Multidimensional Approach to Phenotype Obstructive Sleep Apnea (OSA): Integrating Biological, Clinical and Computational Insights

Background Obstructive Sleep Apnea (OSA) is a common, heterogeneous sleep disorder marked by recurrent upper airway collapse during sleep, leading to systemic consequences. This disorder remains underdiagnosed and suboptimally managed due to the limitations of current diagnostic tools and the one-size-fits-all therapeutic approach. There is a growing recognition of the need for precision medicine strategies, supported by advanced phenotyping methodologies that consider both biological and physiological patient-specific traits. Objective This study aims to develop and apply a multidimensional approach to characterize the phenotypes of OSA by integrating molecular biomarkers, physiological traits, with a focus on the arousal threshold (ArTH), and machine learning techniques. The ultimate goal is to refine risk stratification, improve treatment choices, and optimize prediction of clinical outcomes. Methods Data from separate cohorts were analyzed: (1) OSA and patients with metastatic colorectal cancer (mCRC) undergoing polysomnographic evaluation and miRNA expression; (2) subjects naïve to OSA stratified by obesity and ArTH; and (3) a retrospective cohort with 15 years of follow-up used for predictive mortality analysis by machine learning models. In the first case, circulating levels of six microRNAs (miR-21, miR-23b, miR-26a, miR-27b, miR-145, miR-210) were quantified by qRT-PCR. In the second, ArTH was estimated by a validated clinical model. Finally, analysis of phenotypes and predictors of mortality was conducted by different machine learning models. Results In the mCRC cohort, 37% of patients were diagnosed with OSA. Patients with OSA (ONCO-OSA) showed upregulated expression of miR-21, miR-23b, miR-26a and miR-210, which correlated with poor response to chemotherapy and reduced overall and progression-free survival. In the physiologic cohort, patients with elevated ArTH, especially those who were obese, had more severe OSA, higher nocturnal hypoxemia (T90), and higher prevalence of cardiometabolic comorbidities. Unconventional therapy of OSA was more effective in non-obese subjects with low ARTH. In the long-term cohort (n=402), cluster analysis identified three phenotypes differing in age, BMI, comorbidity burden, gas exchange metrics, and mortality. Notably, apnea-hypopnea index (AHI) was not correlated with survival, while clusters incorporating nocturnal hypoxemia and comorbidity burden were strongly predictive. Machine learning identified age and nocturnal hypoxemia as the most significant predictors of mortality. Conclusion This study demonstrates that a multidimensional phenotyping model, which integrates biomarkers, physiological traits, and computational tools, can better capture the complexity of OSA. MicroRNA profiles, arousal threshold, and obesity emerge as key factors in disease expression and response to treatment. Compared with standard-based metrics, phenotypic clustering more accurately predicts outcomes, reinforcing the need to move away from AHI-centered models to individualized assessment strategies that incorporate dynamic patient profiles for optimized management.