logo

SCIENCE CHINA Information Sciences, Volume 64 , Issue 6 : 162403(2021) https://doi.org/10.1007/s11432-020-3100-8

Detection of the interictal epileptic discharges based on wavelet bispectrum interaction and recurrent neural network

More info
  • ReceivedMay 5, 2020
  • AcceptedSep 3, 2020
  • PublishedApr 27, 2021

Abstract


Acknowledgment

This work was supported in part by National Key Research and Development Program of China (Grant No. 2019YFB2204500), National Natural Science Foundation of China (Grant No. 61874171), and Alibaba Group through Alibaba Innovative Research (AIR) Program.


References

[1] De Cooman T, Varon C, Van de Vel A. Adaptive nocturnal seizure detection using heart rate and low-complexity novelty detection.. Seizure, 2018, 59: 48-53 CrossRef PubMed Google Scholar

[2] P. World Health Organization. Regional Office for the Western. (2004). Epilepsy in the Western Pacific Region: a call to action: Global Campaign Against Epilepsy. Google Scholar

[3] Zacharaki E I, Mporas I, Garganis K. Spike pattern recognition by supervised classification in low dimensional embedding space.. Brain Inf, 2016, 3: 73-83 CrossRef PubMed Google Scholar

[4] Wang S, Pang B, Liu M. A novel compression framework using energy-sensitive QRS complex detection method for a mobile ECG. Sci China Inf Sci, 2019, 62: 069409 CrossRef Google Scholar

[5] Zhang Q, Xie Q, Duan K. A digital signal processor (DSP)-based system for embedded continuous-time cuffless blood pressure monitoring using single-channel PPG signal. Sci China Inf Sci, 2020, 63: 149402 CrossRef Google Scholar

[6] Xu C, Liu Y, Yang Y. An intelligent partitioning approach of the system-on-chip for flexible and stretchable systems. Sci China Inf Sci, 2018, 61: 060415 CrossRef Google Scholar

[7] Bai B, Shu H, Wang X. Towards silicon photonic neural networks for artificial intelligence. Sci China Inf Sci, 2020, 63: 160403 CrossRef Google Scholar

[8] Golmohammadi M, Harati Nejad Torbati A H, Lopez de Diego S. Automatic Analysis of EEGs Using Big Data and Hybrid Deep Learning Architectures. Front Hum Neurosci, 2019, 13 CrossRef Google Scholar

[9] Carey H J, Manic M, Arsenovic P. Epileptic Spike Detection with EEG using artificial Neural Networks. In: Proceedings of the International Conference on Human System Interactions (HSI), 2016. 89--95. Google Scholar

[10] Lodder S S, Askamp J, van Putten M J A M. Inter-ictal spike detection using a database of smart templates.. Clin NeuroPhysiol, 2013, 124: 2328-2335 CrossRef PubMed Google Scholar

[11] Malik M H, Saeed M, Kamboh A M. Automatic threshold optimization in nonlinear energy operator based spike detection. In: Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2016. 774--777. Google Scholar

[12] Tieng Q M, Kharatishvili I, Chen M. Mouse EEG spike detection based on the adapted continuous wavelet transform. J Neural Eng, 2016, 13: 026018 CrossRef PubMed ADS Google Scholar

[13] Liu Y C, Lin C C, Tsai J J. Model-based spike detection of epileptic EEG data.. Sensors, 2013, 13: 12536-12547 CrossRef PubMed Google Scholar

[14] Douget J E L, Fouad A, Filali M M, et al. Surface and intracranial EEG spike detection based on discrete wavelet decomposition and random forest classification. In: Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2017. 475--478. Google Scholar

[15] Quintero-Rinc?r)n A, Muro V, D'Giano C. Spike-and-Wave detection in epileptic signals using cross-correlation and decision trees. Rev Argent de Bioingeniería, 2018, 22: 3--6. Google Scholar

[16] Bou Assi E, Gagliano L, Rihana S. Bispectrum Features and Multilayer Perceptron Classifier to Enhance Seizure Prediction. Sci Rep, 2018, 8: 15491 CrossRef PubMed ADS Google Scholar

[17] Chua C K, Chandran V, Acharya R, et al. Higher order spectral (HOS) analysis of epileptic eeg signals. In: Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology Society, 2007. 6495--6498. Google Scholar

[18] Al-Kadi M, Reaz M, Ali M. Evolution of electroencephalogram signal analysis techniques during anesthesia.. Sensors, 2013, 13: 6605-6635 CrossRef PubMed Google Scholar

[19] Kumar D, Jadeja R, Pande S. Wavelet bispectrum-based nonlinear features for cardiac murmur identification. Cogent Eng, 2018, 5: 1502906 CrossRef Google Scholar

[20] Elsayed M A K. Wavelet bicoherence analysis of wind-wave interaction. Ocean Eng, 2006, 33: 458-470 CrossRef Google Scholar

[21] Hadjileontiadis L J. EEG-Based Tonic Cold Pain Characterization Using Wavelet Higher Order Spectral Features.. IEEE Trans Biomed Eng, 2015, 62: 1981-1991 CrossRef PubMed Google Scholar

[22] Obeid I, Picone J. The Temple University Hospital EEG Data Corpus.. Front Neurosci, 2016, 10 CrossRef PubMed Google Scholar

[23] L?r)pez S, Gross A, Yang S, et al. An analysis of two common reference points for EEGS. In: Proceedings of IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2016. 1--5. Google Scholar

[24] Esteller R, Vachtsevanos G, Echauz J. A comparison of waveform fractal dimension algorithms. IEEE Trans Circuits Syst I, 2001, 48: 177-183 CrossRef Google Scholar

[25] Richman J S, Moorman J R. Physiological time-series analysis using approximate entropy and sample entropy. Am J Physiol-Heart Circulatory Physiol, 2000, 278: H2039-H2049 CrossRef Google Scholar

[26] A N Akansu and R A Haddad. Multiresolution Signal Decomposition. 2nd ed. San Diego: Academic Press, 2001. 391--442. Google Scholar

[27] McDonnell J T E, Bentley P M. Wavelet transforms: an introduction. Electron Communication Eng J, 1994, 6: 175-186 CrossRef Google Scholar

[28] Hou C, Han H, Liu Z J, et al. A wind direction forecasting method based on z-score normalization and long short-term memory. In: Proceedings of IEEE 3rd International Conference on Green Energy and Applications (ICGEA), 2019. 172--176. Google Scholar

[29] Huang D, Fu Y, Qin N. Fault diagnosis of high-speed train bogie based on LSTM neural network. Sci China Inf Sci, 2021, 64: 119203 CrossRef Google Scholar

[30] Zaremba W, Sutskever I, Vinyals O. Recurrent neural network regularization. 2014,. arXiv Google Scholar

[31] Pang Y, Bai X, Zhang G. Special focus on deep learning for computer vision. Sci China Inf Sci, 2019, 62: 220100 CrossRef Google Scholar

[32] Estabrooks A, Jo T, Japkowicz N. A Multiple Resampling Method for Learning from Imbalanced Data Sets. Comput Intell, 2004, 20: 18-36 CrossRef Google Scholar

[33] Hanchuan Peng , Fuhui Long , Ding C. Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy.. IEEE Trans Pattern Anal Machine Intell, 2005, 27: 1226-1238 CrossRef PubMed Google Scholar

[34] Shuo Wang , Xin Yao . Multiclass Imbalance Problems: Analysis and Potential Solutions. IEEE Trans Syst Man Cybern B, 2012, 42: 1119-1130 CrossRef PubMed Google Scholar

[35] Reus E E M, Visser G H, Cox F M E. Determining the Spike-Wave Index Using Automated Detection Software.. J Clin NeuroPhysiol, 2019, Publish Ahead of Print CrossRef PubMed Google Scholar

[36] Zarifia M H, Ghalehjogh N K, Baradaran-nia M. A new evolutionary approach for neural spike detection based on genetic algorithm. Expert Syst Appl, 2015, 42: 462-467 CrossRef Google Scholar

[37] Pfammatter J A, Maganti R K, Jones M V. An automated, machine learning-based detection algorithm for spike?wave discharges (SWDs) in a mouse model of absence epilepsy. Epilepsia Open, 2019, 11: epi4.12303 CrossRef PubMed Google Scholar

  • Figure 1

    (Color online) Six events of TUEV EEG dataset. (a) SPSW; (b) GPED; (c) PLED; (d) EYEM; (e) ARTF; (f) BCKG.

  • Figure 2

    (Color online) Auto-interactions of the whole frequency range of six different IED events using the WBS method.protectłinebreak (a) SPSW; (b) GPED; (c) PLED; (d) EYEM; (e) ARTF; (f) BCKG.

  • Figure 3

    (Color online) Auto-interactions of six different IED events using the WBS method: rows represent the interactions of five frequency-bands of (a) SPSW, (b) GPED, (c) PLED, (d) EYEM, (e) ARTF, and (f) BCKG respectively, while columns represent the interactions between $\delta$ vs. $\delta$, $\theta$ vs. $\theta$, $\alpha$ vs. $\alpha$, $\beta$ vs. $\beta$, and $\gamma$ vs. $\gamma$ respectively.

  • Figure 4

    (Color online) Block diagram of the proposed system for detecting the epileptic IED discharges. The proposed system consists of four steps, namely data preprocessing, feature extraction, feature selection and classification.

  • Figure 5

    (Color online) Feature extraction of the proposed system. The first step is segmenting the 1-s epoch into multiple timesteps (frames), then each timestep is decomposed into five bands and the WBS among the different bands are calculated. 180 features are calculated from the obtained 15 bispectrums and 10 features are extracted from the time domain.

  • Figure 6

    (Color online) The proposed classifier of the proposed system.

  • Figure 7

    (Color online) Preparing data for training and testing the classification network using a 10-fold cross-validation method.

  • Figure 8

    (Color online) Performance results for varying units of the LSTM layers. (a) 2-way classifier; (b) 3-way classifier.

  • Figure 9

    (Color online) Performance results for varying the neurons of the input layer. (a) 2-way classifier; (b) 3-way classifier.

  • Figure 10

    (Color online) Examining the informativeness of the selected features on the performance of the 2-way classifier.protectłinebreak (a) Replacing each feature by zero; (b) considering only one feature at each case.

  • Figure 11

    (Color online) DET curve of the five cases of the proposed system and the Golmohammedi's system.

  • Figure 12

    (Color online) Detection of IEDs for 10-s epoch of EEG recording in TUEV dataset. Red events are the TP detections, Green events are the FP detections, Pink events are the FN detections and Blue events are the TN detections.

  • Figure 13

    (Color online) DET curve of the proposed method and the existing methods.

  • Figure 14

    (Color online) Comparison between the traditional window segmentation method and the NEO-diff segmentation technique. (a) Moving Window segmentation method; (b) NEO-diff segmentation technique.

  • Table 1  

    Table 1State-of-the-art methods for detection of epileptic IED discharges

    Method Dataset DatasetFeatures FeaturesClassifierACCSEN SPEFPR PREC F1-scoreGmean
    availability method type(%) (%) (%) (%) (%) (%)
    Zacharaki et al. [3] One-subject with101 spikes(2 channels)NOLocality preserving projections (LPP)Amplitude-based Supportvectormachine(SVM)97 98.741.26%62.8 76.24 97.87
    Golmoham-medi etal. [8] Publicly largest TUH EEG dataset (390 subjects,22 channels) YesHidden Markov models (HMM)Deep learning90.195.114.89%92.6
    Carey et al.[9] 6 patients(one-channel)NO Artificialneuralnetwork (ANN) 82.68 72.777.37
    Lodder et al. [10] 23 patients with723 IEDs NO Template matching Amplitude-based SVM 902.36 (min) 23.10 36.76
    Malik et al. [11] 13 subjectsNO Gradient-based NEOAmplitude-based 74.1
    Liu et al. [12] 12 epileptic patients (16 channels)NO Morphological-based AdaBoost93.995.5 92.47.6%93.94
    Douget et al. [13] 17 subjects(3 channels) NO Discrete wavelet transform (DWT) Wavelet coefficient-based Randomforest62 2636.64
    Antonioet al. [14] two-subjects with96 spikesNOCross-correlationAmplitude-basedDecision tree97 8698 2%91.80
  • Table 2  

    Table 2Distribution of EEG's events in the TUEV EEG dataset

    Training set Testing set
    Event CountPercentage (%) CountPercentage (%)
    SPSW 6450.91 5671.93
    GPED 70509.54 467715.9
    PLED 41205.58 1998 6.79
    EYEM 9401.27 3291.12
    ARTF 932912.63 22047.49
    BCKG 5179070.11 1964666.78
    Total 73874100 29421 100
  • Table 3  

    Table 3Time-domian based features$^{\rm~a)}$

    NumberFeatureEquationNumberFeatureEquation
    1Mean of absolutevalue(X_rm avg= frac1
    N
    n=1^Nx_f
    )6Kurtosis of data (X_rm kurt=frac1
    NX_σ^4∑
    n=1^N(x_f-μx
    )^4)
    2Maximum value (X_rm max=rm maxx_f)) 7Hjorth mobility(X_rm mob=rm mobx_w (n))=sqrtBig(fracrm varx_f^prime)rm varx_f)Big) )
    3 Sum of logarithmicamplitude (X_rm slog=∑n=1^Nx_f
    )8Hjorth complexity(X_rm comp=frac rm mobx_f^prime)rm mobx_f))
    4 Variance of data (X_σ=rm varx_w (n)=frac1
    N
    n=1^N(x_f-μx
    )²) 9 Fractal dimensionindex (X_rm FD= frac łog_10 (N-1)łog_10d/L)+łog_10 (N-1))
    5Skewness of data(X_rm skew=frac1
    NX_σ³
    n=1^N(x_f-μx
    )³) 10Sample entropy(X_rm SE=-łog(A/B))
  • Table 4  

    Table 4Results of the optimized models

    Classifier modelResults
    (N_f) (N_rm lstm1)(N_rm lstm2) Gmean (%) F1-score (%) SPE (%) SEN (%)
    Model 1 2590902-way classifier95.6291.495.68 95.55
    Model 2256020 95.42 91.1795.48 95.36
    Model 3 25 90 80 3-way classifier75.16 64.16 85.86 67.07
    Model 4 25 60 40 74.28 61.67 84.74 65.88
    Model 5 20 60 40 75.16 63.27 85.4 67.33
  • Table 5  

    Table 5Results and confusion matrix of 2-way classification

    ResultsProposed system Golmohammedi's
    (N_f)=25, (N_rm lstm1)= 60, (N_rm lstm2)=20 system [8]
    Case 1Case 2Case 3Case 4Case 5
    (1 timesteps)(4 timesteps)(9 timesteps)(19 timesteps)(49 timesteps)
    Validating Testing Validating Testing Validating Testing Validating Testing Validating Testing
    Gmean (%)91.6890.8694.2792.1396.2193.9496.6994.5996.33 95.4292.6
    F1-score (%)90.4783.0493.4285.9295.5889.1796.1490.2195.7691.17
    SPE (%)91.0790.8694.0993.1595.6094.9396.2095.3796.2195.4895.11
    SEN (%)92.2990.8694.4691.1296.8492.9697.19 93.8196.4495.3690.1
    FPR (%)8.939.145.916.854.45.073.84.63 3.79 4.52 4.89
    EventENEENEENEENEENEENE
    E90.869.1491.128.88 92.967.0493.816.19 95.364.6490.19.9
    NE9.1490.866.8593.155.0794.964.6395.374.52 95.484.8995.11
  • Table 6  

    Table 6Classification results of the six IED events

    MethodProposed system (case 5)Golmohammedi's system
    EventEYEMARTFBCKGSPSWGPEDPLEDEYEMARTFBCKGSPSWGPEDPLED
    EYEM89.654.761.07 4.5279.312.3017.24 1.1500
    ARTF0.6976.0918.7010.18 14.0472.98 2.8100
    BCKG0.465.48 89.548.93 3.4281.405.95 0.300
    SPSW 4.6451.9915.4927.88 13.3310 33.3333.33100
    GPED8.1666.1421.0500.3 3.6517.63 65.0513.37
    PLED1311.7270.640.49010.7613.699.78 65.28
  • Table 7  

    Table 7Comparison of the proposed system with existing methods

    MethodACC (%)SEN (%)SPE (%)FPRPREC (%) F1-score Gmean
    Zacharaki et al. [3] 9798.741.26%62.876.2497.87
    Golmohammedi et al. [8] 90.195.114.89%92.6
    Carey et al. [9] 82.6872.777.37
    Lodder et al. [10] 902.36 (min)23.1036.76
    Malik et al. [11] 74.1
    Liu et al. [12] 93.995.592.47.6%93.94
    Douget et al. [13] 622636.64
    Antonio et al. [14] 9786982%91.80
    Proposed system95.4595.3695.484.52% 87.33 91.17 95.42
qqqq

Contact and support