Construction and validation of scene data-based classification models for traumatic brain injury
10.3760/cma.j.cn501098-20250414-00210
- VernacularTitle:基于事故现场伤情数据创伤性脑损伤分级模型的构建及验证
- Author:
Jiaming WAN
1
;
Lin YANG
;
Hantao LI
;
Hongpeng YIN
;
Juxiang CHEN
;
Shengqing LYU
Author Information
1. 重庆大学生物工程学院，重庆 400030
- Publication Type:Journal Article
- Keywords:
Wounds and injuries;
Brain injuries;
Progressive patient care;
Artificial intelligence
- From:
Chinese Journal of Trauma
2025;41(6):587-593
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To construct classification models of traumatic brain injury (TBI) based on the injury data collected at the scene of the accidents and validate its efficacy.Methods:A retrospective cohort study was conducted to analyze the pre-hospital treatment data of 368 TBI patients admitted to the Second Affiliated Hospital of Army Military Medical University from January 2019 to December 2023, including 243 males and 125 females, aged 18-82 years [(48.1±20.8)years]. The patients′ Glasgow coma scale (GCS) scores were 3-15 points [11.0(3.0, 15.0)points] at emergency medical service arrival. The patients were randomly assigned to the training set ( n=257) and test set ( n=111) at a ratio of 7∶3. According to the admission diagnosis, the patients fell into the mild TBI group ( n=62), medium TBI group ( n=137), severe TBI group ( n=120), and extremely severe TBI group ( n=49). In the training set, 44 patients fell into mild TBI group, 98 into medium TBI group, 82 into severe TBI group and 33 into extremely severe TBI group, while in the test set, 18 patients fell into mild TBI group, 39 into medium TBI group, 38 into severe TBI group and 16 into extremely severe TBI group. The following 12 kinds of injury data, including MARCH [massive hemorrhage (M), airway obstruction (A), respiratory failure (R), circulatory failure (C) and hypothermia (H)], GCS, pre-hospital index (PHI), shock index (SI), reverse SI multiplied by GCS (rSIG), optic nerve sheath diameter (ONSD) measured by ultrasound, scalp and skull injuries were collected at the scene of the accidents. Three machine algorithm including random forest (RF), support vector machine (SVM) and logistic regression (LR) were used to construct scene data-based TBI classification models. The accuracy rate, precision rate, recall rate, F1 value and area under receiver operating characteristic (ROC) curve (AUC) of the 3 models were used to verify the efficiency of the models for TBI classification. Shapley additive explanations (SHAP) method was used to interpret the results of the optimal model. The 12 kinds of injury data in the models were sorted according to their contribution to the TBI classification and the injury data with greater contribution were selected. Results:In the test set, the accuracy rate of the RF, SVM and LR models was 0.93, 0.92 and 0.87, respectively; the precision rate was 0.93, 0.92 and 0.89, respectively; the recall rate was 0.93, 0.92 and 0.87, respectively; the F1 value was 0.93, 0.92 and 0.87, respectively. In the mild, medium, severe and extremely severe TBI groups in the test set, the AUC of the RF model was 0.96 (95% CI 0.92, 0.98), 0.98 (95% CI 0.94, 0.99), 0.97 (95% CI 0.95, 0.98), and 0.97 (95% CI 0.96, 0.98), respectively; the AUC of the SVM model was 0.90 (95% CI 0.88, 0.94), 0.95 (95% CI 0.92, 0.97), 0.96 (95% CI 0.94, 0.98), and 0.95 (95% CI 0.92, 0.99), respectively; the AUC of the LR model was 0.90 (95% CI 0.83, 0.96), 0.90 (95% CI 0.84, 0.95), 0.96 (95% CI 0.95, 0.98), and 0.95 (95% CI 0.94, 0.97), respectively. The RF model demonstrated optimal discriminative performance for TBI classification. As the SHAP′s interpretation of the RF model indicated, among the 12 kinds of injury data, those with greater contributions to the TBI classification were GCS, rSIG, SI, PHI, respiratory failure, ONSD, and circulatory failure in sequence. Conclusions:Of the scene data-based TBI classification models, the RF model achieves good predictive performance for TBI classification when compared with the SVM model and LR model. Besides, GCS, rSIG, SI, PHI, respiratory failure, ONSD and circulatory failure contribute significantly to the classification of TBI in the RF model, which may assist emergency medical personnel in field triage and management of TBI at accident scenes.
