We developed a simple score (Bleeding Audit and Triage Trauma Score - BATT) to predict death due to bleeding in trauma patients. We conducted an external validation of this score using data from the UK Trauma Audit Research Network (TARN) from 1st January 2017 to 31st December 2018. Finally, we evaluated the impact of TXA treatment thresholds in trauma patients.
Development of the BATT score
We previously developed and validated a prognostic model to predict death due to bleeding in trauma patients. The methods are described in detail elsewhere.[5] Briefly, data on bleeding trauma patients from 298 hospitals in 41 countries were used to derive the model. We validated the model using an internal–external cross-validation method based on data from 41 countries to ensure that the results are widely applicable. The final prognostic model included age, systolic blood pressure, Glasgow Coma Scale, heart rate, respiratory rate and mechanism of injury. To develop the BATT score, we assigned points for each predictor that were proportional to the coefficients of the regression equation. We added the criterion high velocity trauma as the intercept of the regression equation corresponding to the inclusion criteria of the trauma registry used for the development of prognostic model. High velocity trauma is routinely assessed at the scene and corresponds to injury from road traffic crash (with intrusion, ejection, death in same passenger compartment, and motor vehicle versus pedestrian or bicyclist), fall from high height (> 3 meters), blow or blast.[7] An electronic version of the score is available for computer or smartphone: https://www.evidencio.com/models/show/1393
Validation of the BATT score
We used data from the Trauma Audit Research Network (TARN) from 1st January 2017 to 31st December 2018 to validate the BATT score for use in England and Wales. The TARN database includes data on patients with an Injury Severity Score (ISS) of nine or more who are admitted to hospital in England and Wales for at least three nights, died in hospital or were transferred to another hospital for specialist care.[8] The exclusion criteria were isolated mild traumatic brain injury with loss of consciousness, superficial scalp injury , patients 65 years or older with femoral neck or single pubic rami fracture, fracture or dislocation of the foot or hand, closed fracture or dislocation of an isolated limb, simple skin laceration with blood loss < 20%.
Because death due to bleeding is not recorded in the TARN database, we used early deaths and early deaths with evidence of haemorrhage as a proxy for death due to bleeding.[9] Specifically, we included deaths from all cause within 12 hours of injury (excluding asphyxia, drowning, hanging, or massive destruction of skull or brain) and deaths between 12 to 24 hours with evidence of bleeding (activation of massive transfusion protocol or blood within 6 hours or an abbreviated injury scale (AIS) diagnosis associated with haemorrhage supplementary file 1).
We assessed the accuracy, discrimination and calibration of the BATT score. Accuracy was assessed using the Brier score. Because the Brier score depends on the prevalence of the outcome, we also calculated the scaled Brier score to account for the baseline risk of death due to bleeding (supplementary file 2). The scaled Brier score ranges from 0% to 100% and indicates the degree of error in prediction.[10] A scaled Brier score of 0% shows perfect accuracy. Discrimination is the ability of the score to correctly identify patients with the outcome. We estimated the sensitivity, specificity, positive and negative likelihood ratio for each threshold of the BATT score. The likelihood ratio is the likelihood of a positive score in a patient with the outcome compared to the likelihood of a positive score in a patient without the outcome.[11] The positive likelihood ratio is the ratio of sensitivity to 1-specificity. The negative likelihood ratio is the ratio of 1-sensitivity to specificity. A positive likelihood ratio of 10 or above will result in a large increase in the probability of the outcome. A negative likelihood ratio of 0.1 or less will result in a large decrease in the probability of the outcome. We plotted the Receiving Operating Characteristic (ROC) curve which is the sensitivity (true positives) on 1-specificity (false positives) for different threshold of the BATT score.[12] An ideal score will reach the upper left corner (all true positive with no false positive). We estimated the area under the ROC curve (AUROC) that corresponds to the concordance statistic (C-Statistic) for binary outcome. A C-statistic of 1.0 shows perfect discrimination ability. Calibration is the agreement between observed and predicted outcomes. We estimated calibration in the large as the difference between the mean predicted and observed probabilities and the ratio of the predicted and observed number of events (P/O). We also plotted the observed and predicted probabilities of death by decile of the score and with local regression based on LOESS algorithm.[10] We estimated the calibration intercept and slope of the calibration plot as a measure of spread between predicted and observed outcome. Ideally, the intercept would be zero indicating that the predictions are neither systematically too low or too high and the slope would be 1.[13] There were missing value for some predictors but no missing outcome data. To estimate baseline risk for the full dataset, we replaced missing predictors using multiple imputation by chained equations on early death, age, systolic blood pressure, respiratory rate, heart rate, Glasgow coma scale, time for injury, time for prehospital ambulance arrival, and time for hospital admission with 20 imputed datasets.
Evaluation of TXA treatment criteria
We evaluated two different TXA treatment strategies: (1) prehospital treatment of all trauma patients with an ISS ≥9 at the scene of the injury, (2) hospital treatment of all trauma patients with an ISS>9 in the emergency department (ED). We compared each treatment strategy according to different thresholds of the BATT score to assess its clinical usefulness and treatment criteria.
We estimated the impact of TXA treatment for each treatment criteria. Since randomized trials of TXA in trauma patients report no increase in deaths from adverse events, the net impact of TXA was given by the number of deaths due to bleeding avoided by the treatment.[6,14] To estimate the number of deaths avoided by TXA, we predicted the baseline risk of death due to bleeding using our previously published prognostic model.[5] To estimate post-treatment probabilities, we applied the treatment effect to these baseline risks taking into account time to treatment.[4] The risk difference was used to estimate the number of deaths avoided. To account for miscalibration of predicted baseline risks, we conducted a sensitivity analysis using observed early deaths with evidence of haemorrhage as baseline risks. The details of both modelling methods and equations are described in the supplementary file 3. We plotted the cumulative number of death due to bleeding avoided by BATT score threshold in a decision curve analysis as described by Vickers et al.[15] We compared decision curve analysis for each scenario. We estimated the number needed to treat to save one life for each BATT score threshold and each scenario. The registry-based study design predetermines the sample size. All analyses were performed using STATA software (version 16.0; Stata Corp, College Station, TX, USA).