Long-range interdependencies aware signal analysis for rapid and accurate obstructive sleep apnea prediction

Obstructive sleep apnea, characterized by frequent interruptions in breathing and intermittent hypoxia, is a public health concern due to its potential to increase multiple organ system morbidity and mortality. Although polysomnography is the gold standard for diagnosis, it is laborious, costly, and complex. Here, we present a biological geometry-aware artificial intelligence framework that processes physiological signals to predict sleep apnea severity. This framework extracts multifractal coupling features that capture long-range dependencies among physiological processes and integrates them into deep learning for classification. We validate this approach across five heterogeneous cohorts comprising more than 35,000 subjects. The method achieves robust and consistent performance, with area-under-the-curve values exceeding 0.91 and improvements of 8-18% over state-of-the-art models. The extracted coupling representations exhibit rapid convergence and consistent performance across heterogeneous cohorts. This framework provides a robust representation of physiological dynamics and supports future developments in automated analysis of sleep-disordered breathing.