Helicopter Emergency Medical Services attendance is associated with favourable survival outcomes in major trauma: derivation and internal validation of prediction models in a regional trauma system Authors: Joanne Griggs, Jenny Harris, Jack Barrett, Scott Clarke, John O’Neill, Leigh Curtis, Malcolm Tunnicliff, Claire Baker, Duncan Bootland, Emma Ream, Richard Lyon, On behalf of Air Ambulance Charity Kent Surrey Sussex Published: 2026 ABSTRACT AbstractBackgroundSurvival benefit of Helicopter Emergency Medical Services (HEMS) attended major trauma remains inadequately quantified across injury severity. We evaluated HEMS performance and identified predictors of survival.MethodsRetrospective observational analysis of 3225 trauma patients attended by a regional HEMS in South-East England (2013–2022). Survival was assessed using W-statistic (Ws) methodology stratified by probability of survival (Ps) bands. Multivariable logistic regression identified predictors of 30-day mortality in major trauma (injury severity score (ISS) ≥15). Sub-analysis examined unexpected survival predictors and return of spontaneous circulation (ROSC) rates in traumatic cardiac arrest (TCA).ResultsAmong 2125 patients meeting Ws analysis criteria, observed (O) 30-day survival exceeded expected (E) survival (84.7% vs 81.3%; O/E ratio 1.04), yielding adjusted Ws of 5.23 (95% CI 3.27 to 7.19), representing 5.23 excess survivors per 100 patients. Survival benefit was greatest in severely injured patients with moderate survival probability (Ps 25–45%: 3.33 excess survivors per 100, 95% CI 1.37 to 5.29). Among patients with low probability of survival (Ps<50), 38.7% survived unexpectedly; younger ages and higher presenting Glasgow Coma Scale scores were key predictors of unexpected survival. Pre-hospital emergency anaesthesia (PHEA) was independently associated with unexpected survival in this group (adjusted OR 2.01, 95% CI 1.12 to 3.72, p=0.023). TCA ROSC rates demonstrated an annual improvement (6.3% increased odds per year, 95% CI 1.02 to 1.10, p=0.002).ConclusionHEMS attendance to major trauma in this regional service was associated with survival exceeding case-mix adjusted predictions, and was most pronounced in severely injured patients. PHEA was associated with survival benefit in low probability patients, supporting the value of advanced pre-hospital interventions.Keywords: Survival Analysis; Mortality; trauma; Models, Statistical; pre-hospital WHAT IS ALREADY KNOWN ON THIS TOPIC International evidence for Helicopter Emergency Medical Services (HEMS) effectiveness in major trauma remains inconsistent due to methodological heterogeneity, inadequate sample sizes and variable outcome definitions. Previous studies have shown conflicting results regarding survival benefits, with limited granular analysis across injury severity bands to identify which patient populations derive the greatest benefit from this resource-intensive service. International evidence for Helicopter Emergency Medical Services (HEMS) effectiveness in major trauma remains inconsistent due to methodological heterogeneity, inadequate sample sizes and variable outcome definitions. Previous studies have shown conflicting results regarding survival benefits, with limited granular analysis across injury severity bands to identify which patient populations derive the greatest benefit from this resource-intensive service. WHAT THIS STUDY ADDS This large regional cohort study demonstrates that in one UK HEMS, observed survival exceeded case-mix adjusted predictions by 5.23 per 100 major trauma patients, with the greatest effect observed in severely injured patients with moderate survival probability (25–45% Ps band). The study provides robust evidence using adjusted W-statistic methodology and identifies age, Glasgow Coma Scale score and pre-hospital interventions (such as PHEA) as key predictors of survival, while demonstrating significant annual improvement in TCA ROSC rates. This large regional cohort study demonstrates that in one UK HEMS, observed survival exceeded case-mix adjusted predictions by 5.23 per 100 major trauma patients, with the greatest effect observed in severely injured patients with moderate survival probability (25–45% Ps band). The study provides robust evidence using adjusted W-statistic methodology and identifies age, Glasgow Coma Scale score and pre-hospital interventions (such as PHEA) as key predictors of survival, while demonstrating significant annual improvement in TCA ROSC rates. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY These findings provide supportive evidence for continued investment in HEMS, particularly for severely injured patients, though comparative studies with alternative care pathways are needed to establish causal effectiveness. The validated prediction models may encourage real-time performance benchmarking and quality improvement initiatives, while the methodology provides a framework for trauma system evaluation that could inform international HEMS service development and clinical governance. We advocate for collaborative UK-wide HEMS data collection using standardised methodology to enable pooled analysis, to examine rare outcomes with adequate statistical power and establish national performance benchmarks. These findings provide supportive evidence for continued investment in HEMS, particularly for severely injured patients, though comparative studies with alternative care pathways are needed to establish causal effectiveness. The validated prediction models may encourage real-time performance benchmarking and quality improvement initiatives, while the methodology provides a framework for trauma system evaluation that could inform international HEMS service development and clinical governance. We advocate for collaborative UK-wide HEMS data collection using standardised methodology to enable pooled analysis, to examine rare outcomes with adequate statistical power and establish national performance benchmarks. Introduction Trauma is a leading cause of death and disability worldwide, with optimal outcomes dependent on advanced interventions and expedient transfer to specialised trauma care.1,3 Helicopter Emergency Medical Services (HEMS) represent a crucial component of many trauma systems, providing advanced clinical capabilities and expedited transport.2 4 5 Operational demand on HEMS has increased over recent decades, raising questions regarding which trauma patient populations derive the most significant benefit from this resource-intensive, finite service.6,8 Traditional trauma system performance has been evaluated by comparing the ratio of observed (O) outcomes against predicted (E) outcomes, which scoring systems refer to as expected outcomes (O/E).9 Performance is then measured using the W-statistic (Ws) which represents the number of excess survivors per 100 patients.10 However, stratification of performance across different probability of survival bands (Ps) provides a more granular understanding of which patients trauma systems most benefit.11 Of particular interest are patients who survive despite a low predicted probability of survival (either Ps<30% or Ps<50%), often termed ‘unexpected survivors’ (hereafter termed Ps<30 and Ps<50).12 Unexpected survivors may offer insights into modifiable factors that contribute to positive mortality outcomes.13 14 Mortality predictors in trauma populations are well-established,15 16 but conversely, few studies have identified patient and intervention factors associated with unexpected survival in severely injured patients (those with an Injury Severity Score (ISS) ≥15).1 15 16 Identifying mortality predictors in HEMS-attended cohorts validates performance measurement methodology and provides clinical context for interpreting outcomes.17 Further, these factors could enhance trauma outcome benchmarking and performance measurement frameworks, particularly as evidence for clinical effectiveness of advanced trauma interventions delivered by specialist teams remains inconsistent.18 19 We aimed to: (1) evaluate regional trauma system performance by comparing observed-to-expected survival across probability bands; (2) identify independent predictors of 30-day mortality in HEMS-attended trauma sub-grouped by ISS ≥15; and (3) determine factors associated with unexpected survival in patients with low predicted survival probability. Methods Study design and setting A retrospective observational cohort study of all trauma patients attended and conveyed by the Air Ambulance Charity Kent Surrey Sussex (KSS) between 1 January 2013 and 31 December 2022. Patients were included if KSS clinicians attended and conveyed patients with a documented traumatic mechanism of injury. For survival analysis and performance benchmarking, patients were eligible if they met Trauma Audit and Research Network (TARN) inclusion criteria: hospital length of stay (LOS) >72 hours, death in hospital from injury, critical care admission or inter-hospital transfer for specialist care.20 Patients pronounced life extinct (PLE) at the scene were excluded from primary survival analyses as they do not receive TARN Ps scores, but were included in a traumatic cardiac arrest (TCA) sub-analysis. Patients who died at the scene but underwent post-mortem examination were not included as TARN does not calculate Ps scores for pre-hospital deaths. Transparent Reporting of a multivariable Prediction Model for Individual Prognosis of Diagnosis (TRIPOD+AI 2024)21 and the Reporting of studies Conducted using Observational Routinely Collected Health Data (RECORD) Statement were adhered to.22 Major trauma patients attended by KSS are conveyed to one of three major trauma centres (MTCs) in the South-East region. These include Royal Sussex County Hospital, Brighton (Sussex Trauma Network), St George’s Hospital, London (South-West London and Surrey Trauma Network), or King’s College Hospital, London (South-East London, Kent and Medway Trauma Network), covering a region of 7500 km2 with a mixed rural–urban resident population of 4.5 million. Conveyance to an MTC may involve bypassing a trauma unit (TU). KSS operates 24/7 with two Leonardo Augusta Westland 169 (AW169) helicopters and two rapid response vehicles used interchangeably depending on operational need. Operating this model of care aims to deliver a doctor–paramedic team to a patient within 30 mins of an emergency 999/112 call. Advanced interventions and deployment Doctor–paramedic teams complement ground emergency medical service (EMS) provision, including a specialist critical care paramedic (CCP) workforce, by delivering advanced trauma interventions at the scene and during transport. Interventions include pre-hospital emergency anaesthesia (PHEA) and post-PHEA infusion, advanced ventilation strategies, blood component transfusion, advanced vascular access and invasive arterial blood pressure monitoring, and surgical procedures to include lateral canthotomy, thoracostomy and resuscitative thoracotomy (RT). Regional pathways comprise direct to CT, direct transfer for neurological intervention and massive transfusion protocols. Clinicians have 24/7 consultant and operational decision support. From the point of injury, a 999/112 call is made to the emergency operations centre and coded using NHS pathways.23 HEMS deployment is determined through joint decision-making between co-located non-clinical dispatchers and a CCP at the associated EMS. Each trauma call is interrogated, triaged and an appropriate asset tasked according to pre-defined criteria, ensuring clinical oversight of tasking decisions.24 Blood component administration follows a haemostatic resuscitation approach prioritising early and balanced blood product transfusion in patients with suspected major haemorrhage. The standard approach involves balanced transfusion ratios, with packed red blood cells (PRBCs) and plasma (either freeze-dried plasma (FDP) or fresh frozen plasma (FFP)) administered in a 1:1 ratio, with the initial component at the discretion of the attending clinicians. Patient eligibility and selection Major trauma patients were eligible for inclusion if attended by KSS between 1 January 2013 and 31 December 2022. Patients were included if attended by KSS clinicians, regardless of transport modality (helicopter or ground vehicle) and irrespective of whether advanced interventions were delivered. Study size was determined by the number of patients meeting the eligibility criteria during the study duration. Data sources and data acquisition Data were extracted from the Electronic Patient Clinical Record (EPCR) HEMSBase 3.0 (MedicOne Systems) using Zoho Analytics. Notable variables included: baseline demographic descriptors, including age and sex; scene descriptors: critical care paramedic, blunt/penetrating, mechanism of injury (MOI); physiological parameters: Glasgow Coma Scale (GCS) score, respiratory rate (RR), heart rate (HR), oxygen saturation (SpO2), end-tidal carbon dioxide (EtCO2), systolic blood pressure (SBP) and diastolic blood pressure (DBP); injury type: cardiac arrest (CA), injured body region (head, neck, thorax, abdomen/pelvis and limb); trauma interventions: PHEA, blood component transfusion including PRBC, FFP, FDP and anticoagulant reversal with Beriplex. Others included: endotracheal intubation (ETI) placement in CA, supraglottic airway, front-of-neck access, surgical airway, oropharyngeal airway, nasopharyngeal airway, positive pressure ventilation, thoracostomy, intercostal chest drain, thoracotomy, pelvic binder, haemostatic dressing, blast dressing, compression dressing, suture, maxillofacial intervention; patient disposition: conveyance status and injury severity; outcomes: 30-day survival, length of stay (LOS), ISS, and Ps (defined in Definitions and calculation of survival scores). A post-calculated, standardised case-mix adjusted outcome score was included for each trauma patient from a TARN participating centre.25 26 Presenting physiology refers to the first recorded vital signs on HEMS patient contact, before advanced interventions where possible. Final physiology refers to the last recorded measurements before hospital handover. ISS and Ps scores were calculated by TARN using in-hospital data including CT imaging, operative findings and diagnostic codes submitted by receiving trauma centres according to standard methodology.22 25 27 Patients attended by HEMS are subsequently matched to TARN records to obtain retrospectively calculated ISS and Ps values. This approach is standard for HEMS outcome studies as anatomical injury scoring requires complete diagnostic information which is only available after hospital assessment.1 Study data were secured anonymously within the Research and Data Capture (REDCap) Data Management System hosted at KSS. REDCap is a secure, web-based application designed to support data capture for research which holds the required safeguards for data security and privacy.27 Data validation and pre-processing Before data extraction, the primary EPCR database underwent quality assurance processes to ensure data validity and completion, where individual records are checked for errors and logic flaws. Anonymised data were subsequently downloaded into Microsoft Excel (Microsoft Corp, Redmond, WA, USA, version 16.79.1) for pre-processing and coding. Data extraction was performed by one investigator (SC), with a random sample of records independently verified by a second investigator (JG) to ensure accuracy. Discrepancies were resolved through consensus review of source documentation. Automated validation checks flagged logical inconsistencies for manual review. Definitions and calculation of survival scores Survival outcome was defined as alive/dead status at 30 days post-injury. Unexpected survivors were defined as those surviving to 30 days despite low predicted survival probability (Ps<30 and Ps<50). Probability of survival (Ps) represents the expected survival probability for a trauma patient based on their specific combination of demographic, physiological and injury characteristics compared with similar patients in the historical TARN database. Ps is calculated by TARN using the Ps17 model incorporating age, sex, GCS, physiological parameters at hospital arrival and anatomical injuries determined from CT imaging and clinical findings.20 27 The TARN prediction model has been validated in UK trauma populations demonstrating excellent discrimination (AUC 0.90, 95% CI 0.89 to 0.90) and represents the national standard for risk-adjusted trauma outcome benchmarking.20 A Ps score represents retrospective probability; a score of 70% implies that 70 out of every 100 patients with that profile survived historically. Expected outcomes (E) for performance evaluation are derived by summing individual Ps scores across all patients (E=ΣPs). For example, if 100 patients each have Ps=70%, the expected survivors are E=70. Observed outcomes (O) are the actual number of patients surviving to 30 days. Trauma system performance uses both crude (W) and case-mix adjusted (Ws) survival statistics.28,30 Both represent excess survivors per 100 patients compared with expected outcomes. The W score is calculated as: The Ws statistic applies a weighting system using the formula Ws=(∑(O−E)×Weight)/(∑Weight)×100, where Weight=Ps×(1−Ps). This gives greater emphasis to patients with moderate to severe survival probabilities (~50%) and less emphasis to those with very high or very low survival probabilities. A positive Ws indicates more survivors than expected and a negative score indicates fewer than expected survivors. These metrics were subsequently stratified by probability of survival (%) bands (95–100, 90–95, 80–90, 65–80, 45–65, 25–45 and 0–25) using right-inclusive boundary classification common to this database. For TCA secondary analyses, return of spontaneous circulation (ROSC) was defined as sustained return of circulation (palpable pulse, measurable blood pressure) maintained until hospital handover. ROSC represents a pre-hospital performance metric partly attributable to HEMS interventions. Primary and secondary outcomes Primary outcome The primary outcome was standardised mortality ratio using observed-to-expected (O/E) mortality rate (Ws) across predicted survival probability bands (Ps) in HEMS-attended major trauma. Secondary outcome(s) Thirty-day mortality and clinical outcomes were examined across clinically relevant subgroups: 30-day mortality in all HEMS-attended major trauma 30-day mortality in all HEMS-attended major trauma 30-day mortality in severe trauma (ISS ≥15) 30-day mortality in severe trauma (ISS ≥15) Unexpected survival rates in low predicted survival subgroups (Ps<30 and Ps<50) Unexpected survival rates in low predicted survival subgroups (Ps<30 and Ps<50) Hospital and intensive care unit (ICU) LOS in severe trauma (ISS ≥15) Hospital and intensive care unit (ICU) LOS in severe trauma (ISS ≥15) ROSC and 30-day mortality in TCA. ROSC and 30-day mortality in TCA. Statistical analysis Continuous variables are presented as mean±SD or median and IQR. Categorical variables are presented as frequencies and percentages (%). Baseline characteristics were compared using χ2 or Fisher’s exact test for categorical variables and independent t-tests or Mann-Whitney U test for continuous variables, as appropriate. Missing data were handled using multiple imputation (MI) by chained equations (MICE) with predictive mean matching generating five imputed datasets pooled from 50 iterations according to Rubin’s rules.31 32 All regression analyses used pooled MI results. Convergence diagnostics are shown in onlinesupplemental table S4 figure S5. Complete case sensitivity analysis showed consistent effect estimates with <10% difference for most predictors confirming robustness to imputation method (online supplemental table S6). Three complementary analyses were performed: (1) Observed/expected survival analysis: Crude and case-mix adjusted Ws statistics were calculated for each Ps band. Multivariable prediction models with logistic regression identified independent predictors of 30-day mortality. Variable selection for multivariable models was theory-driven, based on established trauma literature and clinical plausibility. Candidate variables were identified a priori and included: patient factors (age, sex), injury characteristics (mechanism, anatomical region, suspected traumatic brain injury (TBI, suspected traumatic haemorrhage), presenting physiology (GCS, SBP, heart rate, EtCO₂), and pre-hospital interventions (PHEA, ETI, thoracostomy, blood component transfusion, pelvic binder). All candidate variables with univariate associations (p<0.10) or strong theoretical rationale were entered into multivariable models. Variables were retained in final models based upon independent significance (p<0.05) or identification as important confounders based on clinical knowledge. Multicollinearity was assessed with correlated blood pressure variables removed to avoid redundancy. Clinically plausible two-way interactions were tested using likelihood ratio tests, with significant interactions retained in the final model. The events-per-variable (EPV) ratio exceeded recommended thresholds at 28:1. Model performance was assessed using McFadden’s pseudo-R² and Akaike Information Criterion (AIC).3 Additional variable selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation, selecting the optimal regularisation parameter (λ) using the one-standard-error rule to balance parsimony with predictive performance.33 Internal validation was undertaken with a 70/30 random data split. The prognostic model aimed to identify independent predictors of 30-day mortality rather than optimise clinical prediction or triage. Performance assessment prioritised overall discrimination (area under the curve (AUC)) and calibration across the full probability range. Calibration was evaluated using calibration plots and the Hosmer-Lemeshow (H-L) test, with discrimination via AUC. For prognostic research identifying risk factors, accurate probability estimates across all risk strata and appropriate risk ranking are essential, rather than optimising binary classification or sensitivity at a fixed threshold. Secondary analyses for LOS (hospital and ICU) were restricted to patients surviving to discharge to avoid survival bias. Due to right-skewed distributions, LOS data were log-transformed and analysed using a negative binomial regression. Results are presented as odds ratios (OR) with 95% confidence intervals (95% CI). Data were pre-processed in Microsoft Excel (Microsoft Corp, Redmond, WA, USA). Statistical significance was pre-determined at p<0.05. Statistical analysis and visualisations were performed using R version 4.3.3 (R Foundation for Statistical Computing, Vienna, Austria). Ethical considerations Data were routinely collected and met Health Research Authority (HRA, UK) criteria for service evaluation. Research ethics committee approval was not required, and the project was approved by the KSS Research and Innovation Committee under a TARN data sharing agreement. Major trauma systems provided letters of support for conducting this research. Study data were collected and managed using REDCap electronic data capture tools hosted securely with anonymised patient identifiers to ensure data protection and confidentiality. Results Baseline demographics and patient characteristics Of 12 361 patients screened, 3225 HEMS attended trauma patients with complete outcome data were included in the primary analysis, with 2125 eligible for Ws analysis. Additional subgroup analyses were conducted in patients with severe trauma (ISS ≥15, n=1984) and patients with low probability of survival (Ps<30, n=156; Ps<50, n=331) (figure 1). Median age was 43 years (IQR 25–61) with a predominance of males (74%). Blunt trauma accounted for 92% of cases, with road traffic collisions (58%) being the most common mechanism of injury. Median ISS was 24 (IQR 14–33), with 1984 patients (72.7%) classified as severe (ISS ≥15) and 605 (22.2%) as moderate (ISS 9–15). Patient demographics, injury characteristics, trauma interventions and clinical outcomes are stratified by survival or mortality at 30 days (table 1). Significant differences were observed across groups for age, mechanism of injury, ISS, presenting physiology, pre-hospital interventions (including PHEA, blood component transfusion and thoracostomy) and outcomes (p<0.001). Non-conveyed patients had significantly higher median Ps (99% vs 95.6%, p<0.001) validating on-scene transport decisions and our subsequent focus analysing conveyed patients (online supplemental table S1). Flow diagram for derivation of study population stratified by survival to 30 days. Derivation of study population. GA, ground assist; HEMS, Helicopter Emergency Medical Services; ISS, Injury Severity Score; MICE, Multiple Imputed Chained Equations; PLE, pronounced life extinct; Ps, predicted survival score; TARN, Trauma Audit and Research Network. Table 1Patient demographics, injury characteristics and trauma interventions stratified by 30-day survivalVariableOverallN=322530-day survivaln=269230-day mortalityn=533P valueBaseline demographics Age (median (IQR)) (N=3225)43 (25–61)40 (25–57)63 (39–78)<0.001 Sex, male (n, %) (N=3225)2434 (75.5)2075 (77)359 (67.3)<0.001 Sex, female (n, %) (N=3225)790 (24.5)616 (22.9)174 (32.6)Injury characteristics ISS (median (IQR)) (N=2729)24 (14–33)20 (13–29)30 (25–41)<0.001 Missing (n)496389107 Penetrating mechanism (n, %) (N=3225)266 (8.2)260 (9.7)6 (1.1)<0.001Anatomical injury site Head injury (n, %) (N=3225)1839 (57.0)1401 (52.0)438 (82.2)<0.001 Neck injury (n, %) (N=3225)284 (8.8)215 (8.0)69 (12.9)<0.001 Thorax injury (n, %) (N=3225)1281 (39.7)1067 (39.6)214 (40.2)0.863 Abdomen/pelvis injury (n, %) (N=3225)1179 (36.6)1044 (38.8)135 (25.3)<0.001 Limb injury (n, %) (N=3225)1398 (43.3)1235 (45.9)163 (30.6)<0.001Presenting physiology RR (median (IQR)) (N=2989)17 (8–24)18 (10–24)15 (0–23)<0.001 Missing (n)23721720 SpO2 (median (IQR)) (N=3084)97 (90–100)98 (92–100)93 (81–99)<0.001 Missing (n)14112219 HR, bpm (median (IQR)) (N=3096)86 (67–107)86 (68–106)86 (60–113)0.567 Missing (n)12911415 SBP, mmHg (median (IQR)) (N=3059)128 (107–146)128 (109–145)125 (90–154)0.174 Missing (n)16613828 DBP, mmHg (median (IQR)) (N=3059)80 (66–93)81 (68–93)78 (58–95)0.002 Missing (n)16613828 MAP, mmHg (median (IQR)) (N=3059)97 (81–110)97 (83–110)95 (70–113)0.031 Missing (n)16613828 SI (median (IQR)) (N=2918)0.70 (0.50–0.90)0.70 (0.50–0.90)0.70 (0.50–1.10)0.058 Missing (n)30724760 Lactate (median (IQR)) (N=364)2.7 (1.9–4.0)2.7 (1.8–3.8)3.2 (2.1–5.1)0.017 Missing (n)28612384477 GCS (median (IQR)) (N=3169)14 (8–15)14 (11–15)4 (3-9)<0.001 Missing (n)56497Final physiology RR (median (IQR)) (N=2989)16 (12–21)17 (12–22)14 (10–18)<0.001 Missing (n)23621818 SpO2 (median (IQR)) (N=3084)98 (95–100)99 (96–100)97 (86–99)<0.001 Missing (n)14112219 HR, bpm (median (IQR)) (N=3096)86 (70–104)86 (70–103)88 (68–109)0.228 Missing (n)12911415 SBP, mmHg (median (IQR)) (N=3059)124 (108–141)124 (110–141)122 (99–145)0.010 Missing (n)16613828 DBP, mmHg (median (IQR)) (N=3059)80 (67–92)80 (68–92)78 (61–92)0.003 Missing (n)16613828 MAP, mmHg (median (IQR)) (N=3059)95 (82–108)96 (83–108)93 (75–110)0.006 Missing (n)16613828 Cardiac arrest and subsequent resuscitation (n, %) (N=3225)164 (5.1)53 (2.0)111 (20.8)<0.001Trauma Interventions ETI (n, %) (N=3225)1244 (38.6)796 (29.6)448 (84.1)<0.001 PHEA (n, %) (N=3225)1095 (34.0)744 (27.6)351 (65.9)<0.001 Thoracostomy (n, %) (N=3225)459 (14.2)259 (9.6)200 (37.5)<0.001 RT (n, %) (N=3225)11 (0.3)1 (0.0)10 (1.9)<0.001 PRBC (n, %)* (N=3225)155 (4.8)118 (4.4)37 (6.9)<0.001 Plasma (n, %)*† (N=3225)218 (6.8)158 (5.9)37 (6.9)<0.001Morbidity and mortality Hospital LOS (median (IQR)) (N=2602)13 (6–25)15 (8–29)3 (1–7)<0.001 Missing (n)623523100 ICU LOS (median (IQR)) (N=2648)2 (0–8)2 (0–8)2 (1–5)0.117 Missing (n)577477100 Ps (median (IQR)) (N=2232)95.6 (74.0–99.0)97.2 (86.9–99.2)42.1 (22.2–63.2)<0.001 Missing (n)993789204Data presented as median (IQR) for continuous variables and n (%) for categorical variables. Statistical tests: Kruskal-Wallis test for continuous variables, χ2 or Fisher’s exact test for categorical variables. n denotes the number of patients with available data for that specific variable.*Blood component presented as binary n (%) of those receiving any units.†Freeze-dried plasma or fresh frozen plasma.bpm, beats/min; DBP, diastolic blood pressure; ETI, endotracheal intubation; GCS, Glasgow Coma Scale; HR, heart rate; ICU, intensive care unit; LOS, length of stay; MAP, mean arterial pressure; PHEA, pre-hospital emergency anaesthesia; PRBC, packed red blood cells; Ps, predicted survival score; RR, respiratory rate; RT, resuscitative thoracotomy; SBP, systolic blood pressure; SI, shock index; SpO2, saturations. Patient demographics, injury characteristics and trauma interventions stratified by 30-day survival Data presented as median (IQR) for continuous variables and n (%) for categorical variables. Statistical tests: Kruskal-Wallis test for continuous variables, χ2 or Fisher’s exact test for categorical variables. n denotes the number of patients with available data for that specific variable. Blood component presented as binary n (%) of those receiving any units. Freeze-dried plasma or fresh frozen plasma. bpm, beats/min; DBP, diastolic blood pressure; ETI, endotracheal intubation; GCS, Glasgow Coma Scale; HR, heart rate; ICU, intensive care unit; LOS, length of stay; MAP, mean arterial pressure; PHEA, pre-hospital emergency anaesthesia; PRBC, packed red blood cells; Ps, predicted survival score; RR, respiratory rate; RT, resuscitative thoracotomy; SBP, systolic blood pressure; SI, shock index; SpO2, saturations. Observed and expected survival of HEMS attended patients Among 2125 patients with complete Ps and outcome data, overall 30-day survival rate was 84.7% compared with expected 81.3%, yielding an adjusted Ws of 5.23 (95% CI 3.27 to 7.19, p<0.001) and O/E ratio of 1.04 (figure 2). This overall Ws exceeds individual band values due to variance-weighting methodology (Ps×(1−Ps)), which gives greater influence on moderate survival probability bands (45–80%). Observed and expected trauma survival across predicted survival categorisation (n=2125). Ws Statistic; adjusted W-statistic accounting for case-mix, calculated as (Observed−Expected)/√Variance. Positive values indicate better than expected performance with a reference line (Ws=0) indicating expected performance based on case-mix adjusted predictions, and error bars showing 95% CI. Black circles represent individual probability of survival bands; red diamond represents overall performance. Sample sizes shown on left (n=patients per band). Overall survival: 84.7% observed versus 81.3% expected (O/E ratio=1.04). Greatest benefit in 25–45% band: 78 actual versus 58 expected survivors (Ws=3.33). Overall Ws (5.23) exceeds individual effects due to weighting that emphasises moderate Ps bands. Statistical significance denoted as *p<0.05, ***p<0.001. Analysis restricted to patients with complete Ps and 30-day outcome data. Ps, predicted survival score. System performance showed positive values across all Ps bands with incremental positive values maintained throughout the study period (annual change: −0.138, 95% CI −0.40 to –0.13, p=0.26). Survival advantage was greatest in patients with moderate to severe injuries (Ps 25–45%, Ws 3.33, 95% CI 1.37 to 5.29, p<0.001), representing a 35% relative increase (49.1% actual vs 36.4% expected). A calibration plot is available in online supplemental figure S2. Predictors of 30-day mortality Multivariable logistic regression identified key independent predictors of 30-day mortality (table 2). Key predictors included: age (adjusted OR (aOR) 1.02 per year, 95% CI 1.01 to 1.04, p<0.001), GCS (aOR 0.63 per point, 95% CI 0.57 to 0.69, p<0.001), and pre-hospital intubation (aOR 3.66, 95% CI 1.91 to 6.99, p<0.001). Cardiac arrest (aOR 2.49, 95% CI 1.53 to 4.08, p<0.001) and traumatic haemorrhage (aOR 1.64, 95% CI 1.12 to 2.40, p=0.010) were associated with increased mortality, while penetrating injury was independently associated with reduced mortality (aOR 0.33, 95% CI 0.11 to 0.87, p=0.037). Thoracic injury demonstrated increased mortality risk in the adjusted model (aOR 2.75, 95% CI 1.27 to 5.95, p=0.010). A significant age×GCS interaction improved model fit (aOR 1.00, 95% CI 1.00 to 1.01, p<0.001). PHEA was associated with reduced mortality (aOR 0.55, p=0.021). The final model achieved excellent discrimination (AUC 0.919, 95% CI 0.907 to 0.930) with adequate calibration (H-L p=0.151). Table 2Univariate and multivariable logistic regression analysis for predictors associated with 30-day mortality (n=3225)VariableUnivariate analysisMultivariable analysisOR (95% CI)P valueaOR (95% CI)P valueDemographics Age (years)1.03 (1.03 to 1.04)<0.0011.02 (1.01 to 1.04)<0.001 Sex (male)0.61 (0.50 to 0.75)<0.0010.88 (0.67 to 1.16)0.373Physiological variables GCS0.73 (0.72 to 0.75)<0.0010.63 (0.57 to 0.69)<0.001 Presenting SBP (mmHg)1.00 (1.00 to 1.00)0.1491.00 (1.00 to 1.00)0.702 Heart rate1.00 (1.00 to 1.00)0.2911.00 (1.00 to 1.00)0.292 EtCO₂ (kPa)0.95 (0.91 to 0.99)0.0120.95 (0.90 to 1.00)0.072Injury characteristics TBI4.25 (3.38 to 5.40)<0.0011.05 (0.75 to 1.48)0.777 Thoracic injury1.02 (0.84 to 1.23)0.8252.75 (1.27 to 5.95)0.010 Abdomen-pelvic injury0.54 (0.43 to 0.66)<0.0010.85 (0.61 to 1.19)0.351 Penetrating injury0.11 (0.04 to 0.22)<0.0010.33 (0.11 to 0.87)0.037 Cardiac arrest13.10 (9.34 to 18.58)<0.0012.49 (1.53 to 4.08)<0.001Intervention variables PHEA5.05 (4.15 to 6.16)<0.0010.55 (0.33 to 0.91)0.021 ETI12.55 (9.87 to 16.15)<0.0013.66 (1.91 to 6.99)<0.001 Thoracostomy5.64 (4.54 to 7.01)<0.0011.25 (0.81 to 1.94)0.310 Pelvic binder0.99 (0.82 to 1.19)0.9101.03 (0.78 to 1.36)0.847 Blood component transfusion3.40 (2.75 to 4.19)<0.0011.64 (1.12 to 2.40)0.010Interactions Age (years)×GCS––1.00 (1.00 to 1.01)<0.001Multivariable analysis of 30-day mortality risk factors.Results from multiple imputation (m=5).aOR, adjusted OR; EtCO2, end-tidal carbon dioxide; ETI, endotracheal intubation; GCS, Glasgow Coma Scale; PHEA, pre-hospital emergency anaesthesia; SBP, systolic blood pressure; TBI, traumatic brain injury. Univariate and multivariable logistic regression analysis for predictors associated with 30-day mortality (n=3225) Multivariable analysis of 30-day mortality risk factors. Results from multiple imputation (m=5). aOR, adjusted OR; EtCO2, end-tidal carbon dioxide; ETI, endotracheal intubation; GCS, Glasgow Coma Scale; PHEA, pre-hospital emergency anaesthesia; SBP, systolic blood pressure; TBI, traumatic brain injury. Missing data patterns and imputation procedures are detailed in onlinesupplemental figure S3 S5, online supplemental table S4. Sensitivity analysis comparing complete case analysis with MI showed consistent effect estimates, with ORs differing by <10% (online supplemental table S6). LASSO regression validated core prognostic factors, with a parsimonious 7-variable model maintaining excellent performance (AUC 0.911, ΔAUC=0.007 vs full model) (online supplemental table S7). Internal model validation Internal model validation (70:30 training: test data split) was performed with stratified sampling to maintain mortality rates, with the final model trained on 2258 (70%) and validated on 967 (30%) patients. The model demonstrated excellent discrimination with a test AUC of 0.922 (95% CI: 0.90-0.94) and no evidence of overfitting (AUC difference: -0.007). Test set performance showed 85.4% accuracy, 88.5% sensitivity, 97.1% specificity, negative predictive value (NPV). The H-L test indicated acceptable calibration (χ² =15.9, p = 0.043) (Supplementary Figure S8) Severely injured (ISS ≥15) subset analysis Among 1984 HEMS-conveyed patients (ISS ≥15) 30-day mortality was 20.0% (n=397). In multivariable analysis, independent predictors were age (aOR 1.02 per year, p<0.001), presenting GCS (aOR 0.61 per point decrease, p<0.001), blood component transfusion (aOR 1.60, p=0.021), cardiac arrest (aOR 1.85, p=0.020) and ETI (aOR 2.80, 95% CI 1.40 to 5.60, p=0.004 (table 3). Abdomen-pelvic injury was reported (aOR 0.79, p=0.193) (table 3). An age×GCS interaction significantly improved prediction (ΔAIC=−21.6, p<0.001) with age effect stronger in patients with higher GCS (9.6% increased risk per year at GCS 15 compared with 3.2% at GCS 3). The model achieved strong discrimination (McFadden R²=0.389). Table 3Univariate and multivariable logistic regression analysis for mortality risk factors in severe trauma (ISS ≥15), n=1984VariableUnivariate analysisMultivariable analysis*OR (95% CI)P valueaOR (95% CI)P valueDemographics Age (years)1.03 (1.03 to 1.04)<0.0011.02 (1.01 to 1.04)<0.001 Sex (male)0.63 (0.51 to 0.78)<0.0010.86 (0.64 to 1.15)0.302Physiological variables Presenting GCS0.75 (0.73 to 0.77)<0.0010.61 (0.55 to 0.68)<0.001 Presenting SBP (mmHg)1.00 (1.00 to 1.00)0.4431.00 (1.00 to 1.00)0.224 Heart rate1.00 (1.00 to 1.00)0.3791.00 (1.00 to 1.01)0.162 Presenting EtCO2 (kPa)0.96 (0.92 to 1.00)0.0530.95 (0.90 to 1.01)0.105Injury characteristics TBI3.00 (2.34 to 3.88)<0.0010.90 (0.62 to 1.31)0.593 Thoracic injury0.89 (0.73 to 1.09)0.2631.01 (0.71 to 1.45)0.947 Abdomen-pelvic injury0.53 (0.42 to 0.66)<0.0010.79 (0.55 to 1.12)0.193 Penetrating injury0.21 (0.07 to 0.47)<0.0010.57 (0.15 to 1.77)0.363 Cardiac arrest8.82 (6.15 to 12.82)<0.0011.85 (1.11 to 3.12)0.020Intervention variables PHEA3.80 (3.08 to 4.70)<0.0010.68 (0.40 to 1.16)0.163 ETI9.03 (6.98 to 11.83)<0.0012.80 (1.40 to 5.60)0.004 Thoracostomy4.05 (3.22 to 5.08)<0.0011.34 (0.85 to 2.10)0.207 Pelvic binder0.83 (0.68 to 1.01)0.0700.96 (0.72 to 1.29)0.788 Blood component transfusion†2.54 (2.03 to 3.17)<0.0011.60 (1.07 to 2.37)0.021Interactions Age (years)×GCS––1.00 (1.00 to 1.01)<0.001Multivariable model including candidate variables + (Age×GCS) interaction. n=1984 conveyed patients with ISS ≥15, mortality rate 397 (20.0%). *Model performance R2 = 0.389, ΔAIC = 21.6.†Blood component transfusion treated as binary variable. Results from multiple imputation (m=5).AIC, Akaike Information Criterion; aOR, adjusted OR; EtCO2, end-tidal carbon dioxide; ETI, endotracheal intubation; GCS, Glasgow Coma Scale; ISS, Injury Severity Score; PHEA, pre-hospital emergency anaesthesia; SBP, systolic blood pressure; TBI, traumatic brain injury. Univariate and multivariable logistic regression analysis for mortality risk factors in severe trauma (ISS ≥15), n=1984 Multivariable model including candidate variables + (Age×GCS) interaction. n=1984 conveyed patients with ISS ≥15, mortality rate 397 (20.0%). Model performance R2 = 0.389, ΔAIC = 21.6. Blood component transfusion treated as binary variable. Results from multiple imputation (m=5). AIC, Akaike Information Criterion; aOR, adjusted OR; EtCO2, end-tidal carbon dioxide; ETI, endotracheal intubation; GCS, Glasgow Coma Scale; ISS, Injury Severity Score; PHEA, pre-hospital emergency anaesthesia; SBP, systolic blood pressure; TBI, traumatic brain injury. Length of hospital and ICU stay in HEMS attended and ISS ≥15 Among 2075 patients, median hospital LOS was 15 days (IQR 8–29), and ICU stay was 2 days (IQR 0–8), increasing to 19 days (IQR 10–36) and 4 days (IQR 0–12), respectively, in severely injured patients (ISS ≥15, n=1397). Multivariable regression identified consistent predictors of longer hospital stay: higher injury severity (incidence rate ratio (IRR) 1.02–1.04) and reduced GCS (IRR 0.87–0.96), at p<0.001. Pre-hospital interventions reflecting disease severity were associated with increased LOS, including PHEA (hospital IRR 1.41, ICU IRR 2.63) and blood component transfusion (hospital IRR 1.44, ICU IRR 1.96), at p<0.001. Penetrating injury was associated with shorter stays (hospital IRR 0.60, ICU IRR 0.52, p<0.001). Analysis of low probability of survival in patients with Ps<30 and Ps<50 Among 156 patients with Ps<30, 41 (26.3%) unexpectedly survived to 30 days, while among 331 patients with Ps<50, 128 (38.7%) survived unexpectedly (figure 3, online supplemental table S9). In adjusted analyses controlling for age and presenting GCS, younger age and higher GCS remained the primary independent predictors of unexpected survival (Ps<50: age aOR 0.98 per year, 95% CI 0.97 to 0.99, p<0.001; GCS aOR 1.19 per point, 95% CI 1.12 to 1.28, p<0.001; Ps<30: age aOR 0.98, 95% CI 0.96 to 0.99, p=0.014; GCS aOR 1.19, 95% CI 1.08 to 1.31, p<0.001). Forest plot showing prognostic factors for unexpected survival in both the Ps<30 (n=156, 41 survivors) and Ps<50 (n=331, 128 survivors) cohorts. Each intervention variable was adjusted separately for age and presenting GCS; age and GCS were mutually adjusted. Confidence intervals exceeding the display range are truncated at OR=5. ETT, endotracheal intubation; GCS, Glasgow Coma Scale; PHEA, pre-hospital emergency anaesthesia; Ps, predicted survival score. SBP, systolic blood pressure. Age: Consistent protective factor with younger age (adjusted OR (aOR) 0.98 in both groups). GCS: Strongest predictor (aOR 1.19*** in both groups). PHEA: Significant protective effect in Ps<50 (aOR 2.01*, p=0.023). Blood transfusion and thoracostomy: Inverse associations (aOR <1) likely reflect confounding by indication as these interventions are performed in patients with more severe physiological derangement. Injury Severity Score (ISS) excluded to prevent circularity with Ps score. Statistical significance denoted by *p<0.05, **p<0.01, ***p<0.001, †p<0.1. Among intervention covariates, PHEA demonstrated a significant protective association with unexpected survival in Ps<50 patients (aOR 2.01, 95% CI 1.12 to 3.72, p=0.023). Blood transfusion (aOR 0.54, 95% CI 0.32 to 0.92, p=0.024) and thoracostomy (aOR 0.52, 95% CI 0.31 to 0.88, p=0.015) were inversely associated with survival, likely reflecting confounding by indication as these variables are preferentially performed in patients with severe physiological derangement. ETT and pelvic binder showed no significant independent associations. Sample size limitations in the Ps subgroup resulted in wide CIs for intervention estimates although thoracostomy (aOR 0.40, 95% CI 0.15 to 0.95, p=0.047) and pelvic binder (aOR 0.42, 95% CI 0.17 to 0.98, p=0.051) showed similar inverse patterns (online supplemental table S10). Traumatic cardiac arrest and return of spontaneous circulation Among 1316 patients presenting in TCA, 356 (27.1%) sustained ROSC to hospital and 960 patients were PLE at scene. Among the 356 ROSC patients, 30-day survival data were available for 185 (52%). Of these, 46 survived to 30 days (24.9% survival among ROSC with data, 12.9% of all ROSC patients). Post-ROSC in-hospital mortality was 75.1% (139/185). TCA ROSC rates demonstrated an annual improvement (6.3% increased odds per year, 95% CI 1.02 to 1.10, p=0.002). Advanced trauma interventions were performed in 93.8% of TCA patients. In multivariable logistic regression adjusting for sex, age and pre-hospital interventions (n=1309), male sex was associated with reduced ROSC (aOR 0.62, 95% CI 0.38 to 1.00, p=0.049). ETI (aOR 7.05, 95% CI 3.23 to 15.38) and plasma units transfused (aOR 1.63 per unit, 95% CI 1.29 to 2.07) were independently associated with achieving ROSC. Pelvic binder application was also associated with ROSC (aOR 1.91, 95% CI 1.11 to 3.28, p=0.019). Thoracostomy was negatively associated with ROSC (aOR 0.13, 95% CI 0.07 to 0.25), likely reflecting confounding by indication. Model discrimination was AUC=0.783 (online supplemental table S11). Discussion This study sought to evaluate the performance of a regional trauma system by analysing observed versus expected survival rates across different probability bands. Our findings indicate that observed survival significantly exceeded case-mix adjusted predictions among 2125 HEMS attended trauma patients, with 5.23 excess survivors per 100 patients (95% CI 3.27 to 7.19). Positive system performance was maintained throughout the study period. Differential performance was observed across injury severity. The 35% relative survival increase in the 25–45% probability band (49.1% actual vs 36.4% expected) suggests that the observed survival advantage over predictions is greatest in patients with moderate to severe injuries. While the adjusted Ws statistic of 5.23 indicates excess survivors per 100 patients compared with expected outcomes, it is important to interpret this in clinical context. Based on the annual volume of HEMS attended trauma patients in our system, this effect size may translate to approximately 115 additional lives saved per year. This estimate should be interpreted cautiously, as it represents excess survival compared with model predictions rather than a demonstrated causal effect of HEMS attendance and assumes consistent performance and case-mix over time. Nevertheless, it illustrates the potential magnitude of clinical benefit, consistent with previous economic and social benefits demonstrated in previous studies.34 35 The precision of our findings is supported by narrow CIs (95% CI 3.27 to 7.19) for the Ws estimate. International studies present conflicting evidence regarding HEMS effectiveness, likely reflecting methodological heterogeneity including inadequate sample sizes and inconsistent outcome definitions rather than true differences in clinical effectiveness.736,39 While some studies report minimal survival advantages,36 37 others demonstrate substantial benefits in heterogenous trauma populations.13 17 35 40,42 43 Notably, studies showing differential effectiveness across injury severity align with our findings. One UK study found minimal overall benefit (0.9 excess survivors per 100 patients, p=0.58) but meaningful improvement in major trauma subgroups (4.5 excess survivors per 100, p=0.03).2 Northern England demonstrates remarkably similar results to ours, with an enhanced care team Ws of 3.22 versus a non-enhanced care Ws of −2.97, yielding a difference of 6.18 (95% CI 3.19 to 9.17),1 supporting our Ws methodology and the magnitude of effect observed. Derivation of prognostic models demonstrates excellent performance for case-mix adjustment. The primary model achieved robust discrimination (AUC 0.919, 95% CI 0.907 to 0.930) with acceptable calibration (H-L p=0.151), and internal validation confirmed no evidence of overfitting (training AUC 0.915, test AUC 0.922). Age, GCS and ETI emerged as the strongest independent predictors of 30-day mortality. The significant age×GCS interaction reveals that older patients with preserved neurological function face disproportionately higher mortality risk, highlighting the complex interplay between physiological reserve and injury severity. LASSO regression validated core prognostic factors, confirming that a parsimonious 7-variable model maintains excellent performance with minimal loss (ΔAUC=0.007). Multivariable analysis in the full cohort shows PHEA associated with reduced mortality (aOR 0.55, 95% CI 0.33 to 0.91). This is consistent with the association observed in the low probability subgroup analysis (aOR 2.01) and likely reflects the clinical benefit of securing definitive airway management despite being performed preferentially in more severely injured patients. Similarly, penetrating injury was associated with reduced mortality (aOR 0.31, 95% CI 0.10 to 0.83), aligning with previous studies demonstrating better outcomes when rapid surgical intervention is available.44 45 46 Alternatively, our trauma system purposefully prioritises the triage of central penetrating trauma to an MTC, which may explain the seemingly protective effect.47,48 Low probability of survival patients (Ps<30 and Ps<50) had unexpected survival of 38.7% and 26.3%, respectively. Younger age (aOR 0.97 per year, 95% CI 0.96 to 0.98) and higher GCS (aOR 1.22 per point, 95% CI 1.15 to 1.30 for Ps<50) emerged as key predictors of unexpected survival, underscoring the prognostic importance of physiological reserve and neurological status beyond what conventional prediction models capture. In contrast, anatomical injury site was not independently associated after adjustment, suggesting that other patient factors and physiological response may be more important determinants in patients with low probability of survival.47 This finding challenges the traditional emphasis on anatomical scoring systems and is suggestive of a holistic physiological approach to trauma prognostication. Expanded analysis of intervention associations in low probability patients revealed that PHEA was independently associated with unexpected survival in Ps<50 patients (aOR 2.01, 95% CI 1.12 to 3.72, p=0.023). This association persisted after adjustment for age and GCS, suggesting clinical benefit beyond confounding by severity and supporting the value of definitive airway management in severely injured patients. In contrast, blood transfusion and thoracostomy showed inverse associations with survival (aOR 0.54 and 0.52, respectively), likely reflecting confounding by indication, as they are performed in patients with severe haemorrhagic and thoracic injuries. The absence of significant associations for ETI and pelvic binder may reflect a lack of independent effect, insufficient statistical power or complex confounding. Propensity-matched or instrumental variable approaches would be required to isolate true intervention effects, but were not feasible given sample size constraints. Intervention effects were clearer in the TCA subpopulation. Among TCA patients, 356 (27.1%) achieved ROSC and were conveyed to hospital. Among ROSC patients, 30-day survival data were available for 185 (52%) with 46 surviving (24.9% survival among ROSC with data, 12.9% of all ROSC patients). Post-ROSC in-hospital mortality was 75.1%. ROSC rates demonstrated annual improvement (6.3% increased odds per year), increasing from 12.5% in 2013 to 27.4% in 2022. These results demonstrate meaningful survival in selected TCA patients. In multivariable analysis (n=1309), male sex was associated with reduced ROSC (aOR 0.62, 95% CI 0.38 to 1.00, p=0.049), warranting further investigation. ETI (aOR 7.05, 95% CI 3.23 to 15.38) and plasma transfusion (aOR 1.63 per unit, 95% CI 1.29 to 2.07) were independently associated with achieving ROSC. Pelvic binder application was also associated with ROSC (aOR 1.91, 95% CI 1.11 to 3.28, p=0.019). Thoracostomy was negatively associated with ROSC (aOR 0.13, 95% CI 0.07 to 0.25). The substantial post-ROSC mortality (75.1%) indicates that while ROSC is necessary for survival, continued definitive hospital care remains critical. However, given advances in early trauma practice, ROSC rates may represent an important metric for evaluating pre-hospital trauma system effectiveness. Phase-specific performance metrics in pre-hospital trauma systems require further exploration. Current predicted survival scores rely on hospital arrival physiology which may already reflect HEMS intervention benefits, potentially confounding the assessment of pre-hospital effectiveness. This limitation is exemplified by GCS measurements; in HEMS cohorts, patients may receive PHEA, making hospital arrival GCS unreliable for baseline assessment.49 25 Missing or altered physiological data may also reduce the precision of survival predictions and limit accurate performance evaluation. Nevertheless, international studies have effectively incorporated risk estimation tools to account for missing variables, evidencing robust methodology to develop phase-specific performance metrics.50 We demonstrate that pre-hospital variables achieve excellent discrimination (AUC 0.919) supporting the feasibility of developing real-time performance metrics that could enable continuous benchmarking and quality improvement. Machine-learning approaches may provide automated, phase-specific performance measurement for operational, governance and strategic oversight.45 Recurrent neural networks analysing continuous monitoring data offer the ability to transform static predictors into dynamic assessments and such systems may encourage a move from retrospective analysis to prospective modelling. Our findings demonstrate that observed survival exceeded case-mix adjusted predictions in HEMS-attended major trauma in our region, particularly in moderately to severely injured patients (Ps 25–45%, Ws 3.33). The identified mortality predictors serve two key purposes. First, they validate our measurement methodology; the associations between age, GCS and pre-hospital interventions with outcomes demonstrate that our models achieve excellent discrimination (AUC 0.919) for case-mix adjustment, enabling fair comparison of observed versus expected survival across different patient populations and time periods. This supports robust benchmarking and detection of temporal trends in system performance. Second, while not directly mandating protocol changes, this knowledge contextualises decision-making and reinforces physiological principles underlying current trauma management. Further, these findings establish a foundation for sophisticated performance monitoring. Limitations inherent to observational studies are evident. First, we demonstrate associations but cannot establish causality between interventions and outcomes. We adopted W-statistic methodology rather than propensity score matching because valid control groups are unavailable and HEMS is routinely dispatched to the most severely injured patients, creating systematic case-mix differences that cannot be adequately matched against ground EMS cohorts. Absence of a ground EMS comparison group prevents determining whether benefits reflect HEMS-specific interventions versus the broader regional trauma system. Our findings demonstrate survival exceeding model predictions but cannot establish a causal relationship, as model calibration imperfections or unmeasured confounders could contribute to observed differences. While a randomised controlled trial would provide definitive evidence, this is neither ethical nor feasible given established HEMS integration into trauma systems. The 10-year period encompasses changes in protocols and technology confounding temporal trends. Additionally, the Ps score uses hospital (post-intervention) arrival data, meaning expected survival already reflects HEMS effects; our observed 5.23 excess survivors per 100 may thus underestimate true benefit. Second, unmeasured confounders (comorbidities, frailty) may influence outcomes, and dispatch involves complex triage introducing potential selection bias. Blood component transfusion suffers confounding by indication, that is, plasma recipients likely represent more severe cases requiring escalation. Third, no a priori sample size was calculated, though the events-per-variable ratios exceeded traditional thresholds (>28:1). Recent guidance suggests larger samples optimise performance and external validation is needed. Our model’s 88.5% sensitivity reflects class imbalance and is adequate for benchmarking, but would require recalibration for further clinical applications. Regarding data completeness, many patients had incomplete data for W-statistic analysis, with attrition primarily through non-TARN submission and not meeting inclusion criteria. Importantly, we cannot definitively establish missing at random assumptions, and if data were missing not at random, our estimates may be biased. Fourth, we could not reliably distinguish from electronic patient records those patients in cardiac arrest on HEMS arrival. This limits interpretation of TCA interventions, as pre-arrest interventions (eg, PHEA, blood transfusion) may differ from peri-arrest interventions. Also, among TCA patients achieving ROSC, 75.1% died after hospital admission, indicating substantial post-resuscitation mortality. We lack data on ethnicity, neurological outcomes and quality of life, limiting our assessment of potential disparities in outcomes across ethnic groups and long-term patient-centred outcomes. Finally, findings from one regional service may not generalise to different pre-hospital configurations. Unexpected survival analysis used conventional Ps thresholds; machine learning or dynamic physiological assessment incorporating frailty and functional status may provide more comprehensive understanding. Clinical and research implications The findings validate current practices including low-threshold ETI and blood component resuscitation in TCA, while highlighting the need for enhanced monitoring in older trauma patients given the age×GCS interaction demonstrating disproportionate mortality risk. The pronounced survival association in moderately to severely injured patients (Ps 25–45%) supports prioritising HEMS resources to this severity band. Beyond local implications, we advocate for collaborative UK-wide HEMS data collection using standardised methodology. Regional studies are limited by sample size, single-service biases and the inability to distinguish service-specific from system-level effects. A coordinated national approach with agreed outcome measures, standardised definitions and data sharing infrastructure would enable pooled analysis and the examination of rare outcomes with adequate power to establish meaningful national benchmarks. This would accelerate evidence generation and inform service development. Conclusion HEMS attendance to major trauma in this regional trauma system was associated with survival exceeding case-mix adjusted predictions (5.23 excess survivors per 100 patients, 95% CI 3.27 to 7.19), with greatest effect in severely injured patients with moderate survival probability (Ps 25–45%, Ws 3.33). PHEA was independantly associated with unexpected survival in Ps <50. Our findings provide supportive evidence for advanced pre-hospital trauma interventions and demonstrate that phase-specific benchmarking can effectively evaluate HEMS performance, though comparative studies are needed to establish causal effectiveness. Supplementary material Acknowledgements The authors gratefully acknowledge the invaluable contributions of the clinical, operational and senior management teams at Air Ambulance Charity Kent Surrey Sussex. We extend particular thanks to Sophy Dunn for her meticulous efforts in data entry and validation processes at KSS. We acknowledge the academic critique provided by colleagues within the Faculty of Health Sciences at the University of Surrey, whose expertise significantly enhanced this study. Special recognition is due to the clinicians, research staff and senior management teams at both South-East Coast Ambulance Trust and across major trauma systems in the South-East region, whose collaboration made this research possible. The authors declare that this work would not have been feasible without the collective expertise and commitment of all individuals and organisations. Footnotes Funding: The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors. Provenance and peer review: Not commissioned; externally peer reviewed. Handling editor: Richard Body Patient consent for publication: Not applicable. Data availability free text: The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Analytical code is available on reasonable request. Patient and public involvement: Representatives from each major trauma system and the internal charity board expressed support for research into trauma outcomes. While patients were not directly involved in study design, recruitment, or conduct, the results and clinical interpretation will be shared with lay representatives and partner organisations within the regional trauma networks to ensure broad stakeholder engagement and knowledge translation. Ethics approval: All data were routinely collected and met Health Research Authority (HRA, UK) criteria as service evaluation. Research Ethics Committee approval was not required. The project was approved by the KSS Research and Innovation Committee and registered with the University of Surrey. Study conduct was in accordance with Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Data availability statement Data are available upon reasonable request. References Associated Data Supplementary Materials Data Availability Statement Data are available upon reasonable request.