基于机器学习的创伤伤员检伤分类预测模型构建及验证

张睿智; 罗瑞虹; 卢志林; 李春平; 卢兵; 邢家溢; 黎檀实

doi:10.12435/j.issn.2095-5227.2024.003

基于机器学习的创伤伤员检伤分类预测模型构建及验证

Construction and validation of a machine learning-based predictive model for trauma casualty triage

摘要

摘要:
背景创伤现场批量伤员的现场检伤分类是现场急救中的关键环节，探索研究如何更加高效准确对伤员进行检伤分类具有重要意义。
目的基于生命体征数据和机器学习算法建立并验证创伤伤员检伤分类预测模型。
方法回顾性分析美国创伤数据库2017至2019年的院前急救创伤伤员数据，采用支持向量机(Support Vector Machine，SVM)、随机森林(Random Forest)、梯度提升决策树(Gradient Boosting Decision Tree，GBDT)、极端梯度上升(eXtreme Gradient Boosting，XGBoost)和多层感知机(Multi-Layer Perceptron，MLP)5种机器学习算法开发创伤伤员检伤分类预测模型并验证。采用准确率(Accuracy)、精准度(Precision)、召回率(Recall)、F1值(F1 Score)和AUC值(ROC曲线下面积)进行结果评价，使用ROC曲线进行可视化，并对最优模型结果在解放军总医院第一医学中心急诊创伤数据集中进行验证。
结果共选取伤员数据24 948条，其中轻伤9 496例，中等伤9 532例，重伤5 496例，危重伤424例。基于ISS分级标准，通过ROC曲线分析显示，相较于其他四种模型，GBDT算法效果最好，准确率82.63%，精确度68.21%，召回率60.92%，F1值61.91%，AUC 90.38%。在解放军总医院第一医学中心急诊创伤数据集中验证结果，准确率83.15%，精确度77.38%，召回率59.89%，F1值55.26%，AUC值90.38%。
结论本研究成功开发并验证了一组检伤分类机器学习预测模型，未来可应用于创伤伤员现场检伤分类辅助决策。

Abstract:
Background On-site triage of lot victims at the scene of the trauma is an essential link in first aid, and it is important to study how to triage casualty more effectively and accurately.
Objective To develop and validate a predictive model for trauma casualty triage based on vital signs data and machine learning algorithms.
Methods A retrospective analysis of pre-hospital emergency trauma casualty data from 2020 to 2021 in the National Trauma Data Bank (NTDB) was performed using five types of models, including Support Vector Machine (SVM), Random Forest, Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP) to develop and validate the predictive model for trauma casualty detection and classification. The results were evaluated using Accuracy, Precision, Recall, F1 Score and AUC (area under the ROC curve), and visualized using the ROC curve. The results of the optimal model were also validated in the trauma database of the Emergency Department of the First Medical Centre of Chinese PLA General Hospital.
Results A total of 24 948 records of the injured were collected, including 9 496 cases of mild injuries, 9 532 cases of moderate injuries, 5 496 cases of serious injuries, and 424 cases of critical injuries. Based on the ISS grading criteria, the ROC curve analysis showed that the GBDT algorithm was the most effective compared to the other four models, with an accuracy of 82.63%, a precision of 68.21%, a recall of 60.92%, an F1 value of 61.91%, and an AUC value of 90.38%. In the validation results in the trauma database of the Emergency Department of the First Medical Centre of Chinese PLA General Hospital, the accuracy reached 83.15%, the precision reached 77.38%, the recall reached 59.89%, the F1 value reached 55.26%, and the AUC value reached 90.38%.
Conclusion We have successfully developed and validated a set of machine learning predictive models for triage of injuries, which can be applied to assist decision-making for on-site triage of trauma injuries in the future.

HTML全文

参考文献(29)

施引文献

资源附件(0)