Automated detection of interictal epileptiform discharges in scalp electroencephalogram of patients with epilepsy
Finding interictal epileptiform discharges (IED) or spikes in the electroencephalogram (EEG) is a part of diagnosing epilepsy. Automated annotation of interictal EEG is important for epilepsy diagnosis and management. Manual annotation of IEDs is tedious and there is substantial disagreement amon...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Theses and Dissertations |
Language: | English |
Published: |
2019
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/89750 http://hdl.handle.net/10220/48049 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Finding interictal epileptiform discharges (IED) or spikes in the electroencephalogram
(EEG) is a part of diagnosing epilepsy. Automated annotation of interictal
EEG is important for epilepsy diagnosis and management. Manual annotation
of IEDs is tedious and there is substantial disagreement among experts. In the
present thesis, various methods contributing to automated IED detection are proposed
and investigated.
We develop a feature-based method, in the form of a cascade of simple
fast classifiers to reject most background waveforms and to speed up IED detection.
We can eliminate 78.41% of background waveforms while retaining 96.93%
of IEDs on our cross-validated dataset comprising 156 subjects. Rejecting background
by our proposed method, speeds up the classification by a factor ranging
from 4.69 to 1.76 for the considered classifiers.
Next, we propose an ensemble algorithm comprising a cascade of SVM
classifiers to enhance the precision of IED detection. At a fixed sensitivity, we
are able to improve precision by 6.78%; and at a fixed false positive rate, the
sensitivity improves by 2.83%.
Additionally, we utilize a dataset annotated by several neurologists to analyze
inter-rater agreement and signal characteristics. We determine features
correlating with expert scoring and contributing to agreement or disagreement
among experts, and fit models to predict expert scores. The features that best
predict inter-rater agreement among experts regarding the presence of IED are
certain wavelet features, using locally standardized EEG. Our models for predicting expert opinion based on EEGer’s scores perform best with a large group
of EEGers (more than 10).
Finally, we develop and implement several statistical measures for evaluating
the performance of a IED detector and comparing it with human experts.
These measures include the mean pairwise sensitivity and false positive rate performance
of the P13 algorithm versus 19 academic EEGers, and a statistical Turing
test using accelerated bootstrap analysis. |
---|