BI Model Deep Learning Architecture for Improved Detection of Seizures from EEG Signal

Abstract

Among the main techniques for the identification and diagnosis of epileptic newlineseizures, Electroencephalography, commonly referred to as EEG, remains the newlinemost popular. The EEG data that is collected from the numerous patients who newlineundergo this medical procedure is now considered to be part of big data as we newlineknow it. EEG analysis and detection of data are finding numerous applications newlinein today s world, and few dedicated algorithms have already been developed newlineto deal with this data. Datamining methods are used to find interesting facts newlinefrom the available dataset. Moving from data mining to machine leaning and newlinethen to deep learning algorithms, is required for handling EEG data as it is newlinenonstationary, voluminous as well as fast accumulating in nature. This work newlineis aimed at developing algorithms, with the sole purpose of comparing the newlineaccuracy in the prediction of the seizures from the EEG data, with data newlinevolume reduction. The endeavour is to filter out the essentials and exploit the newlinecapabilities of the existing algorithms like SVM and ANN, to distil out an newlinealgorithm that can achieve the aim of the project by incorporating automated newlinemachine learning and deep learning models. The feature extraction method newlineproposes the use of sampling using fixed sized windows, to reduce the volume newlineof EEG data that is handled for analysis, and extraction of the top_k amplitude newlinemeasure from each of these samples. These top_k values along with the newlinestatistical features from the selected EEG, is then fed into the classification newlinealgorithms, with the aim of predicting whether it is a seizure or not. Initially, newlinemachine learning algorithms like 3-layer Artificial Neural Network and two newlinedimensional Convolutional Neural Networks are tested against the dataset. newlineThen the entire raw dataset is experimented with Deep Neural Network newlineAlgorithms like Bidirectional Long Short-term memory with additional newlinefunctionalities like Attention Mechanism and Spatial Weight matrix addition. newlineFinally, both 2 D CNN and BiLSTM are applied in parallel with the additional newlinefunctionalities of FFT applied EEG signal. newline