Participant characteristics by data set.

Background
Osteoporosis is a major global health issue, weakening bones and increasing fracture risk. Dual-energy X-ray absorptiometry (DXA) is the standard for measuring bone mineral density (BMD) and diagnosing osteoporosis, but its costliness and complexity impede widespread screening adoption. Predictive modeling using genetic and clinical data offers a cost-effective alternative for assessing osteoporosis and fracture risk. This study aims to develop BMD prediction models using data from the UK Biobank (UKBB) and test their performance across different ethnic and geographical populations.
Methods and findings
We developed BMD prediction models for the femoral neck (FNK) and lumbar spine (SPN) using both genetic variants and clinical factors (such as sex, age, height, and weight), within 17,964 British white individuals from UKBB. Models based on regression with least absolute shrinkage and selection operator (LASSO), selected based on the coefficient of determination (R2) from a model selection subset of 5,973 individuals from British white population. These models were tested on 5 UKBB test sets and 12 independent cohorts of diverse ancestries, totaling over 15,000 individuals. Furthermore, we assessed the correlation of predicted BMDs with fragility fractures risk in 10 years in a case-control set of 287,183 European white participants without DXA-BMDs in the UKBB.
With single-nucleotide polymorphism (SNP) inclusion thresholds at 5×10−6 and 5×10−7, the prediction models for FNK-BMD and SPN-BMD achieved the highest R2 of 27.70% with a 95% confidence interval (CI) of [27.56%, 27.84%] and 48.28% (95% CI [48.23%, 48.34%]), respectively. Adding genetic factors improved predictions slightly, explaining an additional 2.3% variation for FNK-BMD and 3% for SPN-BMD over clinical factors alone. Survival analysis revealed that the predicted FNK-BMD and SPN-BMD were significantly associated with fragility fracture risk in the European white population (P < 0.001). The hazard ratios (HRs) of the predicted FNK-BMD and SPN-BMD were 0.83 (95% CI [0.79, 0.88], corresponding to a 1.44% difference in 10-year absolute risk) and 0.72 (95% CI [0.68, 0.76], corresponding to a 1.64% difference in 10-year absolute risk), respectively, indicating that for every increase of one standard deviation in BMD, the fracture risk will decrease by 17% and 28%, respectively. However, the model’s performance declined in other ethnic groups and independent cohorts. The limitations of this study include differences in clinical factors distribution and the use of only SNPs as genetic factors.
Conclusions
In this study, we observed that combining genetic and clinical factors improves BMD prediction compared to clinical factors alone. Adjusting inclusion thresholds for genetic variants (e.g., 5×10−6 or 5×10−7) rather than solely considering genome-wide association study (GWAS)-significant variants can enhance the model’s explanatory power. The study highlights the need for training models on diverse populations to improve predictive performance across various ethnic and geographical groups.

## 背景
骨质疏松症是全球性重大公共卫生问题，可导致骨量流失、骨折风险显著升高。双能X线吸收法（Dual-energy X-ray absorptiometry, DXA）是目前测量骨密度（bone mineral density, BMD）、诊断骨质疏松症的金标准，但其成本高昂、操作复杂，限制了大规模筛查的推广应用。利用遗传与临床数据构建预测模型，为骨质疏松症及骨折风险评估提供了经济高效的替代方案。本研究旨在基于英国生物库（UK Biobank, UKBB）的数据构建骨密度预测模型，并在不同种族与地理人群中验证其性能。

## 方法与结果
本研究基于英国生物库中17964名英国白人个体的数据，结合遗传变异与临床因素（包括性别、年龄、身高、体重），构建了针对股骨颈（femoral neck, FNK）与腰椎（lumbar spine, SPN）的骨密度预测模型。模型基于最小绝对收缩与选择算子（least absolute shrinkage and selection operator, LASSO）回归构建，通过英国白人亚群中5973名个体的模型选择子集的决定系数（coefficient of determination, R²）进行筛选。随后，我们在5个英国生物库测试集与12个不同血统的独立队列（总样本量超15000）中对上述模型进行了性能验证。此外，我们针对英国生物库中287183名未接受DXA骨密度检测的欧洲白人参与者组成的病例对照队列，评估了预测骨密度与10年内脆性骨折风险的相关性。

当单核苷酸多态性（single-nucleotide polymorphism, SNP）的纳入阈值设为5×10^−6与5×10^−7时，股骨颈骨密度与腰椎骨密度预测模型的最高决定系数分别达到27.70%（95%置信区间[CI]：27.56%~27.84%）与48.28%（95%CI：48.23%~48.34%）。相较于仅使用临床因素的模型，加入遗传因素可小幅提升预测性能，分别使股骨颈骨密度与腰椎骨密度的变异解释度提升2.3%与3%。生存分析结果显示，在欧洲白人人群中，预测的股骨颈骨密度与腰椎骨密度均与脆性骨折风险显著相关（P < 0.001）。预测股骨颈骨密度与腰椎骨密度的风险比（hazard ratio, HR）分别为0.83（95%CI：0.79~0.88，对应10年绝对风险差值为1.44%）与0.72（95%CI：0.68~0.76，对应10年绝对风险差值为1.64%），表明骨密度每升高1个标准差，骨折风险分别降低17%与28%。但该模型在其他种族人群与独立队列中的性能有所下降。本研究存在一定局限性：临床因素的分布存在异质性，且仅使用单核苷酸多态性作为遗传因素。

## 结论
本研究结果显示，相较于仅使用临床因素的模型，结合遗传与临床因素可提升骨密度的预测性能。调整遗传变异的纳入阈值（如5×10^−6或5×10^−7）而非仅考虑全基因组关联研究（genome-wide association study, GWAS）的显著变异，可提升模型的变异解释能力。本研究强调，需在多样化人群中训练预测模型，以提升其在不同种族与地理人群中的预测性能。