SuperLearner ML Improves Cardiovascular Risk Prediction in OSA

A SuperLearner ensemble machine learning framework outperformed conventional single-model approaches in predicting cardiovascular risk in patients with obstructive sleep apnea (OSA) when tested on a geographically distinct held-out population — addressing one of the central limitations that has prevented epidemiologic risk models from translating reliably into clinical practice.

Oren Cohen, MD, assistant professor of medicine in the Division of Pulmonary, Critical Care and Sleep Medicine at the Icahn School of Medicine at Mount Sinai in New York, is co-primary author of the work, presented at the 2026 American Thoracic Society (ATS) International Conference in Orlando, Florida. First author Shishir Adhikari, a postdoctoral researcher in the laboratory of Naomi Shah, MD, MPH, MS, conducted the primary analysis in close collaboration with biostatistician Mayte Suárez-Fariñas, PhD, professor of Population Health Science and Policy at Icahn School of Medicine.

The analysis used data from 2,159 participants in the Multi-Ethnic Study of Atherosclerosis (MESA) sleep ancillary study with no prior cardiovascular events, using exam 5 clinical, sleep, and coronary artery calcium data to predict time-to-cardiovascular event. The dataset was split into training (65%) and test (35%) sets, with participants from Midwestern US sites held out to simulate distributional shift — the scenario in which a model trained on one population is deployed in a meaningfully different one. The SuperLearner combined multiple base learners including random forest and XGBoost, and was compared against gradient boosting and conventional survival models under different missing data assumptions. On held-out test data, the SuperLearner explicitly modeling missing-not-at-random (MNAR) data achieved the highest discrimination (C-index, 0.716), outperforming gradient boosting with Cox loss under MAR assumptions (0.705) and SuperLearners assuming MCAR (0.706) or MAR (0.698).

In an interview with HCPLive, Cohen emphasized the research’s rationale: randomized trial data have not demonstrated that CPAP reduces cardiovascular events, yet clinicians increasingly face patients — often self-referred based on consumer device data — worried about OSA-related cardiac risk. Better risk stratification could identify which asymptomatic patients warrant treatment and in whom CPAP might actually modify cardiovascular outcomes.

"We don't really know which patients with sleep apnea are at highest risk, and more importantly, which ones need to be treated or where treatment with CPAP can modify those cardiovascular outcomes — once we get to that, we'll be able to… be precise in who we treat, so that we're not pushing treatment on the millions of Americans with sleep apnea who may be asymptomatic with no apparent benefit," Cohen said.

Cohen had no relevant disclosures to report.

Reference

Adhikari S, Cohen O, Suárez-Fariñas M, et al. Improving the Generalizability of Sleep-based Cardiovascular Risk Prediction Using Super Learner Ensemble. Presented at: ATS International Conference; Orlando, Florida; May 2026.