基于机器学习的中国绝经后女性骨质疏松筛查模型:全国性研究及前瞻性骨折验证
本研究利用中国骨质疏松患病率研究(n=6,574)训练机器学习模型预测绝经后女性骨质疏松,并在两个独立队列外部验证,AUC为0.744-0.798;模型还显示对5年骨折风险的预后价值,支持在DXA资源有限地区进行人群筛查。
要点
- 模型基于5个易获取预测因子,训练队列AUC为0.798,外部验证AUC为0.750和0.744。
- 排除阈值≥0.098在外部队列中敏感度94.7%和89.3%,阴性预测值91.3%和92.7%。
- 5年随访中高风险组临床骨折累积发生率显著更高(HR 1.78,95%CI 1.06-2.99,P=0.026)。
核心结论
- 该机器学习模型在中国绝经后女性中预测骨质疏松性能良好,并可能促进两步筛查策略。
研究设计与样本
- 模型训练使用中国骨质疏松患病率研究(n=6,574)数据,外部验证使用中国椎体与骨质疏松研究(n=1,758)和北京椎体骨折研究基线(n=1,439)。
- 通过LASSO回归选择5个预测因子训练模型,并在5年随访队列(n=795)中评估骨折预测能力。
主要发现
- 模型在训练队列AUC为0.798,内部验证队列AUC为0.775,两个外部验证队列AUC分别为0.750和0.744,所有队列校准良好。
- 排除阈值≥0.098在两个外部队列中敏感度分别为94.7%和89.3%,阴性预测值分别为91.3%和92.7%。
- 5年随访中高风险组临床骨折累积发生率显著更高(HR 1.78;95%CI 1.06-2.99;P=0.026),所有骨折发生率15.9% vs. 9.5%(P=0.015),非创伤性骨折7.8% vs. 3.9%(P=0.041)。
关键数值与推荐
- 模型使用LASSO回归选择的5个预测因子。
- 排除阈值≥0.098,外部验证敏感度94.7%和89.3%,阴性预测值91.3%和92.7%。
- 高风险组5年临床骨折HR 1.78(95%CI 1.06-2.99;P=0.026),所有骨折15.9% vs. 9.5%(P=0.015),非创伤性骨折7.8% vs. 3.9%(P=0.041)。
临床怎么用
- 在DXA资源有限地区,该模型可用于人群-level筛查,通过排除阈值识别低风险个体,减少不必要的DXA检查。
- 模型预测高风险个体应关注骨折风险,考虑进一步评估或干预。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
UNLABELLED: This study develops a machine learning model to predict osteoporosis in Chinese postmenopausal women. The model was trained using the largest nationwide cohort and externally validated. It showed AUCs of 0.744-0.798 and demonstrated prognostic value for 5-year fracture risk, supporting population-level screening where DXA resources are limited. PURPOSE: Osteoporosis is highly prevalent among Chinese postmenopausal women, yet underdiagnosis remains due to limited DXA availability. We aimed to develop and validate a machine learning-based model for osteoporosis prediction using easily obtainable variables. METHODS: The model was trained using data from the largest nationwide cohort (China Osteoporosis Prevalence Study, n = 6,574) and externally validated in two independent cohorts: the China Vertebral and Osteoporosis Study (n = 1,758) and the Peking Vertebral Fracture Study baseline (n = 1,439). Five predictors selected via LASSO regression were used to train the prediction model. Model performance was evaluated by area under the curve (AUC) and calibration plot. To evaluate its ability to predict future fractures, the model was evaluated in a 5-year follow-up cohort (n = 795). A web-based calculator was developed for public use. RESULTS: The model showed strong discrimination (AUC: 0.798 in the training cohort; 0.775 in the internal validation cohort; 0.750 and 0.744 in the external validation cohorts) and good calibration across all cohorts. A rule-out threshold (≥ 0.098) demonstrated high sensitivity (94.7%, 89.3%) and negative predictive value (91.3%, 92.7%) in two external cohorts. The model also demonstrated prognostic value for fracture risk stratification. In the 5-year follow-up, individuals classified as high risk had a significantly higher cumulative incidence of clinical fractures (HR 1.78; 95% CI, 1.06-2.99; P = 0.026), as well as higher rates of all fractures (15.9% vs. 9.5%; P = 0.015) and non-traumatic fractures (7.8% vs. 3.9%; P = 0.041). CONCLUSIONS: This model demonstrated good performance in predicting osteoporosis among Chinese postmenopausal women and may facilitate two-step screening strategies.
引用
- Sun Q, Yin X, Yu W, et al. Machine learning-based screening model for osteoporosis in Chinese postmenopausal women: a nationwide study with prospective fracture validation. Osteoporos Int. 2026. PMID: 42658240. doi:10.1007/s00198-026-08185-1
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。