CPAP and Cardiovascular Outcomes: Why the Trials Disappoint — and Who Actually Benefits.

The observational signal that started the question

Obstructive sleep apnea (OSA) is independently associated with cardiovascular events. The size of that association is not small. In one widely-cited prospective cohort, adults with severe untreated OSA had roughly three times the rate of fatal cardiovascular events and three to four times the rate of non-fatal events compared with healthy controls, after adjustment for the usual cardiovascular risk factors [1]. Subsequent meta-analyses of population-level cohorts have produced a consistent picture: untreated moderate-to-severe OSA carries a 60 to 80% higher relative risk of major adverse cardiovascular events compared with no OSA, with the strongest signal for stroke and incident hypertension [2].

Mechanistically the story is also tidy. Repeated obstructive events produce intermittent hypoxia, large negative intrathoracic pressure swings, surges in sympathetic outflow, and downstream endothelial dysfunction, hypertension, atrial remodeling, and increased platelet reactivity. Each of those signals individually predicts cardiovascular events. Stack them across years of nightly exposure and the cumulative effect is the kind of biology that should respond to an intervention that abolishes the apneas.

That is the picture that drove a decade of CPAP-prescribing for cardiovascular prevention. Patients with OSA and established cardiovascular disease were given CPAP not only to feel less tired during the day but also — explicitly or implicitly — to lower their risk of a second heart attack or first stroke. By the early 2010s most major society guidelines reflected that assumption.

Then three randomized trials tested it.

SAVE: the trial everyone wanted to be positive

The Sleep Apnea Cardiovascular Endpoints (SAVE) trial, published in the New England Journal of Medicine in 2016, was the largest of the three [3]. SAVE enrolled 2,717 adults aged 45 to 75 with moderate-to-severe OSA (oxygen desaturation index ≥ 12) and established coronary or cerebrovascular disease, drawn from 89 centers across seven countries. Participants were randomized to CPAP plus usual care or usual care alone and followed for a mean of 3.7 years.

The primary composite endpoint — death from cardiovascular causes, myocardial infarction, stroke, or hospitalization for unstable angina, heart failure, or transient ischemic attack — occurred in 229 patients in the CPAP arm and 207 patients in the usual-care arm. The hazard ratio was 1.10 (95% confidence interval 0.91 to 1.32, p = 0.34). No cardiovascular subcomponent showed a statistically significant benefit. CPAP did improve daytime sleepiness as measured by the Epworth Sleepiness Scale, reduced depression and anxiety symptoms, and improved health-related quality of life. But the trial was designed to answer a cardiovascular question, and on the cardiovascular question the answer was flat.

Two design choices in SAVE turned out to matter enormously for interpretation. First, SAVE enrolled non-sleepy patients (Epworth score ≤ 15 at screening). This was deliberate — the trial wanted to test the cardiovascular hypothesis cleanly without the symptomatic-driven CPAP use that confounds usual-care comparisons. Second, mean nightly CPAP adherence was 3.3 hours — below the 4-hour threshold most clinicians associate with meaningful symptomatic and physiological benefit. Both of those choices were defensible. Both of those choices also pushed the result toward null.

ISAACC and RICCADSA: the confirmatory disappointments

ISAACC, a Spanish multicenter trial published in 2020, randomized 1,264 non-sleepy adults with acute coronary syndrome and moderate-to-severe OSA to CPAP or usual care and followed them for a median of 3.35 years [4]. The composite cardiovascular endpoint occurred in 16% of CPAP patients and 17% of controls (hazard ratio 0.89, 95% CI 0.68 to 1.17, p = 0.40). Again, null.

RICCADSA, a Swedish trial published in 2016, enrolled 244 patients with revascularized coronary artery disease and non-sleepy OSA and followed them for a median of 57 months [5]. The intention-to-treat analysis was negative. A post-hoc analysis restricted to patients who used CPAP at least 4 hours per night suggested a benefit, but post-hoc, adherence-conditional analyses always carry a healthier-user-bias asterisk that should be read carefully.

A 2024 meta-analysis pooling SAVE, ISAACC, RICCADSA, and several smaller secondary-prevention trials produced an aggregate hazard ratio for major adverse cardiovascular events of approximately 1.00 — exactly the null [6]. By every conventional standard of clinical-trial inference, the cardiovascular-prevention hypothesis for CPAP in non-sleepy secondary-prevention populations is not supported.

The trials are not wrong. They asked a specific question in a specific population and got a specific answer: at the group-average level, CPAP added to usual care in non-sleepy adults with established cardiovascular disease and moderate-to-severe OSA does not reduce cardiovascular events.

Why the trials went flat — three honest explanations

Three explanations have been offered for the gap between the strong observational signal and the null trial results. All three are partly true.

Selection of non-sleepy patients. The observational signal is largest in patients with severe symptomatic OSA. SAVE, ISAACC, and RICCADSA all enrolled non-sleepy patients, in part for ethical reasons (it is difficult to randomize a severely symptomatic patient to no treatment) and in part to isolate the cardiovascular question from the symptomatic one. Non-sleepy OSA may simply be a milder physiological exposure. The hypoxic burden, autonomic disturbance, and arousal architecture in a non-sleepy patient differ measurably from those of a sleepy patient at the same apnea-hypopnea index.

Adherence. Mean nightly use in SAVE was 3.3 hours. The biological exposure being tested is "CPAP-treated sleep." If patients spend roughly half their sleep period untreated every night, the trial is testing partial treatment, not treatment. The 4-hour-conditional post-hoc analyses in SAVE and RICCADSA do show some cardiovascular signal — but those analyses are no longer randomized comparisons, and the patients who use CPAP for 4+ hours per night are systematically different from those who do not.

The apnea-hypopnea index is a poor exposure metric. OSA severity has traditionally been graded by the apnea-hypopnea index (AHI). AHI counts events. It does not measure the magnitude of oxygen desaturation, the duration of hypoxic burden, the arousal architecture, the autonomic surge, or the negative intrathoracic pressure swings — all of which are mechanistically closer to cardiovascular injury. Two patients with the same AHI can have very different hypoxic burdens and very different cardiovascular trajectories. A trial that enrolls and stratifies by AHI alone is grouping together biologically distinct populations.

Each of these explanations dilutes the average treatment effect. Stacked, they predict roughly what SAVE found: a real biological intervention applied to a partially-treated, mechanistically heterogeneous, symptomatically mild population producing a group-average effect indistinguishable from zero.

The 2026 causal-forest re-analysis

In March 2026, Cohen and colleagues at Mount Sinai published an individualized-treatment-effects analysis of SAVE in Communications Medicine [7]. The method — causal survival forest — is a machine-learning framework designed specifically for the problem of "the average treatment effect is null, but the treatment effect might still be large in subgroups defined by features the trial collected." Rather than estimating one hazard ratio for the whole trial population, causal survival forest estimates a personalized hazard-ratio for each individual based on baseline features, and then groups individuals into predicted-benefit and predicted-harm strata.

The investigators used 2,687 SAVE participants with complete baseline data and 23 baseline features chosen from a larger pool of more than 100 candidates — combinations of demographic data, prior medical conditions, comorbidities, smoking status, baseline sleep-study metrics, and quality-of-life measures. The model produced an individualized treatment effect score for every participant.

Two findings drove the headline coverage. First, the participants whose model-predicted treatment effect placed them in the predicted-benefit subgroup experienced markedly lower cardiovascular event rates on CPAP than on usual care — the kind of difference SAVE was originally powered to detect, but only in this subset. Second, the participants in the predicted-harm subgroup experienced markedly higher cardiovascular event rates on CPAP than on usual care. Two large signals, of opposite sign, averaging to zero at the trial level.

It is worth being clear about what this analysis is and is not. It is a hypothesis-generating re-analysis of a single completed randomized trial. It identifies subgroups whose characteristics are consistent with differential CPAP response. It does not constitute prospective validation. A predicted-benefit subgroup defined on SAVE features will need to be re-derived on an independent cohort, and ideally tested in a prospective enrichment trial, before the framework changes clinical care. But methodologically, this is the kind of analysis that has repeatedly rescued real biological signals from "negative" cardiovascular trials in the past decade — the cardiology literature has multiple examples of subgroup-defined response that survived prospective replication.

Who the modeling says actually benefits

The features that loaded most heavily on predicted CPAP benefit in the Mount Sinai analysis are consistent with what mechanistically-minded sleep-medicine clinicians have been saying for years: the patients with the largest predicted benefit are those with the highest hypoxic burden, the strongest autonomic signature of nighttime stress, and the cardiometabolic profile most likely to respond to abolishing those signals.

That maps onto a few archetypes worth naming. The first is the patient with severe OSA, high oxygen-desaturation index, and a relatively preserved sympathetic-response curve — the body is still reactive to the apneic events, which is exactly the population where removing the signal should change downstream cardiovascular biology. The second is the patient with OSA and concurrent metabolic dysregulation — uncontrolled hypertension, insulin resistance, visceral adiposity — where the nightly sympathetic surge is acting on already-fragile vascular physiology. The third is the patient with frequent oxygen desaturations during REM sleep specifically, a feature increasingly understood as the most cardiovascularly damaging sub-phenotype of OSA.

The predicted-harm subgroup is harder to characterize cleanly and the data are exploratory, but the features that load on predicted harm are broadly consistent with patients whose baseline cardiovascular risk is dominated by non-OSA pathways — established advanced atherosclerosis, multiple comorbidities, low predicted physiological reserve — where the intervention may be displacing other therapeutic energy without offering meaningful biological return, or where adherence problems and mask-related sleep fragmentation may net out negatively. None of this should be interpreted as "CPAP harms these patients"; it should be interpreted as "the average effect of CPAP in this subgroup, in this dataset, was negative, and it needs prospective replication before it changes anything."

Mechanism: why the signal is real even when the average is zero

It is tempting to read the negative trials as "the cardiovascular hypothesis for CPAP was always wrong." The mechanistic literature does not support that reading. The biological signal it pulls — abolishing intermittent hypoxia, eliminating the negative-pressure swings, restoring nocturnal blood-pressure dipping, reducing sympathetic outflow — has reproducible measurable physiological effects.

Blood pressure. CPAP lowers 24-hour ambulatory blood pressure by a modest but real 2 to 3 mmHg on average, with larger effects in patients with resistant hypertension and in patients with high adherence [8]. In the resistant-hypertension population specifically, CPAP-associated blood-pressure reductions are clinically meaningful and would, by any cardiovascular-outcome model, predict a mortality benefit over decades.

Endothelial function. CPAP improves flow-mediated dilation in patients with moderate-to-severe OSA, with effect sizes reproducible across multiple smaller mechanistic trials [9]. Endothelial function is upstream of atherosclerosis progression and downstream cardiovascular events.

Autonomic balance. CPAP shifts heart-rate variability toward greater parasympathetic dominance, reduces overnight sympathetic outflow as measured by microneurography, and reduces nocturnal arrhythmia burden in patients with concurrent atrial fibrillation. The atrial-fibrillation signal in particular has been observed across multiple independent datasets — CPAP-adherent patients with OSA and AF show meaningfully lower recurrence rates after rhythm-control procedures.

Inflammatory and metabolic markers. CPAP reduces high-sensitivity C-reactive protein and several inflammatory cytokines, improves insulin sensitivity modestly in non-obese OSA patients, and reduces overnight cortisol surges. None of these are large effects individually. Stacked, they describe a real biological intervention whose downstream cardiovascular implications are not zero.

The honest synthesis is this: the biology is real, the mechanisms are reproducible, and the cardiovascular pathway is plausible. The randomized trials measured what they were designed to measure — the average effect in a non-sleepy, partially-adherent, mechanistically heterogeneous population — and found it indistinguishable from zero. Those two statements are not contradictory. They are the structure of nearly every cardiovascular-intervention story where the underlying physiology is real but the patient population is heterogeneous and the exposure metric is coarse.

What this changes in practice

Three practical shifts are worth attention, even before the causal-forest framework is prospectively validated.

Stop treating "CPAP for cardiovascular prevention" as a single yes/no question. A patient with severe OSA, high oxygen-desaturation index, REM-predominant events, resistant hypertension, atrial fibrillation, and a willingness to use the device is a fundamentally different cardiovascular-prevention candidate than a patient with mild non-sleepy OSA, normal nocturnal blood-pressure dipping, well-controlled cardiometabolic risk, and a history of declining mask therapy. The first patient has every reason to expect cardiovascular benefit from adherent CPAP. The second does not, and the randomized data correctly say so.

Look beyond AHI when grading severity. Modern sleep-medicine literature increasingly grades OSA by hypoxic burden, autonomic markers, arousal frequency, and event-specific desaturation depth rather than AHI alone. A clinician reading a sleep study should be looking for total time below 90% oxygen saturation, REM-AHI versus NREM-AHI, and the magnitude of nocturnal blood-pressure non-dipping — not just the overall event count. This matters even more when the cardiovascular-prevention question is on the table.

Adherence is the rate-limiting step. If the predicted-benefit phenotype is real, CPAP only delivers in patients who actually use it. Modern auto-titrating devices, heated-humidified circuits, varied mask styles, and the recent generation of more comfortable interfaces have meaningfully improved adherence over the 3.3-hour SAVE-era average. Sleep-medicine programs that aggressively troubleshoot adherence in the first 30 days of therapy produce dramatically better long-term use rates, and adherence is what converts the physiology into outcomes.

Caveats the modeling cannot fix

The causal-forest framework rescues real signal from heterogeneous data. It does not solve every problem.

First, the predicted-benefit subgroup defined on SAVE features needs prospective external validation. A model that perfectly fits its derivation cohort can fail completely on the next dataset. The cardiology literature is full of heterogeneity-defined response subgroups that survived initial discovery but did not survive replication.

Second, the modeling depends on the features SAVE collected. If the true effect-modifying biology runs through features the trial did not measure — REM-specific desaturation depth, specific genotypes, particular cardiac MRI markers — the model can only approximate them through correlated covariates. The next generation of cardiovascular-OSA trials should be designed prospectively around the candidate effect-modifiers identified by this kind of re-analysis, with deeper phenotyping at baseline.

Third, CPAP is not the only OSA therapy. Mandibular advancement devices, hypoglossal nerve stimulation, weight loss, positional therapy, and combination approaches all change the exposure profile in different ways. The cardiovascular-prevention question for each modality has its own trial structure and its own evidence base, and conflating CPAP-specific findings with "OSA treatment" findings broadly is a category error. GLP-1 agonists are now being tested as OSA therapy in their own right, and the cardiovascular outcomes from those trials will be a separate, important readout.

Fourth, the OSA-cardiovascular literature has a real sleep-health story underneath it that does not depend on resolving the CPAP-prevention question. Untreated symptomatic OSA degrades sleep quality, daytime function, mood, driving safety, and quality of life. The case for treating symptomatic OSA stands on those grounds independently of whether the additional cardiovascular signal in non-sleepy populations is large or small.

Frequently asked questions

Did the SAVE trial show CPAP prevents heart attacks and strokes?

No. SAVE randomized 2,717 adults with moderate-to-severe obstructive sleep apnea and known cardiovascular disease to CPAP plus usual care or usual care alone. After 3.7 years of mean follow-up the primary composite cardiovascular endpoint occurred in 17.0% of CPAP patients vs 15.4% of usual-care patients (hazard ratio 1.10, 95% CI 0.91–1.32, p = 0.34). CPAP improved daytime sleepiness, mood, and quality of life but did not reduce the cardiovascular event rate at the group-average level.

If CPAP doesn't reduce cardiovascular events, why is it still prescribed?

CPAP has well-demonstrated benefits for daytime sleepiness, mood, quality of life, modest blood-pressure reduction, and safer driving in patients with moderate-to-severe OSA. The cardiovascular-prevention question is more specifically about whether treating mostly-non-sleepy OSA patients to prevent secondary cardiac events works — and at the group-average level in the trials we have, it does not. Symptomatic OSA treatment and cardiovascular-prevention treatment are different clinical questions with different evidence bases.

What did the 2026 Mount Sinai re-analysis of SAVE find?

Investigators applied causal survival forest modeling to 2,687 SAVE participants and 23 baseline features to estimate individualized treatment effects. They identified a subgroup predicted to benefit from CPAP — and within that subgroup, CPAP assignment was associated with markedly lower cardiovascular event rates compared with usual care. They also identified a subgroup predicted to be harmed by CPAP, in whom assignment was associated with higher cardiovascular event rates. The average treatment effect of zero was hiding two large opposing signals.

Should I use CPAP if I have sleep apnea but no daytime sleepiness?

Discuss the decision with a sleep-medicine clinician. The randomized cardiovascular-prevention evidence is weakest in the non-sleepy OSA population — that is the population SAVE, ISAACC, and RICCADSA all studied. Adherence to CPAP is also lowest in non-sleepy patients because the immediate symptomatic benefit is harder to feel. Modern decisions weigh baseline cardiovascular risk, hypoxic burden, autonomic markers, and patient preference rather than treating the diagnosis automatically.