A Deep Convolutional Neural Network Method to Detect Seizures and Characteristic Frequencies Using Epileptic Electroencephalogram (EEG) Data.

Valsamma Eapen,Mohammad Ali Moni,Salem A. Alyami,Md. Rashed-Al-Mahfuz,Matthew A. Summers,Shahadat Uddin

doi:10.1109/jtehm.2021.3050925

Abstract

Background: Diagnosing epileptic seizures using electroencephalogram (EEG) in combination with deep learning computational methods has received much attention in recent years. However, to date, deep learning techniques in seizure detection have not been effectively harnessed due to sub-optimal classifier design and improper representation of the time-domain signal. Methods: In this study, we focused on designing and evaluating deep convolutional neural network-based classifiers for seizure detection. Signal-to-image conversion methods are proposed to convert time-domain EEG signal to a time-frequency represented image to prepare the input data for classification. We proposed and evaluated three classification methods comprising of five classifiers to determine which is more accurate for seizure detection. Accuracy data were then compared to previous studies of the same dataset. Results: We found our proposed model and signal-to-image conversion method outperformed all previous studies in the most cases. The proposed FT-VGG16 classifier achieved the highest average classification accuracy of 99.21%. In addition, the Shapley Additive exPlanations (SHAP) analysis approach was employed to uncover the feature frequencies in the EEG that contribute most to improved classification accuracy. To the best of our knowledge, this is the first study to compute the contribution of frequency components to target seizure classification; thus allowing the identification of distinct seizure-related EEG frequency components compared to normal EEG measures. Conclusion: Thus our developed deep convolutional neural network models are useful to detect seizures and characteristic frequencies using EEG data collected from the patients and this model could be clinically applicable for the automated seizures detection.

Highlights

Epilepsy is a chronic neurological disorder that affects approximately 50 million people worldwide, with around 2.4 million people newly diagnosed annually [1]
Case A_E shows the classification accuracy of 99.38% obtained from the proposed FT-VGG16 approach; the highest result with perfect accuracy of 100% is achieved in [7], [8] using Deep Pyramidal 1D-CONVOLUTIONAL NEURAL NETWORKS (CNN), and ESD-LSTM, respectively. For this A_E case, methods using extreme learning machine (ELM) [10] and Support Vector Machine (SVM) [2] achieved a classification accuracy of 96.50% and 97.25% respectively, which are lower than the classification accuracy gained by our proposed FT-VGG16
Two different types of signal-to-image conversion techniques, STFT and continuous wavelet transform (CWT), have been applied, and they show the differences in classifier performance

Summary

Introduction

Epilepsy is a chronic neurological disorder that affects approximately 50 million people worldwide, with around 2.4 million people newly diagnosed annually [1]. Visual examination of EEG traces is time-consuming and puts a heavy burden on the treating physician. These issues have inspired significant efforts in the application of automated seizure detection techniques to assist neurologists, speed up the diagnosis process and, thereby, improve the accuracy. To date, deep learning techniques in seizure detection have not been effectively harnessed due to sub-optimal classifier design and improper representation of the time-domain signal. Methods: In this study, we focused on designing and evaluating deep convolutional neural network-based classifiers for seizure detection. Signal-toimage conversion methods are proposed to convert time-domain EEG signal to a time-frequency represented image to prepare the input data for classification. Conclusion: our developed deep convolutional neural network models are useful to detect seizures and characteristic frequencies using EEG data collected from the patients and this model could be clinically applicable for the automated seizures detection

Methods

Results

Discussion

Conclusion