AUTHOR=Wang Fengqin , Ke Hengjin 

TITLE=Global Epileptic Seizure Identification With Affinity Propagation Clustering Partition Mutual Information Using Cross-Layer Fully Connected Neural Network

JOURNAL=Frontiers in Human Neuroscience

VOLUME=Volume 12 - 2018

YEAR=2018

URL=https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2018.00396

DOI=10.3389/fnhum.2018.00396

ISSN=1662-5161

ABSTRACT=A long-standing issue in epilepsy research and practice is to classify synchronization patterns hidden in multivariate electroencephalography (EEG) routinely superimposed with intensive noise. It is essential to select a suitable feature extraction method to achieve high recognition performance. A typical example is to extract the mutual information (MI) between pairs of channels. The effectiveness of its calculation, which considers the differences between the sequence pairs to build reasonable partition, will improve the classification performance. On this basis, it is even more difficult to adaptively classify the synchronization patterns hidden in multivariate EEG data under the circumstance of insufficient a prior knowledge of domain dependency such as denoising, feature extraction on special patient, etc. To address these problems: (1) effectively calculate the mutual information matrix (synchronization pattern), and (2) accurately classify the seizure or non-seizure state, this study first measures the synchronization between channel pairs in terms of Affinity Propagation Clustering Partition  Mutual Information (APCPMI) accurately. The global synchronization measurement is then obtained by organizing APCPMIs of all channel pairs to a correlation matrix. Finally, a cross layer fully connected net is designed to  characterize the synchronization dynamics correlation matrices adaptively and identify the states of seizure or non-seizure automatically. Experiments are performed over the CHB-MIT scalp EEG dataset to evaluate the proposed approach. Seizure states can be identified with an accuracy, sensitivity and specificity of $[97.93\% \pm\ 0.0001\%]$, $[99.42\% \pm\ 0.05\%]$, and $[96.76\% \pm\ 0.08\%]$, respectively; the resulted performance is superior to those of most existing methods over the same dataset. Furthermore, the approach alleviates the necessary for strictly preprocessing (denoising, removing interferences and artefact) the EEG in terms of prior knowledge, which is widely used in existing approaches.