Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
An Epileptic Seizure Detection Method from EEG Signals based on a Classifier-Driven Feature Reduction Technique
Epileptic seizure detection can improve the quality of life of epileptic patients, allow for more accurate medication, and minimize the risk of sudden unexpected death in epilepsy (SUDEP). This thesis work aims to develop a robust and stable algorithm for epileptic seizure detection through the clas...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
AUC Knowledge Fountain
2023
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | Epileptic seizure detection can improve the quality of life of epileptic patients, allow for more accurate medication, and minimize the risk of sudden unexpected death in epilepsy (SUDEP). This thesis work aims to develop a robust and stable algorithm for epileptic seizure detection through the classification of EEG signals. To achieve this aim, a methodology is proposed to develop a classifier that can differentiate between the healthy (normal), interictal, and ictal states of EEG signals, while maximizing the classification accuracy and minimizing the computational redundancy. The main pillar upon which this methodology is designed is using a problem-specific classifier-driven feature reduction technique. This technique involves training a bagged trees ensemble that utilizes the complete set of features extracted from the denoised time-domain and time-frequency domain sub-bands of interest. Based on this ensemble, a predictor importance analysis is conducted to reduce the features fed to the classifier to only those with the highest estimates of importance due to their significant contribution to the classification accuracy. A random forest ensemble is finally trained using the reduced features set to classify the involved signals. The University of Bonn EEG dataset was used for testing and validating the proposed methodology through formulating nine 2-class and two 3-class classification problems using its different signal sets. The classification accuracy achieved on the experimented 11 classification problems ranged between 97.15% ± 1.45% and 100.00% ± 0.00%, and the stability of the developed models were assured through running each model 100 times and analyzing their performance metrics. Developing such an algorithm for the real-time classification would minimize the need for the laborious manual classification of the EEG signals and serve as an accurate seizure detection algorithm for the implantable seizure control devices. |
|---|