Defense of Ms. Gaëlle MILON-HARNOIS thesis2:00 pm | UNIVERSITE CATHOLIQUE DE L’OUEST | Amphi BEDOUELLE | 3, Place André Leroy 49000 ANGERS

The December 19, 2023

Subject : Automatic detection and analysis of High Frequency Oscillations in surface EEG-HD.

Director of thesis : Mr Pierre CHAUVET

Abstract

A third of epileptics fail to improve with current treatments, prompting doctors to consider surgery to remove the seizure-generating area of the brain. High Frequency Oscillations (HFOs) are emerging as a biomarker for locating these epileptogenic areas, but their detection is difficult due to their rarity and brevity. In scalp EEG, the low amplitude of the signal complicates the task. This thesis aims to automate the detection of HFOs in EEG-HD signals recorded at 1 kHz on 256 electrodes in 5 patients. After visual marking of HFOs, classification models between HFOs and background noise were explored. Knowledge of signal processing was exploited to extract time- or frequency-domain features. The most statistically relevant features were selected and subjected to classical supervised algorithms (logistic regression, random forest, MLP, gradient boosting). These methods were compared with deep algorithms (CNN, LSTM, Attention) automatically generating signal features in the 1D time domain or 2D time-frequency maps. All models showed convincing results, with the 1D deep algorithms being the most effective, with 91% sensitivity and 87% specificity, outperforming the published surface HFO detectors. Running the best models on the whole signal to automatically detect HFOs showed promising results, but this part of the work still needs to be improved to compensate for the scarcity of HFOs in the data. Several avenues of research are proposed.