基于三种机器学习方法研究骨质疏松性骨折危险因素及构建列线图预测模型  

Risk factors and nomogram prediction model for osteoporotic fracture based on three machine learning methods

作　　者：云思敏 王雄毅 谢伊代·如则 朱柯雨 翁程伟 郑苗 魏祺 徐又佳[1] YUN Si-min;WANG Xiong-yi;XIEYIDAI Ru-ze;ZHU Ke-yu;WENG Cheng-wei;ZHENG Miao;WEI Qi;XU You-jia(The Second Affiliated Hospital of Soochow University,Suzhou 215000,Jiangsu,China)

机构地区：[1]苏州大学附属第二医院骨科、骨质疏松症临床中心,苏州215000

出　　处：《中华骨质疏松和骨矿盐疾病杂志》2025年第1期36-48,共13页Chinese Journal Of Osteoporosis And Bone Mineral Research

基　　金：国家自然科学基金(82372455);江苏省医学重点实验室建设项目(JSDW202254);中核医疗“核医科技创新”计划资助项目(ZHYLZD2023001);苏州大学研究生教育改革成果奖培育项目;江苏省体医融合促进老年骨骼健康应用工程研究中心项目。

摘　　要：目的运用3种机器学习方法探讨骨质疏松症患者发生骨质疏松性骨折的危险因素,并构建列线图预测模型。方法选取2021年10月至2023年5月苏州大学附属第二医院289例骨质疏松症患者作为研究对象,按是否骨折分为骨折组(93例)和非骨折组(196例)。采用R语言开展相关研究。运用拉索回归(least absolute shrinkage and selection operator,LASSO)、极端梯度提升(extreme gradient boosting,XGBoost)、随机森林(random forest,RF)3种机器学习方法平行评估“变量”与骨质疏松性骨折风险的相关性,分析3种机器学习方法“重叠覆盖”的风险因素,采用多因素Logisitic回归方法验证结果的独立预测性。基于机器学习和多因素Logisitic回归分析筛选出的独立危险因素构建“列线图预测模型”,采用Bootstrap方法进行内部和外部验证,通过受试者工作特征(receiver operating characteristic,ROC)曲线下面积(area under the curve,AUC)、Hosmer-Lemeshow拟合优度检验、校准曲线及决策曲线(decision curve analysis,DCA)评估列线图预测模型的准确性和临床适用性。结果3种机器学习方法各自筛选到不同风险因素,经过“重叠覆盖”分析得到6项共同的骨质疏松性骨折重要风险因素:年龄、甘油三酯、Ca、血清25羟维生素D、股骨颈T值、血清白蛋白。多因素Logistic回归分析显示其中5个因素是脆性骨折的独立危险因素:年龄(OR=1.075,95%CI:1.017~1.136,P=0.011)、甘油三酯(OR=0.207,95%CI:0.103~0.415,P<0.001)、Ca(OR=0.010,95%CI:0.000~0.602,P=0.028)、股骨颈T值(OR=0.443,95%CI:0.245~0.800,P=0.007)、血清25羟维生素D(OR=0.902,95%CI:0.830~0.980,P=0.015)。基于5个预测危险因素构建的列线图预测模型验证结果显示:模型训练集AUC值为0.934(95%CI:0.897~0.972),内部验证集的AUC值为0.893(95%CI:0.802~0.984),外部验证集的AUC值为0.849(95%CI:0.792~0.905),预测效能良好。校准曲线显示预测值与理想曲线有较好一致性。HosObjective To explore the risk factors for osteoporotic fractures in patients with osteoporosis using three machine learning methods and to develop a nomogram prediction model.Methods Totally 289 patients with osteoporosis treated in the Second Affiliated Hospital of Soochow University between October 2021 and May 2023 were selected and divided into a fracture group(n=93)and a non-fracture group(n=196)based on whether they had fractures.The study employed R language for analysis.Three machine learning methods as LASSO regression,extreme gradient boosting(XGBoost),and random forest(RF)were used to evaluate the correlation between various variables and the risk of osteoporotic fractures.The overlapping risk factors identified by these three methods were further validated using multivariate logistic regression to confirm their independent predictive value.A nomogram prediction model was constructed based on the independent risk factors identified through machine learning and multivariate logistic regression,with internal and external validation performed using the Bootstrap method.The accuracy and clinical applicability of the nomogram were assessed using the area under the receiver operating characteristic(ROC)curve,Hosmer-Lemeshow goodness-of-fit test,calibration curve and decision curve analysis(DCA).Results The three machine learning methods identified different risk factors,with six common key risk factors emerging from the overlapping analysis,namely age,triglycerides,serum calcium,serum 25-hydroxyvitamin D,femoral neck T-score,and serum albumin.Multivariate logistic regression analysis revealed that five of these factors were independent risk factors for fragility fractures as age(OR=1.075,95%CI:1.017-1.136,P=0.011),triglycerides(OR=0.207,95%CI:0.103-0.415,P<0.001),calcium(OR=0.010,95%CI:0.000-0.062,P=0.028),femoral neck T-score(OR=0.443,95%CI:0.245-0.800,P=0.007),and 25(OH)D(OR=0.902,95%CI:0.830-0.980,P=0.015).The validation results of the nomogram prediction model constructed based on five predictive risk fa

关 键 词：骨质疏松性骨折 机器学习 列线图 预测模型 

分 类 号：R683[医药卫生—骨科学]