Arousal Detection in Obstructive Sleep Apnea using Physiology-Driven Features

Obstructive sleep apnea (OSA) is a condition in which a person repeatedly stops breathing during sleep due to closure of the upper airway, leading to a cycle of sleep fragmentation and intermittent hypoxia (oxygen deficiency).Conventional methods for detecting and quantifying OSA are largely based on physiological monitoring during sleep followed by manual labeling of sleep stages and arousals.Here there is scope for computerized methodologies that can efficiently and objectively perform this characterization of sleep.As part of the CinC/Physionet 2018 challenge to automatically detect arousals in a large, expert-annotated sleep dataset, we extracted 27 spectral and time domain features, chosen for their physiological relevance, from the available training set and implemented two contrasting methods, Generalized Linear Model (GLM) and Random Forest (RF), to classify arousals and non-arousals.We were able to achieve non-trivial classification accuracy, even in an imbalanced data set with far fewer arousals than non-arousals.This suggests that large machine learning problems can still benefit from physiology-informed feature selection, especially in the biomedical space.