Machine Learning-Based Multidimensional Oximetry for Obstructive Sleep Apnea Screening: Development and External...

Area of Science:

• Sleep Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Obstructive sleep apnea (OSA) affects nearly one billion people globally, representing a significant public health challenge.
• There is a critical need for effective and accessible methods for identifying individuals at risk of OSA.

Purpose of the Study:

• To develop and externally validate a machine learning model using multi-parameter pulse oximetry (SpO2) for OSA screening.
• To evaluate the model's performance, interpretability, and robustness across different sex and age groups.

Main Methods:

• Utilized eight SpO2-derived parameters, including oxygen desaturation index (ODI) and hypoxic burden (HB), from 2195 participants (internal) and 446 (external).
• Trained six machine learning algorithms, prioritizing F1-score and area under the curve, with Shapley additive explanations for interpretability.

Main Results:

• A 4-parameter model (ODI-HB-MinSpO2-ST90) demonstrated optimal performance (F1-score=0.9516, AUC=0.9879), outperforming single-parameter models.
• Categorical boosting algorithm showed superior performance and robustness, with ODI, HB, and MinSpO2 identified as key predictors.
• Sex- and age-specific model configurations improved performance in certain subgroups.

Conclusions:

• A multi-parameter oximetry model using categorical boosting offers a simple, accurate tool for OSA screening.
• Sex- and age-stratified approaches can enhance the clinical utility of this OSA screening tool.