Patients presenting to four English trauma networks by ambulance with non-trivial injury were most commonly elderly females with blunt trauma from ground level falls. The overall prevalence of major trauma was low (3.1%). Observed prehospital provider triage decisions demonstrated overall pooled sensitivity of 46.7% (95% CI 43.5-49.9) and specificity of 94.5% (95% CI 93.4-95.6) for the primary reference standard. However, accuracy was heterogenous with SWAS demonstrating lower sensitivity and higher specificity than other ambulance services. There was a clear trend of decreasing sensitivity and increasing specificity from younger to older ages. Prehospital provider triage decisions did not always follow theoretical triage tool results, with ambulance service triage decisions demonstrating higher specificity. Sensitivity of triage decisions was increased for urgent trauma interventions (80%, 95%CI 75.0-84.0), with specificity remaining high (94%, 95%CI 93.0-95.0).

Statistical measures of diagnostic accuracy are typically reported in terms of sensitivity and specificity [14]. These metrics may be counterintuitive as they are defined retrospectively by disease status, rather than the clinically relevant probability of a condition being present when the test result in known [15]. There is evidence that graphical representations, predictive values, natural frequencies, or using likelihood ratios, promotes better understanding [15,16,17]. A range of visual interpretations of the main study findings are therefore presented in the supplementary materials. Taking a Bayesian system-level approach, on average there is a 3.1% chance of major trauma when an acutely injured patient presents to ambulance services. After a positive triage decision for bypassing/pre-alerting the MTC there will be a 22% chance that patient will turn out to have major trauma. Conversely, if the triage decision is negative then there will be a 98% chance that the patient does not have major trauma [18].

For triage tools structured as a 'checklist', diagnostic accuracy could be manipulated by changing the number, type, and thresholds of included variables. This results in an inevitable trade-off between sensitivity and specificity, where the number of false positive cases increases as false negatives are reduced. Sensitivity is often prioritised, for example the ACS-COT has published a target of >95% for field triage, with a consequent penalty of reduced specificity (ASC-COT targets 65-70%) [19]. However, in low prevalence settings this approach would lead to poor positive predictive values and many false positives [20]. Economic modelling has confirmed that prioritising sensitivity is not cost-effective, and specificity should be favoured [21,22,23,24]. It is interesting that real-life prehospital triage decisions are closely aligned to these empirical cost-effectiveness targets.

The index test definition for positive prehospital triage is open to debate. Our primary definition of a pre-alert call to the MTC, reflects the dual purposes of a major trauma triage tool of selecting patients for bypass to a distant MTC and facilitating emergency department pre-alert calls. It could be argued that transport to a MTC is the most important factor. However, this would not account for delayed resuscitation and treatment in cases not pre-alerted to the ED. It could also be misleadingly influenced by the proportion of patients injured within the MTC catchment area, as these patients' destination is fixed. Alternatively, from a system perspective, the final common pathway for major trauma patients' is reception into a MTC resuscitation area with hospital trauma team activation. However, full hospital trauma team assessment may not be required in stable patients, activation criteria may differ from triage tool variables, and deployment is outside the control of prehospital providers.

The reference standard against which triage decisions should be assessed is also arguable. The traditionally used Injury Severity Score has many limitations, not least its focus on the degree of anatomical trauma per se, rather than the potential to benefit from MTC care [11]. Intervention based reference standards, such as the US consensus definition, may better reflect the need for MTC care, but do not account for benefit arising from MTC supportive care and rehabilitation [10, 25]. Operationalisation in triage research is also potentially challenged by the absence of counterfactual information, and cases may be incorrectly classified as false negatives if they did not receive an intervention that was indicated due to lack of availability or expertise outside the MTC. Evaluation of multiple reference standards will provide a comprehensive and holistic assessment of triage performance.

Differences were apparent in observed prehospital triage, and calculated theoretical triage tool accuracy, across participating trauma networks. The overall pooled accuracy results should therefore be interpreted judiciously. The heterogeneity of results may reflect the use of different ambulance service triage tools. Theoretical accuracy of mandatory triage tool steps (SWAS most specific, LAS most sensitive, WMAS/YAS intermediate), appeared to correlate with real-life triage decision making (SWAS most specific, LAS most sensitive, WMAS/YAS intermediate). Alternatively, the variation could reflect differences in philosophies (inclusive versus exclusive focus), geographies (more urban Vs more rural), or network organisation (e.g., extent of remote decision support available from trauma desks).

Prehospital providers triage decisions were observed to differ from those indicated by triage tools, apparently improving triage performance. This could result from discretionary triage tool use in selected patients, application of additional clinical judgement to over-rule indicated triage tool results, or the influence of applying discretionary triage tool steps. Previous qualitative research has demonstrated that prehospital triage making is often heuristic with triage tools used less commonly as experience increases [26]. It is perhaps not surprising that subjective real-life decisions outperformed objective triage tool accuracy, as clinical judgement has been consistently demonstrated to be superior in multiple studies across many disease areas [27, 28]. This has important implications, as to benefit from better overall triage performance achieved through subjective clinical judgement, some incorrect individual clinical decisions will need to be accepted; and clinical governance strategies to increase triage tool use and adherence may be counterproductive. Spectrum effects were apparent across different age groups with decreasing sensitivity and increasing specificity for triage decisions from younger to older ages for the primary reference standard. Interestingly, the proportion of false negatives requiring urgent trauma interventions remained negligible (<1%), perhaps reflecting provider judgement in selecting patients perceived to benefit most from MTC care.

The four included trauma networks are representative of the broader English population, including a diverse mix of urbanisation, socioeconomic status, geographies, and injury profiles. The results of this study should therefore be generalisable throughout the UK National Health Service. In response to the COVID-19 pandemic, new triage tools have been implemented in LAS and SWAS, which may influence contemporary triage. External validity to other settings is limited. Different injury patterns (e.g., higher numbers of gun-shot wounds), alternative health system models (state Vs insurance Vs private funding), trauma network organisation (level of training, inclusive Vs exclusive), medicolegal risk, and patient demographics (e.g., population age profiles) in other settings could strongly influence triage decision making.

Two recent systematic reviews have examined real-life major trauma triage decisions and compliance with theoretical triage tool results. Van Rein (2018a) evaluated prehospital triage decisions, defined by initial hospital destination, in 33 studies [29]. The findings were limited by poor methodological quality and very heterogenous results were reported with sensitivity ranging from 32% to 99% and specificity from 1% to 99%. Notwithstanding the different populations and trauma system organisation seen in the predominantly US settings, results from the better-quality studies were not disimilar to the current findings. It was also concluded, in common with the current study, that EMS provider judgment added value to triage protocols in the identification of severely injured patients. Van Rein (2018b) also investigated compliance, reviewing 11 studies comparing objective triage tool results with actual triage destinations [30]. The methodological quality of most studies was again poor, with widely disparate compliance rates between 21% and 91% reported. One study with good methodological quality showed, in common with our findings, that the triage protocol identified only a minority of severely injured patients, and a tendency to transport elderly trauma patients to lower-level trauma centres, even if the patient met one or more triage criteria.

This study has several strengths. The risk of information bias was minimised by following recommendations for collection of routine data in retrospective studies [31, 32]. Other common sources of systematic error in diagnostic accuracy studies were avoided, including a 'two-gate' case-control study design; and test, diagnostic review, partial verification, incomplete verification, incorporation, and disease progression biases [33]. However, there are some potential limitations. The use of routine data may have resulted in index test or reference standard misclassification. Selection bias could arise from imperfect matching of ambulance service and TARN data and complete case analyses, omitting cases with missing data. However, the matching rate was high (>95%) and missing data rate was low (<10%). Furthermore, reference standard classification is dependent on encompassing TARN inclusion criteria, complete case ascertainment by TARN, and accurate matching of TARN and prehospital data.