基于机器学习的战现场伤员预警模型建立与评价

Evaluation and Establishment of Casualty Warning Model in Battefield Based on Machine Learning

目的为实现战现场对重伤伤员进行预警的需求,以达到分诊前移并指导救治顺序,提高救治效率。方法本研究从某军队医院急诊数据库和美国建立的一个对外开放的多参数临床重症监护数据库(Medical Information Mart for Intensive CareⅢ,MIMICⅢ)中分别收集1469、1464名患者数据,根据战现场可行性与科学性分析筛选出7个代表性的指标作为训练模型的输入特征,基于XGBOOST进行特征权重排序筛选出最优特征组合,并进一步分析所选取的特征对预测结果的影响程度。利用随机森林、支持向量机、朴素贝叶斯、XGBOOST、多层感知机5种方法构建重伤伤员预警模型,并对预警模型进行性能评价。结果通过实验显示,心率、收缩压与重伤预警组的关联最密切(分别为0.272、0.206),从而揭示相互之间存在阶梯级关系、构建三个阶梯级传递给后续机器学习方法,得到多层感知机性能显著高于其他方法,仅心率、收缩压精确率就达到了91.40%,其他模型精确率均在80%以上,这些表现良好的模型在外部验证中也有一定的泛化能力。结论基于战现场环境,通过采用不同类型的阶梯特征构建模型是可行的,该模型可以在重伤伤员发生前快速预警,并在救治人员到达前辅助做出救援决策,上述研究对降低战现场死亡率、改善预后有一定的实用价值。

Objective To realize the demand of early warning of severe wounded on the battlefield,to achieve the advance of triage and guide the order of treatment,and to improve the efficiency of treatment.Methods In this study,1469 and 1464 data of patients were collected from the emergency department database of a military hospital and an open multi-parameter clinical intensive care database Medical Information Mart for Intensive Care II(MIMIC II)established by the U.S.According to the feasibility and scientific analysis of the battlefield,7 representative indicators were selected as the input features of the training model,the optimal feature combination was selected based on the feature weight ranking based on XGBOOST,and the influence of the selected features on the predictive results was further analyzed.Five methods,random forest,support vector machine,naive Bayes,XGBOOST and multilayer perceptron,were used to construct the early warning model of serious injuries and evaluate its performance.Results The experiments showed that the heart rate and systolic blood pressure were the most closely related to the severe injury warning group(0.272 and 0.206 respectively),thus revealing the existence of a ladder-level relationship between each other and constructing three ladderlevels to pass to the subsequent machine learning methods.The performance of the multilayer perceptron was significantly higher than those of other methods.Only the heart rate and systolic blood pressure reached an accuracy of 91.40%,and the accuracy of other models was above 80%,these well-behaved models also had certain generalization ability in external verification.Conclusion Based on the battlefield environment,it is feasible to construct the model by using different types of ladder features.The model can quickly warn the severe wounded before the occurrence,and assist in making rescue decisions before the arrival of the treatment personnel.The above research has certain practical value for reducing the mortality rate and improving the prognosis.