基于机器学习的郑州地区2型糖尿病患者骨折风险预测  

Machine learning-based prediction of fracture risk in patients with type 2 diabetes in Zhengzhou area

作　　者：殷璐[1] 董其娟 孙晓菲[1] 杨雪[1] 田勇[1] 孙晓利[1] 杨旭光[1] 范慧洁[1] YIN Lu;DONG Qi-juan;SUN Xiao-fei;YANG Xue;TIAN Yong;SUN Xiao-li;YANG Xu-guang;FAN Hui-jie(Department of Endocrinology and Metabolism,People's Hospital of Zhengzhou,Zhengzhou 450002,China)

机构地区：[1]郑州人民医院内分泌代谢科,郑州450002

出　　处：《医药论坛杂志》2024年第18期1939-1942,1947,共5页Journal of Medical Forum

基　　金：河南省医学科技攻关计划联合共建项目(LHGJ20200690,LHGJ20220777)。

摘　　要：目的通过骨折风险因子筛选,将多种机器学习算法应用于2型糖尿病患者骨折风险预测,以实现2型糖尿病患者的骨折风险的早期发现及早期干预。方法使用FRAX和糖尿病所涉及骨折风险因子对郑州人民医院标准化代谢性疾病管理中心收录的2018年1月至2022年12月的795例2型糖尿病合并骨质疏松的患者进行风险预测,构建不同机器学习模型并获得不同的受试者工作特征(receiver operating characteristic,ROC)曲线。结果使用随机森林算法对特征变量的重要性进行排序,得到特征重要性排序。三种机器学习模型对骨折风险预测均具有一定效果,其中随机森林和梯度提升(extreme gradient boosting,XGBoost)模型效果最佳,准确度分别为0.94和0.93,而精确度分别为0.97和0.95;反向传播算法(back propagation,BP)神经网络次之,准确度和精确度分别为0.89和0.91。结论对特征变量的重要性进行排序,发现糖皮质激素应用、吸烟饮酒、家族史占据前列,血糖控制水平凸显其重要性,空腹血糖及糖化血红蛋白对骨折风险占较大权重。三种模型的预测效果均具有较高的预测价值。Objective To screen for fracture risk factors and apply various machine learning algorithms to the prediction of fracture risk in patients with type 2 diabetes,with the aim of achieving early detection and intervention for fracture risk in patients with type 2 diabetes.Methods Using FRAX and fracture risk factors related to diabetes,risk prediction was carried out on 795 patients with type 2 diabetes complicated by osteoporosis from January 2018 to December 2022 included in the Standardised-Metabolic Disease Management Center of People's Hospital of Zhengzhou,and different machine learning models were constructed to obtain different receiver operating characteristic(ROC)curve.Results Using the random forest algorithm to rank the importance of feature variables,the ranking of feature importance was obtained.All three machine earning models had certain effects on fracture risk prediction,among which the random forest and extreme gradient boosting(XGBoost)model performed the best,achieving accuracy of 0.94 and 0.93,respectively,and precision of 0.97 and 0.95,respectively,followed by the back propagation(BP)neural network,achieving accuracy and precision of 0.89 and 0.91,respectively.Conclusion Ranking the importance of feature variables reveal that the application of glucocorticoids,smoking and alcohol consumption,and family history are in the forefront,and the level of blood sugar control highlight its importance,with fasting blood sugar and glycated hemoglobin contributing significantly to fracture risk.The predictive effects of the three models all have high predictive value.

关 键 词：机器学习 2型糖尿病 骨折风险 郑州地区 

分 类 号：R587.2[医药卫生—内分泌]