کنفرانس مهندسی برق ایران

صفحه اصلی / سی و دومین کنفرانس بین المللی مهندسی برق

HFO detection from iEEG signals in epilepsy using time-trained graphs and Deep Graph Convolutional Neural Network

نویسندگان :

Fatemeh Gharebaghi asl¹ Sepideh Hajipour Sardouie²

1- دانشگاه صنعتی شریف 2- دانشگاه صنعتی شریف

کلمات کلیدی :

intracranial Electroencephalography (iEEG)،Epilepsy،(High Frequency Oscillations (HFOs،Deep Graph Convolutional Neural Network (DGCNN)

چکیده :

Intracranial electroencephalography (iEEG) is a type of brain signal widely used to study neurological diseases. Usually, the iEEG signal has frequency components of up to 80 Hz. However, recent studies have shown that, in some conditions, such as epilepsy, the brain signal contains frequency components higher than 80 Hz. These are called high-frequency oscillations (HFOs), and are considered biomarkers for epilepsy. This paper proposes a new methodology for the automated detection of HFOs based on time-domain features of signals and a deep graph convolutional neural network (DGCNN) algorithm. The proposed method was evaluated using the iEEG data of the Fedele’s group from 20 patients with medically intractable epilepsy. The method assumes that the temporal data structure is a graph structure that differs between HFO and non-HFO intervals. By treating the sequence of time samples as the nodes of a graph and training the adjacency matrix of the resulting graph using time data, different graphs are obtained for HFO and non-HFO intervals. Moreover, other features such as RMS, STE, LL, and Teager energy distinguish the intervals. Therefore, these features are considered as node features that help to increase classification accuracy. The DGCNN network is used to classify the time-trained graphs with extracted node features. The proposed methodology has the following significant advantages: 1) it achieves a higher sensitivity than the recently reported HFO detectors using the DGCNN classifier, and 2) it can automatically extract the common features of HFO events from different patients and is more robust, unlike other automated methods in the literature where the features of HFOs were manually extracted based on researchers' knowledge, which may be subject to observer bias. The proposed method achieved 90.7% sensitivity and 93.3% specificity so it has a higher sensitivity than the recently reported HFO detectors.