Exploring phenotypes to improve long-term mortality risk stratification in obstructive sleep apnea through a machine learning approach: an observational cohort study

Background: Obstructive sleep apnea (OSA) is a heterogeneous sleep disorder for which the identification of phenotypes might help for risk stratification for long-term mortality. Thus, the aim of the study was to identify distinct phenotypes of OSA and to study the association of phenotypes features with long-term mortality by using machine learning.

Methods: This retrospective study included patients diagnosed with OSA who completed a 15-year follow-up and were adherent to continuous positive airway pressure (CPAP) therapy. Multidimensional data were collected at baseline and were used to identify OSA phenotypes using the hierarchical approach. Associations between phenotypic features and long-term mortality were assessed using supervised analysis.

Results: A total of 402 patients, predominantly male (70 %), were included. Clustering analysis identified three distinct phenotypes: Cluster 1 (middle-aged, severely obese, very severe OSA with nocturnal hypoxemia), Cluster 2 (young, overweight, moderate OSA with limited nocturnal hypoxemia), and Cluster 3 (elderly, obese, multimorbid, severe OSA with nocturnal hypoxemia). Mortality was significantly higher in Clusters 1 and 3 (p < 0.001). Supervised methods identified eight main features of these clusters, among which nocturnal hypoxemia was found to be the main risk factor for mortality even after confounding factors-adjustment (hazard ratio 2.63, 95 % confidence interval 1.09-6.36, p = 0.032).

Conclusions: This study demonstrated the interest of attributing OSA patients to distinct phenotypes including precise determination of nocturnal hypoxemia to improve mortality risk stratification.