EEG Spike Detection in Epilepsy using Functional Connectivity Networks

Authors

Ahmed Mohammed, Florida International UniversityFollow

Document Type

Dissertation

Degree

Doctor of Philosophy (PhD)

Major/Program

Mechanical Engineering

First Advisor's Name

Naphtali Rishe

First Advisor's Committee Title

Committee Member

Second Advisor's Name

Jean Andrian

Second Advisor's Committee Title

Committee Member

Third Advisor's Name

Alberto Pinzon-Ardila

Third Advisor's Committee Title

Committee Member

Fourth Advisor's Name

Armando Barreto

Fourth Advisor's Committee Title

Committee Member

Fifth Advisor's Name

Mercedes Cabrerizo

Fifth Advisor's Committee Title

Co-Committee Chair

Sixth Advisor's Name

Malek Adjouadi

Sixth Advisor's Committee Title

Committee Chair

Keywords

Functional Connectivity, EEG, Epilepsy, Spike Detection, Graph Neural Networks

Date of Defense

3-3-2023

Abstract

This dissertation explores functional connectivity (FC) networks of electroencephalogram (EEG) recordings of epileptic patients to understand the electrical dynamics changes during the occurrence of interictal epileptiform discharges (IEDs). The visual identification of IEDs in long EEG records is quite tedious. Nevertheless, IEDs play crucial role in the localization of the epileptic focus out of which the hyperactivity originates. These FC networks could serve as a biomarker for identifying IEDs which is a critical step when planning for resection surgery for drug-resistant focal epileptic patients.

This work first introduces an FC-based method that is statistically more precise at defining the region of interest (ROI) which includes the epileptic focus as determined by the neurologist. This is achieved by scaling the standard FC metric with: i) true connectivity estimate, and ii) power penalization factor. Theta frequency wave is found to require the least amount of power penalization to yield significant localization enhancement. We then show that the frontal temporal (FT) lobe that is affected by the IED is characterized by an increased margin of local connectivity for the theta, alpha, and beta waves. The local connectivity within the same region during an IED segment is also found to be higher compared to the most active FT region during a non-IED (NIED) segment for the same waves. Distant connections are also compared for IED vs. NIED segments, and a significant difference is found between them.

A graph neural network (GNN) architecture is developed to exploit this inherent disproportionality for the purpose of IED detection. Initially, candidate IEDs are collected by a highly sensitive rule-based algorithm based on the morphology of the discharges in context of the containing segments. The collected candidates are then validated by an FC-based GNN (FC-GNN) model which combines the graph embeddings of the different waves to produce the final classification. The results show superiority over other existing models. The architecture is further modified granting each node the freedom to choose the other nodes it should attend to as directed by backpropagation, hence bypassing the FC analysis feature engineering step. This also yields a better validation average precision (AP).

Identifier

FIDC011008

ORCID

https://orcid.org/0000-0002-6427-3003

Previously Published In

1) A. H. Mohammed et al., "Penalized Functional Connectivity Maps for Patients With Focal Epilepsy," in IEEE Access, vol. 9, pp. 204-217, 2021, doi: 10.1109/ACCESS.2020.3046851.

2) A. H. Mohammed et al., "Dynamics of Electrical Activity in Epileptic Brain and Induced Changes Due to Interictal Epileptiform Discharges," in IEEE Access, vol. 10, pp. 1276-1288, 2022, doi: 10.1109/ACCESS.2021.3138385.

Recommended Citation

Mohammed, Ahmed, "EEG Spike Detection in Epilepsy using Functional Connectivity Networks" (2023). FIU Electronic Theses and Dissertations. 5315.
https://digitalcommons.fiu.edu/etd/5315