Reading a Trial That Missed Its Endpoint: The Lactate Question in Septic Shock

A clinical trial that misses its primary endpoint has not proven the treatment worthless; it has usually shown that the answer is less certain than a single threshold implies. ANDROMEDA-SHOCK, published in JAMA in 2019, compared two ways to guide resuscitation in septic shock, one steered by capillary refill time and one by blood lactate. The perfusion-guided group had lower 28-day mortality, 34.9 percent versus 43.4 percent, yet with a hazard ratio of 0.75, a 95 percent confidence interval of 0.55 to 1.02, and a P value of .06, the result fell just short of the conventional line for significance. Reading it well means looking past that one number to the confidence interval, the surrogate being chased, the size of the effect the study could detect, and what a larger trial later showed.

What the trial actually tested

ANDROMEDA-SHOCK enrolled 424 patients with early septic shock across 28 intensive care units in five countries: Argentina, Chile, Colombia, Ecuador, and Uruguay. Each patient was assigned to one of two eight-hour resuscitation protocols. One group was steered by capillary refill time, the low-cost bedside check of how quickly color returns to a compressed fingertip, with a target of three seconds or less. The other was steered by serum lactate, aiming to normalize it or cut it by at least 20 percent every two hours. The primary outcome was death at 28 days.

Look past the P value to the confidence interval

By day 28, 34.9 percent of the perfusion group had died versus 43.4 percent of the lactate group. The hazard ratio was 0.75, with a 95 percent confidence interval from 0.55 to 1.02 and a P value of .06. Because the interval crossed 1 and the P value sat above .05, the authors correctly reported that the strategy did not reduce 28-day mortality.

Here is where careful reading begins. A confidence interval is the range of effects reasonably compatible with the data. This one runs from a 45 percent relative reduction in the hazard of death down to a 2 percent increase. The large majority of that range favors the perfusion strategy, and the point estimate, the single most likely value, is a 25 percent reduction. Collapsing all of that into "not significant" discards most of what the trial found. The absolute numbers told the same story: about 8.5 fewer deaths per 100 patients, with an interval spanning roughly 18 fewer to 1 more.

Why lactate is a surrogate, not the target

Lactate earned its place in sepsis care because it correlates with severity and outcome, and clearing it is associated with survival. But correlating with an outcome is not the same as being the thing worth treating toward. Lactate is a downstream chemical signal with several sources. In shock it reflects tissue hypoperfusion, yet it also rises from adrenergic drive, impaired hepatic clearance, and other pathways that fluids will not correct. Chase the number after perfusion is already restored and you risk pushing more fluid and vasoactive drugs into a patient who does not need them. Capillary refill, by contrast, tracks perfusion more directly and tends to normalize faster.

That is the surrogate-endpoint lesson in a single trial. A marker that reliably tracks a disease is not automatically a good target to steer treatment by. The same caution applies far beyond sepsis, to any study where a blood value or an imaging finding stands in for the outcome patients actually feel.

A missed endpoint is often an underpowered one

ANDROMEDA-SHOCK was designed to detect a large effect: a fall in mortality from 45 percent to 30 percent, an absolute reduction of 15 percentage points. Effects that big are uncommon. The effect it actually observed, around 8.5 points, would be a major clinical win, yet the study did not enroll enough patients to call an effect of that realistic size statistically significant. Absence of evidence is not evidence of absence; failing to cross a threshold is not the same as demonstrating that the two strategies are equivalent. A great many trials labeled negative are simply underpowered for the effect that was plausible in the first place.

Two other clues pointed the same direction. The perfusion group had less organ dysfunction at 72 hours, with a mean SOFA score of 5.6 versus 6.6, a difference of about one point (P = .045). That is a secondary outcome and should be read as supportive and hypothesis-generating, not as proof on its own. And the direction of every major finding favored the same arm, which is worth more than any single P value.

A Bayesian lens, and a sequel

In 2020, a Bayesian reanalysis in the American Journal of Respiratory and Critical Care Medicine recomputed the result as the probability that the perfusion strategy was beneficial. Across a spread of prior assumptions, from optimistic to frankly skeptical, the probability that it lowered 28-day mortality exceeded 90 percent. That does not overturn the original conclusion, but it shows how much the significant-or-not dichotomy hides when a P value lands at .06.

The real test of a careful reading is what comes next. In 2025, JAMA published ANDROMEDA-SHOCK-2, a much larger trial of capillary-refill-guided resuscitation: 1,467 patients across 86 ICUs in 19 countries. On its hierarchical composite outcome it favored the perfusion strategy, with a win ratio of 1.16 (95 percent CI, 1.02 to 1.33; P = .04). A reader who had filed the first trial under "negative and forgotten" would have been surprised. A reader who had noted the effect size, the interval, and the consistent direction would not have been.

How to read a trial that misses

A few habits travel well from one study to the next. Find the point estimate and the whole confidence interval, then ask what fraction of that interval would change your decision. Separate surrogate from outcome, and ask whether the endpoint is the thing patients care about or only a proxy for it. Check the effect the trial was powered to find and compare it with the effect it actually saw. Treat .05 as a convention rather than a law of nature; .06 and .04 are neighbors, not opposites. Then weigh direction and consistency across endpoints, and wait for replication before either adopting or discarding.

None of this makes a single trial definitive, and none of it is medical advice. It is a way to keep a negative headline from erasing a signal worth following.