Deep learning for obstructive sleep apnea diagnosis based on single channel oximetry

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Respiratory Medicine

Background:

• Obstructive sleep apnea (OSA) is a prevalent but challenging condition to diagnose.
• Current home sleep testing methods have limitations impacting diagnostic accuracy.
• Accurate OSA diagnosis is crucial for effective patient management and preventing comorbidities.

Purpose of the Study:

• To develop and validate a deep learning model (OxiNet) for estimating the apnea-hypopnea index (AHI) using solely the oximetry signal.
• To assess OxiNet's diagnostic performance across diverse demographic groups and comorbidities.
• To compare OxiNet's efficacy against established diagnostic benchmarks.

Main Methods:

• Retrospective analysis of 12,923 polysomnography recordings from six independent datasets.
• Development of OxiNet, a deep learning algorithm utilizing oximetry data for AHI estimation.
• Rigorous evaluation of OxiNet's performance, including subgroup analyses by ethnicity, age, sex, and comorbidity.

Main Results:

• OxiNet demonstrated high accuracy in estimating the AHI from oximetry signals.
• The model showed consistent performance across various demographic factors and comorbidities.
• OxiNet missed only 0.2% of moderate-to-severe OSA cases, compared to 21% missed by the best benchmark method.

Conclusions:

• OxiNet offers a promising, accurate, and potentially more accessible method for diagnosing obstructive sleep apnea using oximetry.
• The AI model's ability to perform across diverse populations enhances its clinical utility.
• OxiNet represents a significant advancement over current benchmark methods, potentially improving OSA screening and diagnosis.