Mount Sinai scientists have created a machine learning model that predicts whether CPAP therapy will raise or lower heart and stroke risk in people with obstructive sleep apnea. The work could help doctors tailor treatment instead of relying on a one-size-fits-all approach.
For millions of people with obstructive sleep apnea, a nighttime mask attached to a continuous positive airway pressure, or CPAP, machine is the standard treatment. But when it comes to long-term heart health, that one-size-fits-all approach may not be the best fit.
Researchers from the Icahn School of Medicine at Mount Sinai have developed a machine learning model that predicts how CPAP therapy will affect an individual patient’s risk of cardiovascular disease. The tool, described in the journal Communications Medicine, is designed to estimate whether CPAP is likely to lower or raise a person’s chance of future problems such as heart attacks and strokes.
Obstructive sleep apnea is a common disorder in which the airway repeatedly collapses during sleep, causing breathing to stop and start. It affects an estimated 25 million people in the United States and is linked to higher risks of cardiovascular disease, including stroke and heart disease.
CPAP, which delivers a steady stream of pressurized air through a mask to keep the airway open, is considered the most effective treatment for sleep apnea symptoms. Yet large clinical trials have not clearly shown that CPAP reduces cardiovascular events across all patients with the condition.
That disconnect prompted Mount Sinai investigators to ask a more nuanced question: instead of asking whether CPAP helps everyone, could they predict which specific patients are likely to benefit — or be harmed — when it comes to heart and blood vessel health?
To answer it, the team turned to artificial intelligence.
They used a machine learning algorithm to build an analytic model based on data from the Sleep Apnea Cardiovascular Endpoints (SAVE) trial, the largest clinical cohort to evaluate CPAP for preventing cardiovascular disease. The SAVE trial followed more than 2,600 participants at 89 sites in seven countries, providing a rich set of information on sleep patterns, medical history and health outcomes.
From this dataset, the researchers examined more than 100 potential predictors and narrowed them down to 23 key baseline features. These included factors such as prior medical conditions and smoking status, along with detailed sleep and health measures. The model then generated an individualized treatment effect score for each patient, estimating how CPAP would change that person’s cardiovascular risk compared with usual care.
The results suggested that CPAP’s impact on heart health is far from uniform.
The model identified a subgroup of patients for whom CPAP was expected to improve cardiovascular risk. Among participants in this subgroup who were randomly assigned to receive CPAP in the SAVE trial, the researchers observed a 100-fold improvement in future cardiac risk compared with similar patients who received usual care.
At the same time, the tool flagged another subgroup predicted to be harmed by CPAP. In that group, patients who used CPAP experienced more than a 100-fold increase in cardiovascular disease outcomes, including recurrent strokes and heart attacks, compared with those on usual care.
The work points toward a more tailored way to treat sleep apnea, according to co-corresponding author Neomi A. Shah, a professor of medicine (Pulmonary, Critical Care and Sleep Medicine), and artificial intelligence and human health, and an associate chief for academic affairs in the Division of Pulmonary, Critical Care and Sleep Medicine at the Icahn School of Medicine.
“Our findings represent a significant advancement in personalized medicine, moving away from a one-size-fits-all strategy in the treatment of obstructive sleep apnea,” Shah said in a news release. “This underscores the value of new data-driven approaches like our model to assist clinicians in making informed decisions about CPAP treatment recommendations, enhancing personalized care to meet the individual needs of every patient.”
The study highlights a growing trend in medicine: using artificial intelligence not just to spot patterns, but to guide decisions.
Co-primary author Oren Cohen, an assistant professor of medicine (Pulmonary, Critical Care and Sleep Medicine) at the Icahn School of Medicine, emphasized both the promise and the caution needed with such tools.
“These results demonstrate the power of machine learning for prediction of treatment effects in an era of precision medicine; however, such models require careful validation to prove their utility in clinical practice,” Cohen said in the news release.
Unlike many AI systems that simply predict who is at risk for a disease, the Mount Sinai model aims to estimate what will happen if a specific treatment is used. That kind of causal reasoning is crucial if AI is going to help doctors choose therapies, not just label patients.
“Artificial intelligence in medicine must move beyond pattern recognition to causal reasoning,” added co-corresponding author Mayte Suarez-Farinas, a co-director for the Division of Biostatistics and Data Science, and a professor of population health science and policy, and artificial intelligence and human health, at the Icahn School of Medicine. “By estimating individualized treatment effects over time using randomized clinical trial data, we move predictive AI toward decision-support tools grounded in causality and capable of informing real-world treatment decisions and improving outcomes.”
The researchers stress that their model is not ready to replace clinical judgment. It will need to be validated in additional patient groups and tested in real-world settings before it can be used routinely in clinics.
Still, the approach points to a future in which sleep specialists and cardiologists could use AI-driven tools to personalize care. Instead of automatically prescribing CPAP to every eligible patient with obstructive sleep apnea, clinicians might one day be able to weigh a data-driven estimate of how the therapy will affect that individual’s long-term heart and stroke risk.
The work also underscores the value of large, carefully conducted clinical trials like SAVE, which provide the kind of detailed, randomized data that can power more sophisticated AI models. Investigators from the SAVE trial, including those at The George Institute for Global Health in Sydney, the University of New South Wales, the University of Adelaide and Flinders University, contributed to the study.
For patients, the research offers a hopeful message: as data and AI tools improve, treatments for common conditions like sleep apnea may become more precisely matched to each person’s unique health profile. That could mean better outcomes, fewer side effects and a clearer path forward for those trying to protect both their sleep and their hearts.