- Design
- Causal-forest reanalysis of a randomised trial with external validation
- Population
- 704 immunocompromised adults with acute hypoxaemic respiratory failure (HIGH trial)
- Primary outcome
- Invasive mechanical ventilation by day 28
- Effect
- No average effect; predicted benefit ranged from −20.2% to +7.4% absolute risk
This reanalysis applied a causal-forest machine-learning model to 704 patients from the HIGH trial, which randomised immunocompromised adults with acute hypoxaemic respiratory failure to high-flow nasal oxygen or standard oxygen. It asked whether some patients benefited even though the trial found no overall effect, and tested the model in an observational cohort of 818.
Predicted effects on intubation by day 28 ranged widely, from about a 20% absolute reduction with high-flow to a 7% reduction with conventional oxygen. Those predicted to benefit most were older, more often solid organ transplant recipients, less often neutropenic, more hypoxaemic and tachycardic, with less non-respiratory organ failure. The ranking held in the external cohort, though the size of effect was less well calibrated.
This is a hypothesis, not a selection rule. It suggests that high-flow may suit some immunocompromised patients and not others, and that a blanket judgement either way may be wrong.
- Continue to use high-flow nasal oxygen as a reasonable first-line option in hypoxaemic respiratory failure.
- Watch closely for failure in neutropenic patients with multi-organ dysfunction.
- Do not delay intubation in a patient deteriorating on high-flow.
- Treat the predicted-benefit profile as hypothesis-generating.
Why it matters
A neutral trial may hide patients who benefit and patients who are harmed.
Don't overread it
A post hoc machine-learning analysis; it needs a prospective trial before it guides care.
The statistics, in plain English
A causal forest estimates how much each individual might benefit based on their characteristics. Such models can find patterns that are chance, which is why the authors tested it in a second cohort; ranking held up but the absolute numbers did not transfer well.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for emergency & critical care, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free