How to Read ARISTOTLE: Apixaban Versus Warfarin in Atrial Fibrillation

ARISTOTLE randomized 18,201 people with atrial fibrillation to apixaban or warfarin, and apixaban cut stroke or systemic embolism by roughly a fifth, reduced major bleeding by about a third, and lowered deaths from any cause. The trial was built only to show apixaban was not worse than warfarin, so the interesting question is how it earned the right to say something stronger. The answer sits in a pre-planned order of tests that most readers skim past.

Key points#

• ARISTOTLE was designed as a non-inferiority trial, a study that asks whether a new drug can match an old one rather than beat it.
• A pre-specified, step-by-step testing plan let it move from "not worse" to "better" without inflating the odds of a false positive.
• Apixaban lowered stroke or systemic embolism, major bleeding, and death, with the death benefit the most fragile of the three.
• Much of the stroke reduction came from fewer bleeds in the brain, so part of apixaban's edge is really a safety story.
• The trial says nothing about mechanical heart valves or moderate-to-severe mitral stenosis, where warfarin stays standard.

The testing plan that turned "not worse" into "better"#

Start with the part that decides what a trial is allowed to claim, because everything else depends on it. Many newer anticoagulants were first tested only to prove they were not meaningfully worse than warfarin, a modest bar called non-inferiority. The logic is practical. Warfarin already prevents strokes, so it would be hard to justify withholding a working treatment just to see whether a more convenient drug could beat it. The realistic question is whether the easier drug can hold the line.

ARISTOTLE set non-inferiority as its main hypothesis but wrote a hierarchical testing plan into the protocol, and that plan behaves like a ladder with fixed rungs. First it asked whether apixaban was non-inferior to warfarin for the primary outcome, stroke or systemic embolism. Only if that rung held could the trial then test for outright superiority on the same outcome. Only if superiority held could it climb to major bleeding, and then to death from any cause. Testing in a locked order, and stopping the moment a rung fails, is how a single study can report several wins without the statistical sleight of hand that comes from running many comparisons and keeping the flattering ones.

This is the first thing to look for in any similar report. When a headline says a drug was "superior," ask whether that word sat at the top of a protected, pre-planned sequence or was pulled out of the data afterward. In ARISTOTLE it was protected, and apixaban held every rung.

What the trial was built on#

The design was unusually clean for a warfarin comparison. ARISTOTLE (Apixaban for Reduction in Stroke and Other Thromboembolic Events in Atrial Fibrillation) appeared in the New England Journal of Medicine in September 2011 and used a double-blind, double-dummy method. Each person took one active pill plus a dummy version of the other drug, so neither patients nor investigators could tell who was on apixaban and who was on dose-adjusted warfarin. Warfarin trials usually give the assignment away, because the INR blood test that guides warfarin dosing is itself a clue. Hiding it removes a bias that has muddied other anticoagulant studies.

The 18,201 participants all had atrial fibrillation plus at least one more stroke risk factor, such as a previous stroke, older age, diabetes, high blood pressure, or heart failure. This is a moderate-to-high-risk group, not a random sample of everyone with an irregular heartbeat. Apixaban was given at 5 mg twice daily, dropped to 2.5 mg twice daily for people who met certain markers of higher bleeding risk, such as older age, low body weight, or reduced kidney function. Warfarin was adjusted toward an INR of 2.0 to 3.0. Follow-up ran about 1.8 years on average.

Reading the results one rung at a time#

Because the rungs were tested in order, the cleanest way to read the numbers is to climb them in the same order.

For the primary outcome, stroke or systemic embolism, the yearly rate was 1.27 percent with apixaban versus 1.60 percent with warfarin. The hazard ratio was 0.79, with a 95 percent confidence interval of 0.66 to 0.95. That is about a 21 percent relative reduction, and because the whole confidence interval sits below 1.0, the benefit is unlikely to be a fluke.

Major bleeding, scored by International Society on Thrombosis and Haemostasis criteria, ran at 2.13 percent per year with apixaban versus 3.09 percent with warfarin, a hazard ratio of 0.69 (0.60 to 0.80). Bleeding inside the skull, the most dangerous kind, was markedly lower with apixaban.

Death from any cause was 3.52 percent with apixaban versus 3.94 percent with warfarin, a hazard ratio of 0.89 (0.80 to 0.998). Notice how close the upper bound is to 1.0. That result crossed the line into statistical significance, but only just, so it deserves to be read as encouraging rather than settled.

The stroke benefit hides a bleeding story#

One detail is easy to overlook and worth pausing on. Most of the stroke reduction came from fewer hemorrhagic strokes, which are bleeds in the brain, while the gap in ischemic (clot-driven) strokes was smaller and less certain. A drug that both blocks clots and bleeds less is doing part of its stroke-prevention job simply by causing fewer brain bleeds. That is a real benefit for patients, but it reframes apixaban's advantage as much about safety as about extra clot-busting power. Keeping the two apart matters when you counsel a specific patient, because their personal balance of clotting versus bleeding risk decides how much of that advantage applies to them.

What the trial cannot tell you#

Good appraisal ends with the edges of the evidence. The absolute differences here are modest: about a third of a percentage point per year on stroke and roughly one point per year on major bleeding. For you, the payoff scales with your baseline risk, and this population was chosen for elevated risk with reasonably preserved kidney function. ARISTOTLE also excluded people with mechanical heart valves or moderate-to-severe mitral stenosis, for whom warfarin remains the standard, so its results do not transfer to them.

Warfarin quality is another moving part. The average time in the target INR range was about 62 percent, and clinics that keep warfarin in range more consistently may see the gap shrink. And a median of 1.8 years says little about a medication many people take for decades. Cost, sticking to a twice-daily schedule, and the availability of reversal agents all sit outside the trial's figures and belong in your real-world decision.

None of this weakens the finding. It sharpens it. ARISTOTLE reads as one of the more trustworthy large anticoagulation trials precisely because of its blinding and its disciplined order of tests, and its direction of benefit has held up in later real-world analyses. The transferable lesson for reading any comparable study is to find the pre-planned hierarchy first, confirm the confidence intervals actually exclude no effect, and never let a borderline secondary result borrow the confidence of a solid primary one.