Epileptic Zone Localization by Unsupervised Adaptive Graph Convolution

Date:30-11-2023   |   【Print】 【close

A research team led by Prof. ZHAN Yang from the Shenzhen Institute of Advanced Technology (SIAT) of the Chinese Academy of Sciences has recently introduced a novel unsupervised dual-stream model based on adaptive graph convolution, aiming to predict seizure onset zones in epilepsy patients. 

The study was published in Neuroimage on Nov. 7. 

Epilepsy is a chronic brain disorder that affects people in every country in the world. One form of this disorder is refractory epilepsy, characterized by persistent and drug-resistant seizures, which presents a challenging scenario that often necessitates brain surgery for the removal of the epileptic onset zone. In current clinical practices, surgical decisions rely heavily on Stereoelectroencephalography (SEEG) monitoring during seizure episodes, neglecting valuable data from rest and sleep states. 

Additionally, the Cortical-Cortical Evoked Potentials (CCEP) technique, effective in measuring synaptic connections between brain regions, is underutilized in seizure localization. 

To address the clinical challenges, the researchers conducted personalized analyses of SEEG (sleep, rest, and seizure states) and CCEP datasets for each patient, constructing a data-driven dual-stream deep network model. This approach sidesteps the complexities of data labeling with traditional machine learning formulations, providing patient-specific, personalized results for epilepsy focus localization. 

The proposed model abstracts the brain into a graph attribute network, treating each brain region as a node, and the effective connections between areas as edges. Intracranial EEG recordings were used to build features for nodes, while CCEP was employed to establish topological connection relationships. Through time-frequency feature extraction and unsupervised autoencoder, the dimensions of the features were reduced. The adaptive graph convolution network then unsupervisedly aggregated features of adjacent nodes, ultimately classifying the location of epileptic brain regions. 

Furthermore, the researchers delved into group-level network dynamics, examining network characteristics between classified epileptic and non-epileptic brain areas. The results showed that there are brain network disparities among different epilepsy types (frontal lobe epilepsy, temporal lobe epilepsy, parietal lobe epilepsy), aligning with clinical observations. 

"This study autonomously determines seizure onset zones based on individual patients' SEEG data," said Prof. ZHAN, "providing a more precise and personalized treatment approach for refractory epilepsy patients." 


Media Contact:
ZHANG Xiaomin