Machine Learning Model Combining Ventilatory, Hypoxic, Arousal Domains Across Sleep Better Predicts Adverse...

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Medical Diagnostics

Background:

• Obstructive Sleep Apnea (OSA) severity is clinically assessed using the Apnea-Hypopnea Index (AHI).
• AHI shows inconsistent associations with short- and long-term health outcomes.
• Known OSA pathophysiology involves abnormalities in ventilatory, hypoxic, and arousal domains.

Purpose of the Study:

• To develop a more accurate method for predicting adverse OSA consequences.
• To compare a physiology-guided machine learning (ML) approach against the AHI for outcome prediction.
• To leverage AI/ML and OSA pathophysiology for improved clinical decision-making.

Main Methods:

• Utilized a physiology-guided ML approach incorporating features from ventilatory, hypoxic, and arousal domains.
• Employed an XGBoost model for prediction.
• Compared ML model performance against the AHI using Area Under the Receiver Operating Characteristic curve (AUROC).

Main Results:

• The physiology-guided ML approach achieved an AUROC of 0.81 for predicting Excessive Daytime Sleepiness.
• The ML approach achieved an AUROC of 0.93 for predicting All-Cause Mortality.
• The AHI alone resulted in AUROC values below 0.6 for both outcomes.

Conclusions:

• A physiology-guided ML approach demonstrates superior prediction of adverse OSA outcomes compared to the AHI.
• This AI/ML method effectively integrates multiple OSA pathophysiological domains.
• The findings suggest a potential for improved clinical decision-making in OSA management.