Machine learning optimization of obstructive sleep apnea screening: development and validation of a gradient boosting...

Area of Science:

• Sleep Medicine
• Artificial Intelligence in Healthcare
• Diagnostic Tools

Background:

• Obstructive sleep apnea (OSA) diagnosis relies on screening tools with varying accuracy.
• Existing screening methods like STOP-Bang and Berlin questionnaires have limitations in diagnostic performance.
• There is a need for more accurate and efficient OSA screening methods.

Purpose of the Study:

• To develop and validate a machine learning-enhanced screening questionnaire for OSA.
• To create a clinically deployable visual prediction framework with improved diagnostic accuracy.
• To compare the performance of gradient boosting algorithms against existing screening paradigms.

Main Methods:

• A mixed-methods study analyzed polysomnography data from 4,036 participants.
• A 15-item questionnaire integrated ESS and STOP-Bang components.
• XGBoost, SVM, ANN, and multinomial logistic regression were evaluated, with XGBoost selected for its superior performance.

Main Results:

• XGBoost achieved high AUC values (0.92-0.97) for OSA severity, outperforming STOP-Bang (0.68) and Berlin (0.72) questionnaires.
• The clinical nomogram showed excellent calibration (C-index: 0.93).
• Neck circumference, BMI, and witnessed apneas were key predictors; screening costs reduced by 39.7% with increased detection efficiency.

Conclusions:

• The gradient boosting-enhanced OSA screening model advances sleep disorder diagnosis.
• It provides clinically actionable risk stratification via interpretable visualization and maintains implementation feasibility.
• This AI integration offers a framework for clinical decision support with potential applications beyond sleep medicine.