Introduction
Despite medical advances, trauma continues to be the leading cause of mortality and acquired disability in children. In 2015 alone, there were more than 11,000 deaths and over 8 million nonfatal injuries caused by trauma to children between the ages of 1 and 19 [
1]. Of children who die from traumatic injuries, most die within the first 24 h upon hospital arrival [
2]. Therefore, due to its high prevalence, up-to-date information regarding pediatric trauma is continuously being produced and brought to the attention of clinicians. For instance, current literature regarding trauma-related pediatric injury shows that firearm injuries account for more than 25% of all unintentional deaths among children. More alarming, this rate now puts firearm injuries as the number one cause of death for children in the United States, surpassing motor vehicle accidents [
3,
4].
Several prediction models for pediatric mortality outcomes have been developed and can assist in decision making [
5‐
9]. Some of these scoring systems are for triage prior to hospital admission, while others assess injury severity or mortality outcomes in manners that are complex, performed retrospectively, or time-consuming [
5]. The Injury Severity Score (ISS) is a commonly used trauma scoring system that has been validated in the pediatric population. Although the ISS associates well with trauma-related mortality, it does have some setbacks. For instance, it requires specialized training and can only be done for research purposes as the calculation is performed retrospectively [
6]. The Revised Trauma Score (RTS) is another example of a well-established and widely used injury scoring system. This tool leverages respiratory rate, Glasgow Coma Scale (GCS), and systolic blood pressure into consideration, but fails to take into account any non-physiologic variables that may also be central in determining survival. Finally, the pediatric trauma ‘BIG’ score uses base deficit, International Normalized Ratio (INR), and GCS as predictive variables. Despite the novelty, it was developed in a small military population which puts into question the generalizability of the model, as most patients succumbed to a blast injury [
9‐
11]. More importantly, many of these prediction models were built before guidelines for methodologic rigor were established, (e.g. TRIPOD statement).
The purpose of this study was to create a pragmatic mortality prediction model for children treated in a U.S. trauma center. We included all-cause trauma-related injuries and relied on using readily available clinical/physiologic variables. Such a prediction model would quickly provide impactful information to clinicians who can better make treatment decisions and more accurately inform patients’ families. Our secondary objective was to transform this prediction model into a web-based dynamic nomogram that can quantify mortality risk.
Discussion
Pediatric trauma is the number one leading cause of mortality in children in the United States and most deaths occur within 24 h of injury [
20]. Therefore, in this study we opted to accurately predict mortality after hospital admission in pediatric patients suffering a traumatic injury. Age, GCS, systolic blood pressure, mechanism of injury, injury type, sex, temperature, and race were found to be the best predictors of mortality and were significant enough to generate a prediction-based web application. This study highlighted several important trends in mortality rates for pediatric trauma patients. Some of the most notable trends include: black patients faced significantly higher mortality rates than their white counterparts, raising the question of what role race plays in pediatric mortality, and firearm and motor vehicle injuries accounted for most of the traumatic injuries analyzed in this study. These trends are discussed in further detail below. The model used in this study, TRAGIC + , demonstrated higher discrimination, calibration, specificity, and sensitivity when compared to already established models such as the ISS and RTS.
Significant sociodemographic discrepancies among different races were observed in this analysis. Specifically, Black patients admitted to the hospital had a significantly higher mortality rate than that of White patients. Previous studies have also demonstrated the remarkable difference between the mortality outcomes of different races. For example, Haider et al. found not only that minority patients experience less favorable outcomes after a traumatic injury than their White counterparts, but that they also tend to live near trauma centers that have overall worse than expected overall mortality rates, which they argue may explain the racial disparities in trauma related outcomes. This was conducted by analyzing over 500,000 patients with an Injury Severity Score greater than or equal to 9 at Level I and II trauma centers in the National Trauma Bank between 2007–2010 [
21]. Another study also concluded that relative to White patients, Black and Asian patients had a higher risk of death after injury using data from the Healthcare Cost and Utilization Project from 1998 to 2002 while controlling for various variables such as race and gender [
22].
Firearm injuries and motor vehicle collisions have been among the top causes of traumatic injuries for years and as expected, this trend was also detected in our study. Other studies have also observed this pattern including McGough et al. who use data from CDC Wonder 2020 Underlying Cause of Death database and the IHME Global Burden of Disease 2019 study. They not only found that firearms and motor vehicle injuries were the number one and two causes of death in children in the United States respectively, but they also concluded that no other similarly wealthy or large country in the world has firearm deaths even in their top four causes of mortality, let alone the number one cause of death in children [
23]. In additional findings from the Global Burden of Disease and Injuries study, it was found that motor vehicle accidents were responsible for 1.3 million deaths in 2010, making it the leading cause of death. This was about 50% more than what it was two decades earlier [
24]. Most studies have also concluded that firearm and motor vehicle injuries are the most common causes of trauma- related mortality, but those that have looked at too broad or too specific of a population. For instance, Krug et al. conducted an epidemiologic analysis of emergency- based surveillance data for 990 infants less than 12 months old from 1994 to 2000 and found that falls were the leading cause of injury [
25]. Due to the relatively small sample size and specific population, this is not generalizable to all children.
Age, GCS, systolic blood pressure, sex, intent of injury, race and mechanism of injury were found to be the best determinants of mortality in our prediction model tool. Many published and established studies have also incorporated these variables into their prediction models. The Prehospital Injury Mortality Score (PIMS) developed a tool to predict blunt trauma mortality using only prehospital variables such as age, mechanism of injury, sex, and trauma activation criterion. Their validation and derivation groups each consisted of over 160,000 patients and they displayed good discrimination with AUC of 0.79 in both groups [
26]. Driessen et al. also developed a mortality prediction model in all age groups after trauma using variables such as systolic blood pressure, GCS, age, best motor response (BMR), etc. using over 300,000 patients. Therefore, this model and the PIMS model are both well calibrated and demonstrated good discrimination using variables similar to that of our prediction model including age, GCS, mechanism of injury, and systolic blood pressure [
27].
Finally, race proved to be a significant determinant of mortality among pediatric trauma patients and was used in the TRAGIC + prediction model. This is seen in many other studies, one such example being a study conducted by Haider et al. in which a systematic review and meta-analysis including thirty-five studies demonstrated that black race is associated with a higher odds of death in trauma when compared with white race [
28].
We found that firearm injuries and motor vehicle collisions were the leading causes of traumatic injuries, consistent with findings from other studies. Our prediction model incorporated variables that have been widely established in previous models, such as age, GCS, systolic blood pressure, sex, intent of injury, race, and mechanism of injury. The TRAGIC + model demonstrated higher discrimination, calibration, specificity, and sensitivity compared to established models like the ISS and RTS. By outperforming these widely used models, our prediction model presents a significant advancement in risk stratification and prognostication for pediatric trauma patients. Furthermore, the development of a live web application for the model enhances its accessibility and usability for healthcare providers, facilitating its potential adoption in clinical practice. Overall, the unique combination of easily accessible variables, superior predictive performance, and user-friendly implementation distinguishes our prediction model as a potentially valuable tool in improving outcomes for pediatric trauma patients.
Strengths of our project include that the TRAGIC + prediction model was created using a large sample size from a large national data bank, while adhering to all 22 of the TRIPOD guidelines (Supplemental File
1). This model was also successfully internally validated
. Temporal validation is also a strength of the study, the participants of this cohort were not part of the derivation cohort, meaning the temporal validation is an external validation. As was seen in the results, comparing the TRAGIC + prediction model to the ISS and the RTS demonstrated that our model had higher discrimination, calibration, specificity, and sensitivity than even these published and well-known models. The derived model exhibited reasonable calibration, while both the ISS model and the RTS model display varying degrees of underfitting and overfitting, which renders their estimations unstable and less suitable for application in the NTDB. In addition to being statistically superior to these widely used models, we have also created a live web application in order to make our TRAGIC + model an easily usable and accessible tool that will further encourage healthcare providers to adopt this model.
Despite the fact that our model has a lot of potential, limitations to our study were inevitable. We were unable to validate PMIS and BMR because of the low granularity of the NTDB data (e.g., predictors not available). The TRAGIC + model also needs to be validated in other populations to increase its generalizability. TRAGIC + is also a retrospective, rather than prospective study, which can be seen both as an advantage and disadvantage. Clinicians should use their best judgment to make the decisions that are in the best interest of the patient and provide the best treatments possible.
Overall, we were able to develop a prediction model for pediatric mortality following a traumatic injury utilizing easily assessable and universally applicable clinical variables. This prediction model not only proved to be more accurate than already established models, but it was also developed into a user-friendly web application for clinical use:
https://agmoreir.shinyapps.io/TRAGIC/. Future research includes validation of the model in more current NTDB databases and external validation. Conducting prospective studies to validate the prediction model in real-time clinical settings would be the ultimate goal as this can help determine the feasibility and clinical utility of the model and potential implementation process.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.