McCaig Institute for Bone and Joint Health, Cumming School of Medicine, University of Calgary, Calgary, AB T2N 4Z6, Canada.
Department of Cell Biology and Anatomy, Cumming School of Medicine, Alberta Children's Hospital Research Institute, University of Calgary, Calgary, AB T2N 4N1, Canada.
J Bone Miner Res. 2024 May 24;39(5):571-579. doi: 10.1093/jbmr/zjae039.
The continued development of high-resolution peripheral quantitative computed tomography (HR-pQCT) has led to a second-generation scanner with higher resolution and longer scan region. However, large multicenter prospective cohorts were collected with first-generation HR-pQCT and have been used to develop bone phenotyping and fracture risk prediction (μFRAC) models. This study establishes whether there is sufficient universality of these first-generation trained models for use with second-generation scan data.
HR-pQCT data were collected for a cohort of 60 individuals, who had been scanned on both first- and second-generation scanners on the same day to establish the universality of the HR-pQCT models. These data were each used as input to first-generation trained bone microarchitecture models for bone phenotyping and fracture risk prediction, and their outputs were compared for each study participant. Reproducibility of the models were assessed using same-day repeat scans obtained from first-generation (n = 37) and second-generation (n = 74) scanners.
Across scanner generations, the bone phenotyping model performed with an accuracy of 93.1%. Similarly, the 5-year fracture risk assessment by μFRAC was well correlated with a Pearson's (r) correlation coefficient of r > 0.83 for the three variations of μFRAC (varying inclusion of clinical risk factors, finite element analysis, and dual X-ray absorptiometry). The first-generation reproducibility cohort performed with an accuracy for categorical assignment of 100% (bone phenotyping) and a correlation coefficient of 0.99 (μFRAC), whereas the second-generation reproducibility cohort performed with an accuracy of 96.4% (bone phenotyping) and a correlation coefficient of 0.99 (μFRAC).
We demonstrated that bone microarchitecture models trained using first-generation scan data generalize well to second-generation scans, performing with a high level of accuracy and reproducibility. Less than 4% of individuals' estimated fracture risk led to a change in treatment threshold, and in general, these dissimilar outcomes using second-generation data tended to be more conservative.
高分辨率外周定量计算机断层扫描（HR-pQCT）的不断发展催生了第二代分辨率更高、扫描区域更长的扫描仪。然而，第一代 HR-pQCT 采集了大量的多中心前瞻性队列数据，并已用于开发骨表型和骨折风险预测（μFRAC）模型。本研究旨在确定这些第一代训练模型是否具有足够的通用性，可用于第二代扫描数据。
本研究共纳入 60 名个体，这些个体在同一天分别使用第一代和第二代扫描仪进行 HR-pQCT 扫描，以确定 HR-pQCT 模型的通用性。将这些数据分别作为输入输入第一代训练的骨微观结构模型，用于骨表型和骨折风险预测，并比较每个研究参与者的输出结果。使用第一代（n=37）和第二代（n=74）扫描仪获得的同一天重复扫描来评估模型的重现性。
在不同扫描仪世代中，骨表型模型的准确性为 93.1%。同样，μFRAC 对 5 年骨折风险的评估与三种不同μFRAC 变异（临床危险因素、有限元分析和双能 X 线吸收法的不同纳入）的皮尔逊（r）相关系数 r>0.83 高度相关。第一代重现性队列的分类赋值准确性为 100%（骨表型），相关系数为 0.99（μFRAC），而第二代重现性队列的准确性为 96.4%（骨表型），相关系数为 0.99（μFRAC）。
我们证明了使用第一代扫描数据训练的骨微观结构模型可以很好地推广到第二代扫描，具有高度的准确性和重现性。不到 4%的个体的估计骨折风险导致治疗阈值发生变化，并且通常情况下，第二代数据的这些不同结果趋于更保守。
Zotero 插件
