Machine Learning Model Combining Ventilatory, Hypoxic, Arousal Domains Across Sleep Better Predicts Adverse Consequences of Obstructive Sleep Apnea

Obstructive sleep apnea(OSA) severity is currently assessed clinically using the apnea-hypopnea index (AHI), which is inconsistently associated with short- and long-term outcomes. Ventilatory, hypoxic, and arousal domains are known to exhibit abnormalities in OSA. Using the same set of features across these three domains, albeit using different models, we show that a physiology-guided ML approach can better predict adverse consequences of OSA compared to the AHI. The proposed approach utilizes known pathophysiology of OSA with the power of AI/ML to transform our understanding of OSA and its consequences such as Excessive Daytime Sleepiness and All-Cause Mortality to guide clinical decision-making. Using an XGBoost-model, the proposed approach obtained an AUROC of 0.81 for Excessive Daytime Sleepiness and 0.93 for All-Cause Mortality. In contrast, the model with AHI alone achieved AUROC values of less than 0.6 for either outcome suggesting that a physiology-guided ML approach may be better at combining ventilatory/hypoxic/arousal domains than AHI.