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

Abstract

Background: Recent research has revealed the potential value of machine learning (ML) models in improving prognostic prediction for patients with trauma. ML can enhance predictions and identify which factors contribute the most to posttraumatic mortality. However, no studies have explored the risk factors, complications, and risk prediction of preoperative and postoperative traumatic coagulopathy (PPTIC) in patients with trauma.

Objective: This study aims to help clinicians implement timely and appropriate interventions to reduce the incidence of PPTIC and related complications, thereby lowering in-hospital mortality and disability rates for patients with trauma.

Methods: We analyzed data from 13,235 patients with trauma from 4 medical centers, including medical histories, laboratory results, and hospitalization complications. We developed 10 ML models in Python (Python Software Foundation) to predict PPTIC based on preoperative indicators. Data from 10,023 Medical Information Mart for Intensive Care patients were divided into training (70%) and test (30%) sets, with 3212 patients from 3 other centers used for external validation. Model performance was assessed with 5-fold cross-validation, bootstrapping, Brier score, and Shapley additive explanation values.

Results: Univariate logistic regression identified PPTIC risk factors as (1) prolonged activated partial thromboplastin time, prothrombin time, and international normalized ratio; (2) decreased levels of hemoglobin, hematocrit, red blood cells, calcium, and sodium; (3) lower admission diastolic blood pressure; (4) elevated alanine aminotransferase and aspartate aminotransferase levels; (5) admission heart rate; and (6) emergency surgery and perioperative transfusion. Multivariate logistic regression revealed that patients with PPTIC faced significantly higher 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), and urinary tract infection (1.95-fold), along with longer hospital and intensive care unit stays. Random forest was the most effective ML model for predicting PPTIC, achieving an area under the receiver operating characteristic of 0.91, an area under the precision-recall curve of 0.89, accuracy of 0.84, sensitivity of 0.80, specificity of 0.88, precision of 0.88, F₁-score of 0.84, and Brier score of 0.13 in external validation.

Conclusions: Key PPTIC risk factors include (1) prolonged activated partial thromboplastin time, prothrombin time, and international normalized ratio; (2) low levels of hemoglobin, hematocrit, red blood cells, calcium, and sodium; (3) low diastolic blood pressure; (4) elevated alanine aminotransferase and aspartate aminotransferase levels; (5) admission heart rate; and (6) the need for emergency surgery and transfusion. PPTIC is associated with severe complications and extended hospital stays. Among the ML models, the random forest model was the most effective predictor.

Trial registration: Chinese Clinical Trial Registry ChiCTR2300078097; https://www.chictr.org.cn/showproj.html?proj=211051.

Keywords: Medical Information Mart for Intensive Care; machine learning models; postoperative; preoperative; random forest; traumatic coagulopathy.

Figure 1

The data processing. MIMIC: Medical Information Mart for Intensive Care; ML: machine learning; PLA: People’s Liberation Army.

Figure 2

The AUROC curves of 10 machine learning models in the (A) training, (B) test, and (C) external validation set for predicting PPTIC in patients with trauma. The ROC curve for predicting PPTIC in patients with trauma is defined by the false positive rate on the x-axis, representing the proportion of negative cases misclassified as positive, and the true positive rate on the y-axis, reflecting the proportion of correctly identified positive cases (sensitivity), with both axes ranging from 0 to 1. The AUROC is a key metric for assessing the model’s predictive performance, where a higher AUROC signifies superior discriminatory ability in predicting PPTIC and improved generalizability. AUROC: area under the receiver operating characteristic curve; PPTIC: preoperative and postoperative traumatic coagulopathy; ROC: receiver operating characteristic curve; SVM: support vector machine.

Figure 3

The precision-recall curve and train data calibration curve of 10 machine learning models in (A-B) train, (C-D) internal test, and (E-F) external validation sets predicted patients with PPTIC. PPTIC: preoperative and postoperative traumatic coagulopathy; SVM: support vector machine.

Figure 4

SHAP values of (A) GB, (B) XGBoost, and (C) RF for predicting the risk of TIC before and after surgery in patients with trauma. We selected the three best-performing ML models (GB, XGBoost, and RF) from a set of 10 to predict PPTIC outcomes and evaluated their SHAP values. The SHAP plot highlights the top 15 features with the most significant impact on model predictions, ranked in descending order of importance from top to bottom. The x-axis represents the SHAP value, while the y-axis lists the corresponding feature names. Red points represent higher feature values, while blue points represent lower feature values. If red points are skewed toward the positive direction (higher SHAP values), it indicates that higher feature values contribute positively to the prediction outcome (increasing the predicted value). Conversely, if blue points are skewed toward the negative direction (lower SHAP values), it suggests that lower feature values contribute negatively to the prediction outcome (decreasing the predicted value). ALT: alanine aminotransferase; APTT: activated partial thromboplastin time; AST: aspartate aminotransferase; DBP: diastolic blood pressure; GB: gradient boosting; HR: heart rate; PPTIC: preoperative and postoperative traumatic coagulopathy; PT: prothrombin time; RBC: red blood cell; RF: random forest; SBP: systolic pressure; SHAP: Shapley additive explanation; TIC: traumatic coagulopathy.

Figure 5

Multivariate logistic regression of postoperative complications in patients with trauma who developed TIC both before and after surgery. LOS: length of stay; ICU: intensive care unit; OR: odds ratio; TIC: traumatic coagulopathy. P value˂.05 was considered statistically significant.