Abstract Body (Do not enter title and authors here): Background: Obstructive sleep apnea (OSA) affects about 25% of middle-aged adults and more than doubles cardiovascular disease (CVD) mortality via sympathetic activation, oxidative stress, and metabolic derangements. Clinical risk stratification still relies almost solely on the Apnea-Hypopnea Index (AHI), overlooking heterogeneity in multimorbidity and therapeutic response.

Hypothesis: To test whether an unsupervised machine learning pipeline can reveal stable, clinically distinct OSA-CVD phenotypes with different comorbidity burdens and temporal trajectories.

Methods: The Wisconsin Sleep Cohort provided 2,570 polysomnographic assessments from 1,123 adults across up to five visits. A two-stage workflow was applied. Stage one used Ward linkage clustering on z-standardized variables, trained an extreme-gradient-boosted tree to rank predictor importance and repeated clustering until fifteen high-yield predictors remained. Stage two reclustered the reduced dataset to assign final labels. Five-fold cross-validation assessed reproducibility. Comorbidity patterns were contrasted, and first-order Markov chains generated transition matrices describing stability.

Results: Unsupervised clustering produced four reproducible phenotypes with cross-validated accuracy at 0.82 and adjusted Rand at 0.96. The healthy sleeper group (mean AHI = 6.9 events/h) showed only 2.7% prevalent CVD, whereas the severe OSA (AHI = 52) presented 20.3% CVD. Two mild-OSA groups shared a mean AHI around 11 but differed metabolically: the metabolically healthy phenotype had 4.8% diabetes and 16.5% CVD, while the metabolically healthy phenotype had 54.5% diabetes and 20.7% CVD. Markov modeling revealed strong year-to-year stability for all phenotypes except the severe OSA group, from which 44% of participants migrated to the metabolically unhealthy pattern. Smaller bidirectional flows (10–15%) between the two mild phenotypes suggest gradual, reversible cardiometabolic drift.

Conclusions: A data-driven unsupervised framework delineated four reproducible OSA-CVD phenotypes and mapped their temporal evolution. Integrating these phenotypes into routine care may enable proactive surveillance and personalized intervention that lower CVD morbidity among adults with OSA.