Table 1 Summary of seizure detection and prediction methods
From: EEG seizure detection and prediction algorithms: a survey
Method | Domain and algorithm classification | Single or multi-channel | Database | Frame length | Features | Classifier | Performance metrics | Hardware implementation |
|---|---|---|---|---|---|---|---|---|
Runarsson and Sigurdsson[9] | Time, detection | Single channel | Self-recorded data | Variable length frames | Histogram bin amplitudes for amplitude difference and separation time between peaks and minima | SVM | Sensitivity: 90% | No |
Yoo et al.[14] | Time, detection | Multi-channel | MIT database (scalp EEG) | 2 s | Energies of sub-bands | SVM | Accuracy: 84.4%, classification energy 2.03 μJ | Yes |
Dalton et al.[16] | Time, detection | Single channel | Dataset of 21 seizures | 12 to 25Â s | Signature of seizure | Template matching | Sensitivity: 91 %, specificity: 84 %, battery lifetime: 10.5Â h | Yes |
Time, prediction | Multi-channel | 561Â h of scalp EEG with 86 seizures for 20 patients | 15Â s | Combined index | Non-linear SVM with Gaussian classifier | Sensitivity: 88.34%, false prediction rate: 0.155Â h- 1, average prediction time: 22.5Â min | No | |
Aarabi and He[19] | Time, prediction | Single channel | 316Â h of iEEG data containing 49 seizures for 11 patients from Freiburg database | 10Â s | Correlation dimension, correlation entropy, noise level, Lempel-Ziv complexity, largest Lyapunov exponent | No classifier | Average sensitivity: 79.9%, 90.2% with average false prediction rate: 0.17 and 0.11/h | No |
Schelter et al.[25] | Time, prediction | Single channel | Eight patients with focal epilepsy | 30Â s | Probabilistic threshold | No classifier | Sensitivity: 60% | No |
Wang et al.[26] | Time, prediction | Single channel | Data for five patients | 10Â min | Area under the ROC curve | KNN | Average accuracy: 70% | No |
Bedeeuzzaman et al.[27] | Time, prediction | Single channel | 21 patients from Freiburg database (scalp and iEEG) | 1Â min | MAD, IQR | Binary linear classifier | Sensitivity: 100%, FPR: 0 (for 12 patients), average prediction time: 51 to 96Â min | No |
Li et al.[28] | Time, prediction | Multi channel | 21 patients from Freiburg database (scalp and iEEG) | 5Â s | Spike rate | No classifier | Accuracy: 75.8%, false prediction rate: 0.09 | No |
Chisci et al.[30] | Time, prediction | Multi channel | 21 patients from Freiburg database (scalp and iEEG) | 2Â s | AR coefficients | SVM | Sensitivity: 100%, average prediction time: 60Â min | No |
Rana et al.[32] | Frequency, detection | Multi-channel | EEG and ECoG time series, BW: 0.5 to 100Â Hz | 10 to 60Â s | Phase slope index | No classifier | Detection accuracy: 100% | No |
Khamis et al.[33] | Frequency, detection | Single channel | 12 patients with six data records (R1 to R6) | 32 s | Range of frequencies, range of frequencies used for background subtraction, number of points, hemisphere parameters, moment parameters, artifact removal parameters, p-norm power, cutoff point | Powell’s direction set method | Sensitivity: 91%, false positives per hour: 0.02 | No |
Acharya et al.[34] | Frequency, detection | Single channel | Self-recorded data | 23.6Â s | Phase entropy 1 (S1), phase entropy 2 (S2), approximate entropy (ApEn), and sample entropy (SampEn) | SVM, FSC, PNN, KNN, NBC, DT, and GMM | Average accuracy: 98.1% | No |
Zhou et al.[12] | Wavelet, detection | Multi-channel | Freiburg database of 21 patients; six channels: channels 1, 2, and 3 near the epileptic focus, and channels 4, 5, and 6 in remote locations | 4Â s without overlapping | Lacunarity and fluctuation index on wavelet scales | Bayesian linear discriminant analysis (BLDA) classifier | Sensitivity: 96.25%, false detection rate: 0.13/h, mean delay time: 13.8Â s | No |
Liu et al.[36] | Wavelet, detection | Multi-channel | Dataset of 509Â h from 21 epileptic patients | 4Â s | Relative energy, relative amplitude, coefficient of variation, fluctuation index | SVM | Sensitivity: 94.46%, specificity: 95.26%, false detection rate: 0.58/h | No |
Panda et al.[37] | Wavelet, detection | Single channel | 500 epochs of EEG data from five different brain activities (100 signals per epoch) | 0.5Â s | Energy, entropy, standard deviation | SVM | Accuracy: 91.2% | No |
Khan et al.[38] | Wavelet, detection | Single channel | Self-recorded data for five patients | 25Â s | Energy and normalized coefficient of variation | Simple LDA classifier | Sensitivity: 83%, specificity: 100%, accuracy: 92%, overall precision: 87% | No |
Wang et al.[56] | Wavelet, detection | Single channel | Changhai Hospital database (scalp EEG) | 70 to 200Â ms | Approximate entropy | Neyman-Pearson criteria and SVM | Detection accuracy: 98%, false detection rate: 6% | No |
Zainuddin et al.[40] | Wavelet, detection | Single channel | University of Bonn database (scalp EEG) | 23.6Â s | Maximum, minimum, standard deviation of absolute wavelet coefficients | WNN | Sensitivity: 98%, accuracy: 98% | No |
Niknazar et al.[42] | Wavelet, detection | Single channel | 100 single-channel segments from Bonn University Database (scalp EEG) | 23.6Â s | Time delay, embedding dimension | ECOG | Accuracy: 98.67% | No |
Daou and Labeau[43] | Wavelet, detection | Multi-channel | Nine patients from MIT database, and nine patients from Montreal Neurological Institute database (scalp EEG) | 1Â s | SPIHT codes | No classifier | Accuracy: 90% | No |
Mehta et al.[44] | Wavelet, detection | Single channel | Self-recorded | 10Â s | Slope of the regression line on a logarithmic plot for the wavelet scales 6 to 2 | No classifier | Slope decreases with onset | No |
Shoaib et al.[45] | Wavelet, detection | Multi-channel | MIT database (scalp EEG) | 2Â s | Wavelet energy | SVM | Sensitivity: 91% to 96%, latency: 4.7 to 5.3Â s, false-alarm rate: 0.17 to 0.3/h | Yes |
Zandi et al.[46] | Wavelet, detection | Multi-channel | EEG recordings from 14 patients approximately 75.8Â h with 63 seizures | 10 to 40Â s | Combined seizure index | No classifier | Sensitivity: 90.5%, false detection rate: 0.51Â h-1, median detection delay: 7Â s | No |
Paul et al.[47] | Wavelet, prediction | Single channel | Data for rats | 5Â s | RSWE | No classifier | Detection: 90% | No |
Hung et al.[48] | Wavelet, prediction | Multi-channel | 11 patients from Freiburg database (scalp and iEEG) | 1Â h | Chaos features, wavelet features, correlation dimension | No classifier | Accuracy: 87%, false prediction rate: 0.24/h, average warning time: 27Â min ahead the ictal | Yes |
Chiang et al.[49] | Wavelet, prediction | Single channel | Freiburg database, CHB-MIT database (eight patients), National Taiwan University Hospital database (one patient) | 60Â s | Wavelet coherence | SVM | 74.2% and 52.2% sensitivity on intracranial and scalp databases, sensitivity improvement by 29.0% and 17.4% for both databases | No |
Rojas et al.[52] | Wavelet, prediction | Single channel | 20 patients (267 seizures, 3,400Â h) | Up to 22Â h | Cross-frequency coupling | No classifier | Sensitivity: 100% | No |
Gadhoumi et al.[54] | Wavelet, prediction | Multi-channel | 1,565Â h iEEG, 17 patients with mesial temporal lobe epilepsy with 175 seizures | 1Â min with 75% overlap | Distance, inclusion, persistence | SVM | Sensitivity: >85%, warning rate: < 0.35/h, proportion of time under warning: <30% | No |
Gadhoumi et al.[55] | Wavelet, prediction | Multi-channel | Self-recorded dataset for six patients | 2Â s, disjoint | Wavelet energy and entropy | Binary classifier | Average sensitivity: >80%, FP rate: from 0.09 to 0.7 FP/h | No |
Wang et al.[56] | Wavelet, prediction | Multi-channel | Ten patients (self-recorded data) | 10Â min with 50% overlap | Lyapunov exponent, correlation dimension, Hurst exponent, entropy | KNN | Sensitivity: 73%, specificity: 67% | No |
Soleimani et al.[57] | Wavelet and time, prediction | Single channel | 21 patients Freiburg database (scalp and iEEG) | 10Â s | Time- and wavelet-domain features | Neuro-fuzzy | Prediction: 99.52%, FPR: 0.1417/h | No |
Moghim and Corne[59] | Wavelet and time, prediction | Single channel | One patient from Freiburg database (135Â min, three seizures, 10Â min following each seizure) (scalp and iEEG) | Up to 5Â min | Time-domain energies, wavelet-domain energies, correlation dimension, max Lyapunov exponent | MC-SVM and EANN | Sensitivity and specificity with different values according to simulation assumptions | No |
Tafreshi et al.[61] | EMD, detection | Single channel | Five patients of Freiburg database (scalp and iEEG) | 4 to 6Â s | Mean of the absolute of each IMF, wavelet features | Neural network | Accuracy: 95% | No |
Orosco et al.[62] | EMD, detection | Single channel | 90 EEG segments acquired for nine patients | 1Â h | Energies of the IMFs | No classifier | Sensitivity: 56%, specificity: 75%, positive predictive value: 61%, negative predictive value: 72% | No |
Guarnizo and Delgado[63] | EMD, detection | Single channel | Five groups with 100 single-channel registers sampled at 173.61 Hz | Whole signal | Instantaneous frequency, amplitude of each EMD component, skewness, kurtosis, Shannon’s entropy | Linear Bayes classifier | Accuracy: 98% | No |
EMD, detection | Single channel | Bonn University database (scalp EEG) | 23.6Â s | Skewness, kurtosis, variance, largest Lyapunov exponent, correlation dimension, approximate entropy from intrinsic modes | ANN | Accuracy: 100% | No | |
Bajaj and Pachori[69] | EMD, detection | Single channel | Freiburg database for 21 patients (scalp and iEEG) | 15Â s | Modified central tendency measure | No classifier | Sensitivity: 90%, specificity: 89.31%, error detection rate: 24.25% | No |
Qi et al.[71] | EMD, prediction | Single channel | Pre-surgical epilepsy monitoring database with 80.4Â h of EEG data with ten seizures of four patients | Whole signal | IMF-VoE | No classifier | Sensitivity: 100%, false detection rate: 0.16/h, average time delay: 10.7 to 19.4Â s are 19.4Â s | No |
Vanrumste et al.[74] | SVD, detection | Single channel | Simulated EEG signals | 10Â s | Dipole parameters, relative residual energy (RRE) | No classifier | Visual inspection of parameters | No |
Shahid et al.[76] | SVD, detection | Single channel | Recordings of four pediatric patients with 20 seizures | 1Â s | Largest r-singular values, Euclidean distance from the r-singular values of a baseline window located 1Â h before the seizure | No classifier | Visual inspection of results | No |
Xie and Krishnan[79] | PCA in wavelet domain, detection | Single channel | Data for eight patients | 4,097Â s | Adjusted random index (ARI) | Empirical classifier | ARI up to 1 | No |
Miri and Nasrabadi[80] | PCA, prediction | Multi-channel | Six patients of Freiburg database (scalp and iEEG) | 5 min | Zero-crossing rate, a statistical index | No classifier | Prediction time: 20 ± 5 min, sensitivity 86.6%, false-alarm rate: 0.067% | No |
Williamson et al.[81] | PCA, FFT, prediction | Multi-channel | Freiburg database (scalp and iEEG) | 15Â s | Eigen spectral features | SVM | Sensitivity: 85.5%, false prediction rate: 0.033/h | No |