Abstract
Abstract
Objective: Sleep apnea (SA) diagnosis remains challenging due to the cost and limited accessibility of polysomnography. Most automated screening approaches estimate the Apnea–Hypopnea Index (AHI) from oximetry-derived features and do not provide temporally resolved characterization of respiratory events. This study investigates the feasibility of dense temporal respiratory event profiling and clinically meaningful SA severity estimation from single-channel photoplethysmography (PPG).
Approach: A two-stage framework was developed using single-channel PPG. A fully convolutional neural network operating on PPG-derived waveforms outputs time-resolved probabilities for normal breathing, central apnea, obstructive apnea, hypopnea, and desaturation-only states. These predictions are aggregated via a machine learning metamodel to estimate clinically relevant indices including AHI, Obstructive – Central AHI difference, and the Apnea-to-Apnea-Hypopnea ratio. Models were trained and evaluated on 2029 polysomnography studies from the Multi-Ethnic Study of Atherosclerosis.
Main results: Fine-grained discrimination between respiratory event subtypes remained challenging, particularly for rare classes like central apnea, whereas aggregated respiratory disturbance detection was more reliable (apnea AUPRC = 0.28, prevalence = 1.9%, AUROC = 0.92). In the PPG-only setting, AHI estimation explained 76% of variance; inclusion of an auxiliary SpO₂-derived desaturation feature improved this to 83%. SpO₂ assisted SA severity classification achieved Macro F1 of 0.71, with only 1.18% of apnea cases misclassified as non-apneic.
Significance: These findings support the feasibility of dense temporal respiratory disturbance profiling and SA severity estimation from single-channel PPG, while auxiliary oximetry can improve severity estimation when available. They also indicate that fine-grained subtype discrimination remains challenging from PPG alone.