Liu Bo, Zhang Qiang, Li Xin
Department of Orthopaedics, Beijing Ditan Hospital, Capital Medical University, Beijing, China.
National Center for Infectious Diseases, Beijing, China.
Front Cell Infect Microbiol. 2025 Mar 14;15:1461740. doi: 10.3389/fcimb.2025.1461740. eCollection 2025.
This study aimed to develop and validate a novel web-based calculator using machine learning algorithms to predict fragility fracture risk in People living with HIV (PLWH), who face increased morbidity and mortality from such fractures.
We retrospectively analyzed clinical data from Beijing Ditan Hospital orthopedic department between 2015 and September 2023. The dataset included 1045 patients (2015-2021) for training and 450 patients (2021-September 2023) for external testing. Feature selection was performed using multivariable logistic regression, LASSO, Boruta, and RFE-RF. Six machine learning models (logistic regression, decision trees, SVM, KNN, random forest, and XGBoost) were trained with 10-fold cross-validation and hyperparameter tuning. Model performance was assessed with ROC curves, Decision Curve Analysis, and other metrics. The optimal model was integrated into an online risk assessment calculator.
The XGBoost model showed the highest predictive performance, with key features including age, smoking, fall history, TDF use, HIV viral load, vitamin D, hemoglobin, albumin, CD4 count, and lumbar spine BMD. It achieved an ROC-AUC of 0.984 (95% CI: 0.977-0.99) in the training set and 0.979 (95% CI: 0.965-0.992) in the external test set. Decision Curve Analysis indicated clinical utility across various threshold probabilities, with calibration curves showing high concordance between predicted and observed risks. SHAP values explained individual risk profiles. The XGBoostpowered web calculator (https://sydtliubo.shinyapps.io/cls2shiny/) enables clinicians and patients to assess fragility fracture risk in PLWH.
We developed a web-based risk assessment tool using the XGBoost algorithm for predicting fragility fractures in HIV-positive patients. This tool, with its high accuracy and interpretability, aids in fracture risk stratification and management, potentially reducing the burden of fragility fractures in the HIV population.
本研究旨在开发并验证一种基于网络的新型计算器，该计算器使用机器学习算法来预测感染人类免疫缺陷病毒（HIV）的人群（PLWH）发生脆性骨折的风险，这类人群因此类骨折面临更高的发病率和死亡率。
我们回顾性分析了2015年至2023年9月北京地坛医院骨科的临床数据。数据集包括用于训练的1045例患者（2015 - 2021年）和用于外部测试的450例患者（2021 - 2023年9月）。使用多变量逻辑回归、LASSO、Boruta和RFE - RF进行特征选择。六个机器学习模型（逻辑回归、决策树、支持向量机、K近邻、随机森林和XGBoost）通过10折交叉验证和超参数调整进行训练。使用ROC曲线、决策曲线分析和其他指标评估模型性能。将最优模型集成到在线风险评估计算器中。
XGBoost模型显示出最高的预测性能，关键特征包括年龄、吸烟、跌倒史、替诺福韦酯（TDF）使用情况、HIV病毒载量、维生素D、血红蛋白、白蛋白、CD4细胞计数和腰椎骨密度。在训练集中其ROC - AUC为0.984（95% CI：0.977 - 0.99），在外部测试集中为0.979（95% CI：0.965 - 0.992）。决策曲线分析表明在各种阈值概率下均具有临床实用性，校准曲线显示预测风险与观察到的风险之间具有高度一致性。SHAP值解释了个体风险概况。由XGBoost驱动的网络计算器（https://sydtliubo.shinyapps.io/cls2shiny/）使临床医生和患者能够评估PLWH发生脆性骨折的风险。
我们开发了一种基于网络的风险评估工具，使用XGBoost算法预测HIV阳性患者的脆性骨折。该工具具有高准确性和可解释性，有助于骨折风险分层和管理，可能减轻HIV人群中脆性骨折的负担。
