Background
Prehospital airway management seems to still be controversial for many. I’m obviously quite happy to support it, and I genuinely believe that if a patient needs an RSI for trauma, then it’s intuitively better to do it as early as possible, but not everyone agrees. I remember a recent case where I was questioned as to why we had done a prehospital RSI when only 10 mins away from a hospital….., I think I can answer that one with some decent evidence, but the point that it’s still asked is important. It’s still the case that one of the most controversial questions is whether prehospital emergency anaesthesia (PHEA) and tracheal intubation genuinely improve outcomes when delivered before hospital arrival. Airway control is obviously fundamental to trauma care, but whether that intervention is best delivered on scene, in transit, or in the emergency department remains uncertain. That said, any prehospital clinician reading this will recognise patients where the the decision can be difficult. Often because it’s not just physiology, access, entrapment, weather, transport time, team composition, competing priorities and cognitive load all make some of these situations incredibly complex.
The evidence base has lagged behind practice and policy. Observational studies are conflicting and the only major randomised controlled trial, focused on traumatic brain injury, suggested improved neurological outcomes without a mortality benefit. That matters because UK PHEA is generally delivered by physician-led critical care teams, usually through HEMS services. Access to those teams is variable, resource intensive, and dependent on system design. Questions around dispatch, training, governance and cost-effectiveness therefore become just as important as the airway intervention itself.
The challenge is methodology. Randomised trials are incredibly hard to deliver in critically injured patients. Consent, logistics and equipoise all create major barriers. As a result, we are usually left interpreting observational data that are highly vulnerable to confounding by indication, the sickest patients are both more likely to need intubation and more likely to die.
This paper attempts to tackle that problem using causal inference techniques and machine learning applied to a large UK trauma dataset. Whether it succeeds completely is debatable, but it is certainly an interesting approach. The abstract is below, but as always….., do go and read the full paper yourself.
Abstract
Background
Trauma is among the top 10 most common causes of disability worldwide, accounting for more than 100 million disability-adjusted life-years (DALYs) per year. Although the value of prompt management in severe injury is undisputed, the benefit of prehospital emergency anaesthesia with intubation remains uncertain. We aimed to estimate the causal effect of prehospital intubation on survival using doubly robust joint modelling of outcome and allocation, a task facilitated by machine learning.
Methods
For this causal modelling study we analysed the prehospital clinical characteristics of 6467 patients admitted to a UK major trauma centre for the periods Feb 23, 2012, to March 31, 2017 (n=3882), and April 1, 2017, to Nov 13, 2019 (n=2585). We built machine-learning models to predict survival and stratify patients by propensity for early anaesthesia and intubation. To estimate the causal survival effect of prehospital intubation, we used doubly robust estimation with inverse probability weighting by predicted intubation.
Findings
Early intubation and 30-day mortality were highly predictable from early prehospital data alone (AUCs 0·943 and 0·867, respectively). Patients who were predicted to require prehospital intubation but did not receive it showed substantially lower survival than those who were not (66·8% [95% CI 61·3–71·7] vs 93·6% [92·5–94·6], log-rank p<0·0001). The conditional average treatment effect of prehospital intubation was –0·103 (95% CI –0·119 to –0·087), corresponding to a 10·3% reduction in 30-day mortality. Scaled UK-wide, this policy is estimated to prevent 170 (95% CI 148 to 191) deaths per year, an effect comparable to the benefit of major trauma centres in England, with an annual cost-effectiveness value of £101 million (95% CI 93 to 111).
Interpretation
Guiding prehospital intubation by a machine-learning stratification model of prehospital data is predicted to improve 30-day survival of major trauma patients. This is, to our knowledge, the highest level of evidence to date on prehospital intubation efficacy in major trauma and could inform policy discussions on funding specialist prehospital critical care teams to consider improving access to this intervention.
What kind of study is this?
This is a retrospective observational study using modern causal modelling techniques. Not something I’ve come across before, but interesting approach
The authors analysed data from a UK major trauma centre across two time periods. One dataset was used to train machine learning models and the second acted as a prospective validation cohort.
The analysis is technically complex, but the underlying clinical question is straightforward:
If we had intubated the patients who appeared to need intubation, would they have done better?
To answer that, the authors developed predictive models for both early intubation and 30-day mortality. They then applied a doubly robust causal inference framework to estimate treatment effects.
That phrase sounds intimidating, but the principle is actually fairly straightforward (I think). The model attempts to account for both:
- the likelihood of a patient receiving intubation, and
- the likelihood of death
If either of those models is correct, the causal estimates should theoretically remain valid. The model they created is the Intub-8 model. https://intub-8.streamlit.app/
This is not the same as an RCT, and it cannot remove all bias, but it is an attempt to approximate causal inference from observational data.
Tell me about the patients
The study included 6467 patients admitted to a single UK major trauma centre. The cohort looks reasonably representative of contemporary UK trauma populations:
- mean age in the early 60s,
- predominantly blunt trauma,
- around 10% mortality,
- and prehospital intubation rates of roughly 10%.
The inclusion criteria were broad and the authors excluded relatively few patients beyond interhospital transfers and cases with missing prehospital data. Both adult and paediatric patients were included.
However, and this is a problem with a lot of trauma research, patients who died before reaching hospital were excluded. That matters in a study examining airway interventions because those are potentially the very patients most likely to benefit, or indeed be harmed, by PHEA/intubation attempts.
What were the measured outcomes?
The primary outcome was: 30-day mortality.
However, the study also explored a range of secondary analyses including:
- prediction of early intubation,
- subgroup survival analyses,
- conditional average treatment effects,
- and health economic modelling.
The authors are not simply identifying who is sick. They are trying to estimate which patients may benefit from intervention.
What are the main results?
The predictive models performed impressively well. The model predicting early intubation achieved an AUROC of 0.943, while the 30-day mortality model achieved an AUROC of 0.867, suggesting good discrimination between high- and low-risk patients.
The clinically interesting findings emerged when the authors stratified patients according to predicted need for early intubation.
Patients identified as high risk had substantially lower survival overall:
- 66.8% versus 93.6%.
More importantly, among patients predicted to require intubation, those who did not receive prehospital intubation appeared to fare worse:
- 71.4% survival compared with 94.4% in lower-risk non-intubated patients.
Using their causal inference model, the authors estimated the conditional average treatment effect (CATE) of prehospital intubation.
In the highest-risk subgroup (n=229), prehospital intubation was associated with:
- an absolute mortality reduction of 10.3%,
- approximately 28 additional survivors,
- and a relative survival improvement of around 12%.
By contrast, the low-risk cohort showed little apparent benefit, with an estimated mortality reduction of only 0.7%.
That distinction is important. The signal here is not that all trauma patients benefit from PHEA, but rather that benefit may be concentrated in a relatively small, high-risk subgroup.
The subgroup analyses suggested that patients most likely to benefit were those with:
- low GCS,
- physiological instability,
- and markers of severe injury.
The survival curves also separated early, raising the possibility that timing of intervention may matter as much as the intervention itself.
The authors extrapolated these findings nationally, estimating that targeted prehospital intubation could potentially save around 170 additional lives per year in the UK. Health economic modelling suggested this would remain cost effective, with gains of approximately 1540 QALYs annually and an estimated societal value of around £101 million per year.
Those extrapolations should be interpreted cautiously, but they help frame the potential scale of effect if the findings are genuine.
What do we think about the methodology?
This is an interesting study with a new approach in this area. The authors are tackling a genuinely difficult clinical question using methods that are considerably more advanced than we usually see in trauma observational research. The use of a separate validation cohort strengthens the credibility of the predictive modelling, and the causal inference framework seems appropriate for the problem being addressed.
The findings are also biologically plausible. It makes intuitive sense that patients with severe physiological compromise may benefit from earlier airway control delivered by experienced teams. However, none of that removes the fundamental limitations of observational research.
There are many potential confounders here that are difficult to model:
- operator skill,
- scene complexity,
- dispatch accuracy,
- team dynamics,
- and real-time clinical judgement.
Those factors do, in practice, influence both the decision to intubate and the patient outcome. There is also a degree of circularity here. The model predicts “need for intubation” based on historical clinician behaviour and then uses that prediction to infer whether different decisions should have been made.
Generalisability is another issue. This is a single-centre UK dataset derived from a physician-led system. Whether the same findings apply in paramedic-led systems, shorter transport environments or non-UK healthcare structures is uncertain.
Lastly, there is little here about neurological outcome beyond death. For head injured patients we really want to know about longer term morbidity/disability as well, perhaps even more so and that data is just not here.
Should this change practice?
Probably not though it is very interesting, but it is not definitive. It should not trigger indiscriminate expansion of prehospital intubation, nor should it be interpreted as proof that PHEA universally improves trauma survival. I’ve looked at the Intub-8 tool online and it clearly uses the same sort of information that I do when making real clinical decisions. That said, I think it’s a bit more conservative than me. I put in a GCS 6 patient who was tachy and a bit hypoxic, and that did not meet threshold for more than 50% likely to get a PHEA….., so maybe it does know more than me!
That said the data does strengthen the argument that:
- patient selection matters,
- advanced prehospital critical care may provide meaningful benefit in specific groups,
- and system design is crucial.
It also reinforces something most experienced prehospital clinicians probably already believe intuitively: missed opportunities for timely intervention may matter just as much as inappropriate intervention. The idea of machine-learning supported decision tools is interesting, particularly if simplified models such as Intub-8 prove externally valid. But these tools should support clinical judgement, not replace it.
For now, this paper is probably best viewed as hypothesis-strengthening rather than practice-changing.
Summary
Using machine learning and causal inference methods, the authors suggest that carefully selected high-risk trauma patients may derive significant survival benefit from prehospital intubation, while lower-risk patients probably do not.
The study is relevant, but despite some really interesting methodological approaches, it remains observational and therefore vulnerable to residual confounding.
vb
S
References and further reading
- Nelson APK, et al. Survival effect of prehospital emergency anaesthesia with intubation in risk-stratified patients with major trauma: a causal modelling study. Lancet Respir Med. 2026.
- Bernard SA, Nguyen V, Cameron P, et al. Prehospital rapid sequence intubation improves functional outcome for patients with severe traumatic brain injury: a randomized controlled trial. Ann Surg. 2010;252(6):959–965.
- Heritage D, et al. Helicopter emergency medical services demonstrate reduced time to emergency anaesthesia compared with emergency department RSI in major trauma. Scand J Trauma Resusc Emerg Med. 2024.
- Price J, Lachowycz K, Steel A, et al. Intubation success in prehospital emergency anaesthesia: a retrospective observational analysis of the interchangeable operator model. Scand J Trauma Resusc Emerg Med. 2022;30:44.
- Hodkinson M, et al. Pre-hospital emergency anaesthesia in 2025. BJA Educ. 2025.
- Lossius HM, Søreide E, Hotvedt R, et al. Prehospital advanced airway management in Oslo: an audit of outcome and practice. Acta Anaesthesiol Scand. 2002;46(7):835–842.
