SIAM News Blog

Detection of Epilepsy in Children Using Neural Network Methods

By Ming Yan and Yi Pan

“Children” is a beautiful word that for many people implies the emotional fountain of happiness in our families. Unfortunately, special pediatric epilepsy—termed benign epilepsy with centrotemporal spikes (BECT)—occurs in approximately 10-20 of every 100,000 children and plagues thousands of families every year [3].

BECT is one of the most common types of epilepsy in children aged four to 13 and accounts for 15 to 25 percent of pediatric epilepsy cases [2]. Fortunately, most children outgrow the disease, hence the use of “benign” in the name. However, recent research suggests that BECT could lead to verbal dysfunction, attention-deficit/hyperactivity disorder, and language impairment [4, 6]. Treatment is also subjective, not well-defined, and dependent upon the pediatrician and the syndrome’s severity.

Electroencephalograms (EEGs)—tests that detect electrical activity in the brain via electrodes—are a cornerstone for BECT diagnosis. BECT patients always have centrotemporal spikes in their EEGs that manifest during sleep. Yet these spikes are not always a clear indication of BECT, as healthy children also have them [5]. As a result, the cause of BECT is still unclear.

Because researchers have found that magnetic resonance imaging (MRI) and functional magnetic resonance imaging (fMRI) reveal some abnormalities in the brain areas or connection activities of BECT patients [1, 7], our group builds a multi-view deep neural network model to identify the real BECT samples.

Our Method

In the eye of a data scientist, all information is derived from data. Our team constructs a deep neural network method from a neuroimaging view that is different from the EEG-based BECT diagnosis methods. Figure 1 shows the architecture of our multi-view model with three different views: the handcrafted feature view, the MRI view, and the fMRI view. These views complement each other and offer different portraits of the same disease.

Figure 1. The architecture of our multi-view deep neural network model.

For the handcrafted feature view, we choose 51 handcrafted features from brain areas including the hippocampus, cerebral white matter, cortex, and center front. We show all of the features correlated to BECT. First we indicate the probability density function in Figure 2 that demonstrates feature similarities between BECT patients and healthy patients. Next we evaluate their importance to the classifier, which demonstrates that none of them contribute significantly to it. For this view, we performed classification on the prior features and found that a support vector machine yielded the best performance.

Figure 2. Probability density function of random selected features in BECT patients and healthy patients.

For the MRI and fMRI viewpoints, all instances of neuroimaging are high-dimension data. The MRI view has a larger spatial resolution than the fMRI view, which reflects more brain details. We employ a specially-designed three-dimensional (3D) convolutional neural network to reduce the problem of overfitting on the small training dataset. fMRIs generate four-dimensional data that reveals the temporal variation of the 3D brain on its external dimension. In the fMRI view, use of a 3D convolutional neural network first reduces data dimension. Then a recurrent neural network that benefits from the temporal learning capability of its dynamic system learns the sequential variation of BECT.

With a decision neural network, the multi-view model aims to combine all three views to merit and boost the final performance. As the decision neural network performs another learning process, it learns information complementarily from different views. Figure 3b shows that our multi-view model overperforms single-view methods in five-fold evaluations.

Figure 3. 3a. The importance of the handcrafted features in the classifier. 3b. The performance comparison of multi-view to single-view.

Future Research

Neuroimaging technology directly reflects the brain structure information from a 3D view, which makes it valuable. To fully utilize the spatial and temporal information, we want to not only perform classification but also directly detect BECT’s origin site with our visualization technology for improved diagnosis.

[1] Kim, E.-H., Yum, M.-S., Shim, W.H., Yoon, H.-K., Lee, Y.-J, & Ko, T.-S. (2015) Structural abnormalities in benign childhood epilepsy with centrotemporal spikes (BCECTS). Seizure — Eur. J. Epilepsy, 40(6).
[2] Kramer, U. (2008). Atypical presentations of benign childhood epilepsy with centrotemporal spikes: a review. J. Child Neurol., 23(7), 785-90.
[3] Larsson, K, & Eeg-Olofsson, O. (2006). A population based study of epilepsy in children from a Swedish county. Eur J. Paediatr. Neurol., 10(3), 107-13.
[4] Liasis, A., Bamiou, D.E., Boyd, S., & Towell, A. (2006). Evidence for a neurophysiologic auditory deficit in children with benign epilepsy with centro-temporal spikes. J. Neural. Transm., 113(7), 939-949.
[5] Okubo, Y., Matsuura, M., Asai, T., Asai, K., Kato, M, Kojima, T., & Toru, M. (1994). Epileptiform EEG discharges in healthy children: prevalence, emotional and behavioral correlates, and genetic influences. Epilepsia, 35(4), 832-841.
[6] Parakh, M., & Katewa, V. (2015). A review of the not so benign-benign childhood epilepsy with centrotemporal spikes. J. Neurol Neurophysiol., 6(4).
[7] Zeng, H., Ramos, C.G., Nair, V.A., Hu, T., Liao, J., La., C.,…Prabhakaran, V. (2015). Regional homogeneity (ReHo) changes in new onset versus chronic benign epilepsy of childhood with centrotemporal spikes (BECTS): a resting state fMRI study. Epilepsy Res., 116, 79-85.

Further Reading
[1] Yan, M., Liu, L., Chen, S., & Pan, Y. (2018). A Deep Learning Method for Prediction of Benign Epilepsy with Centrotemporal Spikes. In Int. Symp. Bioinform. Res. Appl. 2018 (pp. 253-258). Beijing, China.
[2] Yan, M., Liu, L., Basodi, S., & Pan, Y. (2018). Multi-view Learning for Benign Epilepsy with CentroTemporal Spikes. IET Comp. Vis.

Ming Yan is a Ph.D. student at Sichuan University and was a visiting Ph.D. student at Georgia State University. Epilepsy research is a project in his group. Yan’s research interests include medical data analysis and neural networks.
Yi Pan is Regents’ Professor of Computer Science and chair of the Department of Computer Science at Georgia State University. He has served as an editor-in-chief or editorial board member for 20 journals, including seven IEEE Transactions. His research covers parallel and cloud computing, wireless networks, and bioinformatics
blog comments powered by Disqus