Clustering of High Frequency Oscillations HFO in Epilepsy Using Pretrained Neural Networks

EasyChair Preprint 9167

Abstract

High frequency oscillations (HFOs) have been presented as a promising clinical biomarker of regions responsible of epileptic seizure onset zone (soz) and thus a potential aid to guide epilepsy surgery. Visual identification of HFOs in long-term continuous intracranial EEG (iEEG) is cumbersome, due to their low amplitude and short duration. The objective of our study is to improve and automate HFO detection by developing analysis tools based on an unsupervised clustering method. First, we used a temporal basis set from Jmail et al 2017 while exploiting the time-frequency content of iEEG data. Subsequently, we used a CNN (resnet 18) feature extractor. Then, we applied the clustering method based on reducing the events dimension per frame while preserving the distance between points when displaying from high-dimensional space to a low-dimensional one. The clustering method (Deep Cluster) is based on a standard k-means clustering algorithm. This algorithm successfully isolated HFOs from artifacts, peaks and peaks with ripples. Using this algorithm, we were able to locate the seizure onset area.

Keyphrases: CNN, Epileptic Seizure, K-means, high frequency oscillations (HFOs), intracranial EEG (iEEG)

@booklet{EasyChair:9167,
  author    = {Zayneb Sadek and Abir Hadriche and Nawel Jmail},
  title     = {Clustering of High Frequency Oscillations HFO in Epilepsy Using Pretrained Neural Networks},
  howpublished = {EasyChair Preprint 9167},
  year      = {EasyChair, 2022}}