Off-campus WSU users: To download campus access dissertations, please use the following link to log into our proxy server with your WSU access ID and password, then click the "Off-campus Download" button below.
Non-WSU users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Access Type
WSU Access
Date of Award
January 2023
Degree Type
Dissertation
Degree Name
Ph.D.
Department
Electrical and Computer Engineering
First Advisor
Mohammad Alhawari
Abstract
This research aims to develop a computationally-efficient model for automatic patientspecificseizure prediction using a two-layer LSTM from multichannel intracranial electroencephalogram (iEEG) time-series data. We attempted to decrease the number of parameters by employing a smaller input size, fewer electrodes, and managing the power based on the patient-specific epilepsy pattern to make our model a viable alternative for wearable or implantable devices. To get started with the suitable dataset, we provided a comprehensive analysis of various EEG datasets, which can be used for epilepsy prediction which includes the following datasets: Melbourne, CHB-MIT, American Epilepsy Society, Bonn, and European Epilepsy. In terms of denoising the EEG data, we also analyzed and evaluated various denoising approaches for removing artifacts and noise from sEEGs and iEEGs. Five denoising techniques are analyzed and simulated using MATLAB, including the moving average method and multiple methods based on the WT: autocorrelation threshold, SURE Shrink algorithm, universal threshold, and statistical threshold. We suggested a generalized seizure prediction algorithm based on a two-layer LSTM of 16 memory units followed by dropout, dense, and output layers. The model utilized the Swish activation function, which does not suffer from the vanishing gradient problem. Our proposed prediction model was tested on the three drug-resistance patients from the Melbourne dataset and 26 patients from the European iEEG dataset, which is the largest dataset among the available datasets for epileptic seizure applications. We proposed an automatic preprocessing technique based on a common average reference (CAR) to remove artifacts from the European iEEG dataset. The simulation results on the European iEEG dataset showed that the model with the simple structure and the mean post-processing procedure performed the best, with an average AUC of 0.885. We also investigated the patient-specific details such age and gender, seizure types, and surgery results on the system’s performance. Finally, we evaluated different hardware-based epileptic seizure systems and their challenges and providing possible solutions for lowering power consumption and extending battery life. Additionally, we provided a power management approach that regulates the system’s total power usage by using patient-specific seizure patterns. When the chance of a seizure occurring is zero or extremely low, our model calculates the patient-specific seizure pattern and switches the machine to low-power mode. Our investigation demonstrated that, when compared to the complicated model, the suggested power management model can lower the power consumption by 49% with a performance drop of less than 1%.
Recommended Citation
Maleki Varnosfaderani, Shiva, "Epilepsy Prediction Based On Deep Learning Techniques" (2023). Wayne State University Dissertations. 3956.
https://digitalcommons.wayne.edu/oa_dissertations/3956