Ten Machine Learning Models for Predicting Preoperative and Postoperative Coagulopathy in Patients With Trauma: Multicenter Cohort Study

Introduction

Severe trauma is a significant global public health burden and a leading cause of reduced life expectancy. The Global Burden of Diseases, Injury, and Risk Factors Study 2017 estimates that trauma accounts for 8% of total deaths annually [GBD 2017 Causes of Death Collaborators. Global, regional, and national age-sex-specific mortality for 282 causes of death in 195 countries and territories, 1980-2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet. 2018;392(10159):1736-1788. [FREE Full text] [CrossRef] [Medline]1]. Traumatic coagulopathy (TIC) is a systemic inflammatory state accompanied by coagulation dysfunction, acidosis, and hypothermia that occurs after traumatic injury [Kornblith LZ, Moore HB, Cohen MJ. Trauma-induced coagulopathy: the past, present, and future. J Thromb Haemost. 2019;17(6):852-862. [FREE Full text] [CrossRef] [Medline]2]. TIC is observed in one-quarter to one-third of patients with trauma [Brohi K, Singh J, Heron M, Coats T. Acute traumatic coagulopathy. J Trauma. 2003;54(6):1127-1130. [CrossRef] [Medline]3,Maegele M, Lefering R, Yucel N, Tjardes T, Rixen D, Paffrath T, et al. Early coagulopathy in multiple injury: an analysis from the German Trauma Registry on 8724 patients. Injury. 2007;38(3):298-304. [CrossRef] [Medline]4] and is associated with increased rates of massive transfusion and multiple organ failure (MOF), prolonged intensive care unit (ICU) stay and length of stay (LOS), and a 4-fold increase in mortality [Brohi K, Singh J, Heron M, Coats T. Acute traumatic coagulopathy. J Trauma. 2003;54(6):1127-1130. [CrossRef] [Medline]3]. Harhangi et al [Harhangi BS, Kompanje EJO, Leebeek FWG, Maas AIR. Coagulation disorders after traumatic brain injury. Acta Neurochir. 2008;150(2):165-175. [CrossRef] [Medline]5] reported that TIC is linked to a 9-fold increase in mortality and a 36-fold increase in the likelihood of adverse outcomes [Harhangi BS, Kompanje EJO, Leebeek FWG, Maas AIR. Coagulation disorders after traumatic brain injury. Acta Neurochir. 2008;150(2):165-175. [CrossRef] [Medline]5]. Early endogenous TIC is associated with increased bleeding, higher red blood cell (RBC) transfusion rates, and a greater risk of secondary complications like multiple organ dysfunction syndrome (MODS) and thromboembolism [Duque P, Calvo A, Lockie C, Schöchl H. Pathophysiology of trauma-induced coagulopathy. Transfus Med Rev. 2021;35(4):80-86. [CrossRef] [Medline]6]. Cohen and Christie [Cohen MJ, Christie SA. Coagulopathy of trauma. Crit Care Clin. 2017;33(1):101-118. [CrossRef] [Medline]7] also reported that patients with TIC have worse prognoses, including higher rates of infection, thromboembolism, acute lung injury, MOF, and mortality. Without proper and timely diagnosis and treatment, posttraumatic bleeding and associated TIC remain potentially preventable causes of MOF [Cole E, Weaver A, Gall L, West A, Nevin D, Tallach R, et al. A decade of damage control resuscitation: new transfusion practice, new survivors, new directions. Ann Surg. 2021;273(6):1215-1220. [CrossRef] [Medline]8]. Therefore, the latest European guidelines indicate that the immediate identification and management of TIC can improve outcomes for severely injured patients [Rossaint R, Afshari A, Bouillon B, Cerny V, Cimpoesu D, Curry N, et al. The European guideline on management of major bleeding and coagulopathy following trauma: sixth edition. Crit Care. 2023;27(1):80. [FREE Full text] [CrossRef] [Medline]9].

In recent years, artificial intelligence (AI) has received significant attention for its potential use in various aspects of human activities, including health care [National Academy of Medicine. Whicher D, Ahmed M, Thadaney Israni S, Matheny M, editors. Artificial Intelligence in Health Care: The Hope, the Hype, the Promise, the Peril. Washington, DC. National Academies Press; 2023. 10]. Sidey-Gibbons and Sidey-Gibbons [Sidey-Gibbons JAM, Sidey-Gibbons CJ. Machine learning in medicine: a practical introduction. BMC Med Res Methodol. 2019;19(1):64. [FREE Full text] [CrossRef] [Medline]11] and Liu and Salinas [Liu NT, Salinas J. Machine learning for predicting outcomes in trauma. Shock. 2017;48(5):504-510. [CrossRef] [Medline]12] suggested that machine learning (ML) excels over traditional methods in handling large, unstructured, nonlinear, or incomplete datasets, improving outcome predictions for patients with trauma. With the advent of ML and improved computational analysis methods, recent approaches have aimed to enhance predictions and identify which factors are most important in contributing to posttraumatic mortality [Seki T, Kawazoe Y, Ohe K. Machine learning-based prediction of in-hospital mortality using admission laboratory data: a retrospective, single-site study using electronic health record data. PLoS One. 2021;16(2):e0246640. [FREE Full text] [CrossRef] [Medline]13]. Rashidi et al [Rashidi HH, Bowers KA, Reyes Gil M. Machine learning in the coagulation and hemostasis arena: an overview and evaluation of methods, review of literature, and future directions. J Thromb Haemost. 2023;21(4):728-743. [FREE Full text] [CrossRef] [Medline]14] emphasized the critical importance of rapid diagnosis and intervention for managing hemostatic disorders, which can lead to life-threatening bleeding or clotting. AI or ML models can significantly enhance turnaround times, accuracy, and the timeliness of diagnoses and interventions, potentially improving patient care and outcomes.

Our data show that despite surgical interventions such as hemostasis, vascular reconstruction, fracture reduction, organ repair, blood transfusion, and the correction of electrolyte imbalances in patients who experienced preoperative TIC, 50% of these patients still had TIC postoperatively [Johnson A, Bulgarelli L, Pollard T, Gow B, Moody B, Horng S, et al. MIMIC-IV. PhysioNet. 2021. [CrossRef]15]. These patients were more prone to develop complications during hospitalization, and preoperative and postoperative traumatic coagulopathy (PPTIC) was identified as an independent risk factor for these complications. Therefore, we plan to explore the risk factors for PPTIC in patients with trauma and its relationship with in-hospital complications. We will use admission indicators from patients with trauma to predict the risk of PPTIC, applying 10 different ML models to estimate the probability of PPTIC occurrence and conducting a comprehensive evaluation of each model. This predictive approach can aid clinicians in implementing timely and appropriate interventions to reduce the incidence of PPTIC and related complications, thereby lowering in-hospital mortality and disability rates for these patients.

Methods

Ethical Considerations

Data from 4 medical centers were anonymized and integrated. Ethical approvals were granted by the respective ethics committees: Daping Hospital, Medical Research Ethics (2023) 261; Chongqing Emergency Medical Center, 2023 ethical review (48); the People’s Liberation Army Rocket Force Characteristic Medical Center (KY 2023037); and the General Hospital of Southern Theatre Command of People’s Liberation Army (NZLLKZ2024021). As this was a retrospective study, patient informed consent was waived by the ethics committees. It was registered with the Chinese Clinical Trial Registry (ChiCTR2300078097) and adhered to the Declaration of Helsinki. This study adheres to 2 key guidelines to ensure rigorous reporting: the TRIPOD-AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis—Artificial Intelligence) guidelines for the development and reporting of multivariable ML predictive models in biomedical research (

Multimedia Appendix 1

The TRIPOD-AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis—Artificial Intelligence) guidelines.

Inclusion and Exclusion Criteria

The inclusion and exclusion criteria are as follows:

• Inclusion criteria: Patients who underwent surgical treatment after trauma from January 2017 to February 2023.
• Exclusion criteria: Patients without trauma (snake bites, bee stings, and burns); patients who did not undergo surgery or stopped treatment; patients younger than 18 years of age; patients with hematological diseases; patients with a history of liver cirrhosis; pregnant women; and patients without coagulation-related tests results either preoperatively or postoperatively.

Data Collection

We retrieved patients who underwent trauma or concurrent surgeries using the MIMIC IV (Medical Information Mart for Intensive Care) 2.2 database, with the detailed data extraction process outlined in

Multimedia Appendix 2

A comprehensive description of the methodology.

Process of ML Modeling

Handling Missing Data

We processed the data using Python (version 3.9; Python Software Foundation), discarding variables with over 40% missing data (

Multimedia Appendix 3

The amount of data missing and the coding process.

Data Normalization and Feature Selection

Data Normalization

We created a StandardScaler object to standardize the data. Both the training and test data were standardized via the same procedure.

Outlier Handling

We calculated z scores for the training data, considering values above 3 as outliers, and replaced these with the mean using np.where.

Feature Selection

We assessed continuous variables with the Pearson correlation coefficient and converted categorical variables to numerical ones via label encoding. Then, the Pearson correlation coefficient was calculated for each feature relative to the target variable, and the top 50 features with the highest correlations were selected. These selected features were then refined using least absolute shrinkage and selection operator regression, resulting in a final set of 25 features.

Dataset Splitting

We modeled data from 10,023 patients with trauma in MIMIC IV, splitting it into a training set of 7631 cases and a test set of 1908 cases (7:3 ratio) using the train_test_split function. Additionally, we used data from 3212 patients from 3 other medical centers for external validation (Figure 1).

‎

Building the ML Models

We used Python 3.9 to build the ML models, using a total of 10 models (

Multimedia Appendix 4

Sklearn packages for machine learning models.

Model Performance and Evaluation

In this study, we used “grid search” for hyperparameter optimization. We conducted 5-fold cross-validation by dividing the dataset into 5 subsets, using each subset once for testing and the remaining four for training. This process, repeated 5 times, helps reduce overfitting, stabilize performance evaluation, and address sample selection bias [Yu M, Wang S, He K, Teng F, Deng J, Guo S, et al. Predicting the complexity and mortality of polytrauma patients with machine learning models. Sci Rep. 2024;14(1):8302. [FREE Full text] [CrossRef] [Medline]20]. We measured accuracy, precision, recall, and F₁-score.

To evaluate the robustness and generalization of 10 ML models, we used bootstrap sampling with 1000 iterations. Each iteration involved retraining the models on a bootstrap sample and calculating accuracy, precision, recall, and F₁-score. We then estimated the distributions and 95% CIs for these metrics, producing 12 key indicators: accuracy mean, accuracy lower or upper 95% CI, precision mean, precision lower or upper 95% CI, recall mean, recall lower or upper 95% CI, F₁-score mean, and F₁-score lower or upper 95% CI.

To assess model calibration, we used the Brier score (BS), which ranges from 0 to 1, with lower scores indicating better accuracy of predicted probabilities. We trained the model, evaluated predictions on the test set, computed the BS, and plotted a calibration curve to visualize the model’s performance.

Interpretation of the ML

We use Shapley additive explanation (SHAP) values to assess the significance of variables in the model by evaluating their impact on predictions across all feature subsets and averaging these effects [Wang K, Tian J, Zheng C, Yang H, Ren J, Liu Y, et al. Interpretable prediction of 3-year all-cause mortality in patients with heart failure caused by coronary heart disease based on machine learning and SHAP. Comput Biol Med. 2021;137:104813. [FREE Full text] [CrossRef] [Medline]21]. We used the Python package SHAP to visualize and rank the top 15 variables with the highest SHAP values related to PPTIC. These variables were integrated into the model, and SHAP values ranging from 0 to 1 were used for risk interpretation. We also conducted univariate logistic regression on these top 15 variables.

Statistics

Statistical analysis was performed using SPSS Statistics (version 27.0; IBM Corp). Patients were divided into PPTIC and non-PPTIC groups, with PPTIC as the independent variable and other factors as dependent variables. Categorical data are presented as percentages, and continuous data as mean (SD). Depending on the distribution, continuous variables were assessed via 2-tailed t tests or Mann-Whitney U tests, and categorical variables via chi-square tests or exact tests. Univariate analysis identified risk factors for persistent trauma-induced coagulopathy, and multivariable logistic regression analyzed complications during hospitalization in the PPTIC group. A P value less than .05 was considered statistically significant.

Results

Basic Information of the Patients

We ultimately included 13,237 patients with trauma, among whom 6586 (49.8%) experienced PPTIC. The mean age of our patients was 62.45 (SD 17.85) years. In the PPTIC group, the average age was 63.6 (SD 17.4) years, whereas in the non-PPTIC group, it was 61.3 (SD 18.2) years, with P<.001 (Table 1). Among the patients, there were 5590 (42.2%) male individuals and 7646 (57.8%) female individuals, and this difference was significant (P<.001). The racial distribution included 3452 (26.1%) Asian individuals, 7402 (55.9%) White individuals, 342 (2.6%) Black individuals, 940 (7.1%) Latino individuals, and 1101(8.3%) individuals of other races.

Results of the ML Models

Training Results of the ML Models

We developed 10 ML risk models for predicting PPTIC. In the training set, all 10 models achieved an AUROC of 0.83 or higher, with RF, DT, and XGBoost reaching AUROC values of 1. In the internal testing set, RF achieved an excellent AUROC of 0.92, whereas the AUROC of SVM was 0.9, that of GB was 0.93, that of NN was 0.9, that of AdaBoost was 0.91, and XGBoost performed well (Figure 2). RF, GB, and XGBoost demonstrated superior performance in both the training and testing sets, with AUROC, accuracy, sensitivity, specificity, precision, and F₁-score all above 0.8 (

Multimedia Appendix 6

Ten machine learning models predicting patients with trauma who have preoperative and postoperative traumatic coagulopathy and related performance evaluation.

‎

Evaluating the ML Models

Results of the 5-Fold Cross-Validation

The results from the 5-fold cross-validation (

Multimedia Appendix 7

Results of 5-fold cross-validation for 10 machine learning models.

Bootstrap Results

The bootstrap analysis results indicated that RF, DT, GB, AdaBoost, and XGBoost achieved average accuracy, precision, recall, and F₁-score all exceeding 0.91, accompanied by narrow 95% CIs (Table 2). This suggests consistently high accuracy across various sample resampling conditions, underscoring the robustness of these models. Moreover, RF, DT, GB, AdaBoost, and XGBoost demonstrated substantial stability and reliability across their respective performance metrics.

Calibration and Precision-Recall of the ML Models

In the training set, RF, DT, GB, and XGBoost achieved BSs of 0.01, 0.00, 0.05, and 0.00, respectively, with precision-recall values above 0.9, demonstrating robustness. In the internal testing set, RF, GB, and XGBoost had BSs of 0.1, 0.09, and 0.10, and precision-recall values of 0.87, 0.87, and 0.86. During external validation, they recorded BSs of 0.13, 0.13, and 0.15, with precision-recall values around 0.88, 0.84, and 0.87. The BS and precision-recall values of RF, GB, and XGBoost in the external validation set closely mirrored those in the internal testing set (Figure 3), underscoring the consistency and reliability of these models.

‎

Results of the External Validation of the ML Models

In the external validation set, RF, GB, NN, and XGBoost achieved AUROC values of 0.91, 0.89, 0.89, and 0.89, respectively (Figure 2), while other models attained approximately 0.87. RF showed the highest performance with AUROC 0.91 (95% CI 0.90-0.913), AUPRC 0.89 (95% CI 0.88-0.90), accuracy 0.84 (95% CI 0.83-0.85), sensitivity 0.80 (95% CI 0.79-0.81), specificity 0.88 (95% CI 0.87-0.89), precision 0.88 (95% CI 0.87-0.89), and F₁-score 0.84 (95% CI 0.83-0.84). Despite being slightly lower than in the training set, RF’s performance remained stable, demonstrating strong generalizability. Thus, RF was identified as the optimal model based on 5-fold cross-validation, bootstrap, calibration, precision-recall, and external validation results.

Interpretation of the ML Models

Analysis of SHAP values highlighted the key features influencing model predictions. For RF, GB, and XGBoost (Figure 4), as well as the other 7 ML models (

Multimedia Appendix 8

SHAP values of AdaBoost, DT, KNN, LR, NB, NN, and SVM for predicting the risk of TIC before and after surgery in patients with trauma. DT: decision tree; KNN: k-nearest neighbor; LR: logistic regression; NB: naive Bayes; NN: neural network; SVM: support vector machine; SHAP: Shapley additive explanation; TIC: traumatic coagulopathy.

Table 3

‎

Complications During Hospitalization in Patients With PPTIC

Multivariate logistic regression analysis showed that compared with patients with non-PPTIC, those with PPTIC had significantly higher postoperative risks of the following complications: sepsis (1.75-fold, 95% CI 1.18-2.6; P=.006), heart failure (1.5-fold, 95% CI 1.24-1.82; P<.001), delirium (3.08-fold, 95% CI 1.62-5.83; P=.001), abnormal coagulation (3.57-fold, 95% CI 2.26-5.63; P<.001), tracheostomy (2.76-fold, 95% CI 2.02-3.77, P=.001), in-hospital mortality (2.19-fold, 95% CI 1.84-2.61; P<.001), urinary tract infection (1.95-fold, 95% CI 1.46-2.59; P<.001), and longer LOS and ICU stay (Figure 5).

‎

Discussion

ML Models for Predicting PPTIC in Patients With Trauma

Our study found that RF, GB, and XGBoost had AUROC values above 0.89 for predicting PPTIC, with RF performing best at 0.91. RF showed excellent performance across training, testing, and external validation sets. RF is an ML algorithm capable of capturing potential nonlinear relationships between risk factors and outcomes [Uddin S, Khan A, Hossain ME, Moni MA. Comparing different supervised machine learning algorithms for disease prediction. BMC Med Inform Decis Mak. 2019;19(1):281. [FREE Full text] [CrossRef] [Medline]22]. It operates by creating hundreds or thousands of smaller decision trees and combining their outputs for prediction [Badillo S, Banfai B, Birzele F, Davydov II, Hutchinson L, Kam-Thong T, et al. An introduction to machine learning. Clin Pharmacol Ther. 2020;107(4):871-885. [FREE Full text] [CrossRef] [Medline]23]. This approach helps overcome the tendency of decision trees to overfit [Badillo S, Banfai B, Birzele F, Davydov II, Hutchinson L, Kam-Thong T, et al. An introduction to machine learning. Clin Pharmacol Ther. 2020;107(4):871-885. [FREE Full text] [CrossRef] [Medline]23] while enhancing model generalizability [Alderden J, Pepper GA, Wilson A, Whitney JD, Richardson S, Butcher R, et al. Predicting pressure injury in critical care patients: a machine-learning model. Am J Crit Care. 2018;27(6):461-468. [FREE Full text] [CrossRef] [Medline]24]. Therefore, we recommend the use of the RF to predict the occurrence of PPTIC in patients with trauma. The prediction variables were primarily continuous laboratory indicators, eliminating the need for conversion into binary or categorical forms and facilitating practical use. Physicians can directly input these indicators into the RF model to assess the risk of PPTIC in patients with trauma. Recent guidelines highlight the importance of early detection and management of TIC to improve outcomes for severely injured patients [Rossaint R, Afshari A, Bouillon B, Cerny V, Cimpoesu D, Curry N, et al. The European guideline on management of major bleeding and coagulopathy following trauma: sixth edition. Crit Care. 2023;27(1):80. [FREE Full text] [CrossRef] [Medline]9]. Our study aims to predict PPTIC using preoperative indicators to facilitate early intervention and thereby improve clinical outcomes in such patients.

Most robust PPTIC prediction models in our study are tree-based, such as RF, DT, GB, and XGBoost, a trend also seen in other coagulopathy-related ML models. Perkins et al [Perkins ZB, Yet B, Marsden M, Glasgow S, Marsh W, Davenport R, et al. Early identification of trauma-induced coagulopathy: development and validation of a multivariable risk prediction model. Ann Surg. 2021;274(6):e1119-e1128. [CrossRef] [Medline]25] used BN to predict early TIC, defined as an INR>1.2. Their model, trained on 600 patients with trauma, achieved an AUROC of 0.93 with a BS of 0.06 in the development dataset and an AUROC of 0.95 with a BS of 0.05 in the validation set [Perkins ZB, Yet B, Marsden M, Glasgow S, Marsh W, Davenport R, et al. Early identification of trauma-induced coagulopathy: development and validation of a multivariable risk prediction model. Ann Surg. 2021;274(6):e1119-e1128. [CrossRef] [Medline]25]. In comparison, Our RF model achieved an AUROC of 1 in training, 0.92 in validation, and 0.91 in external validation, with BS values of 0.01, 0.10, and 0.13, respectively. Li et al [Li K, Wu H, Pan F, Chen L, Feng C, Liu Y, et al. A machine learning-based model to predict acute traumatic coagulopathy in trauma patients upon emergency hospitalization. Clin Appl Thromb Hemost. 2020;26:1076029619897827. [FREE Full text] [CrossRef] [Medline]26] defined acute traumatic coagulopathy (ATC) as an INR>1.5 and collected data from 818 patients with trauma admitted to the emergency department. They used LR and RF to predict ATC occurrence, with LR achieving an AUROC of 0.858 and RF achieving an AUROC of 0.830 in the internal validation set [Peltan ID, Rowhani-Rahbar A, Vande Vusse LK, Caldwell E, Rea TD, Maier RV, et al. Development and validation of a prehospital prediction model for acute traumatic coagulopathy. Crit Care. 2016;20(1):371. [FREE Full text] [CrossRef] [Medline]27]. Their study used only 2 ML models, with RF and LR achieving moderate predictive performance for ATC. Peltan et al [Peltan ID, Rowhani-Rahbar A, Vande Vusse LK, Caldwell E, Rea TD, Maier RV, et al. Development and validation of a prehospital prediction model for acute traumatic coagulopathy. Crit Care. 2016;20(1):371. [FREE Full text] [CrossRef] [Medline]27] predicted the occurrence of prehospital ATC in 285 severely injured patients, defining coagulopathy as an INR>1.5, with an AUROC value of only 0.73. Previous studies had small sample sizes and moderate predictive power, and were limited to single centers without external validation. In contrast, our multicenter study employed 10 ML models, with 3 demonstrating excellent performance. It included diverse medical centers and populations (White, Asian, Black, and Latino) and featured external validation, showcasing the models’ broad applicability and generalizability.

Preoperative Risk Factors and In-Hospital Complications of PPTIC in Patients With Trauma

The definition of TIC relies primarily on coagulation tests, drawing from extensive literature on the subject. Brohi et al [Brohi K, Singh J, Heron M, Coats T. Acute traumatic coagulopathy. J Trauma. 2003;54(6):1127-1130. [CrossRef] [Medline]3] recommended that patients with a PT > 18 s, an INR > 1.5, an APTT>60 s, or a TT>15 s could be diagnosed with the hypocoagulable phenotype of TIC, and this diagnosis was not affected by the use of anticoagulant medications or abnormal blood samples. MacLeod et al [MacLeod JBA, Lynn M, McKenney MG, Cohn SM, Murtha M. Early coagulopathy predicts mortality in trauma. J Trauma. 2003;55(1):39-44. [CrossRef] [Medline]28] predicted early coagulopathy in 20,103 patients with trauma and reported a higher incidence after trauma, which was associated with mortality. They defined early coagulopathy as a PT>14 s, an APTT>35 s, and a PLT<100×10^9/L, highlighting its role as an independent predictor of death despite other risk factors [MacLeod JBA, Lynn M, McKenney MG, Cohn SM, Murtha M. Early coagulopathy predicts mortality in trauma. J Trauma. 2003;55(1):39-44. [CrossRef] [Medline]28]. Currently, the accepted definition of TIC utilizes the INR, with an INR>1.2 being the threshold for detecting TIC [Cole E, Weaver A, Gall L, West A, Nevin D, Tallach R, et al. A decade of damage control resuscitation: new transfusion practice, new survivors, new directions. Ann Surg. 2021;273(6):1215-1220. [CrossRef] [Medline]8]. Yuan et al [Yuan Q, Yu J, Wu X, Sun YR, Li ZQ, Du ZY, et al. Prognostic value of coagulation tests for in-hospital mortality in patients with traumatic brain injury. Scand J Trauma Resusc Emerg Med. 2018;26(1):3. [FREE Full text] [CrossRef] [Medline]18] defined TIC for predicting mortality among patients with trauma as a PLT<100×10⁹/L, an INR>1.25, a PT>14 seconds, and an APTT>36 seconds [Yuan Q, Yu J, Wu X, Sun YR, Li ZQ, Du ZY, et al. Prognostic value of coagulation tests for in-hospital mortality in patients with traumatic brain injury. Scand J Trauma Resusc Emerg Med. 2018;26(1):3. [FREE Full text] [CrossRef] [Medline]18]. We defined TIC based on coagulation test reference ranges from various centers and literature, using slightly lower standards than Brohi et al [Brohi K, Singh J, Heron M, Coats T. Acute traumatic coagulopathy. J Trauma. 2003;54(6):1127-1130. [CrossRef] [Medline]3]. The main reasons for adopting these criteria are as follows: (1) lower diagnostic criteria alert clinicians to take early preventive measures to prevent PPTIC from progressing to disseminated intravascular coagulation; and (2) patients with preexisting TIC have a 50% chance of persisting with TIC postoperatively.

The prolongation of the preoperative APTT, PT, and INR in patients with trauma is understandable as a risk factor for PPTIC. The decreases in hemoglobin, hematocrit, and RBC levels and lower admission DBP are associated mainly with acute anemia due to preoperative bleeding in patients with trauma. Ryan et al [Ryan ML, Thorson CM, Otero CA, Vu T, Schulman CI, Livingstone AS, et al. Initial hematocrit in trauma: A paradigm shift? J Trauma Acute Care Surg. 2012;72(1):54-59. [CrossRef] [Medline]29] noted a close correlation between initial low hematocrit or hemoglobin levels in patients with trauma and hemorrhagic shock. RBC and hematocrit levels are key factors in platelet transport and adhesion to endothelial injury. Platelet adhesion increases 5-fold when hematocrit rises from 10% to 40% [Turitto VT, Weiss HJ. Red blood cells: their dual role in thrombus formation. Science. 1980;207(4430):541-543. [CrossRef] [Medline]30]. Therefore, the RBC level is closely related to the platelet level. Steele et al [Steele T, Kolamunnage-Dona R, Downey C, Toh C, Welters I. Assessment and clinical course of hypocalcemia in critical illness. Crit Care. 2013;17(3):R106. [FREE Full text] [CrossRef] [Medline]31] reported an independent correlation between hyponatremia at admission and hypocalcemia. Ionic calcium is crucial not only for the formation and stability of fibrin polymerization sites but also for many platelet-related functions [Lier H, Krep H, Schroeder S, Stuber F. Preconditions of hemostasis in trauma: a review. The influence of acidosis, hypocalcemia, anemia, and hypothermia on functional hemostasis in trauma. J Trauma. 2008;65(4):951-960. [CrossRef] [Medline]32]. Platelet activation and aggregation are also calcium-dependent [Matthay ZA, Fields AT, Nunez-Garcia B, Patel MH, Cohen MJ, Callcut RA, et al. Dynamic effects of calcium on in vivo and ex vivo platelet behavior after trauma. J Trauma Acute Care Surg. 2020;89(5):871-879. [FREE Full text] [CrossRef] [Medline]33] and play crucial roles in the pathophysiology of TIC [Vulliamy P, Gillespie S, Armstrong PC, Allan HE, Warner TD, Brohi K. Histone H4 induces platelet ballooning and microparticle release during trauma hemorrhage. Proc Natl Acad Sci U S A. 2019;116(35):17444-17449. [FREE Full text] [CrossRef] [Medline]34]. Elevated ALT and AST levels indicate liver damage, which impacts coagulation factor production and the coagulation process.

Hess et al [Hess JR, Brohi K, Dutton RP, Hauser CJ, Holcomb JB, Kluger Y, et al. The coagulopathy of trauma: a review of mechanisms. J Trauma. 2008;65(4):748-754. [CrossRef] [Medline]35] reported that patients with TIC experience more blood loss, higher transfusion needs, and a greater incidence of MODS compared to those without TIC. Brohi et al [Brohi K, Singh J, Heron M, Coats T. Acute traumatic coagulopathy. J Trauma. 2003;54(6):1127-1130. [CrossRef] [Medline]3] highlighted that TIC is associated with a 4-fold increase in mortality, increased rates of massive transfusion and MODS, and prolonged ICU and LOS. Smalls et al [Smalls N, Obirieze A, Ehanire I. The impact of coagulopathy on traumatic splenic injuries. Am J Surg. 2015;210(4):724-729. [CrossRef] [Medline]36] reported that patients with TIC due to splenic injury face 2.4 times more complications, with increased mortality (1.3-fold), sepsis (2-fold), acute respiratory distress syndrome (2.6-fold), acute renal failure (1.5-fold), and cardiac arrest (1.5-fold). These findings align with our study, where patients with PPTIC had increased risks of sepsis (1.75-fold), heart failure (1.5-fold), delirium (3.08-fold), abnormal coagulation (3.57-fold), tracheostomy (2.76-fold), mortality (2.19-fold), urinary tract infection (1.95-fold), and longer LOS and ICU stay.

Limitations

This retrospective study focused primarily on predicting PPTIC in patients with trauma via admission laboratory indicators and medical records. In this study, we addressed missing data to varying extents by excluding variables with more than 40% missing values and applying distinct imputation methods for variables with less than 40% missing data (KNN for continuous variables and multiple imputation for categorical variables). While multiple imputation is an effective method to mitigate systematic bias from nonrandom missing data, some data remained missing, which could affect the model’s predictive accuracy for specific patient characteristics. Due to missing values exceeding 40% for certain indicators, such as height, admission peripheral capillary oxygen saturation, temperature on admission, pre-lymphocyte, pre-total protein, pre-globulin, pre-direct bilirubin, pre-lactate dehydrogenase, pre-D-dimer, and post-D-dimer (

Multimedia Appendix 3

The amount of data missing and the coding process.

Conclusions

Our study, using data from 4 medical centers, identified risk factors for PPTIC in patients with trauma. Significant factors included (1) prolonged preoperative APTT, PT, and INR; (2) decreased hemoglobin, hematocrit, RBC, calcium, and sodium levels; (3) lower admission DBP; (4) elevated ALT and AST levels; (5) increased admission HR; and (6) the need for emergency surgery or perioperative transfusion. PPTIC was associated with higher rates of sepsis, heart failure, delirium, abnormal coagulation, tracheostomy, mortality, and extended hospital and ICU stays. Of the 10 ML models evaluated for predicting PPTIC, RF demonstrated the best performance with an AUROC of 0.91, AUPRC of 0.89, accuracy of 0.84, and a BS of 0.13.