Do GLP-1 drugs get people off CPAP? A 74,000-patient emulation says: barely.

Since a GLP-1/GIP agonist was first approved for obstructive sleep apnea, the practical question for patients has been narrower than the trial evidence: does taking one of these drugs actually reduce a person's reliance on their CPAP machine? A study published in JAMA Network Open on 22 December went looking for the answer in electronic health records [s1].

The finding is a small effect, sitting next to a large one that ought to change how you read the small one.

What was done

The researchers used a target trial emulation — an observational design that attempts to mimic the structure of a randomised trial by specifying eligibility, treatment assignment, and follow-up in advance. Data came from the TriNetX research network, which aggregates deidentified electronic health records from US healthcare organisations, covering January 2021 to September 2025 [s1].

Adults with diabetes, obesity, and obstructive sleep apnea who started an incretin receptor agonist were matched against those who started an SGLT2 inhibitor, producing 36,981 matched pairs [s1]. The comparator choice matters: SGLT2 inhibitors are an active treatment for a similar population, which guards against comparing treated patients against untreated ones who differ in every way.

The matched cohort had a mean age of 62.3 years (SD 11.3), was 39.7% female, had a mean BMI of 38.5 (SD 7.2), and was 70.5% White, 18.6% Black, and 6.7% Hispanic [s1]. Mean follow-up was 332.2 days (SD 116.3) [s1].

What was found

CPAP use was recorded in 2,459 incretin initiators (6.6%) against 2,677 SGLT2 inhibitor initiators (7.2%) — a hazard ratio of 0.92 (95% CI, 0.87–0.97) [s1].

Two other outcomes were reported [s1]:

• All-cause mortality: HR 0.68 (95% CI, 0.60–0.77)
• All-cause hospitalisation: HR 0.90 (95% CI, 0.86–0.95)

Tirzepatide showed greater risk reductions than GLP-1 receptor agonists relative to SGLT2 inhibitors [s1].

The mortality number is a warning label

A 32% reduction in all-cause mortality over a mean follow-up of under a year is a very large effect. Randomised cardiovascular outcome trials of these drug classes have produced nothing on that scale over far longer periods.

There are two ways to read it. Either these drugs are dramatically more life-saving than randomised evidence has shown — implausible — or the two matched groups differ in ways the matching did not capture. The second reading is the standard one, and it has a name: healthy-user bias. People who start and stay on an incretin agonist tend to be more engaged with healthcare, better able to tolerate treatment, and less frail than those who do not, and mortality is exquisitely sensitive to exactly those unmeasured differences.

If residual confounding is inflating the mortality estimate, it is almost certainly inflating the CPAP estimate too — which means the true effect on CPAP use may be smaller than an 8% relative reduction, or absent.

That is not a criticism of the study, which is competently designed and reports its comparator, matching, and follow-up transparently. It is a note about how to read the numbers it produces.

What "CPAP use" means here

The authors flag it directly: procedural records for CPAP may not reflect actual use, adherence, or apnea severity [s1]. The outcome is a billing or procedural signal, not a device-recorded hours-of-use figure and not an apnea-hypopnea index.

That distinction matters because the clinical question people care about is whether their breathing improved enough to need less support. A procedure code cannot answer that. It can tell you whether a claim was generated.

The authors also name unmeasured confounding by socioeconomic status and by apnea severity — no apnea-hypopnea index was available [s1]. Baseline severity is a strong predictor of who stays on CPAP, and its absence from the model is a real gap.

What the tirzepatide comparison does and does not show

That tirzepatide outperformed GLP-1 receptor agonists on these outcomes [s1] is consistent with what randomised evidence shows about relative weight loss, and with tirzepatide being the agent studied in randomised trials for sleep apnea specifically. But this is a within-observational-study comparison of non-randomised groups, and people prescribed tirzepatide in 2021–2025 differ systematically from people prescribed older agents — in insurance, in prescriber, in recency of diagnosis. The comparison is suggestive, not decisive.

What would settle it

A randomised trial with CPAP discontinuation, or objectively recorded CPAP adherence hours, as a prespecified outcome. Nothing in an EHR-based emulation substitutes for that, however large the sample.

Until then, the honest summary is: in a large matched observational cohort, starting an incretin agonist was associated with a modestly lower rate of recorded CPAP procedures over roughly eleven months, in a dataset that also produced a mortality estimate large enough to suggest the comparison is not fully clean.

What to watch

Whether device-manufacturer adherence data — which record actual hours of use — get linked to prescribing data anywhere. Whether the randomised sleep apnea programme for these drugs reports on CPAP discontinuation rather than apnea-hypopnea index alone. And whether replication in other EHR networks reproduces both the CPAP estimate and the implausible mortality estimate, which would confirm the pattern is a property of the design rather than of this database.

This article describes observational research on prescription drugs. It is not medical advice, and no one should change CPAP or medication use on the basis of it.

Sources

1. Incretin Receptor Agonists and CPAP Use in Adults With Diabetes, Obesity, and Obstructive Sleep Apnea — JAMA Network Open, 22 December 2025