Death is not a successful wean

I have a small, slightly irritating observation about hospital scorecards, and it would not leave me alone until we wrote it down.

Intensive care units get compared on how long their patients stay on a ventilator. Shorter looks better. Faster wean, fewer complications, lower cost. The trouble is that death also ends the clock. If someone dies on day two, the usual metric treats that as a short ventilation, which is to say a good one. That is a bit like timing a race and handing out medals when the runner collapses at the first bend.

A competing risk is just a grown-up name for the other thing that can happen first. You can get off the ventilator alive. You can die. Those two cannot both be true. Ignore the second one, and the units that lose people early start to look efficient.

So we rebuilt the question. Not how many hours of ventilation, but this: by day 7, were you off the machine and still alive? Death gets no credit. We used routinely collected records from 43,880 adults starting invasive ventilation at 114 US hospitals in the eICU Collaborative Research Database. That is 2014 and 2015 practice, not a current league table of American ICUs.

By day 7, about three in four patients were liberated. About one in nine had died. The rest were still ventilated. After we accounted for how sick people were on arrival, hospital liberation rates still ran from roughly a quarter to more than four fifths. How ill the patients were explained almost none of that spread. How hospitals are built, and a handful of things they actually do (sedation, lung-protective volumes, tracheostomy), explained a lot more. Even then, most of the difference stayed unexplained, so this is not a ranking of who weans well. It is a warning about the ruler.

Then we put the old duration score next to ours, using the same statistical fence for both. The two lists sorted hospitals only moderately alike. In the tails, where quality programs actually act, they agreed on nobody. The duration metric praised 26 hospitals and flagged 11 as worse than expected. Ours almost reversed that: 3 praised, 26 flagged. Sixteen hospitals with significantly worse liberation never showed up on the old score at all. Those same hospitals had higher risk-standardized ICU mortality.

That is the whole argument. The conventional metric is not secretly handing trophies to deadly units. The hospitals it commends do have lower mortality. What it does is miss the ones that are failing to get people off the ventilator, and those units are the ones with more deaths. For a tool meant to point quality improvement somewhere useful, missing the problem is the more expensive error.

A few honest limits, because this is still observational data with an inferred ventilation-end time, from a decade ago, and day 7 is a choice. This construction is for comparing units, not for deciding when to extubate the person in front of you. Mortality, reintubation, and whether a unit actually does awakening and breathing trials still matter. I would rather see a dashboard than a single rank: how many people were off the ventilator alive by a stated day, with the mix of liberated, dead, and still ventilated sitting next to it so a low number can be read as delayed weaning or as early death.

The paper is in the Journal of Critical Care, written with Murishi Onesphore, Victor Okpanachi, and January G. Msemakweli. The code is on GitHub and Zenodo. If you work in ICU quality and you still score hospitals on raw duration, this is the bit I would like you to sit with: a short stay is not always a good stay.