Predicting on-going hemorrhage and transfusion requirement after severe trauma: a validation of six scoring systems and algorithms on the TraumaRegister DGU^®

Trauma is the leading cause of death world-wide in persons under the age of 40 years and accounts for approximately 10% of all deaths in general [1, 2]. Almost 50% of patients with lethal outcome within the first 48 hours after trauma die from uncontrolled exsanguinating hemorrhage [2‐4]. On average, one in four trauma patients arriving in the trauma bay is already coagulopathic upon admission and about one third of transfused patients require massive transfusion [5, 6]. Meanwhile, it has been shown that the early activation of massive transfusion (MT) protocols including balanced ratios of blood products and the early substitution of coagulation factors may improve survival [7‐9]. However, the timely identification of patients in need for aggressive hemostatic resuscitation remains challenging.

Over the past years, a lot of effort has been put into identifying parameters to analyze and measure pathophysiological pathways occurring in the setting of the acutely bleeding trauma patient and to mirror the coagulation state. However, parameters like the international normalized ratio (INR) or the activated partial thromboplastin time (aPPT) by themselves have been assessed with only bounded significance [10, 11]. Recent publications have shown that ROTEM^® testing might be superior to analyze coagulopathic states in the early phase after trauma [12‐14]. On the clinical level, several scoring systems and algorithms have been suggested for the early stratification of trauma patients at risk for massive transfusion. These systems have been developed on the basis of retrospective datasets from trauma patients derived from different trauma databases and registries. Usually developed and validated on one registry only using a split-data approach, their validity when subjected to external validation has only been assessed to a limited extent.

Several scoring systems and algorithms to predict the risk for MT have been introduced [15‐22]. In principle, these systems have been developed and validated using retrospective datasets from civilian and combat trauma patients. In the present work, our aim was to compare six frequently used scoring systems and algorithms with the potential to early identify trauma patients at risk for massive transfusion and to validate all six scores on one dataset, including severely injured trauma patients derived from the TraumaRegister DGU^® (TR-DGU, Trauma registry of the German Trauma Society).

In the present study, we queried MEDLINE (1969-July 2011) and The Medical Algorithms Project for scoring systems and algorithms with the potential to identify patients at risk for MT at a very early stage after trauma [23]. Only scores including parameters that could be reproduced by the data captured in the TraumaRegister DGU^® were considered. MT was defined by administration of ≥ 10 units of packed red blood cells (pRBC) between arrival at the emergency room (ER) and the intensive care unit (ICU). Patients who had received hemostatic agents such as fibrinogen, prothrombin complex concentrate (PCC), recombinant activated factor VII (rFVIIa) or any antifibrinolytics with potential influence on the amount of administered pRBCs were excluded from the study. Furthermore, the TraumaRegister DGU^® captures information on additional hemostatics dichotomously, without capturing information about the amount or the time of administration.

A total of 5,147 (9%) patients were extracted from a complete dataset of 56,573 patients derived from the TraumaRegister DGU^®. The mean age of patients was 45.7 ± 19.3 years with a mean ISS of 24.3 ± 13.2. The overall MT rate was 5.6% (n = 289), and 95% patients (n = 4,889) had sustained blunt trauma. The mean time from emergency department (E) admission to ED discharge for the patients analyzed in this cohort was 194.4 ± 146.3 minutes. Table 2 provides an overview of demographics of the study cohort, the baseline physiological data and details of transfused blood products.

Figure 1 depicts the AUC for all six scores and algorithms assessed in this study. The TASH score had the highest overall accuracy as reflected by an AUC of 0.889, followed by the PWH score (0.860) and the score developed by Vandromme and colleagues (0.840). The ABC score of Nunez and colleagues performed less accurately than all other scores as reflected by an AUC of 0.763. The performance of each score is summarized in Table 3.

Table 3 shows the sensitivity, specificity, positive predictive values (PPV) and negative predictive values (NPV) at defined cut-off values for each score based upon the best relationship between sensitivity and specificity. When considering these cut-offs the highest sensitivity was observed for the Schreiber-Score (85.8%) but also the lowest specificity (61.7%). The TASH score, at the defined cut-off ≥ 8.5, showed a sensitivity of 84.4% and also a high specificity (78.4%). The PWH score had a lower sensitivity (80.6%) with comparable specificity. The Larson score showed the lowest sensitivity (70.9%) at a specificity of 80.4%. These results mirror the results from the AUC calculations in that the TASH score, was generally statistically superior, followed by the PWH scores, with the a statistically significant difference between these two best-performing scores (P = 0.0413).

The early aggressive management of the acute coagulopathy of trauma via activation of massive transfusion (MT) protocols including balanced ratios of blood products and the early substitution of coagulation factors may improve survival in the trauma population [7‐9]. However, the timely identification of lethal exsanguination remains challenging. This study validated six scoring systems and algorithms to risk-stratify patients for MT at a very early stage after trauma, using a single dataset of severely injured patients derived from the civilian TraumaRegister DGU^® database. This dataset included data from primary patients with blunt trauma (95%) of relevant magnitude (ISS > 9) and with the need for intensive care.

Our results show, that the performance of the TASH score was superior with an AUC of 0.889, followed by the PWH score (AUC 0.860). It is unsurprising that the PWH score performed with similar accuracy, as to some extent it incorporates almost identical components and also applies weights to selected individual parameters, albeit following a different pattern to the TASH score. All other scores with fewer parameters and absence of weighting performed less accurately. The score published by Vandromme and colleagues (AUC 0.840) performs almost as well as the PWH score by only using five parameters, from which four are available on clinical investigation (measurement of heart rate and systolic blood pressure) or using point-of-care devices (for measurement of haemoglobin and lactate). It is a not weighted score, which makes it easy to use. The inferior performance of both military scores when applied to our dataset calls into question their applicability for use in the civilian trauma setting, in which the majority of patients suffer blunt trauma. In our dataset the rate of blunt trauma was 95%.

The TASH score has previously been developed and validated in the TraumaRegister DGU^® database, and has now been internally revalidated using a different patient cohort. Obviously, this gives the TASH score a somewhat unfair advantage over other scores and may bias the results. In contrast, the other five scores and algorithms have been developed and validated in other databases and were thus externally validated by the present study. The two military scores, Schreiber and Larson scores, that had initially been developed using data derived from combat databases comprising mostly penetrating trauma, were here externally validated on a civilian dataset who mostly suffered blunt trauma. The PWH score, initially developed and validated in the setting of the New Territories of Hong Kong where 95% of the population are Chinese, was validated here on a Caucasian population.

In the past, a few other studies have used a similar approach in externally validating selected scores by their application in other databases, but comparing a maximum of only three scores at a time and including only limited numbers of patients. Here we report the results for six of the most commonly used and cited scores in the literature applied to a total of 5,147 patients from a single trauma database.

Several authors have compared their own scoring system to other existing scores and algorithms [18, 30]. For example, Nunez and co-workers have developed and validated their ABC score on their local database and compared its performance to TASH and the Mc-Laughlin score when applied to the same dataset. The differences between the three scores were not statistically significant [18]. In an external validation step, Mitra and colleagues applied three different scores, including the PWH, TASH, and the ABC score, in 1,234 patients derived from The Alfred Trauma Registry, a single center Australian trauma database [17]. The AUC for the PWH score was significantly less than that for the TASH score and was significantly greater than for the ABC score. In another approach, Cotton and co-workers validated the ABC score across multiple demographically diverse level I trauma centers and reported a sensitivity and specificity for the ABC score predicting MT from 75% to 90% and 67% to 88%, respectively [30]. In our study, the sensitivity and the specificity of this score was only 76% and 70%, respectively.

For the present analysis, only scoring systems and algorithms predicting MT with requirement for blood products that could be reproduced by the given TraumaRegister DGU^® data extract were considered. Some scores were intentionally excluded from the present analysis due to the data not being captured in the TraumaRegister DGU^® to reproduce these scores. These scores and algorithms were, for example, the McLaughlin score, the coagulopathy of severe trauma (COAST) score and the Baker score. McLaughlin published a predictive model for MT based on data from combat casualties [16]. Mitra and co-workers developed the COAST score to accurately identify patients with acute traumatic coagulopathy (ATC) [17], while Baker and colleagues suggested an algorithm to identify risk factors for blood transfusion including four risk factors such as systolic blood pressure, GCS, heart rate and high-risk injury [15]. In the latter case, for example, we were not able to reproduce the high-risk injury based on the data available from the TraumaRegister DGU^®. The prognostic emergency trauma score (EMTRAS), which has also been developed from data derived from the TraumaRegister DGU^® and which may also work for MT, was left out due to mortality as its endpoint [31].

With regard to the clinical applicability of a predictive score in the ED, timely availability of the required parameters is of fundamental importance. Time consuming laboratory analysis or radiographic examination may be good predictors for the need of MT but are not available within appropriate time frames. The TASH score has been proven to be calculated within 10 minutes after the patient's arrival in the ED. This score, however, also relies on laboratory diagnostics for hemoglobin and base excess (BE) but not on global coagulation tests. Most time-consuming assays usually include the assessment of coagulation parameters. Davenport, for example, reported turn-around times for prothrombin time at a median of 78 (62 to 103) minutes [14]. Vice-versa, point-of-care devices as alternatives to conventional laboratory assays, may be inaccurate as point-of-care prothrombin time-ratio has poor agreement with laboratory prothrombin time-ratio in patients with acute traumatic coagulopathy, with 29% false-negative results [14].

In order to quickly assess the coagulation status of a bleeding patient in the ED, rotational thromboelastometry has been advocated to identify acute traumatic coagulopathy at 5 minutes and to predict the need for MT. Davenport and colleagues used rotational thromboelastometry to identify an appropriate diagnostic tool for the early diagnosis of acute traumatic coagulopathy and validated this modality through prediction of transfusion requirements in traumatic hemorrhage. They found that in acute traumatic coagulopathy, the rotational thromboelastometry clot amplitude at 5 minutes was diminished by 42%, and this persisted throughout clot maturation. Finally they pointed out that the clot amplitude at 5 minutes could identify patients who would require MT (detection rate of 71%, vs. 43% for prothrombin time ratio > 1.2, P < 0.001) [14]. Schöchl and colleagues used whole-blood thromboelastometry (ROTEM^®) tests to provide immediate information about the coagulation status of patients with acute traumatic bleeding [32]. They found the best predictive values for MT were provided by hemoglobin and the Quick value (AUC 0.87 for both parameters). However, they also they found similarly high predictive values for FIBTEM MCF (AUC 0.84) and FIBTEM A10 (clot amplitude at 10 minutes, AUC 0.83). They conclude that FIBTEM A10 and FIBTEM MCF provide similar predictive values for MT in trauma patients to the most predictive laboratory parameters [32]. To date, a score using rotational thromboelastometry as a predictor for MT is unavailable, although in some specialist centers thromboelastometry is an instrument that is already used for the prediction of MT in the ED to detect the need for MT and for early application of hemostatic agents.

Cotton and co-workers presented their results of a prospective pilot study (n = 272) comparing conventional coagulation testing (CCT) with rapid thrombelastography (r-TEG) showing that early r-TEG values (ACT, k-time, and r-value) were available within 5 minutes, late r-TEG values (maximal amplitude and angle) within 15 minutes, and CCTs within 48 minutes (P < 0.001). They conclude that r-TEG results correlate with the CCTs that are not as rapidly available, and are predictive of early transfusions of pRBC, plasma, and platelets [33]. However, this advanced technology may not be available in most trauma centers and therefore easy-to-use and quickly applicable strategies for early assessment and risk stratification, for example via clinical scoring, might be the more logical approach for broad clinical use in daily practice. It is not yet known whether one (ROTEM^®) is superior over the other (scores/algorithms) and maybe a combination of both is worthwhile until a prospective clinical evaluation has been made.

The present study has certain limitations. A rather substantial number of patients had to be excluded from the analysis due to missing data, and this might have biased the results. Some laboratory parameters, with particular reference to MT (e.g., fibrinogen, protein C) are not documented at all in the TraumaRegister DGU^®. The data are furthermore limited as the TraumaRegister DGU^® has been designed to register severely and/or multiply injured patients requiring ICU admission only. Patients who received hemostatic agents such as fibrinogen, PCC, rFVIIa or any antifibrinolytics with potential influence on the amount of administered pRBCs were excluded, because some of the hemostatic agents (such as fibrinogen and rFVIIa) are not routinely captured in the database. For the present analysis, a single dataset from the TraumaRegister DGU^® had to be produced containing all variables for the six scoring systems and algorithms that were assessed. Furthermore, detailed data on transfusion practice and use of blood products had to be available. There may also have been significant bias, as the TASH score was validated on the same database that was used for the initial development. It would be advantageous to validate all scores using a multi-center database, in which none of the scores have previously been developed or validated. Finally, a major and common limitation to all models is related to their retrospective nature and prospective validation of these scores is urgently needed.

The timely identification of lethal exsanguination remains challenging. This study validated six scoring systems and algorithms to stratify patients at risk for MT at a very early stage after trauma. We found the TASH (AUC 0.889) and the PWH (AUC 0.860) scores perform superiorly to other tested scores. In general, weighted and more sophisticated systems including higher numbers of variables perform better than simple non-weighted models. However, scores for prediction of MT can only guide a clinician in the decision-making process and should not be used alone. Thrombelastography and thromboelastometry devices seem to be promising tools in the treatment and care of patients with severe bleeding. More research is needed to implement the use of these devices as a standard operating procedure.