A study of more than 132,000 veterans on GLP-1s found an 18 percent lower risk of heart attack, stroke, or death compared with "an older diabetes treatment."

Both numbers are exact. The 132,551 is the GLP-1 arm; the full cohort was 333,687. The WashU release translates the 18 percent into roughly four fewer events per 100 people over three years, which is the more useful number and the one the article skips.

The older treatment is a sulfonylurea. Sulfonylureas cause hypoglycemia and weight gain and have carried a cardiovascular question mark since the 1970s. An 18 percent advantage over sulfonylureas is real, and it is also a comparison against a drug that may itself be pushing risk up. Against metformin or an SGLT2 inhibitor the gap would probably be smaller.

Risk rose every six months off the drug, back toward original levels. Al-Aly's line: three years to build an 18 percent reduction, 18 months of stopping to undo it.

Close. People who used the drug for 18 months or less and then stopped ended the three years with no measurable benefit. Two years off the drug carried an IRR of 1.22 against staying on, and 0.82 times 1.22 lands almost exactly at the sulfonylurea baseline. The shape of the curve is the paper's main finding and the article reports it fairly.

The quote compresses two different analyses into one symmetric slogan. "Eighteen months to undo" is Al-Aly's gloss, not a reported estimate. The data say roughly two years off returns you to baseline. Both are memorable. Only one is in a table.

The protective effects "may go away when the drug itself does," and "metabolic whiplash" is the proposed mechanism. Gerstein's objection (people who stop are different from people who don't) gets one paragraph.

The outcome includes death from any cause. People stop medications when they enter hospice, when they are admitted for something else, when they stop eating, when nausea becomes intolerable during an unrelated illness, and when they lose the job that carried the insurance. Each of those raises the risk of death independently of any vascular biology. Target trial emulation adjusts for what is in the record. It cannot adjust for the decline that prompted the stop and was not yet coded. This is reverse causation, and it would produce exactly the dose response the paper reports: longer time off, higher risk.

The direction is plausible and fits the trial data on weight and glucose rebound. The size of the effect is an upper bound. The randomized test of this question (assign people to stop) has not been done and probably will not be, which means this study is close to the best evidence we will get.

The range is lifted from the introduction of the 2026 JMCP survey paper, which cites prior studies reporting one year discontinuation rates of 36 to 65 percent, higher in people without diabetes.

"At some point" is a different claim from "within a year," and the lifetime number is surely higher. Meanwhile persistence is improving: a Prime Therapeutics analysis of 33,607 non-diabetic starters found one year persistence rose from 33 percent in 2021 to 61 percent in early 2024. The range is real, partly dated, and mislabeled.

In a 2026 survey, more than half stopped because of cost or lack of insurance, a third because of side effects, about a quarter after reaching their goal.

Ipsos national panel, fielded June 2025, published in JMCP in 2026. 440 discontinuers. Cost 36.1 percent, side effects 33.3, no insurance coverage 28.2, reached weight goal 26.1. Respondents could pick more than one. More than half quit within six months, and four in five within a year. Among people with diabetes, side effects drove half of the stops (50.1 percent against 28.1).

"More than half for cost or insurance" requires the union of two overlapping categories, which the abstract does not report. It may well be true. It cannot be checked from what was published. The diabetes finding is the more important one for this article, because the cardiovascular study was entirely in people with diabetes, and in that group the leading reason to stop is that the drug made them feel sick.

The prespecified SELECT analysis in The Lancet (Deanfield and colleagues, 2025): semaglutide cut MACE across every baseline weight category, early weight loss did not predict who benefited, and only about a third of the benefit was mediated through waist circumference. In the placebo arm, weight loss was paradoxically associated with higher risk.

SELECT enrolled people without diabetes. The VA cohort was entirely diabetic. The mechanism probably carries across, but the article treats the two populations as one.

SEMA-VR CardioLink-15, European Heart Journal: 46 people with diabetes or obesity plus cardiovascular risk, randomized to semaglutide or usual care for six months. Circulating vascular regenerative cells rose about 35 percent with semaglutide versus about 1 percent with usual care, and pro-inflammatory granulocyte precursors fell by half. Funded by the Canadian Institutes of Health Research.

David Hess, who supplies the firefighter metaphor, is co-senior author of this trial. That is fine and it should have been said. The endpoint is a cell count by flow cytometry, a surrogate, in 46 people, unblinded to treatment. It is a mechanism worth testing, not a mechanism shown to cause fewer heart attacks. And no one has measured what happens to these cells when the drug stops, which is the question the article is asking.

Scolaro and colleagues (Journal of Clinical Investigation, 2025): in obese hypercholesterolemic mice, caloric restriction promoted plaque resolution, and weight regain accelerated progression and reprogrammed immune progenitors toward a hyperinflammatory state. A 2026 JCI Insight research letter found that cycling GLP-1 agonist treatment on and off in mice produced drug resistance and more fat.

A 2024 FASEB study in a different mouse model found repeated weight cycling did not affect atherosclerosis and reduced liver inflammation compared with continuous high fat feeding. The mouse literature agrees that cycling is worse than staying thin. It disagrees about whether cycling is worse than staying fat. Neither weight regain model used a GLP-1 drug to cause the loss.

The two thirds figure is the STEP 1 extension (Wilding 2022): after 68 weeks of semaglutide and a 17.3 percent loss, participants regained 11.6 points within a year of stopping, and most cardiometabolic improvements drifted back toward baseline. The plateau comes from a 2026 eClinicalMedicine meta-regression of six randomized trials (3,236 people): 60 percent regained at one year, with regain projected to level off at 75.3 percent and a half-life of 23 weeks.

The plateau is extrapolated. No trial has followed people far enough past one year to observe it. Real world data are kinder: a 2026 Cleveland Clinic cohort of people who stopped early regained very little at one year, though many had switched drugs or restarted. The quarter you keep is a model output, and a reasonable one.

The best match is a Cureus case series (Kubota 2023) of nine Japanese patients who finished the SURPASS J-mono trial. A1c fell by more than two points on tirzepatide, then rose at two, four, and six months after stopping in every group. Two years out, the 5 mg group was still 1.6 points below baseline. The 15 mg group was back to exactly where it started.

Nine patients, split across four arms, with whatever diabetes treatment their physicians chose afterward. "Not all the way back" is true for the lower doses and false for the highest one, and "small study" undersells how small. The larger signal comes from the meta-analytic literature, which finds a clear HbA1c rebound after GLP-1 discontinuation.