应用机器学习方法优化住院创伤患者VTE风险预测

Applying Machine Learning Methods to Optimize VTE Risk Prediction in Hospitalized Trauma Patients

目的:筛选住院病历系统中可能与静脉血栓栓塞症相关的特征,与Caprini评分相结合,采用机器学习方法进行建模,在量表评估的基础上进一步提升VTE预测性能。方法:回顾了2019年9月至2020年3月在四川大学华西医院创伤医学中心住院的903例创伤患者,基于Lasso回归方法进行特征筛选后,结合随机森林算法构建VTE预测模型,并与单独使用Caprini评分的预测效果进行比较。结果:Caprini评分在预测静脉血栓栓塞事件上表现出了较好的区分度(TPR=0.667,FPR=0.227,AUC=0.773),最终构建的机器学习模型在Caprini评分的基础上,增加4个预测特征,可进一步提升预测效果(TPR=0.757,FPR=0.290,AUC=0.799)。结论:结合Caprini评分及生理生化特征的机器学习模型适用于对住院创伤患者的静脉血栓栓塞症风险预测。

Objective:To screen features in the inpatient medical record system that may be associated with venous thromboembolism,combine them with Caprini scores,and use machine learning methods for modeling to further improve VTE prediction performance based on scale assessment.Methods:903 trauma patients hospitalized at the Trauma Medical Center of West China Hospital of Sichuan University from September 2019 to March 2020 were reviewed,and after feature screening based on the Lasso regression method,a VTE prediction model was constructed in combination with the random forest algorithm and compared with the predictive effect of the Caprini score alone.Results:The Caprini score showed good discrimination in predicting venous thromboembolic events(TPR=0.667,FPR=0.227,AUC=0.773),and our final constructed machine learning model could further improve the prediction effect by adding 4 predictive features to the Caprini score(TPR=0.757,FPR=0.290,AUC=0.799).Conclusion:A machine learning model combining Caprini score and physiological and biochemical features is suitable for predicting the risk of venous thromboembolism in hospitalized trauma patients.