Author ORCID Identifier

https://orcid.org/0000-0002-6578-2379

Date Available

5-17-2021

Year of Publication

2020

Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation

College

Engineering

Department/School/Program

Computer Science

First Advisor

Dr. Guo-Qiang Zhang

Second Advisor

Dr. Jin Chen

Abstract

The 2016 Epilepsy Innovation Institute (Ei2) community survey reported that unpredictability is the most challenging aspect of seizure management. Effective and precise detection, prediction, and localization of epileptic seizures is a fundamental computational challenge. Utilizing epilepsy data from multiple epilepsy monitoring units can enhance the quantity and diversity of datasets, which can lead to more robust epilepsy data analysis tools. The contributions of this dissertation are two-fold. One is the implementation of a temporal query for epilepsy data; the other is the machine learning approach for seizure detection, seizure prediction, and seizure localization. The three key components of our temporal query interface are: 1) A pipeline for automatically extract European Data Format (EDF) information and epilepsy annotation data from cross-site sources; 2) Data quantity monitoring for Epilepsy temporal data; 3) A web-based annotation query interface for preliminary research and building customized epilepsy datasets. The system extracted and stored about 450,000 epilepsy-related events of more than 2,497 subjects from seven institutes up to September 2019. Leveraging the epilepsy temporal events query system, we developed machine learning models for seizure detection, prediction, and localization. Using 135 extracted features from EEG signals, we trained a channel-based eXtreme Gradient Boosting model to detect seizures on 8-second EEG segments. A long-term EEG recording evaluation shows that the model can detect about 90.34% seizures on an existing EEG dataset with 961 hours of data. The model achieved 89.88% accuracy, 92.32% sensitivity, and 84.76% AUC based on the segments evaluation. We also introduced a transfer learning approach consisting of 1) a base deep learning model pre-trained by ImageNet dataset and 2) customized fully connected layers, to train the patient-specific pre-ictal and inter-ictal data from our database. Two convolutional neural network architectures were evaluated using 53 pre-ictal segments and 265 continuous hours of inter-ictal EEG data. The evaluation shows that our model reached 86.79% sensitivity and 3.38% false-positive rate. Another transfer learning model for seizure localization uses a pre-trained ResNext50 structure and was trained with an image augmentation dataset labeling by fingerprint. Our model achieved 88.22% accuracy, 34.99% sensitivity, 1.02% false-positive rate, and 34.3% positive likelihood rate.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2020.182

Download

Share

COinS