Epileptic Signal Detection Using Quilted Synchrosqueezing Transform Based Convolutional Neural Networks

Abstract

This work proposes a convolutional neural networks-based algorithm to classify electroencephalographic signals (EEG) in normal, preictal and ictal classes to supporting to the physicists to diagnose the epilepsy condition. EEG signals are preprocessed through the application of the synchrosqueezing transform based on the quilted short time Fourier transform (SS-QSTFT) to generate a time-frequency representation, which is the input to the convolutional neural network (CNN). CNN based classifiers are traine dusing the EEG database of the University of Bonn, which have five sets identified as A, B, C, D and E. Normal, preictal and ictal classes were composed with the combination of the sets A-B, C-D and E, respectively. Accuracy, sensitivity and specificity of the best CNN-based classifier were 99.61, 99.10 and 98.99, respectively. Furthermore, another support vector machines (SVM)-based classifier was developed using the previous CNN model as feature extractor, which last output layer was removed. Input features to the SVM were taken from the fully-connected layer of the CNN. SVM were trained using the same data (time-frequency representation) utilized to train the previous CNN, and their performance in accuracy, sensitivity and specificity were 100% for training and testing sets.