Risk factors for pulmonary mortality in patients with obstructive sleep apnea: a competing risk analysis integrating...

Obstructive sleep apnea (OSA) is associated with increased all-cause mortality, yet the specific determinants of pulmonary mortality in this population remain poorly characterized.

To identify independent risk factors for pulmonary-specific death in OSA patients using complementary statistical and machine-learning approaches within a competing risk framework.

We analyzed 4368 patients consecutively referred to a tertiary sleep center (assessed between January 2005 and December 2025), including 2304 with moderate-to-severe OSA (apnea-hypopnea index ≥15 events/h). Pulmonary mortality (International Classification of Diseases, Tenth Revision, J codes) was modeled using cause-specific Cox regression and Fine-Gray subdistribution hazard models. A random survival forest (RSF) competing risk model quantified variable importance, and Gower-distance hierarchical clustering identified phenotypes.

Over a median (interquartile range) follow-up of 8.6 (6.7-12.5) years (range, 0.04-20 years), 101 pulmonary deaths occurred. In multivariable cause-specific Cox models, age (hazard ratio [HR], 1.11; 95% CI, 1.09-1.13), nocturnal desaturation (HR, 3.09; 95% CI, 1.88-5.09), active smoking (HR, 2.13; 95% CI, 1.37-3.30), body mass index (HR, 1.06; 95% CI, 1.02-1.10), and male sex (HR, 1.61; 95% CI, 1.01-2.56) were independent predictors. Fine-Gray models yielded concordant results. The RSF achieved a C-index of 0.84. Clustering identified 4 phenotypes; the highest-risk cluster reached a 20-year cumulative pulmonary mortality approaching 8%.

Nocturnal hypoxemic burden, active smoking, age, body mass index, and cardiometabolic comorbidities are the principal correlates of pulmonary mortality in OSA. To our knowledge, this is the first large-scale competing-risk analysis dedicated specifically to pulmonary mortality in OSA, integrating cause-specific Cox, Fine-Gray, RSF, and clustering in a single cohort. Machine-learning-based stratification identifies clinically distinct phenotypes that may guide personalized intervention; these findings are hypothesis-generating and require prospective validation.