. (2019) | Bispectral | LSTM | iEEG.org | 86.29 % |
Zhao et al. (2021) | Entropy & STFT | FCNN | Barcelona | 93.44 % |
Daoud and Bayoumi (2019) | DCAE & MLP | Barcelona | 93.21 % | |
Li et al. (2019) | 1D-CNN | Barcelona | 85.14 % | |
Lu and Triesch (2019) | CNN | Barcelona | 91.8 % | |
Fraiwan and Alkhodari (2020) | Bi-directional LSTM | Barcelona | 99.60 % |
Predicted positive | Predicted negative | |
---|---|---|
Actual positive | TP: True Positive | FN: False Negative |
Actual negative | FP: False Positive | TN: True Negative |
Evaluation criteria
Segment-wise criteria
-
Accuracy (ACC):$$\begin{aligned} Accuracy=\frac{TP+TN}{TP+FP+TN+FN}\times 100\%, \end{aligned}$$(5)
-
Sensitivity (SEN) or recall:$$\begin{aligned} SEN=\frac{TP}{TP+FN}\times 100, \end{aligned}$$(6)
-
Specificity (SPE):$$\begin{aligned} SPE=\frac{TN}{TN+FP}\times 100, \end{aligned}$$(7)
-
Precision or postitive predictive value (PPV):$$\begin{aligned} Precision=\frac{TP}{TP+FP}\times 100, \end{aligned}$$(8)
-
Fall-out or false positive rate (FPR):$$\begin{aligned} FPR_{nfocal}=\frac{FP}{TN+FP}\times 100, \end{aligned}$$(9)
-
and F\(_{1}\) score, which is the harmonic mean of preision and sensitivity defined as:$$\begin{aligned} F_{1}-score=\frac{2}{\frac{1}{Recall}+\frac{1}{precision}}, \end{aligned}$$(10)
Electrode-wise evaluation criteria
Discussion and open problems
-
In our survey, recent studies that utilize engineering solutions to identify SOZ channels have shown promising results. Different ages and pathological types with an increased number of patients should be considered for future studies.
-
Most of the studies in this survey focused on developing patient-dependent methods to improve computer-aided systems, not on a patient-independent system (PID). For real-world applications, indeed, the patient-independent design (PID) is preferable because epileptologists require some EEG data to label focal and non-focal electrodes used in the system to hypothesis the possible SOZ channels. However, the design of a patient-independent system for identifying SOZ channels is challenging due to the very different electrodes and subject-specific nature of EEG signals. The most promising directions, those that allow for adaptation to different distributions, could be transfer learning and domain adaptation (Pan and Yang 2009; Lotte and Guan 2010; Azab et al. 2018).
-
For designing a supervised computer-aided system (either patient-dependent or patient-independent), the major limitation is the necessity to use SOZ as a prior-basis ground truth for the classifier training stage. Therefore, designing an unsupervised computer-aided system for identification of the SOZ can provide a great facility without prior-basis information of ground truth to take a medical decision.
-
Data recording for clinical protocol depends on the patient’s conditions. In particular, it is challenging to collect enough data to apply to machine learning. Data augmentation is another hot topic in AI design, as it could be used to improve system usability and reduce the training set. Some attempts at data augmentation of focus detection have been reported (Akter et al. 2020a, b).
-
The selection of influential parameters is another critical factor to the design of a computer-aided system. Parameters with more intelligent signal processing and feature-extraction methods are required to further improve focus identification performance.
-
Statistical and information-theoretic features in high-frequency components are promising in this application; however, the interpretation of these features in terms of clinical neurophysiology are still in question.