Prehospital prediction of severe injury in road traffic injuries: A multicenter cross-sectional study

Open AccessPublished:May 26, 2019DOI:https://doi.org/10.1016/j.injury.2019.05.028

      Highlights

      • This study developed and validated risk prediction scores of prehospital death and severe injury for road traffic injury.
      • Ten predictors, which could easily be assessed at scene by emergency medical service personnel, were included in the prediction scores.
      • These risk prediction scores revealed good calibration and discrimination performances for internal/external validations.
      • These scores could classify subjects into low/moderate/high risks of death/SI during prehospital operation.
      • Applying these scores could identify and prioritize RTI patients for appropriate patient transport to hospital.

      Abstract

      Background

      To develop and validate a risk stratification model of severe injury (SI) and death to identify and prioritize road traffic injury (RTI) patients for transportation to an appropriate trauma center (TC).

      Methods

      A 2-phase multicenter-cross-sectional study with prospective data collection was collaboratively conducted using 9 dispatch centers (DC) across Thailand. Among the 9 included DC, 7 and 2 DCs were used for development and validation, respectively. RTI patients who were treated and transported to hospitals by advanced life support (ALS) response units were enrolled. Multiple logistic regression was used to derive risk prediction score of death in 48 h and SI (new injury severity score ≥ 16). Calibration/discrimination performances were explored.

      Results

      A total of 5359 and 2097 RTIs were used for development and external validation, respectively. Seven and 9 predictors among demographic data, mechanism of injury, physic data, EMS operation, and prehospital managements were significant predictors of death and SI, respectively. Risk prediction models fitted well with the developed data (O/E ratios of 1.00 (IQR: 0.69, 1.01) and 0.99 (IQR: 0.95, 1.05) for death and SI, respectively); and the C statistics of 0.966 (0.961, 0.972) and 0.913 (0.905, 0.922). The risk scores were further stratified as low, moderate and high risk. The derive models did not fit well with external data but they were improved after recalibrating the intercepts. However, the model was externally good/excellent discriminated with C statistics from 0.896 (0.871, 0.922) to 0.981 (0.971, 0.991).

      Conclusion

      Risk prediction models of death and SI were developed with good calibration and excellent discrimination. The model should be useful for ALS response units in proper allocation of patients.

      Keywords

      Introduction

      Road traffic injury (RTI) is the most common event for emergency medical services [
      • Key C.B.
      Vehicle related injuries.
      ], accounting for approximately 23% of injury related deaths worldwide [
      • Peden M.
      • Scurfield R.
      • Sleet D.
      • Mohan D.
      • Hyder A.A.
      • Jarawan E.
      • et al.
      World report on road traffic injury prevention.
      ]. In Thailand, RTI was responsible for half of all injuries (49.4%), and accounted for 64.3% of injury related deaths in 2005 [
      • Chadbunchachai W.
      • Suphanchaimaj W.
      • Settasatien A.
      • Jinwong T.
      Road traffic injuries in Thailand: current situation.
      ]. Despite the implementation of several programs (e.g., traffic law enforcement, traffic calming intervention, etcetera.) the high mortality rate remains [
      • Toroyan T.
      • Peden M.M.
      • Iaych K.
      WHO launches second global status report on road safety.
      ].
      An emergency medical service (EMS) system has been set up and integrated into existing health care systems to minimize morbidity and mortality by providing pre-hospital treatments and transportation to the most appropriate hospital [
      • Blomberg H.
      • Svennblad B.
      • Michaelsson K.
      • Byberg L.
      • Johansson J.
      • Gedeborg R.
      Prehospital trauma life support training of ambulance caregivers and the outcomes of traffic-injury victims in Sweden.
      ]. Current evidence indicates that severely injured patients should be transferred to a high-level rather than low-level trauma center (TC) [
      • Haas B.
      • Stukel T.A.
      • Gomez D.
      • Zagorski B.
      • De Mestral C.
      • Sharma S.V.
      • et al.
      The mortality benefit of direct trauma center transport in a regional trauma system: a population-based analysis.
      ], but there are few such centers in Thailand and other developing countries.
      Currently, EMS system in Thailand is developing and there is no standardized triage tool implemented for identifying severity of RTI patients. Knowing their severity of injuries (SI) would aid to properly prioritize and thus lead to allocate treatment management and appropriately transfer of RTI victims to hospitals. Several risk prediction scores have been developed to classify severity of injured victims in general trauma patients, (e.g., revised trauma score (RTS) trauma injury severity score (TRISS), and the field triage decision scheme) mainly considering physiologic factors (i.e., systolic blood pressure (SBP), respiratory rate (RR) and Glasgow coma scale (GCS)) [
      • Sasser S.M.
      • Hunt R.C.
      • Faul M.
      • Sugerman D.
      • Pearson W.S.
      • Dulski T.
      • et al.
      Guidelines for field triage of injured patients: recommendations of the National Expert Panel on Field Triage, 2011.
      ,
      • Champion H.R.
      • Copes W.S.
      • Sacco W.J.
      • Lawnick M.M.
      • Keast S.L.
      • Bain Jr., L.W.
      • et al.
      The Major Trauma Outcome Study: establishing national norms for trauma care.
      ,
      • Offner P.J.
      • Jurkovich G.J.
      • Gurney J.
      • Rivara F.P.
      Revision of TRISS for intubated patients.
      ,
      • Champion H.R.
      • Sacco W.J.
      • Copes W.S.
      • Gann D.S.
      • Gennarelli T.A.
      • Flanagan M.E.
      A revision of the trauma score.
      ]. The field triage decision scheme was firstly recommended by the American College of Surgeons Committee on Trauma (ACS-COT) in 1999 [
      • Sasser S.M.
      • Hunt R.C.
      • Faul M.
      • Sugerman D.
      • Pearson W.S.
      • Dulski T.
      • et al.
      Guidelines for field triage of injured patients: recommendations of the National Expert Panel on Field Triage, 2011.
      ] and later revised in 2006 considering vital signs, anatomical involvement, mechanism of injury and special circumstances in order, which mainly aimed to prioritize high risk of severe trauma patients. However, it was low to fair in discriminating high risk from low risk RTI patients (C statistic 0.55 to 0.65) [
      • Scheetz L.J.
      Trends in the accuracy of older person trauma triage from 2004 to 2008.
      ]. Although some risk prediction models for RTI patients exist, their diagnostic accuracies varied greatly [
      • Newgard C.D.
      • Lewis R.J.
      • Jolly B.T.
      Use of out-of-hospital variables to predict severity of injury in pediatric patients involved in motor vehicle crashes.
      ,
      • Newgard C.D.
      • Hui S.H.J.
      • Griffin A.
      • Wuerstle M.
      • Pratt F.
      • Lewis R.J.
      Prospective validation of an out-of-hospital decision rule to identify seriously injured children involved in motor vehicle crashes.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.E.
      Using crash scene variables to predict the need for trauma center care in older persons.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.
      Classification tree modeling to identify severe and moderate vehicular injuries in young and middle-aged adults.
      ]. Given the high incidence of RTI, a risk prediction score should be able to stratify a patient’s severity and prioritize them properly.
      We conducted a multi-center cross-sectional study, which aimed to develop a risk stratification model of severe injury (SI) and death to identify and prioritize RTI patients for transportation to an appropriate trauma center (TC). This study was approved by the ethics committee of the Faculty of Medicine Ramathibodi Hospital, Mahidol University and applied waiver of informed consent due to urgent situation.

      Methods

       Study design and setting

      The study design was a multi-center cross-sectional study with prospective data collection, which composed of development and validation phases following suggestions for risk prediction model developments by Moons et al [
      • Moons K.G.
      • Kengne A.P.
      • Grobbee D.E.
      • Royston P.
      • Vergouwe Y.
      • Altman D.G.
      • et al.
      Risk prediction models: II. External validation, model updating, and impact assessment.
      ,
      • Moons K.G.
      • Kengne A.P.
      • Woodward M.
      • Royston P.
      • Vergouwe Y.
      • Altman D.G.
      • et al.
      Risk prediction models: I. Development, internal validation, and assessing the incremental value of a new (bio)marker.
      ]. We enrolled RTI subjects under provincial dispatch centers (DC) during July 2015 to February 2017.
      The EMS system in Thailand consists of basic and advanced life support units (BLS, ALS), which are provided by non-government foundations and provincial/regional hospitals, respectively, under regulations of provincial DCs and the National Institute of Emergency Medicine. All EMS personnel must be certified before practicing under regulation of the National Institute of Emergency Medicine. All traffic victims have to be treated and transported by ALS units, except minor injury (e.g., pure laceration wound or mild head injury) which can be looked after by the BLS teams.
      Because prehospital care and the EMS information system in Thailand have been recently developed, we purposively selected 9 DCs across the country based on following three criteria: high density of RTI cases treated by the ALS response unit, an emergency physician (EP) as medical director, and having an EMS information system. Among 9 DCs, 7 DCs (i.e., Saraburi, Ayutthaya, Chiang Mai, Nakhon Ratchasima, Khon Kaen, Nakhom Si Thammarat, and Chonburi) were used to develop and internally validate the risk prediction model whereas two DCs (i.e., Ubonratchathani and Trang) were selected to externally validate the risk model. This study followed transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) [
      • Moons K.G.
      • Altman D.G.
      • Reitsma J.B.
      • Ioannidis J.P.
      • Macaskill P.
      • Steyerberg E.W.
      • et al.
      Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration.
      ].

       Selection of participants

      Subjects were eligible if they were aged 15 years or older, experienced a RTI, and were transported to hospital by ALS response unit in the studied DCs. Patients were excluded if they had at least one sign of irreversible death (i.e., decapitation, incineration, separation, or destruction of heart or brain, rigor mortis, and lividity), declined EMS treatment or transportation to hospital, or could not be assessed for eligibility due to an unsafe accident scene.

       Outcomes

      The primary outcome was death within 48 h of the RTI occurrence. The secondary outcome was SI, defined as a new injury severity score (NISS) ≥ 16. [
      • Kashani A.T.
      • Mohaymany A.S.
      Analysis of the traffic injury severity on two-lane, two-way rural roads based on classification tree models.
      ,
      • Ayoung-Chee P.
      • Mack C.D.
      • Kaufman R.
      • Bulger E.
      Predicting severe injury using vehicle telemetry data.
      ,
      • Newgard C.D.
      • Hui S.H.J.
      • Griffin A.
      • Wuerstle M.
      • Pratt F.
      • Lewis R.J.
      Prospective validation of an out-of-hospital decision rule to identify seriously injured children involved in motor vehicle crashes.
      ,
      • Scheetz L.J.
      Trends in the accuracy of older person trauma triage from 2004 to 2008.
      ,
      • Newgard C.D.
      • Lewis R.J.
      • Jolly B.T.
      Use of out-of-hospital variables to predict severity of injury in pediatric patients involved in motor vehicle crashes.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.E.
      Using crash scene variables to predict the need for trauma center care in older persons.
      ] For individual subjects, NISS was assessed using an abbreviated injury scale (AIS), according to AIS 2005 (update 2008) dictionary [
      • Gennarelli T.A.
      • Wodzin E.
      Abbreviated injury scale 2005 (update 2008): association for the Advancement of Automotive Medicine (AAAM).
      ], for all injury related diagnoses. NISS was then estimated by the sum of squares of the first three highest AIS diagnoses.

       Methods of measurement

      Predictors were classified into 7 domains as follows:
      • Demographic data included age (year), sex, body mass index (BMI). Age was categorized based on field triage decision scheme [
        • Sasser S.M.
        • Hunt R.C.
        • Faul M.
        • Sugerman D.
        • Pearson W.S.
        • Dulski T.
        • et al.
        Guidelines for field triage of injured patients: recommendations of the National Expert Panel on Field Triage, 2011.
        ].
      • The crash characteristics data included types of road user (pedestrian/bicyclist or motorcyclist/4, or more wheels vehicles), and total number of victims.
      • EMS operation domain included response time, on scene time, transportation time, distance from the EMS base to scene and scene to destination hospital, and out of hospital management (i.e., intravenous fluid administration, airway management). Response time, on scene time, and transportation time were defined as the time since EMS was requested to the time at scene arrival, time since the response unit arrival to departure from the scene, and departure to receiving hospital, respectively. Response time was also categorized into ≤ 8 min or longer [
        • Blomberg H.
        • Svennblad B.
        • Michaelsson K.
        • Byberg L.
        • Johansson J.
        • Gedeborg R.
        Prehospital trauma life support training of ambulance caregivers and the outcomes of traffic-injury victims in Sweden.
        ]. Out of hospital management referred to receiving any intravenous fluid administration or airway management (not receiving/ clear airway/ assisted ventilation) before hospital arrival.
      • Mechanism of injury consisted of presence of burn, blunt, and penetrating injuries.
      • Physiological data were results of the first physical examinations at scene, i.e., SBP, RR (breaths/minute), and GCS. These variables were also categorized based on RTS [
        • Champion H.R.
        • Sacco W.J.
        • Copes W.S.
        • Gann D.S.
        • Gennarelli T.A.
        • Flanagan M.E.
        A revision of the trauma score.
        ].
      • Environmental domain included time (day/ night) and location of incidence (highway/non-highway).
      • Risk behavior domain included alcohol consumption before the accident.
      EMS personnel (paramedics/ pre-hospital nurse/ doctors) in study provinces were trained about variable definition and data collection processes before fieldwork. All predictors were prospectively collected at scene during pre-hospital management, except EMS operation and physiological factors which were recorded by EMS operation forms according to EMS system in Thailand and subsequently extracted to case record form. Death and trauma related diagnosis were extracted from medical record. Predictors and outcomes were collected by trained EMS personnel. All data were checked and enquired for completeness, correctness, and inconsistency before data entry. The two databases were validated and rechecked before final analysis.

       Sample size estimation

      The pooled prevalence of SI was 12.5% (95% CI: 10.8%–14.3%) from our literature review [
      • Kashani A.T.
      • Mohaymany A.S.
      Analysis of the traffic injury severity on two-lane, two-way rural roads based on classification tree models.
      ,
      • Ayoung-Chee P.
      • Mack C.D.
      • Kaufman R.
      • Bulger E.
      Predicting severe injury using vehicle telemetry data.
      ,
      • Scheetz L.J.
      Trends in the accuracy of older person trauma triage from 2004 to 2008.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.E.
      Using crash scene variables to predict the need for trauma center care in older persons.
      ]. A sample size was calculated based on estimating this proportion with setting 5% type one error, adjusting for design effect of clustering and 10% missing data; which indicated 5111 subjects were required for the derivation phase and an additional 1022 subjects were enrolled for the external validation. Therefore, 6133 subjects were required for this study.

      Data analysis

       Multiple imputation

      Multiple imputation (MI) with chained equations was used to impute missing data [
      • Moons K.G.
      • Donders R.A.
      • Stijnen T.
      • Harrell Jr., F.E.
      Using the outcome for imputation of missing predictor values was preferred.
      ]. Logistic regression and interval regressions with 40 imputations were used to predict binary and continuous variables, respectively. Complete data of predictors (e.g., age, sex, mean arterial pressure (MAP), GCS, PR) and outcome (i.e., NISS and death in 48 h) were used to predict missing data for weight, height, and RR. These predictors plus types of road users, time of incidence, total number of victims, burn, penetration, and blunt injury were used to predict alcohol consumption. Performance of the MI was explored and assessed using relative variance increase (RVI) and fraction of missing information (FMI). The largest FMI of missing variable was used to suggest the adequate number of imputations according to the rule of thumb, i.e., FMI x100. For instance, if FMI was 0.39, number of imputations was at least 39.

       Derivation phase

      Simple logistic regression was used to screen predictors of death and SI. Variables with a p value less than or equal to 0.1 were simultaneously considered in a multivariable logit model using forward selection. Likelihood ratio test were used to select only significant variables in the final model. Model performance was assessed as follows: Calibration was assessed using Hosmer-Lemeshow goodness of fit (HL-GOF) test by equally dividing predicted probability into 10, 30, 45, 65 and 110 groups where appropriate [
      • David W.
      • Hosmer J.
      • Lemeshow Stanley
      • Sturdivant R.X.
      Assessing the fit of the model.
      ,
      • Paul P.
      • Pennell M.L.
      • Lemeshow S.
      Standardizing the power of the Hosmer-Lemeshow goodness of fit test in large data sets.
      ]. In addition, observed (O) and expected (E) numbers, and O/E ratio along with its interquartile range (IQR) were estimated. The O/E ratio close to 1 indicated the predicted values close to observed values. A calibration plot was constructed by plotting O values on X-axis and E values on and Y-axis. Finally, discrimination was assessed using receive operating characteristic (ROC) curve analysis; and a concordance statistic (C statistic) was estimated.
      All coefficients in the final model were used to construct a risk prediction score. The score cutoff was then calibrated based on its distribution and ROC analysis. Diagnostic parameters of moderate and high-risk groups compared to the rest (i.e., sensitivity, specificity, positive predictive value (PPV), positive likelihood ratio (LR+) were then estimated.

       Internal validation

      A bootstrap with 1000 repetitions was used to assess internal performances. For the calibration coefficient, Somers’ D rank correlations of derived model (Dorig) and each set of bootstrapping (Dboot) were estimated [
      • Harrell Jr., F.E.
      • Lee K.L.
      • Mark D.B.
      Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors.
      ]. Then, optimism (Oc) was calculated by Dorig- Dboot. Finally, the bootstrap corrected calibration coefficient was estimated by Dorig- Mean Oc. For discrimination, C statistic of the derived model was subtracted by C statistic of each bootstrapping (Cboot); and the bootstrapping corrected discrimination coefficient was then estimated.

       External validation

      The risk prediction score from the development phase (called M0) was calculated in the validation data, along with estimated probability of outcome occurrence. Calibration and discrimination performance were explored as previously mentioned. If the risk prediction model did not fit well with the external data, model recalibration was performed by recalibrating the intercept with/without the updated model where appropriate [
      • Toll D.B.
      • Janssen K.J.
      • Vergouwe Y.
      • Moons K.G.
      Validation, updating and impact of clinical prediction rules: a review.
      ]. Model revision by recalibration of the intercept (Called M1) [
      • Moons K.G.
      • Kengne A.P.
      • Grobbee D.E.
      • Royston P.
      • Vergouwe Y.
      • Altman D.G.
      • et al.
      Risk prediction models: II. External validation, model updating, and impact assessment.
      ], and its performance was then re-explored. The scores were categorized into low, moderate and high risk groups, based on the cut off points of derivative phase. Diagnostic parameters of moderate and high-risk groups compared to the rest (i.e., sensitivity, specificity, positive predictive value (PPV), positive likelihood ratio (LR+) were then estimated.
      All analyses were performed using STATA version 15.0 (Stata Corp, College Station, Texas, USA), based on mi estimation commands. P value less than 0.05 was considered as statistically significant.

      Results

       Characteristics of study subjects

      A total of 7456 subjects were enrolled, 5359 and 2097 subjects for derivation and external validation, respectively. Among 5359 subjects, mean age (SD) was 35.1 (16.0) years, 3824 subjects (71.4%) were male, and 4380 (81.7%) subjects rode 2–3 wheel vehicles. A total of 1472 (27.5%; 95% CI: 26.3%, 28.7%) subjects were classified as SI and 696 (13.0%; 95% CI: 12.1%, 13.9%) died within 48 h. The characteristics of subjects are described by province in Supplemental Table 1.
      Table 1Predictors of death: A multiple logistic regression.
      PredictorsCoefficients95%CISEPOR95%CI
      Intercept−6.620−7.280, -5.9601.34<0.0010.0011.01, 1.02
      Age0.0180.011, 0.0260.00<0.0011.011.01, 1.02
      Blunt injury
      Yes0.8350.524, 1.1470.15<0.0012.301.68, 3.15
      No01
      RR 4 groups
      <61.0320.582, 1.4830.22<0.0012.801.79, 4.40
      6-91.0210.380, 1.6620.320.0022.771.46, 5.27
      >291.0510.509, 1.5920.27<0.0012.861.66, 4.91
      10-2901
      SBP 4 groups
      <502.2511.809, 2.6940.22<0.0019.506.10, 14.80
      50-751.6611.103, 2.2200.28<0.0015.263.01, 9.20
      76-891.2340.728, 1.7400.25<0.0013.432.07, 5.70
      >8901
      GCS 5 groups
      32.6022.028, 3.1750.29<0.00113.407.60, 23.90
      4-52.6141.939, 3.2890.34<0.00113.606.95, 26.8
      6-81.5390.967, 2.1110.29<0.0014.662.63, 8.26
      9-121.3360.754, 1.9190.29<0.0013.802.12, 6.81
      13-1501
      Airway management
      Assisted ventilation1.2170.474, 1.9600.370.0013.371.60, 7.10
      Open/clear airway0.419−0.25, 1.0920.340.2231.520.77, 2.98
      No supplement01
      IV fluid administration
      YES0.6110.082, 1.1410.270.0241.841.08, 3.13
      No01
      CI: confidence interval; GCS: Glasgow coma scale; IV: intravenous; OR: odds ratio; P: p value; RR: respiratory rate; SBP: systolic blood pressure; SE: standard error.

      Main results

       Imputation

      There were 13 (0.17%), 28 (0.37%), 40 (0.53%), and 922 (12.36%) observations where data for RR, height, weight, and alcohol consumption were missing respectively. These data were therefore imputed with estimated FMI and RVI of 0.001 to 0.3968 and 0.001 to 0.3226, respectively. Alcohol consumption contributed the largest FMI (0.3968) and RVI (0.3226) (see Supplemental Table 2).
      Table 2Model performance for predicting death and SI in RTI subjects.
      PredictorsPhasesProvincesModelCalibrationDiscrimination
      HLdfPO/EC statistic
      Chi2(IQR)(95% CI)
      DeathDerivation24.43280.661.000.966
      (0.69, 1.01)(0.961, 0.972)
      ExternalUbonratchathaniM03.4580.901.000.981
      Validation(0.57, 1.01)(0.971, 0.991)
      M13.5080.890.990.981
      (0.31, 1.01)(0.971, 0.991)
      TrangM04.4880.811.000.947
      (0.89, 1.09)(0.922, 0.973)
      M15.7180.681.000.947
      (0.97, 1.36)(0.922, 0.973)
      SIDerivation13.8280.090.990.913
      (0.95, 1.05)(0.905, 0.922)
      ExternalUbonratchathaniM028.78<0.0011.000.909
      Validation(0.71, 1.03)(0.885, 0.932)
      M17.280.511.000.909
      (0.95, 1.05)(0.885, 0.932)
      TrangM021.080.0070.990.896
      (0.78, 1.04)(0.871, 0.922)
      M17.560.2810.896
      (0.94, 1.04)(0.871, 0.922)
      C statistic: concordance statistic; df: degree of freedom; P: p value; O/E: observed over expected value; HL Chi2: Hosmer-Lemeshow Chi2; RTI: road traffic injury; SI: severe injury.

       Prehospital prediction of death in 48 h

       Development phase

      Of 20 potential predictors, 17 were candidates in the multivariable analysis but only 16 (80%) were included because direction of daytime effect was counter-intuitive which might be due to survival bias (see Supplemental Table 3). Finally, 7 were kept including age, physical examination (i.e., SBP, RR, and GCS), blunt injury and EMS operation, i.e., IV fluid administration, and airway managements (Table 1). The risk prediction score was calculated following the equation described in Supplemental Figure 1.
      Table 3Diagnostic performance of moderate and high risk compared with low risk groups of derived and external validated models for predicting death and SI.
      OutcomesPhasesProvincesModelRiskDeathAlive%PPV%Sens%SpecLR+
      groups(95% CI)(95% CI)(95% CI)
      DeathDerivationLow932070.3
      Moderate428984.598.768.83.2
      (97.6, 99.4)(67.4, 70.1)(3.0, 3.3)
      High64555853.692.788.07.7
      (90.5, 94.5)(87.1, 89.0)(7.1, 8.4)
      ExternalUbonratchathaniM0Low17690.1
      validationModerate51533.298.976.04.1
      (94.1, 100)(73.2, 78.6)(3.7, 4.6)
      High869048.993.491.110.5
      (86.2, 97.5)(89.2, 92.8)(8.6, 12.9)
      M1Low17670.1
      Moderate51553.198.975.84.1
      (94.1, 100)(73.0, 78.4)(3.7, 4.6)
      High869048.993.591.110.5
      (86.3, 97.6)(89.1, 92.8)(8.6, 12.9)
      TrangM0Low36840.4
      Moderate91635.295.373.63.6
      (86.9, 99.0)(70.7, 76.4)(3.2, 4.1)
      High528238.881.391.29.2
      (69.5, 89.9)(89.2, 92.9)(7.3 11.7)
      M1Low57730.6
      Moderate81017.392.283.25.5
      (82.7, 97.4)(80.6, 85.6)(4.7, 6.4)
      High515548.179.794.113.5
      (67.8, 88.7)(92.4, 95.5)(10.1, 17.9)
      OutcomePhasesProvincesModelRiskSINon-SI%PPV%Sens%SpecLR+
      groups(95% CI)(95% CI)(95% CI)
      SIDerivationLow14429494.7
      Moderate23156229.190.275.93.7
      (88.6, 91.7)(74.5, 77.2)(3.5, 4.0)
      High109737674.574.590.37.7
      (72.2, 76.7)(89.4, 91.2)(7.0, 8.5)
      ExternalUbonratchathaniM0Low247043.3
      validationModerate2613116.687.176.73.7
      (81.4, 91.6)(73.8, 79.4)(3.3, 4.3)
      High1368362.173.1918.1
      (66.1, 79.3)(88.9, 92.7)(6.5, 10.1)
      M1Low317733.9
      Moderate309324.483.384.25.3
      (77.2, 88.4)(81.7, 86.5)(4.5, 6.2)
      High1255270.667.294.311.9
      (60.0, 73.9)(92.6, 95.7)(8.9, 15.7)
      TrangM0Low236213.6
      Moderate4214422.686.9763.6
      (81, 91.5)(72.9, 78.9)(3.2, 4.2)
      High1115268.163.193.69.9
      (55.5, 70.2)(91.7, 95.2)(7.4, 13.2)
      M1Low336614.8
      Moderate3811025.781.380.94.3
      (74.7, 86.7)(78.0, 83.5)(3.6, 5.0)
      High1054669.559.794.410.6
      (52.0, 67.0)(92.6, 95.8)(7.8, 14.4)
      CI: confidence interval; LR+: likelihood ratio positive; Non-SI: non -severe injury; PPV: positive predictive value; Sens: sensitivity; SI: severe injury; Spec: specificity.
      The C statistic of this model was 0.966 (95% CI: 0.961, 0.972) indicating excellent performance in discriminating death from survival in RTI subjects. The HL-GOF test revealed the model fitted well with the data (Chi-square = 24.43, df = 28, P =  0.66) with O/E ratio of 1.00 (IQR: 0.69, 1.01), see Table 2 and Supplemental Table 4. The calibration plot indicated that the observed and predicted values were very close, see Supplemental Figure 2.
      Table 4Predictors associated with SI: Multiple logistic regression.
      PredictorsCoefficients(95%CI)SEPOR(95%CI)
      Intercept−3.934−4.201, -3.6770.131<0.0010.02(0.02, 0.03)
      Age
      >55 years0.3510.105, 0.5970.1250.0051.42(1.11, 1.82)
      ≤ 55 years01
      SBP
      >500.7010.245, 1.1560.2320.0032.021.28, 3.18
      50-750.7900.24, 1.3390.2810.0052.201.27, 3.82
      76-890.5810.152, 1.0090.2190.0081.791.17, 2.74
      >8901
      RR
      <100.208−0.223, 0.6380.2200.341.230.81, 1.89
      >290.6460.175, 1.1150.2400.0071.911.20, 3.05
      10-2901
      GCS
      32.2501.867, 2.6330.195<0.0019.496.47, 13.91
      4-52.5531.988, 3.1170.288<0.00112.847.30, 22.58
      6-81.4761.162, 1.7890.160<0.0014.373.20, 6.00
      9-121.1370.858, 1.4140.142<0.0013.122.36, 4.12
      13-1501
      Blunt injury
      Yes0.6990.512, 0.8840.095<0.0012.011.67, 2.42
      No01
      Type of road user
      Pedestrian0.7800.263, 1.2960.2640.0032.181.30, 3.66
      4 or more wheels0.079−0.15, 0.3070.1170.501.080.86, 1.37
      Bicycle or motorcycle01
      Response time ≤ 8 minutes
      >80.1890.011, 0.3650.0900.041.211.01, 1.45
      ≤801
      Airway management
      Assisted ventilation1.2190.844, 1.5940.191<0.0013.382.33 4.93
      Open/clear airway0.6710.403, 0.9390.137<0.0011.961.50, 2.56
      No supplement01
      IV fluid administration
      YES1.2130.971, 1.4550.124<0.0013.372.64, 4.39
      No01
      CI: confidence interval; GCS: Glasgow coma scale; IV: intravenous; OR: odds ratio; P: p value; RR: respiratory rate; SBP: systolic blood pressure; SE: standard error.
      The risk scoring scheme was constructed, which ranged from -6.319 to 3.635 with a median of -4.723. The score was then categorized into <-4.282, -4.282 to -2.212, and >-2.212 for low, moderate, and high risk groups with the corresponding PPV of 0.3%, 4.5%, and 53.6%, respectively. LR + of moderate and high risk group were 3.2 and 7.7 compared to the rest group (see Table 3).
      A 1000-replication bootstrap yielded calibration and discrimination biases of -0.00127 (95% CI: -0.00162, 0.00092) and -0.00066 (95% CI: -0.00084, -0.00049) indicating good calibration and discrimination (see Supplemental Table 5).

       External validation

       Ubonratchathani

      There were 1104 subjects recruited from Ubonratchathani DC. Death rate was 8.3%, which was lower than in the development data. In addition, the distribution of predictors were different, with a higher proportion of SBP > 90 mmHg (90.6% vs 86.8%), RR 10–29 breaths/minute (92.7% vs 88%), GCS 13–15 (78.1% vs 69.2%), and airway management (59.1% vs 56.4%), whereas blunt injury (37.6% vs 62.3%) and IV fluid administration (42.1% vs 47.6%) were lower (see Supplemental Table 6). All predictors were significantly associated with death without change in the direction of association (see Supplemental Table 7).
      The risk prediction score was calculated using the equation from the derivation phase, with a median of -5.063 (range: -6.338, 3.521). HL-GOF test indicated good calibration (Chi-square = 3.45, df = 8, P =  0.90). The estimated O/E was 1.00 with a wide range IQR of 0.57, 1.01, see Table 2. However, calibration plot showed deviation from the perfect line (see Supplemental Figure 3). Therefore, recalibration of the intercept was performed and indicated M1 was better in calibration plot (see Supplemental Figure 4). The C statistic was 0.981 (95% CI: 0.971, 0.991), see Table 2. According to M0, the PPV of low, moderate and high risk groups were 0.1%, 3.2% and 48.9%, whereas those of M1 were 0.1%, 3.1% and 48.9%, respectively (see Table 3).

       Trang

      Among 993 subjects recruited from Trang DC, 6.3% of subjects died, which was lower than in the development data. Subjects were older (mean age 37 vs 35.5 years), had a higher SBP > 90 mmHg (92.9% vs 86.8%), RR 10–29 breaths/minute (93.8% vs 88%), and GCS 13–15 (80.7% vs 69.2%), but had a lower proportion with blunt injury (37.6% vs 62.3%), IV fluid administration (42.1% vs 47.6%), and airway management (54.8% vs 56.4%) (see Supplemental Table 6). All predictors were significantly associated with death, except age (see Supplemental Table 7). The median risk score was -4.931 (range: -6.319, 3.388). HL-GOF test indicated good calibration (Chi-square = 4.88, df = 8, P =  0.81). The estimated O/E was 1.00 with a wide range IQR of 0.89, 1.09. Calibration plot indicating deviation of predicted value from observed value (see Supplemental Figure 3). Therefore, recalibration of the intercept was performed and indicated M1 was better in calibration plot (see Supplemental Figure 5). The C statistic was 0.947 (95% CI: 0.922, 0.973), see Table 2. According to M0, the PPV of low, moderate and high risk groups were 0.4%, 5.2% and 38.8%, whereas those of M1 were 0.6%, 7.3% and 48.1%, respectively (see Table 3).

       Prehospital prediction of SI

      The NISS cut-off threshold was calibrated by dividing into minor (1–3), moderate (4–8), serious (9–15), severe (16–24), and critical (25–75) injury and their diagnostic performances were described, see Supplemental Table 8.

       Development phase

      Of 20 potential predictors, 9 were kept in the final model of SI (NISS ≥ 16 vs <16) (see Table 4) and prediction score was calculated following the equation in Supplemental Figure 6. The C statistic was 0.913 (95% CI: 0.905, 0.922), indicating excellent discrimination. The calibration plot showed no deviation from the reference line (see Supplemental Figure 6 and 7) with a corresponding HL-GOF test (Chi-square = 13.8, df = 8, P =  0.09), O/E of 0.99 (IQR: 0.95, 1.05), see Table 2 and Supplemental Table 4.
      The median risk score was 2.296 (range: -3.933, 3.488). This was then categorized into low, moderate, and high risk groups and their diagnostic accuracy was then estimated, see Table 3. The PPV for these corresponding groups were 4.7%, 29.1%, and 74.5%, respectively.
      A 1000-replication bootstrap yielded calibration and discrimination biases of 0.0009 (95% CI: -0.0239, 0.0281) and 0.0004 (95% CI: -0.0005, 0.00002), respectively (see Supplemental Table 9).

       External validation

       Ubonratchathani

      Compared with derivation data, the prevalence of SI was lower (16.9% vs 27.5%). The distribution of predictors were compared by SI groups (see Supplemental Table 10) and all, except age > 55 years and response time ≤ 8 min, were significantly associated with SI (see Supplemental Table 11).
      The calibration plot showed deviation from a perfect-fitted line (see Supplemental Figure 8), corresponding to HL-GOF (Chi-square = 28.7, df = 8, P <  .001), O/E ratio (O/E 1.00; IQR: 0.71, 1.03), see Table 2. Recalibration of the intercept was performed and indicated M1 fitted well with the data (Chi-square = 7.2, df = 8, P <  .51) with O/E ratio 1.00 (95% CI: 0.95, 1.05), and an improved calibration test/plot, see Supplemental Figure 9. According to M0, the PPV of low, moderate and high risk groups were 3.3%, 16.6% and 62.1%, whereas those of M1 were 3.9%, 24.4% and 70.6%, respectively (see Table 3).

       Trang

      Compared with development data, the prevalence of SI was lower (17.7% vs 27.5%) and all, except type of road user and response time of ≤ 8 min, were significantly associated with SI (see Supplemental Table 10 and 11). Although discrimination was good (C statistic = 0.896; 95% CI: 0.871, 0.922, see Table 2), calibration plot showed poor fit (see Supplemental Figure 8) corresponding to HL-GOF (Chi-square = 21, df = 8, P =  0.007) with O/E 0.99 (IQR: 0.78, 1.04). Recalibration of the intercept indicated M1 had a better O/E ratio (1.00; 95% CI: 0.94, 1.04) and fitted well in Trang (Chi-square = 7.5, df = 6, P =  0.28). In addition, there was improvement of the calibration plot (see Supplemental Figure 10). According to M0, the PPV of low, moderate and high risk groups were 3.6%, 22.6% and 68.1%, whereas those of M1 were 4.8%, 25.7% and 69.5%, respectively (see Table 3).

      Discussion

      We have derived and validated risk prediction scores for death and SI for RTI patients treated by ALS response units. The scores allow subjects to be classified into low, moderate and high risks of SI and death during prehospital operations and may lead to improved allocation of patients to an appropriate hospital. The results yielded 10 predictors of death and SI including age, blunt injury, RR, SBP, GCS, incidence time, type of road users, response time of ≤ 8 min, prehospital airway management, and IV fluid administration. The risk score performed fairly in validation.

       Cut off point of NISS

      Our results found that NISS and ISS were not much different for predicting early death (i.e., 25 vs 20), which was similar to previous studies [
      • Mica L.
      • Rufibach K.
      • Keel M.
      • Trentz O.
      The risk of early mortality of polytrauma patients associated to ISS, NISS, APACHE II values and prothrombin time.
      ], but the NISS performed better when death within 30 days was considered [
      • Lavoie A.
      • Moore L.
      • LeSage N.
      • Liberman M.
      • Sampalis J.S.
      The New Injury Severity Score: a more accurate predictor of in-hospital mortality than the Injury Severity Score.
      ,
      • Roy N.
      • Gerdin M.
      • Schneider E.
      • Kizhakke Veetil D.K.
      • Khajanchi M.
      • Kumar V.
      • et al.
      Validation of international trauma scoring systems in urban trauma centres in India.
      ]. This might be explained by the fact that subjects who died within 24–48 h were more likely to have major injuries distributed in different body regions and thus experience multi-organ system failure sooner.

       Predictors, model performance and comparison to previous risk prediction scores

      Prevalences of SI and death in this study were higher than previous reports [
      • Kashani A.T.
      • Mohaymany A.S.
      Analysis of the traffic injury severity on two-lane, two-way rural roads based on classification tree models.
      ,
      • Ayoung-Chee P.
      • Mack C.D.
      • Kaufman R.
      • Bulger E.
      Predicting severe injury using vehicle telemetry data.
      ,
      • Scheetz L.J.
      Trends in the accuracy of older person trauma triage from 2004 to 2008.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.E.
      Using crash scene variables to predict the need for trauma center care in older persons.
      ]. These might be due to the fact that we considered only patients being treated by an ALS, which might bias to vector to moderate to severe victims according to the DC protocol. In addition, previous studies [
      • Ayoung-Chee P.
      • Mack C.D.
      • Kaufman R.
      • Bulger E.
      Predicting severe injury using vehicle telemetry data.
      ,
      • Scheetz L.J.
      Trends in the accuracy of older person trauma triage from 2004 to 2008.
      ,
      • Blomberg H.
      • Svennblad B.
      • Michaelsson K.
      • Byberg L.
      • Johansson J.
      • Gedeborg R.
      Prehospital trauma life support training of ambulance caregivers and the outcomes of traffic-injury victims in Sweden.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.E.
      Using crash scene variables to predict the need for trauma center care in older persons.
      ] were conducted in developed countries where road quality and traffic rules compliance are better than in Thailand.
      Our findings indicated that RTIs commonly occurred in male, middle age, and motorcyclists/bicyclists, which was similar to previous global reports [
      Global status report on road safety 2013.
      ,
      • Peden M.
      • Scurfield R.
      • Sleet D.
      • Mohan D.
      • Hyder A.A.
      • Jarawan E.
      • et al.
      World report on road traffic injury prevention.
      ]. Type of road user was also a significant risk factor for both SI and death in RTI, but it was not considered in previous models. Three EMS operations (i.e., response time, prehospital IV administration, and airway management) were also significant predictors. Response time > 8 min was positively associated with SI, i.e, longer waiting time led to progression of injury such as expansion of intracranial hematoma, or tension pneumothorax. Receiving IV fluids, and airway management, due to poor conditions at scene, (i.e., lower RR, SBP, and GCS), were also significantly associated with increased odds of death and SI.
      Discrimination performance of previous risk prediction scores varied because some included only physical examination [
      • Champion H.R.
      • Copes W.S.
      • Sacco W.J.
      • Lawnick M.M.
      • Keast S.L.
      • Bain Jr., L.W.
      • et al.
      The Major Trauma Outcome Study: establishing national norms for trauma care.
      ,
      • Offner P.J.
      • Jurkovich G.J.
      • Gurney J.
      • Rivara F.P.
      Revision of TRISS for intubated patients.
      ,
      • Champion H.R.
      • Sacco W.J.
      • Copes W.S.
      • Gann D.S.
      • Gennarelli T.A.
      • Flanagan M.E.
      A revision of the trauma score.
      ,
      • Stewart T.C.
      • Lane P.L.
      • Stefanits T.
      An evaluation of patient outcomes before and after trauma center designation using trauma and injury severity score analysis.
      ,
      • Scheetz L.J.
      Trends in the accuracy of older person trauma triage from 2004 to 2008.
      ] whereas some others included only crash characteristics [
      • Newgard C.D.
      • Lewis R.J.
      • Jolly B.T.
      Use of out-of-hospital variables to predict severity of injury in pediatric patients involved in motor vehicle crashes.
      ,
      • Newgard C.D.
      • Hui S.H.J.
      • Griffin A.
      • Wuerstle M.
      • Pratt F.
      • Lewis R.J.
      Prospective validation of an out-of-hospital decision rule to identify seriously injured children involved in motor vehicle crashes.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.E.
      Using crash scene variables to predict the need for trauma center care in older persons.
      ,
      • Scheetz L.J.
      • Zhang J.
      • Kolassa J.
      Classification tree modeling to identify severe and moderate vehicular injuries in young and middle-aged adults.
      ]. However, this research simultaneously considered age, mechanism of injury, physical examination, crash characteristics, and additional EMS operations in the model. Including these important variables enabled the models to reach excellent discrimination for predicting death and SI [
      • Moons K.G.
      • Kengne A.P.
      • Woodward M.
      • Royston P.
      • Vergouwe Y.
      • Altman D.G.
      • et al.
      Risk prediction models: I. Development, internal validation, and assessing the incremental value of a new (bio)marker.
      ]. Although our risk score only calibrated fairly with validation data, where death and SI were lower, recalibrating the intercept (M1) improved calibration.
      Our risk prediction scores should be useful for developing prehospital care in Thailand and other countries where numbers of TCs are limited. Applying our risk prediction scores is simple and straight forward. For instance, one traffic accident occurred on the highway road and was reported to DC at 10.00 a.m.. The ALS response unit arrived at scene 7 min later after being notified, finding two victims, i.e., male motorcyclist aged 35 years and female pedestrian aged 25 years. The first examination of motorcyclist reveals RR 8 breaths/minute, SBP 80 mmHg and GCS 8, blunt contusion on his right flank. He is urgently intubated, open venous with saline solution. Whereas, evaluation of female pedestrian reveals RR 20 breaths/minute, SBP 110 mmHg, and GCS 15, small abrasion on her shoulder and no need for further immediate treatment. The risk scores for death and SI of male motorcyclist are -0.286 and 1.364 indicating moderate risk for death and high risk for SI. As a result, the PPVs of this motorcyclist are 4.5% and 74.5% for death and SI. The risk scores of female pedestrian for death and SI are -5.162 and -2.455 with PPVs of 0.3% and 4.7%, indicating low risk of both death and SI. After receiving treatment management at scene, the motorcyclist should be prioritized to directly transport to a TC nearby, whereas the pedestrian should be initially sent to the nearest lower facilities hospital because of low risk.
      Our study has a number of strengths. We complied with recommendations for the development of clinical prediction rules with adequate number of subjects. [
      • Guyatt G.
      Determining prognosis and creating clinical decision rules.
      ] The model has been internally and externally validated with good discrimination and calibration after recalibration of the model [
      • Toll D.B.
      • Janssen K.J.
      • Vergouwe Y.
      • Moons K.G.
      Validation, updating and impact of clinical prediction rules: a review.
      ]. Predictors in the models are simple to measure and available in a routine prehospital practice. We recruited RTI subjects from various regions across the country, which reflected a wide coverage of RTI subjects. Our risk prediction models might be more readily applicable in the form of a mobile application given the widespread availability of mobile devices. Its impact on real practice should be further evaluated.

       Limitations

      There were some limitations in our study. Although data collection was standardized, prospectively corrected by well-trained personnel, and closely monitored; some missing data were unavoidable, in which 4 variables were missing which ranged from 0.17% to 12.36%. Therefore, we applied MI with chain equations to impute those missing values [
      • Moons K.G.
      • Donders R.A.
      • Stijnen T.
      • Harrell Jr., F.E.
      Using the outcome for imputation of missing predictor values was preferred.
      ].
      Apart from the concern of missing data, studied DCs were purposively selected based on availability of EP and a well-developed EMS information system. Results of our study might be less applicable to the DCs where their information system and service are less well organized and also patients were less severely injured. Thus, selection bias might be present. However, we attempted to include a representative sample of RTI subjects for the whole country by selection of subjects stratified by region (i.e., North, Northeast, East, Middle, and South). In addition, numbers of subjects for each center were proportional to the size of their RTI population treated by ALS unit/year. Given our strict inclusion criteria, applying our risk prediction scores may be limited in some settings that have following characteristics: EMS team leader consists of nurse/paramedic and/or EP; RTI subjects are not irreversible death; prehospital managements (e.g. assisting ventilation or IV fluid administration) should be standardized; death rate and severity of RTI subjects are similar to our setting, otherwise, external validation should be performed before applying.

      Conclusions

      In summary, our study has provided prehospital risk prediction scores of death and SI for RTI subjects. The models have fair calibration and excellent discrimination in development and internal validation. The risk score was categorized into low, moderate, and high risk groups. Threshold probabilities of 0.05 and 0.1 were suggested to treat subjects. Although, the model fit with external data was only fair, recalibration of original intercepts indicated improvement of calibration without changes of the discrimination power.

      Conflict of interest statement

      The authors have no conflict of interest.

      Acknowledgement

      This study was supported by Mahidol University, and National Institute for Emergency Medicine, Ministry of Public Health, Thailand . They had no involvement with study’s activities.

      Appendix A. Supplementary data

      The following is Supplementary data to this article:

      References

        • Key C.B.
        Vehicle related injuries.
        in: Krohmer J.R. Sahni R. Schwartz B. Wang H.E. Clinical aspects of prehospital medicine. the United States of America: National Association of EMS Physicians, 2009: 81-89
        • Peden M.
        • Scurfield R.
        • Sleet D.
        • Mohan D.
        • Hyder A.A.
        • Jarawan E.
        • et al.
        World report on road traffic injury prevention.
        World Health Organization, 2004 (Available at:http://whqlibdoc.who.int/publications/2004/9241562609.pdf(Accessed 25 February))
        • Chadbunchachai W.
        • Suphanchaimaj W.
        • Settasatien A.
        • Jinwong T.
        Road traffic injuries in Thailand: current situation.
        J Med Assoc Thai. 2012; 95: S274-81
        • Toroyan T.
        • Peden M.M.
        • Iaych K.
        WHO launches second global status report on road safety.
        Inj Prev. 2013; 19: 150
        • Blomberg H.
        • Svennblad B.
        • Michaelsson K.
        • Byberg L.
        • Johansson J.
        • Gedeborg R.
        Prehospital trauma life support training of ambulance caregivers and the outcomes of traffic-injury victims in Sweden.
        J Am Coll Surg. 2013; 217: 1010-1019
        • Haas B.
        • Stukel T.A.
        • Gomez D.
        • Zagorski B.
        • De Mestral C.
        • Sharma S.V.
        • et al.
        The mortality benefit of direct trauma center transport in a regional trauma system: a population-based analysis.
        J Trauma Acute Care Surg. 2012; 72: 1510-1517
        • Sasser S.M.
        • Hunt R.C.
        • Faul M.
        • Sugerman D.
        • Pearson W.S.
        • Dulski T.
        • et al.
        Guidelines for field triage of injured patients: recommendations of the National Expert Panel on Field Triage, 2011.
        MMWR Recomm Rep. 2012; 61: 1-20
        • Champion H.R.
        • Copes W.S.
        • Sacco W.J.
        • Lawnick M.M.
        • Keast S.L.
        • Bain Jr., L.W.
        • et al.
        The Major Trauma Outcome Study: establishing national norms for trauma care.
        J Trauma. 1990; 30: 1356-1365
        • Offner P.J.
        • Jurkovich G.J.
        • Gurney J.
        • Rivara F.P.
        Revision of TRISS for intubated patients.
        J Trauma. 1992; 32: 32-35
        • Champion H.R.
        • Sacco W.J.
        • Copes W.S.
        • Gann D.S.
        • Gennarelli T.A.
        • Flanagan M.E.
        A revision of the trauma score.
        J Trauma. 1989; 29: 623-629
        • Scheetz L.J.
        Trends in the accuracy of older person trauma triage from 2004 to 2008.
        Prehosp Emerg Care. 2011; 15: 83-87
        • Newgard C.D.
        • Lewis R.J.
        • Jolly B.T.
        Use of out-of-hospital variables to predict severity of injury in pediatric patients involved in motor vehicle crashes.
        Ann Emerg Med. 2002; 39: 481-491
        • Newgard C.D.
        • Hui S.H.J.
        • Griffin A.
        • Wuerstle M.
        • Pratt F.
        • Lewis R.J.
        Prospective validation of an out-of-hospital decision rule to identify seriously injured children involved in motor vehicle crashes.
        Acad Emerg Med. 2005; 12: 679-687
        • Scheetz L.J.
        • Zhang J.
        • Kolassa J.E.
        Using crash scene variables to predict the need for trauma center care in older persons.
        Res Nurs Health. 2007; 30: 399-412
        • Scheetz L.J.
        • Zhang J.
        • Kolassa J.
        Classification tree modeling to identify severe and moderate vehicular injuries in young and middle-aged adults.
        Artif Intell Med. 2009; 45: 1-10
        • Moons K.G.
        • Kengne A.P.
        • Grobbee D.E.
        • Royston P.
        • Vergouwe Y.
        • Altman D.G.
        • et al.
        Risk prediction models: II. External validation, model updating, and impact assessment.
        Heart. 2012; 98: 691-698
        • Moons K.G.
        • Kengne A.P.
        • Woodward M.
        • Royston P.
        • Vergouwe Y.
        • Altman D.G.
        • et al.
        Risk prediction models: I. Development, internal validation, and assessing the incremental value of a new (bio)marker.
        Heart. 2012; 98: 683-690
        • Moons K.G.
        • Altman D.G.
        • Reitsma J.B.
        • Ioannidis J.P.
        • Macaskill P.
        • Steyerberg E.W.
        • et al.
        Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration.
        Ann Intern Med. 2015; 162: W1-73
        • Kashani A.T.
        • Mohaymany A.S.
        Analysis of the traffic injury severity on two-lane, two-way rural roads based on classification tree models.
        Saf Sci. 2011; 49: 1314-1320
        • Ayoung-Chee P.
        • Mack C.D.
        • Kaufman R.
        • Bulger E.
        Predicting severe injury using vehicle telemetry data.
        J Trauma Acute Care Surg. 2013; 74: 190-194
        • Gennarelli T.A.
        • Wodzin E.
        Abbreviated injury scale 2005 (update 2008): association for the Advancement of Automotive Medicine (AAAM).
        2008
        • Moons K.G.
        • Donders R.A.
        • Stijnen T.
        • Harrell Jr., F.E.
        Using the outcome for imputation of missing predictor values was preferred.
        J Clin Epidemiol. 2006; 59: 1092-1101
        • David W.
        • Hosmer J.
        • Lemeshow Stanley
        • Sturdivant R.X.
        Assessing the fit of the model.
        in: David W. Hosmer J. Lemeshow Stanley Sturdivant R.X. Applied logistic regression. the United States of America: John Wiley & Sons, Inc., 2013: 153-226
        • Paul P.
        • Pennell M.L.
        • Lemeshow S.
        Standardizing the power of the Hosmer-Lemeshow goodness of fit test in large data sets.
        Stat Med. 2013; 32 (Epub 2012Jul26): 67-80https://doi.org/10.1002/sim.5525
        • Harrell Jr., F.E.
        • Lee K.L.
        • Mark D.B.
        Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors.
        Stat Med. 1996; 15: 361-387
        • Toll D.B.
        • Janssen K.J.
        • Vergouwe Y.
        • Moons K.G.
        Validation, updating and impact of clinical prediction rules: a review.
        J Clin Epidemiol. 2008; 61: 1085-1094
        • Mica L.
        • Rufibach K.
        • Keel M.
        • Trentz O.
        The risk of early mortality of polytrauma patients associated to ISS, NISS, APACHE II values and prothrombin time.
        J Trauma Manag Outcomes. 2013; 7: 6
        • Lavoie A.
        • Moore L.
        • LeSage N.
        • Liberman M.
        • Sampalis J.S.
        The New Injury Severity Score: a more accurate predictor of in-hospital mortality than the Injury Severity Score.
        J Trauma. 2004; 56: 1312-1320
        • Roy N.
        • Gerdin M.
        • Schneider E.
        • Kizhakke Veetil D.K.
        • Khajanchi M.
        • Kumar V.
        • et al.
        Validation of international trauma scoring systems in urban trauma centres in India.
        Injury. 2016; 47: 2459-2464
      1. Global status report on road safety 2013.
        World Health Organization, 2013 (Available at:http://www.who.int/violence_injury_prevention/road_safety_status/2013/en/(Accessed 22 February))
        • Stewart T.C.
        • Lane P.L.
        • Stefanits T.
        An evaluation of patient outcomes before and after trauma center designation using trauma and injury severity score analysis.
        J Trauma. 1995; 39: 1036-1040
        • Guyatt G.
        Determining prognosis and creating clinical decision rules.
        in: Brian Haynes R. Sackett David L. Guatt Gordon H. Tugwell P. Clinical epidemiology: how to do clinical practice research. United States: Lippincott Williums & Wilkins, 2006: 323-356