Screening for Obstructive Sleep Apnea Risk by Using Machine Learning Approaches and Anthropometric Features.

Area of Science:

• Medical Informatics
• Sleep Medicine
• Machine Learning

Background:

• Obstructive sleep apnea (OSA) is a widespread health issue.
• Current diagnosis via polysomnography (PSG) is time-consuming and resource-intensive.

Purpose of the Study:

• Develop machine learning models for screening moderate to severe and severe OSA risk.
• Utilize easily accessible anthropometric features for risk assessment.

Main Methods:

• Collected anthropometric and PSG data from 3503 Taiwanese patients.
• Developed and compared six machine learning models (logistic regression, k-NN, Naïve Bayes, RF, SVM, XGBoost).
• Validated models using 80/20 training/testing split and analyzed feature importance with Shapley values.

Main Results:

• The Random Forest (RF) model demonstrated the highest accuracy: 84.74% for moderate to severe OSA and 72.61% for severe OSA.
• Visceral fat level emerged as the most significant predictor for OSA risk screening.
• Models showed potential for screening OSA risk in Taiwanese populations and similar craniofacial groups.

Conclusions:

• Machine learning models based on anthropometric data provide an efficient method for OSA risk screening.
• Visceral fat is a critical factor in identifying individuals at risk for moderate to severe and severe OSA.
• These models offer a practical alternative to traditional PSG for OSA screening in specific populations.