Deep learning for epileptic spike detection

epileptic 
based on the LOO-CV. In each observation, the spikes, shape waves and emotion in EEG data in 
best DBN’s configuration is fitted using a [1, 29] and [25] respectively. All the models are 
training data composed of 18 patients and then trained and tested on the same above EEG 
tested by a remaining patient. The measurement dataset. 
is repeated until the last patient is done. 
 The experimental results are shown 
statistically in the Tab 3. It can be clearly that, 
the estimation of emphspecificity is stable in all 
tests which is reasonable at 95% to 100% due to 
the fact that the number of non spikes for testing 
are large compared with the testing epileptic 
spike, meanwhile the sensitivity seems to be 
different in patients. Accordingly, among the 
observations, the patient number 7 and 8 reach 
the highest sensitivity of 100%; whereas the 
DBN can not detect any epileptic spikes of 
patient number 13 and 15 leading to the lowest 
result at 0% or the model returns a sensitivity of 
50% from patient number 5. It may be caused by 
the fact that the patients have a few spike which 
can be considered as anomalies, so it is hard to 
capture them. In addition, the statistics indicate 
that the more epileptic spike we obtain from the Figure 7. ROC curves for some learning models 
testing patient, the higher accuracy the DBN can trained on the EEG data. 
 Table 3. A performance comparison between the 
predict at. For examples, 622 spikes of patient 
 DBN and other learning models 
number 10 are correctly detected over the total 
number of 635 spikes with a precision of Model SEN SPE AUC 
97.95%; and in the case of the patient number 14, DBN 87.35% 97.89% 0.9597 
the experimental results are very high when the DAE 0% 100% 0.5232 
percentage of epileptic spikes and non spikes ANN 65.74% 91.72% 0.8918 
 SVM 58.64% 92.53% 0.8815 
detected correctly is 95.68% and 99.47% 
 kNN 28.40% 95.42% 0.8058 
respectively. In other cases, the outputs returned 
from patients with more than 20 spikes are quite The results are show statistically and 
good and stable in the range sensitivity of 80% graphically in Tab 4 and Fig 8. It is clear that all 
to 86%. the quality evaluation including sensitivity 
 Finally, a performance comparison between (SEN); emphspecificity (SPE) and area 
using the DBN and other learning models was undercurve (AUC) of the DBN are better than 
provided via numerical study by simulation. In that of other models. Moreover, using DBN 
this work, there are the ANN, deep autoencoder consumes less training time than using others for 
(DAE), support vector machine (SVM) and the reason which the training time of DBN can 
K-nearest neighbor (kNN). In particular, the be reduced by the decreasing the number of 
ANN is organized by an input layer, two hidden iterations to convergence in CD algorithm while 
layers and an output layer followed the way of SVM, kNN and ANN are very time-consuming 
Liu [23] and Dao [5]. The DAE which is a deep in the training process due to the high-
generative model is modified into a dimensional input vector space. Specifically, the 
discriminative model to be aiming to predict SEN, SPE of the DBN classifier are 87.35%, 
epileptic spikes that is composed of three stages 97.89% respectively and better 20% than the 
including encoder, decoder and softmax layer classifier ANN, meanwhile, only 58.64% and 
[6]. The SVM and kNN, which are well-know 28.40% of true spikes are correctly detected by 
 L.T. Xuyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 33, No. 2 (2017) 1-13 11
SVM and kNN. It may be caused by the fact that Acknowledgments 
the EEG dataset used in this work is raw without 
filtering and removing artifacts. Therefore, using This work was supported by Project 
shallow architectures are not useful for this CN.16.07 granted by the University of 
work. Surprisingly, the deep DAE model can not Engineering and Technology, Vietnam National 
detect any spikes and provide a worthless result University, Hanoi. Part of this work was 
with very low AUC of approximately 0.5. It presented in the bachelor graduation thesis of Le 
indicates that not all deep learning models are Trung Thanh [40]. The EEG data used in this 
suitable for this problem. In addition, the work were part of the EEG epilepsy database 
experiments show that the DBN reaches the constructed within the framework of Project 
biggest AUC of 0.9597 representing an excellent QG.10.40 funded by Vietnam National 
system which providing better performance than University Hanoi. 
other models. Once again, this emphasizes the 
advantage and efficiency of DBN in epileptic 
spikes detection. References 
 [1] Nurettin Acr and Cüneyt Güzeli s . Automatic 
4. Conclusions spike detection in EEG by a two-stage procedure 
 based on support vector machines. Computers in 
 Biology and Medicine, 34(7):561–575, 2004. 
 In conclusion, we have applied the DBN 
 [2] David H Ackley, Geoffrey E Hinton, and Terrence 
model as a classifier to detect epileptic spikes in J Sejnowski. A learning algorithm for boltzmann 
EEG signal. The training process show that the machines. Cognitive science, 9(1):147–169, 1985. 
DBN can learn hidden features in EEG data [3] Alaa M Al-kaysi, Ahmed Al-Ani, and Tjeerd W 
which distinguish between epileptic spikes and Boonstra. A multichannel deep belief network for 
non spikes group with high accuracy. The the classification of eeg data. In International 
experiment results not only indicates that Conference on Neural Information Processing, 
learning high-level representations of EEG data pages 38–45. Springer, 2015. 
can be achieved successfully for spike detection, [4] Christophe Andrieu, Nando De Freitas, Arnaud 
but also emphasizes the advantage and Doucet, and Michael I Jordan. An introduction to 
 mcmc for machine learning. Machine learning, 
efficiency of DBN in epileptic spikes detection. 50(1-2):5–43, 2003. 
In addition, we also compare the performance of [5] Nguyen Thi Anh Dao, Nguyen Linh Trung, 
the detection system between using the DBN and Nguyen Van Ly, Tran Duc Tan, Nguyen The 
other learning models like SVM, kNN, ANN, Hoang Anh, and Boualem Boashash. A multistage 
DAE. Accordingly, the results returned by the system for automatic detection of epileptic spikes. 
kind of deep learning models are better than REV Journal on Electronics and Communications, 
those earlier methods. 2017 (submitted). 
 In the near future, we would like to build a [6] Pierre Baldi. Autoencoders, unsupervised 
new model to get more suitable features for deep learning, and deep architectures. In Proceedings 
 of ICML Workshop on Unsupervised and Transfer 
networks with better classification result. We Learning, pages 37–49, 2012. 
will also continue to complete the DBN model [7] Irad Ben-Gal. Bayesian networks. Encyclopedia 
and try to use the other state of the art deep of statistics in quality and reliability, 2007. 
learning models, adjust the parameters of these [8] Christine Fredel Boos, Maria do Carmo Vitarelli 
networks to determine which is the best model Pereira, Fernanda Isabel Marques Argoud, and 
for detecting spikes in EEG signal. Moreover, to Fernando Mendes de Azevedo. Morphological 
get higher quality, we will consider improving descriptors for automatic detection of epileptiform 
preprocessing with more appropriate design of events. In Engineering in Medicine and Biology 
 Society (EMBC), 2010 Annual International 
filters, perceptrons to get clearer data before Conference of the IEEE, pages 2435–2438. 
training deep learning models. IEEE, 2010. 
12 L.T. Xuyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 33, No. 2 (2017) 1-13 
 [9] Li Deng, Dong Yu, et al. Deep learning: methods [21] Alex Krizhevsky and Geoffrey E Hinton. Using 
 and applications. Foundations and Trends in very deep autoencoders for content-based image 
 Signal Processing, 7(3–4):197–387, 2014. retrieval. In ESANN, 2011. 
[10] Pedro Guedes De Oliveira, Carlos Queiroz, and [22] Martin Längkvist, Lars Karlsson, and Amy Loutfi. 
 Fernando Lopes Da Silva. Spike detection based Sleep stage classification using unsupervised 
 on a pattern recognition approach using a feature learning. Advances in Artificial Neural 
 microcomputer. Electroencephalography and Systems, 2012:5, 2012. 
 clinical neurophysiology, 56(1):97–103, 1983. [23] He Sheng Liu, Tong Zhang, and Fu Sheng Yang. 
[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, A multistage, multimethod approach for automatic 
 Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron detection and classification of epileptiform EEG. 
 Courville, and Yoshua Bengio. Generative Biomedical Engineering, IEEE Transactions on, 
 adversarial nets. In Advances in neural information 49(12):1557–1566, 2002. 
 processing systems, pages 2672–2680, 2014. [24] Yung-Chun Liu, Chou-Ching K Lin, Jing-Jane 
[12] Jean Gotman and Li-Yan Wang. State dependent Tsai, and Yung-Nien Sun. Model-based spike 
 spike detection: validation. detection of epileptic eeg data. Sensors, 
 Electroencephalography and clinical 13(9):12536–12547, 2013. 
 neurophysiology, 83(1):12–18, 1992. [25] Raja Majid Mehmood and Hyo Jong Lee. Emotion 
[13] J Gotman and P Gloor. Automatic recognition and classification of eeg brain signal using svm and 
 quantification of interictal epileptic activity in the knn. In Multimedia & Expo Workshops (ICMEW), 
 human scalp eeg. Electroencephalography and 2015 IEEE International Conference on, 
 clinical neurophysiology, 41(5):513–529, 1976. pages 1–5. IEEE, 2015. 
[14] Geoffrey E Hinton, Simon Osindero, and Yee- [26] Piotr Mirowski, Deepak Madhavan, Yann LeCun, 
 Whye Teh. A fast learning algorithm for deep and Ruben Kuzniecky. Classification of patterns of 
 belief nets. Neural computation, 18(7): EEG synchronization for seizure prediction. Clinical 
 1527–1554, 2006. neurophysiology, 120(11):1927–1940, 2009. 
[15] Geoffrey E Hinton. Training products of experts [27] Ö Özdamar and T Kalayci. Detection of spikes 
 by minimizing contrastive divergence. Training, with artificial neural networks using raw EEG. 
 14(8), 2006. Computers and Biomedical Research, 31(2): 
[16] Geoffrey Hinton and Miguel A Carreira-Perpinan. 122–142, 1998. 
 On Contrastive Divergence Learning. In [28] C. C. C. Pang, A. R. M. Upton, G. Shine, and M. 
 AISTATS, volume 10, pages 33–40. Citeseer, 2005. V. Kamath. A comparison of algorithms for 
[17] Geoffrey Hinton. A practical guide to training detection of spikes in the electroencephalogram. 
 restricted Boltzmann machines. Momentum, IEEE Transactions on Biomedical Engineering, 
 9(1):926, 2010. 50(4):521–526, April 2003. 
[18] Alexander Rosenberg Johansen, Jing Jin, Tomasz [29] Y. Pan, S. S. Ge, F. R. Tang, and A. Al Mamun. 
 Maszczyk, Justin Dauwels, Sydney S Cash, and M Detection of epileptic spike-wave discharges using 
 Brandon Westover. Epileptiform spike detection svm. In 2007 IEEE International Conference on 
 via convolutional neural networks. In Acoustics, Control Applications, pages 467–472, Oct 2007. 
 Speech and Signal Processing (ICASSP), 2016 [30] G Pfurtscheller and G Fischer. A new approach to 
 IEEE International Conference on, pages 754–758. spike detection using a combination of inverse and 
 IEEE, 2016. matched filter techniques. 
[19] Ron Kohavi. A study of cross-validation and Electroencephalography and clinical 
 bootstrap for accuracy estimation and model neurophysiology, 44(2):243–247, 1978. 
 selection. In Proceedings of the 14th International [31] Sergey M Plis, Devon R Hjelm, Ruslan 
 Joint Conference on Artificial Intelligence - Salakhutdinov, Elena A Allen, Henry J Bockholt, 
 Volume 2, IJCAI’95, pages 1137–1143, San Jeffrey D Long, Hans J Johnson, Jane S Paulsen, 
 Francisco, CA, USA, 1995. Morgan Kaufmann Jessica A Turner, and Vince D Calhoun. Deep 
 Publishers Inc. learning for neuroimaging: a validation study. 
[20] Cheng-Wen Ko and Hsiao-Wen Chung. Frontiers in neuroscience, 8, 2014. 
 Automatic spike detection via an artificial neural [32] R Quian Quiroga, Zoltan Nadasdy, and Yoram 
 network using raw eeg data: effects of data Ben-Shaul. Unsupervised spike detection and 
 preparation and implications in the limitations of sorting with wavelets and superparamagnetic 
 online recognition. Clinical neurophysiology, clustering. Neural computation, 16(8): 
 111(3): 477–481, 2000. 1661–1687, 2004. 
 L.T. Xuyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 33, No. 2 (2017) 1-13 13
[33] Ruslan Salakhutdinov and Geoffrey Hinton. Deep [40] Le Trung-Thanh. Eeg epileptic spike detection 
 boltzmann machines. In Artificial Intelligence and using deep belief network, 2016. 
 Statistics, pages 448–455, 2009. 
 [41] Elif Derya Übeyli and I nan Güler. Features 
[34] Ruslan Salakhutdinov. Learning Deep Generative extracted by eigenvector methods for detecting 
 Models. Annual Review of Statistics and Its variability of EEG signals. Pattern Recognition 
 Application, 2:361–385, 2015. Letters, 28(5):592–603, 2007. 
[35] Tsu-Wang Shen, Xavier Kuo, and Yue-Loong Hsin. [42] WRS Webber, Brian Litt, K Wilson, and RP 
 Ant K-Means Clustering Method on Epileptic Spike Lesser. Practical detection of epileptiform 
 Detection. In Natural Computation, 2009. ICNC’09. discharges (eds) in the eeg using an artificial neural 
 Fifth International Conference on, volume 6, network: a comparison of raw and parameterized 
 pages 334–338. IEEE, 2009. eeg data. Electroencephalography and clinical 
[36] Vairavan Srinivasan, Chikkannan Eswaran, and Neurophysiology, 91(3):194–204, 1994. 
 Natarajan Sriraam. Approximate entropy-based [43] DF Wulsin, JR Gupta, R Mani, JA Blanco, and B 
 epileptic eeg detection using artificial neural Litt. Modeling electroencephalography 
 networks. IEEE Transactions on information waveforms with semi-supervised deep belief nets: 
 Technology in Biomedicine, 11(3):288–295, 2007. fast classification and anomaly measurement. 
[37] V Srinivasan, C Eswaran, Sriraam, and N. Journal of neural engineering, 8(3): 036015, 2011. 
 Artificial neural network based epileptic detection [44] Wei-Long Zheng and Bao-Liang Lu. Investigating 
 using time-domain and frequency-domain features. critical frequency bands and channels for 
 Journal of Medical Systems, 29(6):647–660, 2005. eeg-based emotion recognition with deep neural 
[38] Sebastian Stober, Avital Sternin, Adrian M Owen, networks. IEEE Transactions on Autonomous 
 and Jessica A Grahn. Deep feature learning for eeg Mental Development, 7(3):162–175, 2015. 
 recordings. arXiv preprint arXiv:1511.04306, 2015. [45] Epilepsy Foundation. URL 
[39] Yichuan Tang. Deep learning using linear  
 support vector machines. arXiv preprint [46] WHO. URL 
 arXiv:1306.0239, 2013.  
 D 
 d