In recent years, Stereo-Electroencephalography (SEEG) has become the gold standard for the pre-surgical evaluation of patients with drug-resistant epilepsy. Its ultimate goal is to identify the Epileptogenic Zone (EZ), i.e., the brain region responsible for seizure generation, a task that in clinical practice relies on visual inspection and is therefore prone to significant human error. Addressing this limitation calls for quantitative approaches capable of reliably capturing the complex dynamics of epileptic signals. The present master’s thesis proposes a novel approach to investigate how seizure dynamics can be described through Hidden Markov Models (HMMs). We expect this technique to be particularly well-suited for EZ definition, as it can model the latent state structure underlying neural signals and track how these states evolve over time. Our central hypothesis is that epileptogenic channels exhibit the most pronounced changes at seizure onset, not only at isolated time points, but also in terms of temporal evolution. After a general introduction to both clinical and electrophysiological aspects of SEEG and the theory behind HMMs, the focus then shifts to the description of the proposed modeling pipeline. To identify the EZ, per-channel transition matrices are estimated during both baseline and early ictal periods, capturing the transition from resting state to seizure onset. Afterwards, the Kullback–Leibler (KL) divergence is used as a tool to quantify how much each channel’s ictal dynamics differ from its own baseline. Here, two main approaches based on both local transition matrices and state occupancy distributions are used. Finally, the framework is evaluated over a dataset of 20 seizure recordings against clinically annotated ground truth using AUROC, AUPRC, and F1 score, examining the impact of model complexity and feature selection on EZ identification performance.
Latent State Modeling of Seizure Dynamics for Epileptogenic Zone Identification in SEEG
TACCONI, GIULIA
2025/2026
Abstract
In recent years, Stereo-Electroencephalography (SEEG) has become the gold standard for the pre-surgical evaluation of patients with drug-resistant epilepsy. Its ultimate goal is to identify the Epileptogenic Zone (EZ), i.e., the brain region responsible for seizure generation, a task that in clinical practice relies on visual inspection and is therefore prone to significant human error. Addressing this limitation calls for quantitative approaches capable of reliably capturing the complex dynamics of epileptic signals. The present master’s thesis proposes a novel approach to investigate how seizure dynamics can be described through Hidden Markov Models (HMMs). We expect this technique to be particularly well-suited for EZ definition, as it can model the latent state structure underlying neural signals and track how these states evolve over time. Our central hypothesis is that epileptogenic channels exhibit the most pronounced changes at seizure onset, not only at isolated time points, but also in terms of temporal evolution. After a general introduction to both clinical and electrophysiological aspects of SEEG and the theory behind HMMs, the focus then shifts to the description of the proposed modeling pipeline. To identify the EZ, per-channel transition matrices are estimated during both baseline and early ictal periods, capturing the transition from resting state to seizure onset. Afterwards, the Kullback–Leibler (KL) divergence is used as a tool to quantify how much each channel’s ictal dynamics differ from its own baseline. Here, two main approaches based on both local transition matrices and state occupancy distributions are used. Finally, the framework is evaluated over a dataset of 20 seizure recordings against clinically annotated ground truth using AUROC, AUPRC, and F1 score, examining the impact of model complexity and feature selection on EZ identification performance.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tacconi_Giulia.pdf
accesso aperto
Dimensione
27.55 MB
Formato
Adobe PDF
|
27.55 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/116037