Explainable AI for EEG Biomarkers Identification in Obstructive Sleep Apnea Severity Scoring Task

Abstract

The assessment of Obstructive Sleep Apneas and hypopneas (OSAs) severity has known an increasing interest over the last decade with the use of Apnea-Hypopnea Index (AHI) being highly criticized by the majority of sleep scientists. To go beyond the single AHI, alternative metrics such as hypoxic burden, arousal intensity, odds ratio product, and cardiopulmonary coupling have been investigated in the literature. However, no consensus has currently been found for a common efficient metric. In this paper, we propose a novel architecture of deep learning model aiming at discovering an objective metric for OSAs severity assessment. We demonstrate the efficiency of this method by identifying features of interest in the Electroencephalographic (EEG) signals while training the model based on biomarkers not or indirectly derived from the EEG, i.e. the desaturation area, the arousal events and the respiratory event duration. By inspecting what the model looks for to make the different classifications, we identified that EEG signals from posterior and medial regions in low frequency bands (0–8 Hz) are highly affected by the apnea-hypopnea severity. With this proof of concept, we pave the way towards the use of Explainable Artificial Intelligence (xAI) to make OSAs severity assessment more objective and find a consensus metric adopted across the community of sleep scientists as well as to boost EEG biomarkers discovery in multiple tasks.