Machine learning models for predicting the risk of venous thromboembolism in ICU sepsis patients: a retrospective study

Sepsis is a complex, life-threatening syndrome associated with multiple organ dysfunction and high mortality. Patients with sepsis admitted to the intensive care unit (ICU) are at risk of various complications, among which venous thromboembolism (VTE) is one of the most common and severe. Traditional risk assessment tools have limited effectiveness in sepsis patients. In recent years, machine learning has been used to improve prediction accuracy in healthcare. Therefore, this study aims to develop a machine learning-based approach to better predict the risk of VTE in ICU sepsis patients. Data from ICU sepsis patients were retrospectively collected from the clinical database of the First Affiliated Hospital of Wenzhou Medical University (FAHWMU). Patients from 2021 to 2023 were divided into a training cohort (70%) and a validation cohort (30%), while patients from January to June 2024 were used as the test cohort. Univariate logistic regression was used to screen predictive factors for VTE, followed by Akaike Information Criterion (AIC)-based stepwise regression and least absolute shrinkage and selection operator (Lasso) for feature selection. Synthetic Minority Over-sampling Technique (SMOTE) was applied to balance the training cohort. Models constructed included logistic regression (LR), decision tree (DT), random forest (RF), support vector machine (SVM), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost). Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, positive predictive value (PPV), negative predictive value (NPV), true positive rate (TPR), true negative rate (TNR), accuracy (ACC), F1-score and Brier score. Out of 1,824 patients, 235 sepsis patients developed VTE during their ICU stay. Age, ICU length of stay, pre-ICU hospitalization duration, glucocorticoid use, analgesic use, and D-dimer levels were selected as features for predicting VTE. A comparison of the performance of six models (LR, DT, RF, SVM, GBM, XGBoost) in the validation and test cohorts indicated that the XGBoost model had superior discriminatory power. In the validation and test cohorts, the AUCs of XGBoost were 0.837 (95% CI 0.79–0.883) and 0.792 (95% CI 0.716–0.869), respectively. In the validation cohort, the TPR of XGBoost was 0.733, TNR was 0.759, PPV was 0.324, NPV was 0.948, ACC was 0.755, and F1-score was 0.449. In the test cohort, the TPR was 0.824, TNR was 0.702, PPV was 0.228, NPV was 0.974, ACC was 0.714, and F1-score was 0.357. The XGBoost model demonstrated superior discriminatory ability compared to other models and has the potential to assist clinical healthcare professionals in identifying high-risk VTE patients among those with sepsis in the ICU. Not applicable.