首页|期刊导航|诊断学理论与实践|基于机器学习算法的骨质疏松性骨折预测模型构建

基于机器学习算法的骨质疏松性骨折预测模型构建OA

Research on construction of osteoporotic fracture prediction models based on machine learning algorithms

中文摘要英文摘要

目的:基于骨质疏松患者的临床资料,采用5种机器学习算法分别构建骨质疏松性骨折风险预测模型,并对模型的效能进行评价及比较.方法:根据纳排标准,选取2020年1月至2025年12月于上海交通大学医学院附属松江医院就诊的333例骨质疏松症连续病例(男性66例,女性267例;中位年龄69岁),根据有无脆性骨折史分为骨折组(n=146)和未骨折组(n=187).收集患者基线资料及实验室指标,缺失数据采用链式方程多重插补处理.采用R语言开展相关研究,通过逻辑回归、拉索回归筛选特征变量,随机森林进行重要性排序.按7∶3比例将数据分为训练集和测试集,采用逻辑回归、决策树(decision tree,DT)、随机森林、支持向量机和极限梯度提升(extreme gra-dient boosting,XGBoost)算法分别构建模型,采用受试者操作特征曲线的曲线下面积等指标评估模型预测骨质疏松性骨折风险的效能,并应用SHAP(SHapley Additive exPlanations)值对最优模型进行解释.结果:单因素逻辑回归分析显示,性别、甲状旁腺激素(parathyroid hormone,PTH)、N端中段骨钙素(N-middle osteocalcin,N-MID-OT)、碱性磷酸酶、白蛋白、钙、平均红细胞血红蛋白含量(mean corpuscular hemoglobin,MCH)、平均红细胞血红蛋白浓度(mean corpuscular hemoglobin concentration,MCHC)、血红蛋白、磷与骨质疏松性骨折显著相关.拉索回归筛选出性别及PTH、N-MID-OT、白蛋白、钙、MCHC、磷水平共7个关键特征,进行随机森林重要性排序后,纳入性别、PTH、N-MID-OT、钙、磷、MCHC这6项指标进行机器学习建模.利用测试集数据作为输入,比较每种机器学习模型预测骨质疏松骨折中的效能,结果显示,其中XGBoost模型表现最佳,预测患者发生骨质疏松性骨折风险的曲线下面积为0.974(95%CI为0.950~0.998),其他4种机器学习模型也取得了良好的预测效能,曲线下面积介于0.891~0.964之间.SHAP结果显示,N-MID-OT、钙、PTH在模型中的权重较大,与临床认知一致.结论:基于常规临床指标构建的XGBoost模型在骨质疏松性骨折风险预测中表现出良好的区分度和稳定性,具有潜在的临床应用价值,适用于无法完整开展骨代谢标志物检测的基层医院.

Objective Based on clinical data from patients with osteoporosis,five machine learning algorithms are used to construct osteoporotic fracture risk prediction models,and their performance is evaluated and compared.Methods A total of 333 consecutive patients with osteoporosis(66 males and 267 females;median age,69 years)who attended Song-jiang Hospital Affiliated to Shanghai Jiao Tong University School of Medicine from January 2020 to December 2025 were selected according to inclusion and exclusion criteria.Patients were divided into a fracture group(n=146)and a non-frac-ture group(n=187)based on the presence or absence of a history of fragility fractures.Baseline patient data and laboratory indicators were collected,and missing data were handled using multiple imputation by chained equations.Relevant studies were conducted using the R programming language.Feature variables were screened using logistic regression and lasso regression,and variable importance was ranked using random forest.Data were split into a training set and a test set at a 7∶3 ratio.Models were constructed using logistic regression,decision tree(DT),random forest,support vector machine,and extreme gradient boosting(XGBoost)algorithms.The performance of the models in predicting osteoporotic fracture risk was evaluated using indicators such as the area under the receiver operating characteristic curve(AUC),accuracy,sensitivity,and specificity.The optimal model was interpreted using SHAP(SHapley Additive exPlanations)values.Results Univaria-te logistic regression analysis showed that sex,parathyroid hormone(PTH),N-terminal mid-fragment osteocalcin(N-MID-OT),alkaline phosphatase,albumin,calcium,mean corpuscular hemoglobin(MCH),mean corpuscular hemoglobin concen-tration(MCHC),hemoglobin,and phosphorus were significantly associated with osteoporotic fractures.Lasso regression identified seven key features:sex,PTH,N-MID-OT,albumin,calcium,MCHC,and phosphorus.After random forest impor-tance ranking,six variables-sex,PTH,N-MID-OT,calcium,phosphorus,and MCHC-were included in machine learning modeling.Using the test set data as input,the performance of each machine learning model in predicting osteoporotic fractures was compared.The results showed that the XGBoost model performed best with an AUC of 0.974(95%CI:0.950-0.998).The other four machine learning models also achieved good prediction performance,with AUCs ranging from 0.891 to 0.964.SHAP results indicated that N-MID-OT,calcium,and PTH had relatively large weights in the model,which was consistent with clinical cognition.Conclusions The XGBoost model constructed based on routine clinical indicators demonstrates good discriminative ability and stability in predicting the risk of osteoporotic fractures,has potential clinical application value,and is suitable for primary hospitals that are unable to conduct complete bone metabolism marker testing.

马善源;彭亮;侯彦强;刘维薇

南京医科大学上海十院临床医学院,上海 200040||上海交通大学医学院附属松江医院检验科,上海 201600上海交通大学医学院附属松江医院检验科,上海 201600上海交通大学医学院附属松江医院检验科,上海 201600南京医科大学上海十院临床医学院,上海 200040||上海中医药大学附属龙华医院检验科,上海 200032

医药卫生

机器学习骨质疏松性骨折预测模型

Machine learningOsteoporotic fracturePrediction model

《诊断学理论与实践》 2026 (4)

467-475,9

教育部中国高校产学研创新基金(2024DR031)

10.16150/j.1671-2870.2026.04.010

评论