Deep Learning Classifies Sleep Apnea from EEG Signals

Diagnosing sleep apnea traditionally relies on polysomnography, a resource-intensive process involving manual data analysis and scoring. Recent advancements have shown that the central nervous system effects of sleep apnea events can be detected through electroencephalogram (EEG) signals, offering a less invasive and potentially automated diagnostic pathway. This research presents a thorough comparison of various deep learning architectures and feature representations specifically for automated sleep apnea detection. The study focused on multichannel EEG signals from a single dataset of pediatric subjects, evaluating Vision Transformers and Graph Attention Networks across different signal representations, including raw temporal signals, spectrograms, coherence-based graphs, and two types of topological data analysis (TDA) features. The most promising result was a test AUC of 0.750 achieved by a Vision Transformer model trained on TDA features. While demonstrating the feasibility of EEG-based automated Obstructive Sleep Apnea (OSA) screening, the study also revealed significant performance variations across patient demographics (age, sex, AHI severity) and sleep stages. These findings underscore both the potential and the remaining challenges for integrating such automated systems into clinical practice.