Machine learning models for early mortality prediction in trauma patients using public data: a nationwide retrospective study

Abstract

Background

Trauma is a leading cause of morbidity and mortality worldwide, particularly in younger populations. Early identification of high-risk trauma patients is critical for timely interventions and improved outcomes. Although artificial intelligence and machine learning have demonstrated promise in healthcare, their application in trauma mortality prediction has been limited.

Methods

This study developed and validated machine learning models to predict mortality in trauma patients using a large public dataset from the National Community-Based Critical Injury Survey (South Korea, 2016–2020). Overall, 207,012 cases were analyzed. Six machine learning algorithms, including logistic regression, k-nearest neighbor, decision tree, random forest (RF), extreme gradient boosting (XGB), and multi-layer perceptron, were trained and evaluated. Their performance was assessed using the areas under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC), and other metrics. Shapley additive explanations (SHAP) scores were used to interpret feature importance.

Results

The XGB model demonstrated the highest performance (AUROC 0.985; AUPRC 0.957), followed closely by the RF model (AUROC 0.984; AUPRC 0.956). Performance remained stable during the COVID-19 period, supporting the model’s temporal robustness under systemic disruption. SHAP analysis identified clinically actionable features such as out-of-hospital cardiac arrest, injury severity score, age, and time to transfusion. Unlike many prior studies based on small or single-center datasets, our model was developed using a nationally representative cohort and prioritized interpretability, scalability, and generalizability.

Conclusions

This study presents a high-performing, interpretable machine learning framework for early mortality risk stratification in trauma patients using nationwide registry data. The strong discrimination and temporal robustness of the model support its value as a system-level prediction tool; however, further calibration analyses, external validation, and prospective implementation studies are required before integration into clinical workflows.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13017-026-00706-x.

Background

Trauma is the leading cause of morbidity and mortality worldwide, particularly in younger populations. The World Health Organization reports that injuries result in over 5 million deaths annually, accounting for nearly 9% of the world’s mortality [1]. Despite advancements in trauma care, the prediction of mortality among patients who experienced trauma remains a significant challenge owing to the complexity and variability of injury mechanisms, patient physiology, and the multifactorial nature of trauma outcomes [2].

Artificial intelligence (AI) and machine learning (ML) have shown promise in improving clinical decision-making by enabling accurate risk prediction from large-scale data [3–6]. In trauma care, ML-based models may facilitate early identification of high-risk patients, potentially improving outcomes through timely intervention and resource allocation [7, 8].

Several studies have explored the application of AI in trauma care. For instance, algorithms have been developed to predict the need for massive transfusion [9], the likelihood of sepsis [10], and patient discharge destinations [11]. However, existing models often face limitations such as using small, nonrepresentative datasets, lack of real-time applicability, and limited integration with clinical workflows [12]. Previous machine-learning models for trauma mortality prediction have shown that data-driven algorithms can outperform traditional scoring systems such as Trauma and Injury Severity Score or the Injury Severity Score. Nevertheless, most of these studies were limited by single-center data, absence of external or temporal validation, and insufficient model interpretability. In addition, many approaches prioritized accuracy over clinical integration, making real-time application in trauma systems difficult. These limitations highlight the need for robust, explainable, and temporally validated models developed from nationwide registries. Most importantly, while numerous models exist for predicting various outcomes in patients who experienced trauma, there is a notable gap in the literature concerning robust, early prediction models specifically targeting mortality [13, 14].

Accurate early prediction of mortality in patients who experienced trauma is critical. Early identification of high-risk patients can guide clinical decision-making, optimize resource allocation, and improve patient outcomes through timely and appropriate medical interventions. Furthermore, an effective predictive model can assist in stratifying patients based on their risk, thereby enhancing the efficiency of trauma systems and potentially reducing healthcare costs [15, 16].

This study aimed to develop and temporally validate an interpretable machine-learning model for early mortality prediction in trauma patients using a large, nationally representative dataset from the National Community-Based Critical Injury Survey, the Korean nationwide trauma registry managed by the Korea Disease Control and Prevention Agency. Although conceptually similar to the U.S. National Trauma Data Bank [17], this registry reflects the structure and operation of the Korean trauma system and enables system-level evaluation of early mortality prediction.

Methods

Study population and data collection

A schematic flow of the study is shown in Fig. 1. We analyzed the National Community-Based Critical Injury Survey, which included data from 237,616 incidents collected from all residents of South Korea between January 1, 2016, and December 31, 2020. The data are publicly available from the National Injury Information Portal of Korea Disease Control and Prevention Agency [18]. The survey includes a wide range of variables associated with catastrophic injury to support the establishment and evaluation of injury-prevention and management policies. It records comprehensive details of each incident, including whether multiple casualties or severe injuries occurred, mechanism and type of damage, contributing environmental or human factors, prehospital and emergency-department interventions, and injury-severity scoring parameters. In addition, sociodemographic information such as age, sex, insurance type, and area of residence was collected for every case. These diverse variables were used as input features for the machine-learning models after preprocessing, and complete lists of all variables are presented in Supplementary Tables S1 and S2.

Fig. 1.

The schematic flow of the study. The process of data collection, exclusion, and the final dataset used for model development and validation are illustrated, with the initial dataset of 237,616 incidents; 30,604 cases were excluded owing to data loss, leading to a final dataset of 207,012 cases used for the analysis and model development

All trauma cases recorded in the National Community-Based Critical Injury Survey between 2016 and 2020 were included in the analysis. Cases with missing or inconsistent data for key outcome variables—such as survival status, injury severity score, or major clinical information—were excluded, resulting in a final analytic cohort of 207,012 patients as illustrated in Fig. 1.

The National Community-Based Critical Injury Survey collects both prehospital and immediate in-hospital information within the first 24 h of case registration. Accordingly, the present study defines “early” prediction as mortality estimation during this initial phase of trauma management, encompassing data available from scene reporting through the first 24 h of hospital care. Within this context, the model is intended for early risk stratification during the initial phase, when both prehospital information and early in-hospital data are available to support triage and clinical decision-making.

Data preprocessing

The primary outcome of this study was all-cause mortality after registry entry, encompassing both prehospital and in-hospital deaths as defined in the National Community-Based Critical Injury Survey. No secondary outcomes were analyzed because the registry did not include morbidity or long-term follow-up data.

Data were analyzed to develop a mortality prediction model. After the analysis, 30,604 cases were excluded due to data loss, making it difficult to define the outcome condition (Fig. 1). The excluded records were evenly distributed across years and regions, with demographic characteristics comparable to the retained population, suggesting that the missingness was unlikely to introduce systematic bias. We used 207,012 cases to develop and validate the prediction model. Data were divided according to the year of the survey. We split the data in a non-random manner to evaluate how well the model performs [19]. We also aimed to determine if the model's performance was consistent before and after the impact of coronavirus disease 2019 (COVID-19) on emergency medical protocols in South Korea. Data from 2016 to 2018 (n = 119,075) were used for model development, and data from 2019 to 2020 (n = 87,937) were used for validation.

To maintain the realism of prehospital triage conditions, variables describing prehospital cardiac arrest and prehospital death were retained in the dataset. These variables represent real-time clinical information recorded at the time of emergency report registration and are routinely available to emergency medical services. Mortality was defined as any death occurring after case registration, thereby encompassing patients who presented with prehospital cardiac arrest to capture the entire continuum of trauma severity within the national registry.

Feature engineering

The dataset contained some missing variables, particularly in patients who did not undergo surgery or embolization. For patients who underwent surgery and embolization, we calculated the time elapsed from the incident report time. Then, we created new derived features based on whether this time difference was less than or equal to 24 h. To handle missing cases, we imputed missing values using the median values. We also created a missing indicator, coded as 1 if a value was missing and 0 if the value was not. For time-to-event variables such as transfusion timing, missing values were imputed using the median, and a corresponding missing-indicator variable (“not reported”) was included to explicitly model the presence of missingness and to distinguish documentation-related effects from observed event timing. Time-to-event variables such as transfusion timing were recorded relative to the incident report time and may encompass a wide range of values depending on the clinical course and applicability. These variables exhibited skewed distributions and were therefore interpreted as process-related indicators rather than precise measures of early intervention timing. Additionally, we used one-hot encoding for the categorical variables to prepare them for analysis. To reduce the impact of outliers, we standardized all continuous variables using the median and interquartile range.

Modeling

We developed prediction models to estimate mortality risk in major trauma cases using six different ML algorithms: logistic regression [20], k-nearest neighbor [21], decision tree classifier [22], random forest classifier (RF) [23], extreme gradient boosting (XGB) [24], and multi-layer perceptron classifier [25]. We developed six prediction models using derived features based on the time difference within 24 h from the incidence report time, along with additional feature-engineered variables.

Performance evaluation

Hyperparameter tuning was conducted within the development dataset (2016–2018) using a grid search combined with stratified fivefold cross-validation. Each fold was randomly stratified while remaining confined to the 2016–2018 period to maintain temporal independence. The optimal hyperparameters derived from this process were then used to retrain each model on the entire 2016–2018 dataset. Final evaluation was performed on the independent 2019–2020 dataset, which was reserved as a temporally distinct hold-out set. To further assess temporal robustness, performance metrics were also compared between the non-COVID-19 (2019) and COVID-19 (2020) sub-periods. For evaluation, we calculated various metrics, including the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), accuracy, specificity, precision, positive predicted value, negative predictive value, and F1 score.

Model interpretation

We used SHapley Additive exPlanations (SHAP) scores to evaluate the significance of the features in our model predictions. SHAP provides a unified measure of feature importance by attributing each feature's contribution to the overall prediction [26]. For this analysis, we calculated SHAP scores using two high-performing models, the XGB and RF models. These models were selected based on their top-performance metrics. To visualize and interpret the results, we employed SHAP beeswarm plots, which graphically displayed the distribution of SHAP values for each feature across the dataset [27].

Statistical analysis

Data are presented as numbers with percentages in parentheses for categorical variables and means (standard deviations) for continuous variables. Categorical variables were compared using Pearson's χ² test, and continuous variables were compared using the t-test. All tests were two-sided, and a P-value less than 0.05 was considered statistically significant. Statistical analyses were performed using Python (version 3.11.0; Python Software Foundation). The SHAP values were generated using Python Shap version 0.35.0, and the results of the model interpretation are shown graphically.

Results

Baseline characteristics

Tables S1 and S2 summarize the baseline characteristics of the model development and validation sets, respectively. Overall, the analyzed predictors encompassed four major domains: prehospital variables (e.g., out-of-hospital cardiac arrest, prehospital transfusion), demographic and sociodemographic characteristics (e.g., age, sex, insurance type, region), injury-related factors (e.g., injury severity score, injury mechanism, intentionality), and early care process variables (e.g., transfusion, surgery, or embolization within 24 h).

Overall mortality prevalence was low in both cohorts, reflecting substantial class imbalance. In the development set, mortality accounted for approximately 21% of cases, of which about 69% involved out-of-hospital cardiac arrest or prehospital death. Similarly, in the validation set, mortality accounted for approximately 18% of cases, with approximately 72% attributable to out-of-hospital cardiac arrest or prehospital death.

In the development set, most variables were significantly (P < 0.001) different between the survived and non-survived groups, except for two variables: time from report to emergency department arrival (P = 0.358) and time from report receipt to embolization (P = 0.064). In contrast, in the validation set, both variables were significantly different (P < 0.001 and P = 0.026, respectively), along with the other variables.

Model performance

Figure 2 shows the overall performance of the developed models. While the XGB model demonstrated the highest AUROC at 0.985 (95% CI, 0.979–0.991), we acknowledge that the AUROC may overestimate performance in imbalanced datasets such as ours. Therefore, we emphasize the AUPRC, which better reflects a model's ability to identify true positives in rare-event settings like trauma-related mortality.

Fig. 2.

Model performance metrics. A Receiver operating characteristic (ROC) curve comparing the performance of the six machine learning models: logistic regression (LR), k-nearest neighbor (KNN), decision tree classifier (DT), random forest classifier (RF), extreme gradient boosting (XGB), and multi-layer perceptron (MLP). B Precision-recall curve for the same models, showing the balance between precision and recall across different thresholds, with XGB and RF models achieving the highest performance in both area under the ROC curve (AUROC) and area under the precision-recall curve (AUPRC)

Table 1 lists the detailed model performance using various metrics. All six models demonstrated high accuracy, specificity, and F1 score in predicting mortality. The XGB model also achieved the highest AUPRC of 0.957 (95% CI 0.943–0.971), followed closely by the RF model with an AUPRC of 0.956 (95% CI 0.942–0.970). These values highlight their superior precision-recall balance, a key consideration in high-stakes triage where minimizing false negatives is critical. Both models also maintained high F1 scores (XGB: 0.895; RF: 0.893) and strong precision (XGB: 0.930; RF: 0.932), outperforming simpler models such as logistic regression, decision tree, and k-nearest neighbor, which showed modest drops in sensitivity and precision.

Table 1.

Predictive model evaluation: data from 2016 to 2018 applied to 2019–2020 for predictive performance analysis

Model	AUROC (95% CI)	AUPRC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	F1 Score (95% CI)	PLR (95% CI)	NLR (95% CI)
2019–2020
LR	0.982 (0.975–0.989)	0.952 (0.937–0.966)	0.964 (0.956–0.972)	0.863 (0.829–0.899)	0.986 (0.980–0.991)	0.928 (0.902–0.954)	0.971 (0.964–0.978)	0.894 (0.871–0.918)	59.625 (42.370–96.054)	0.139 (0.103–0.174)
DT	0.915 (0.897–0.934)	0.748 (0.709–0.789)	0.937 (0.927–0.947)	0.869 (0.834–0.904)	0.952 (0.941–0.962)	0.795 (0.762–0.831)	0.971 (0.964–0.979)	0.831 (0.805–0.857)	17.932 (14.796–22.650)	0.137 (0.100–0.173)
KNN	0.958 (0.944–0.972)	0.904 (0.880–0.929)	0.953 (0.945–0.963)	0.816 (0.778–0.857)	0.983 (0.977–0.989)	0.913 (0.884–0.941)	0.961 (0.954–0.969)	0.862 (0.837–0.890)	48.450 (35.242–74.157)	0.187 (0.146–0.225)
RF	0.984 (0.977–0.990)	0.956 (0.942–0.970)	0.963 (0.956–0.971)	0.856 (0.820–0.893)	0.987 (0.981–0.991)	0.932 (0.906–0.956)	0.969 (0.962–0.977)	0.893 (0.870–0.916)	63.539 (44.591–101.274)	0.146 (0.108–0.182)
MLP	0.983 (0.976–0.990)	0.953 (0.935–0.968)	0.963 (0.956–0.971)	0.876 (0.843–0.910)	0.982 (0.975–0.988)	0.913 (0.884–0.940)	0.973 (0.967–0.981)	0.894 (0.872–0.917)	48.566 (35.328–72.283)	0.127 (0.092–0.160)
XGB	0.985 (0.979–0.991)	0.957 (0.943–0.971)	0.964 (0.956–0.972)	0.862 (0.829–0.899)	0.986 (0.980–0.991)	0.930 (0.903–0.955)	0.971 (0.964–0.978)	0.895 (0.871–0.919)	61.257 (43.145–98.998)	0.140 (0.102–0.175)
2019 (non-COVID-19 period)
LR	0.983 (0.976–0.990)	0.951 (0.934–0.967)	0.967 (0.960–0.975)	0.863 (0.826–0.899)	0.987 (0.981–0.992)	0.924 (0.894–0.952)	0.975 (0.968–0.981)	0.893 (0.868–0.918)	65.310 (45.131–104.679)	0.139 (0.102–0.177)
DT	0.918 (0.899–0.937)	0.740 (0.695–0.785)	0.942 (0.932–0.952)	0.872 (0.835–0.908)	0.955 (0.945–0.965)	0.785 (0.750–0.822)	0.975 (0.968–0.982)	0.826 (0.798–0.856)	19.496 (15.986–24.660)	0.134 (0.096–0.174)
KNN	0.961 (0.946–0.975)	0.903 (0.875–0.929)	0.959 (0.951–0.967)	0.822 (0.782–0.861)	0.985 (0.979–0.990)	0.912 (0.878–0.942)	0.967 (0.960–0.974)	0.864 (0.835–0.890)	55.170 (38.188–86.290)	0.181 (0.141–0.222)
RF	0.985 (0.977–0.991)	0.954 (0.936–0.969)	0.967 (0.960–0.975)	0.858 (0.820–0.896)	0.987 (0.982–0.992)	0.927 (0.900–0.956)	0.974 (0.967–0.981)	0.891 (0.866–0.916)	67.953 (48.140–116.808)	0.144 (0.106–0.183)
MLP	0.985 (0.978–0.990)	0.950 (0.931–0.967)	0.966 (0.959–0.974)	0.878 (0.842–0.915)	0.983 (0.977–0.989)	0.906 (0.876–0.936)	0.977 (0.971–0.984)	0.892 (0.866–0.916)	51.182 (37.574–78.541)	0.124 (0.087–0.160)
XGB	0.986 (0.979–0.992)	0.955 (0.939–0.971)	0.967 (0.960–0.975)	0.864 (0.826–0.899)	0.987 (0.980–0.992)	0.924 (0.895–0.953)	0.975 (0.968–0.981)	0.893 (0.867–0.918)	65.339 (45.213–107.767)	0.138 (0.102–0.176)
2020 (COVID-19 period)
LR	0.980 (0.973–0.987)	0.954 (0.941–0.967)	0.959 (0.951–0.967)	0.863 (0.830–0.893)	0.984 (0.978–0.989)	0.932 (0.908–0.955)	0.965 (0.957–0.973)	0.896 (0.874–0.917)	53.082 (38.348–81.213)	0.140 (0.108–0.174)
DT	0.912 (0.895–0.928)	0.758 (0.723–0.797)	0.930 (0.919–0.940)	0.867 (0.835–0.898)	0.946 (0.934–0.957)	0.806 (0.774–0.838)	0.965 (0.957–0.973)	0.836 (0.811–0.858)	16.108 (13.206–20.061)	0.140 (0.108–0.175)
KNN	0.955 (0.941–0.968)	0.905 (0.883–0.927)	0.946 (0.936–0.955)	0.811 (0.774–0.847)	0.980 (0.974–0.987)	0.914 (0.889–0.941)	0.953 (0.944–0.961)	0.860 (0.834–0.883)	41.330 (30.927–61.311)	0.193 (0.156–0.231)
RF	0.982 (0.975–0.989)	0.958 (0.946–0.970)	0.958 (0.950–0.967)	0.854 (0.822–0.888)	0.985 (0.979–0.991)	0.938 (0.914–0.961)	0.963 (0.955–0.972)	0.894 (0.872–0.916)	58.209 (41.004–96.109)	0.148 (0.113–0.181)
MLP	0.981 (0.974–0.987)	0.955 (0.942–0.967)	0.959 (0.950–0.968)	0.873 (0.839–0.905)	0.981 (0.974–0.987)	0.921 (0.895–0.946)	0.968 (0.959–0.976)	0.896 (0.874–0.919)	45.319 (32.901–67.467)	0.130 (0.097–0.164)
XGB	0.984 (0.978–0.989)	0.960 (0.948–0.971)	0.959 (0.951–0.968)	0.860 (0.827–0.893)	0.985 (0.978–0.991)	0.936 (0.911–0.959)	0.965 (0.957–0.973)	0.896 (0.874–0.918)	56.298 (39.344–90.733)	0.142 (0.109–0.176)

AUROC area under the receiver operating characteristic curve, AUPRC area under the precision-recall curve, PPV positive predictive value, NPV negative predictive value, PLR positive likelihood ratio, NLR negative likelihood ratio, LR logistic regression, DT decision tree, KNN k-nearest neighbor, RF random forest, MLP multilayer perceptron, XGB extreme gradient boosting, CI confidence interval

The superior performance of the XGB and RF models is likely due to their ability to model complex, nonlinear interactions and handle variable importance effectively through ensemble learning. In contrast, models like the decision tree and KNN struggled with generalization and class imbalance, as reflected in their lower AUPRC and F1 scores.

Temporal validation further confirmed the robustness of the XGB and RF models. Figure 3 shows that both models maintained high AUROC and AUPRC values across pre-COVID and COVID-19 periods. The XGB model yielded an AUROC of 0.986 and AUPRC of 0.955 before the pandemic, and 0.984 and 0.960 during it. Similarly, the RF model maintained an AUROC of 0.985 and AUPRC of 0.954 pre-pandemic, and 0.982 and 0.958 during the pandemic. These results suggest that models trained on historical data generalize well even under major healthcare system disruptions.

Fig. 3.

Performance of models before and during COVID-19. A Receiver operating characteristic (ROC) curve for the extreme gradient boosting (XGB) model performance using pre-COVID-19 data (2016–2019). B ROC curve for the XGB model performance using COVID-19 period data (2020). C Precision-recall curve for the XGB model using pre-COVID-19 data. D Precision-recall curve for the XGB model using COVID-19 period data, indicating consistent model performance across both periods

Model interpretation via feature importance

To assess which features had the most impact on predicting mortality, we analyzed the top 15 features with the highest SHAP values for the mortality prediction model using XGB and RF (Fig. 4). Among these features, out-of-hospital cardiac arrest, prehospital death, and injury severity scores significantly influenced all model SHAP values. The injury severity score and the occurrence of out-of-hospital cardiac arrest or prehospital death were positively correlated with patient mortality. In the XGB model, age, time from reporting to first transfusion, and time from reporting to first transfusion (not reported) were important variables related to model outputs. Lower values of the time from the report to the first transfusion showed a positive relationship. Not-reported time from report to first transfusion also showed a positive relationship. In the RF model, prehospital cardiac arrest, injury severity score-score reported, and age were substantially related factors with outputs. Prehospital cardiac arrest and age are positively related to patient outcomes, which increase patient mortality. Score reported-injury severity score showed a negative relationship with the outcome.

Fig. 4.

Shapley additive explanations (SHAP) analysis for feature importance in mortality prediction. A SHAP beeswarm plot for the extreme gradient boosting (XGB) model, displaying the top 15 features with the highest SHAP values. Key features include out-of-hospital cardiac arrest, injury severity score, and age, among others. B SHAP beeswarm plot for the random forest classifier (RF) model, highlighting similar key features with minor differences in feature importance rankings compared with the XGB model

Discussion

The prediction of mortality in patients who experienced trauma is crucial for guiding timely and appropriate interventions, optimizing resource allocation, and ultimately improving patient outcomes. While the AUROC is a commonly used metric, it may overestimate predictive performance in imbalanced datasets. Therefore, we report and interpret the AUPRC alongside the AUROC to provide a more clinically meaningful assessment of model quality, especially considering the low incidence of mortality events in our cohort. The results demonstrated that ML, particularly XGB and RF models, can effectively predict mortality with high accuracy, even across different temporal contexts, such as the pre- COVID-19 and COVID-19 periods. The high AUPRC values in this study should be interpreted in the context of the national trauma registry, which encompasses the full continuum of injury severity, including out-of-hospital cardiac arrest and prehospital death. Because these cases constitute a substantial portion of the mortality group and represent genuine events within the national trauma reporting system, their inclusion contributes to strong overall discrimination. Accordingly, the reported performance reflects system-level prediction across a comprehensive nationwide registry rather than mortality modeling confined to hospital-based cohorts.

The XGB and RF models exhibited superior performance in predicting mortality, as evidenced by their high AUROC and AUPRC values. In particular, the XGB model achieved an AUROC of 0.985 (95% CI 0.979–0.991) and an AUPRC of 0.957 (95% CI 0.943–0.971) for the overall dataset. These metrics reflect excellent discriminative ability, which is essential for clinical decision support. Furthermore, the XGB model maintained robust performance across both the pre-COVID-19 and COVID-19 periods (AUROC: 0.986 and 0.984, respectively), underscoring its resilience despite healthcare system disruptions. This consistency is critical in trauma care, where real-time performance must be preserved under unpredictable conditions. Beyond these top-line metrics, a closer look at other performance indicators—such as sensitivity, precision, and F1 score—reveals important trade-offs between models. For instance, the XGB model showed high sensitivity (0.862) and precision (0.930), resulting in a balanced F1 score (0.895), which is ideal for identifying high-risk patients while minimizing false alarms. In contrast, the MLP model had slightly higher sensitivity (0.876) but lower precision (0.913), potentially favoring detection at the cost of more false positives. Logistic regression, while interpretable, yielded a slightly lower sensitivity (0.863), which might risk missing critical cases. These nuances are important in clinical deployment: higher sensitivity models may be favored when the cost of missing a mortality case is high, while higher precision models might be better suited for environments where overtriage could burden limited resources. Thus, the balance achieved by XGB and RF models makes them particularly appropriate for real-world trauma applications. We prioritized XGB and RF because they not only demonstrated top-tier predictive performance but also offered resilience across time periods, compatibility with SHAP-based interpretability, and fast computational speed, making them highly suitable for clinical integration.

The high performance of the XGB model is likely attributable to its ability to capture complex interactions between variables, which are often nonlinear and multidimensional in trauma cases [3, 28]. SHAP analysis showed that the model's predictions were strongly influenced by key factors such as out-of-hospital cardiac arrest, injury severity score, and time to first transfusion. A shorter time to first transfusion was associated with a lower predicted mortality risk, suggesting that early hemorrhage control and hemodynamic stabilization are critical determinants of survival. This is consistent with trauma resuscitation principles that prioritize timely transfusion to reduce shock-related complications and improve outcomes. These factors are well established in the trauma literature as key indicators of patient outcomes, thereby supporting the clinical relevance of the model's features [10, 29].

The consistent performance of the models across different time periods further supports their robustness and adaptability, suggesting potential utility in dynamic and resource-constrained environments such as pandemics [11]. SHAP analysis also reinforced the operational significance of these predictors in emergency settings. Out-of-hospital cardiac arrest and prehospital death—top predictors in both models—highlight the importance of early recognition and resuscitation in the field. Injury severity score underscores the value of early trauma scoring systems in triage and prioritization, while age may inform resource-intensive decisions such as ICU admission or geriatric trauma pathways. Time to first transfusion emerged as a modifiable factor, emphasizing the importance of rapid hemorrhage control in trauma care.

The predictive importance of the "time not reported" indicator for transfusion should be interpreted as reflecting early care processes or documentation patterns rather than a biological effect. In a nationwide trauma registry, absent or delayed documentation of transfusion timing may be associated with early mortality, rapid clinical deterioration, or workflow constraints during acute resuscitation, and thus represents a system-level signal. The wide range of observed values for transfusion timing reflects the heterogeneity of clinical trajectories and registry recording practices. Therefore, these variables should be interpreted cautiously as indicators of care processes and documentation patterns rather than exact measures of intervention timing. Given its prominence in the SHAP rankings, it is important to emphasize that transfusion timing in this context should not be over-interpreted as a direct mechanistic indicator of early treatment efficacy. This feature, rather, likely represents a composite surrogate reflecting documentation practices, care process efficiency, and clinical acuity; its predictive contribution should be interpreted within these operational constraints. These findings collectively suggest that ML-based tools have the potential to inform early risk stratification during the initial phase of trauma management, pending further calibration and prospective validation. Accordingly, this model is intended for use during the early in-hospital phase rather than at a single fixed time point, reflecting the dynamic nature of trauma care.

SHAP analysis provided insights into the variables most influential in predicting mortality. Both XGB and RF models highlighted out-of-hospital cardiac arrest, prehospital death, and injury severity scores as the top contributors to the model outputs. This finding aligns with existing clinical knowledge that severe physiological derangements and critical injuries are strong predictors of mortality [14]. Notably, the importance of prehospital factors, such as out-of-hospital cardiac arrest, suggests that early intervention strategies, potentially even before hospital arrival, could be pivotal in reducing trauma mortality [3, 30]. Age was another significant factor across the models, with older age being associated with higher mortality. This reinforces the need for tailored trauma care strategies for older adults who may have reduced physiological reserves and increased vulnerability to complications [15]. Time to first transfusion also emerged as a critical factor, indicating the importance of prompt hemodynamic stabilization in trauma management. The finding that missing data on the time to first transfusion was associated with increased mortality highlights the potential risks of incomplete documentation and the need for robust data collection in trauma care settings [17].

From the learning algorithm perspective, the RF and XGB models consistently identified the top two features as the most important contributors to model predictions, indicating a robust agreement in their interpretation of key features within the dataset. This consistency suggests that these features possess robust predictive power across different machine-learning algorithms. However, beyond these top two features, the models diverged significantly in their rankings of the remaining features. This variation likely stems from fundamental differences in the internal mechanisms of the two algorithms. RF employs a bagging approach, relying on ensemble decision trees with bootstrapped samples and feature randomness to reduce variance [31]. In contrast, XGB uses a boosting approach, iteratively reweighting features, and retraining weak learners to minimize error, resulting in high bias reduction and fine-tuned model performance [31]. This discrepancy underscores the importance of model selection and the value of employing multiple algorithms in ML to gain diverse insights, especially in complex feature spaces where different models may capture unique patterns [32].

Because prehospital cardiac arrest and prehospital death are highly correlated with mortality, their inclusion as predictors may partially contribute to the observed high discrimination. However, these factors represent real-time clinical information routinely documented during trauma case reporting and are integral to the decision-making context of emergency medical services. The model therefore incorporated them to emulate real operational conditions rather than retrospective classification. In addition, other independent variables such as injury severity score, age, and transfusion-related parameters remained strong determinants of outcome, suggesting that model performance is not solely driven by these prehospital variables. Model performance in the present study should therefore be interpreted in the context of this case mix. Importantly, because prehospital cardiac arrest and prehospital death account for the majority of non-survivors and function as near-deterministic predictors, the reported AUROC and AUPRC primarily reflect discrimination within this combined case mix. Predictive performance in the more clinically ambiguous subgroup of undifferentiated, potentially salvageable trauma patients, where early risk stratification has the greatest clinical value, may be materially lower than these headline metrics suggest. Future studies specifically evaluating model performance after excluding prehospital fatal events are, therefore, needed to determine the practical utility of such models for bedside decision-making.

This study had several strengths. First, it emphasizes the early prediction of patient outcomes during trauma. By accurately identifying high-risk patients shortly after the trauma, our model enables healthcare providers to initiate timely interventions, which can be crucial in improving survival rates [3]. Second, the model's robustness in predicting mortality during the COVID-19 pandemic highlights its applicability in crises [7]. In such scenarios, where healthcare resources are stretched thin, the ability to accurately triage and prioritize patients based on predicted outcomes is invaluable. This capability positions our model as a potentially vital tool for managing future pandemics and other large-scale healthcare emergencies [33].

Beyond these strengths, the novelty of this study lies in its comprehensive approach to predict mortality in patients who experienced trauma. Unlike previous studies, our research leveraged a large, nationally representative dataset [17, 34]. However, since the data are derived solely from the Korean trauma system, the generalizability of our findings to other healthcare systems remains uncertain and should be validated externally. Additionally, the use of sophisticated ML algorithms such as XGB and RF, which are capable of capturing complex, nonlinear relationships among variables, sets our study apart from more traditional approaches [3]. Moreover, the temporal analysis comparing pre- and post-pandemic performance is a novel aspect that has not been widely explored in trauma-related predictive modeling [11]. This analysis provides critical insights into how such models can maintain efficacy despite significant external disruptions, underscoring their potential for long-term integration into clinical practice.

The findings of this study are consistent with previous research that has utilized AI and ML for mortality prediction in trauma patients [10]. However, this study stands out due to its use of a large, nationally representative dataset and the application of sophisticated algorithms like XGB and RF, which have not been extensively explored in prior trauma studies [4]. Additionally, the study's focus on model validation across pre-COVID-19 and COVID-19 periods provides a unique perspective on the resilience and applicability of AI models in the face of external disruptions [14].

Despite these promising results, this study had limitations that warrant further consideration. First, the retrospective nature of the data introduces inherent limitations, including the risk of unmeasured confounding, selection bias, and variability in data completeness, all of which may affect the internal validity of the findings and limit their generalizability [18]. Although the models performed well in both pre-COVID-19 and COVID-19 periods, future studies should validate these findings in prospective settings and across different geographical regions to ensure their generalizability. Additionally, while the models identified key predictors of mortality, they did not provide causal explanations, which could limit their interpretability in clinical decision-making. Moreover, morbidity-related outcomes such as complications or long-term functional impairments were not included in the registry and thus were not analyzed in this study. Future research should incorporate morbidity and recovery outcomes to enable a more comprehensive understanding of trauma prognosis.

Beyond performance and feasibility, the deployment of machine learning in trauma triage raises important ethical considerations. Algorithmic predictions in life-threatening scenarios must be carefully evaluated for fairness across age, sex, socioeconomic status, and geographic location. Biases in historical data or differential access to care could unintentionally propagate health disparities if not appropriately addressed. While our study did not explicitly analyze subgroup performance, future work will include fairness audits and subgroup-specific validation to ensure that model predictions do not disproportionately disadvantage vulnerable populations. Additionally, transparency, explainability, and clinician oversight are essential to maintain trust and accountability in clinical decision-making supported by AI tools.

Moreover, although the models demonstrated strong discrimination, the lack of formal calibration analyses and clinically defined operating thresholds indicates the present results should be interpreted as evidence of model development and temporal validation, rather than clinical readiness. Translation into clinical practice will require calibration assessment, threshold optimization tailored to local workflows, and prospective implementation studies to evaluate real-world safety and effectiveness.

Additionally, the exclusion of 30,604 cases with missing data may introduce selection bias. If the missingness is not random—for example, if it disproportionately affects older adults or severely injured patients—this could influence model training and limit generalizability. The direction and magnitude of any resulting selection bias cannot be determined from the available data; therefore, its impact on model performance and generalizability remains uncertain. Future studies should evaluate patterns of missingness and consider techniques such as multiple imputation to minimize potential bias.

Finally, while median imputation and missing value indicators were chosen for their simplicity and compatibility with tree-based models, we acknowledge that this approach may not fully leverage the richness of our dataset. More advanced imputation strategies—such as multiple imputation, k-nearest neighbors, or model-based methods—could potentially improve data recovery, calibration, and generalizability, and should be explored in future research. In addition, formal calibration analyses and threshold-based evaluations were not performed in this study. Because clinically meaningful alert thresholds depend on local workflows, resource constraints, and risk tolerance, future prospective and implementation-focused studies should assess calibration, Brier scores, and decision thresholds to optimize real-world clinical utility. These additional analyses will be necessary to determine appropriate operating thresholds and to support safe and effective clinical deployment in real-world settings. We also note that our dataset lacked potentially important predictors such as socioeconomic status, comorbidities, and prior functional status. The absence of these features may introduce unmeasured confounding and bias into both model predictions and SHAP-based feature importance rankings. Future research should consider incorporating a broader range of clinical and social variables, and conduct sensitivity analyses or bias assessments to evaluate the robustness and fairness of the models. Furthermore, our dataset is limited to a single national system (South Korea), which may restrict the generalizability of the findings to other healthcare environments. Differences in trauma systems, triage protocols, and population characteristics across countries could impact model performance. Accordingly, this model was developed using data from a single national trauma system, several core predictors—including age, injury severity score, and prehospital cardiac arrest—are universally collected across trauma registries and are likely transferable to other healthcare environments. In contrast, variables reflecting early treatment processes may be influenced by local trauma workflows and resource availability, necessitating contextual adaptation and external validation before international application. Because the present model relies on structured variables derived from a nationwide trauma registry, its performance is inherently dependent on the completeness, consistency, and coding practices of the underlying data source. In data-sparse settings or healthcare environments with less comprehensive registry capture, predictive performance would be expected to decline. Therefore, external validation in independent trauma systems and prospective evaluation are essential prerequisites before clinical implementation. Addressing these methodological and practical considerations is essential to translate our predictive models into usable tools that improve trauma care in real-world settings.

This study provides theoretical and practical contributions to trauma care and ML-based clinical prediction. Theoretically, we built upon prior work by incorporating explainable AI techniques, such as SHAP, to enhance transparency in high-stakes clinical environments. Practically, we demonstrated that tree-based models can achieve strong discriminative performance using real-world trauma registry data, even under pandemic conditions. These findings provide a foundation for future research aimed at integrating AI-assisted risk stratification into emergency medical services and trauma systems. However, such integration will require additional steps, including calibration, external validation, and prospective implementation trials, before consideration of clinical application.

Conclusion

This retrospective study demonstrates the feasibility of using ML models, particularly XGB and RF, for system-level early mortality risk stratification in trauma patients based on large-scale national registry data. The high discrimination and temporal stability of the model, including consistent performance during the COVID-19 period, support the validity of this approach as a foundation for further development. However, the present study should be regarded as a model development and temporal validation study. Future research must include calibration analyses, external validation across multiple regions and healthcare systems, and prospective implementation trials before the integration of these models into clinical practice.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AI

Artificial Intelligence

AUROC

Area Under the Receiver Operating Characteristic Curve

AUPRC

Area Under the Precision-Recall Curve

CI

Confidence Interval

COVID-19

Coronavirus Disease 2019

DT

Decision Tree

EMS

Emergency Medical Services

ICU

Intensive Care Unit

KNN

K-Nearest Neighbor

LR

Logistic Regression

ML

Machine Learning

MLP

Multilayer Perceptron

NLR

Negative Likelihood Ratio

NPV

Negative Predictive Value

PLR

Positive Likelihood Ratio

PPV

Positive Predictive Value

RF

Random Forest

SHAP

Shapley Additive Explanations

XGB

Extreme Gradient Boosting

Author contributions

SMB: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing—original draft, Writing—review and editing. JGL: Validation, Writing—review and editing. HS: Validation, Writing—review and editing. SK: Validation, Writing—review and editing. KSH: Validation, Writing—review and editing. HY: Validation, Writing—review and editing. KHL: Data curation, Methodology, Validation, Visualization, Writing—review and editing. JML: Conceptualization, Supervision, Validation, Writing—review and editing.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(RS-2023-00240766).

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The Ewha Womans University Mokdong Hospital Institutional Review Board deemed this study exempt from ethics review and waived the informed consent requirement.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.