Andersen-Ranberg NC, Poulsen LM, Perner A, et al; AID-ICU Trial Group. Haloperidol for the Treatment of Delirium in ICU Patients.
Haloperidol is the most commonly used ICU delirium drug in the world. Until AID-ICU, the evidence for it was thin.
Delirium affects an estimated 30–50% of ICU patients and is associated with longer ICU stay, longer mechanical ventilation, and worse long-term cognitive outcomes. Haloperidol has been the default pharmacologic treatment for decades largely on the strength of clinical tradition rather than trial data. Two prior trials had already cast doubt: HOPE-ICU found early prophylactic haloperidol did not prevent delirium from developing, and MIND-USA (2018, n=566) found no difference in days alive without delirium or coma with haloperidol or ziprasidone versus placebo — and haloperidol nearly reached statistical significance for worse 90-day mortality (38% vs. 34%, OR 1.17, 95% CI 0.99–1.40). AID-ICU was designed as a larger, dedicated test of haloperidol specifically for treating delirium once it's already present, using CAM-ICU or ICDSC to confirm the diagnosis first — the same algorithm shown in the companion post.
| Design | Multicenter, blinded, placebo-controlled RCT |
| Population | 1,000 adults admitted to ICU for an acute condition, screened positive for delirium by CAM-ICU or ICDSC |
| Intervention | IV haloperidol 2.5 mg TID + up to 2.5 mg PRN, max 20 mg/day, for as long as delirium persisted or recurred |
| Control | Matched placebo |
| Primary outcome | Days alive and out of hospital at 90 days |
| Analyzed | 987 of 1,000 randomized (501 haloperidol, 486 placebo) |
Treatment with haloperidol did not lead to a significantly greater number of days alive and out of hospital at 90 days than placebo. On the standard frequentist analysis that made the NEJM headline, this was a negative trial.
A pre-planned secondary analysis, published a year later, tells a meaningfully different story from the same dataset.
The AID-ICU investigators had pre-specified a Bayesian reanalysis of their own trial, published in Intensive Care Medicine in 2023. Using weakly informative priors, adjusted Bayesian regression on the same 987 patients found:
| Outcome | Result |
|---|---|
| Days alive and out of hospital (primary outcome) | Mean difference +2.9 days (95% CrI −1.1 to 6.9); 92% probability of any benefit, 82% probability of clinically important benefit |
| Mortality (risk difference) | −6.8 percentage points (95% CrI −12.8 to −0.8); 99% probability of any benefit, 94% probability of clinically important benefit |
| Serious adverse reactions | +0.3 percentage points (95% CrI −1.3 to 1.9); 98% probability of no clinically important difference |
The frequentist analysis says "no significant difference." The pre-planned Bayesian analysis of the identical dataset says "high probability of clinically important benefit, including on mortality." Both are valid statistical approaches to the same 987 patients — the discrepancy comes from how each framework treats uncertainty and prior information, not from a data error. This is exactly the kind of nuance a single-slide "how effective is haloperidol" caption can't hold, and it's the single most important thing missing from the original post.
| Trial | N | Question | Result |
|---|---|---|---|
| HOPE-ICU | — | Early prophylactic haloperidol | Did not prevent delirium |
| MIND-USA, 2018 | 566 | Haloperidol or ziprasidone vs. placebo, treatment | No benefit on delirium/coma-free days; near-significant mortality signal against haloperidol (OR 1.17, 95% CI 0.99–1.40) |
| AID-ICU, 2022 (frequentist) | 1,000 | Haloperidol vs. placebo, treatment | No significant difference, days alive out of hospital |
| AID-ICU, 2023 (Bayesian reanalysis) | 987 | Same trial, same data | High probability of benefit, including on mortality |
Read across all four, the picture is genuinely unsettled rather than clearly negative: the largest, best-designed trial to date produced a null frequentist result but a favorable Bayesian one, while the prior trial (MIND-USA) leaned toward possible harm. That's not a clean story, and any single-post summary claiming certainty in either direction is oversimplifying.
The RebelEM post correctly identifies AID-ICU as the trial to know here and correctly links the CAM-ICU algorithm as the diagnostic step that has to happen first. What it leaves as an exercise for the reader is the actual finding — and the honest answer is two-part: the primary, published frequentist result was negative, but the trial's own pre-planned Bayesian reanalysis of the same patients found a high probability of meaningful benefit, including on mortality. Combined with MIND-USA's near-significant harm signal beforehand, haloperidol for ICU delirium remains a genuinely open question rather than a settled negative or positive — treat any confident claim in either direction with the same skepticism this audit is applying to the source post.