Detection of Epilepsy Seizures in Neo-Natal EEG Using LSTM Architecture

Abstract

Epilepsy is the most unpredictable and recurrent disease among neurological diseases. Early detection of epileptic seizures can play a critical role in providing timely treatment to patients especially when a patient is in a remote area. This paper uses deep learning framework to detect epilepsy in the Electroencephalography (EEG) signal. The dataset used is publicly available and has a recording of three kinds of EEG signals: pre-ictal, inter-ictal (seizure-free epileptic) and ictal (epileptic with seizure). The proposed Long Short-Term Memory (LSTM) classifier classifies these three kinds of signals with up to 95% accuracy. For binary classification such as detection of inter-ictal or ictal only, its accuracy increases to 98%. The EEG signal is modelled as wide sense non-stationary random signal. Hurst Exponent and Auto-regressive Moving Average (ARMA) features are extracted from each signal. In this work, two different configurations of LSTM architecture: single-layered memory units and double-layered memory units are also modelled. After standardising the features, double-layered LSTM approach gives the highest accuracy in comparison to previously used Support Vector Machine (SVM) classifier and proved to be computationally efficient at Graphics Processing Unit (GPU).

Keywords

• Deep learning
• neo-natal EEG
• LSTM architecture
• desnoising
• biomedical signal processing