A comparative study on early prediction of venous thromboembolism in patients with traumatic brain injury by machine learning model.

objectiveWe aimed to evaluate the predictive value of the post-injury D-dimer decrease rate for venous thromboembolism (VTE) in patients with traumatic brain injury (TBI). Additionally, we sought to establish a practical and efficient prediction model using a machine-learning algorithm to facilitate the early identification of high-risk individuals for VTE following TBI.

methodsThis study encompassed patients over the age of 18 with TBI who were admitted to our trauma center, between May 2018 and December 2021. The participants were allocated into training (70%) and validation (30%) cohorts. Within the training cohort, predictive models were developed using the generalized linear model (GLM), least absolute shrinkage and selection operator model (LSM), and random forest model (RFM), based on the clinical characteristics of the patients. The predictive accuracy of these models was assessed through the area under the receiver operating characteristic curve (AUROC). The stability and clinical practicability of the models were evaluated using a calibration curve and a clinical impact curve. The repeatability and reliability of the models were confirmed through a validation dataset.

resultsA total of 1,108 patients aged over 18 years with TBI who met the inclusion criteria were included in this study. Post-injury D-dimer on the third day (PDD3) and the post-injury D-dimer decreasing rate on the third day (PDDR3) were common predictors across the three models and were closely related to VTE for patients with TBI. The area under the receiver operating characteristic curve (AUROC) for the GLM, LSM, and RFM in the training cohort were 0.84 (95% confidence interval [CI]: 0.80-0.87), 0.85 (95% CI: 0.82-0.88), and 0.82 (95% CI: 0.78-0.86), respectively. In the verification cohort, the AUROC values were 0.85 (95% CI: 0.79-0.90), 0.85 (95% CI: 0.79-0.91), and 0.79 (95% CI: 0.73-0.86), respectively. The calibration curves of the three prediction models agree well with the actual observed results. All models showed a high clinical net income in the decision and clinical impact curves.

conclusionPDD3 and PDDR3 emerged as effective indices for predicting VTE in patients with TBI. We formulated a practical predictive model based on PDDR3, demonstrating satisfactory performance in forecasting VTE, which will assist clinicians in the early identification and initiation of PTP treatment for TBI patients.