Bispectrum and Recurrent Neural Networks: Improved Classification of Interictal and Preictal States

Laura Gagliano,Mohamad Sawan,Elie Bou Assi,Dang K Nguyen

doi:10.1038/s41598-019-52152-2

Laura Gagliano, Mohamad Sawan + Show 2 more

Open Access

https://doi.org/10.1038/s41598-019-52152-2

Copy DOI

Abstract

This work proposes a novel approach for the classification of interictal and preictal brain states based on bispectrum analysis and recurrent Long Short-Term Memory (LSTM) neural networks. Two features were first extracted from bilateral intracranial electroencephalography (iEEG) recordings of dogs with naturally occurring focal epilepsy. Single-layer LSTM networks were trained to classify 5-min long feature vectors as preictal or interictal. Classification performances were compared to previous work involving multilayer perceptron networks and higher-order spectral (HOS) features on the same dataset. The proposed LSTM network proved superior to the multilayer perceptron network and achieved an average classification accuracy of 86.29% on held-out data. Results imply the possibility of forecasting epileptic seizures using recurrent neural networks, with minimal feature extraction.

Highlights

Epilepsy is one of the most prevalent neurological conditions in the world affecting about 69 million people of all ages (World Health Organization)
Similar to recent seizure forecasting investigations based on canine intracranial electroencephalography (iEEG), this database was chosen for this study because canine epilepsy has been demonstrated as a suitable model for human epilepsy[9,14,23,24] and because it provides longer recordings than human databases which generally include short-term recordings (∼2 weeks) from patients admitted for epilepsy surgery evaluation
In a proof-of-principle study, our group has previously demonstrated that there exists a statistically significant change in bispectrum measures prior to seizure onset and that a simple multilayer perceptron (MLP) neural network classifier is capable of learning to distinguish between preictal and interictal 30-second iEEG recordings based on single iEEG bispectral features from 16 channels[14]

Summary

Introduction

Epilepsy is one of the most prevalent neurological conditions in the world affecting about 69 million people of all ages (World Health Organization). In the case of seizure prediction, artificial neural networks are trained to represent raw iEEG or iEEG-extracted features from short segments of interictal (non-seizure activity) and preictal recordings and learn to map these representations to their class label (preictal or interictal)[6] This type of approach allows for prediction algorithms to have a general pipeline while they are trained in a subject-specific manner (meaning that they learn personal preictal patterns), an important attribute for accurate prediction[10]. In line with recent studies and literature on seizure prediction, a neural network classifier was trained and tested on data from a single animal to account for the high inter-subject variability (e.g. different seizure types, onset patterns, etc.)[4,5,6,7,8,10,16]. We conclude with an interpretation of the current findings and proposal of prospective studies

Methods

Results

Conclusion