The Added Value of High-Resolution Peripheral Quantitative Computed Tomography in Fracture Risk Prediction

Modern assessment of fracture risk in clinical practice involves both considerations of clinical risk factors and measurement of bone mineral density (BMD) using dual X-ray absorptiometry (DXA). Although DXA-derived BMD is robustly associated with fracture, with every standard deviation decrease in the femoral neck (FN) BMD increases the risk of hip fracture by two to three times, the sensitivity and specificity of FN BMD alone are insufficient. Consideration of clinical risk factors, such as previous fracture, oral glucocorticoid use, heredity, and low body mass index, used in combination with FN BMD improves model performance in clinically used risk prediction tools, such as the fracture risk assessment tool FRAX, QFracture, and the Garvan calculator. The FRAX tool has been widely implemented globally and performs well in the prediction of hip fractures, while receiver operating characteristic (ROC) curve performance is lower for the prediction of major osteoporotic fracture (MOF). Especially for any fracture and MOF, novel methods to achieve improved fracture risk prediction are warranted. It is known that bone microstructural parameters obtained by high-resolution peripheral quantitative computed tomography (HR-pQCT) predict fracture risk independently of FN BMD and clinical risk factors, but any interactions between microstructural measures and how they each contribute to fracture risk have primarily been studied using traditional methods, such as Cox regression. Indeed, leveraging the information reflected by a large number of highly correlated bone microarchitecture and strength outcomes from across multiple anatomic locations remains a significant challenge to the goal of synthesizing a fracture risk metric from HR-pQCT or similar modalities. In this issue of the JBMR, Whittier et al. present a novel fracture risk assessment tool (μFRAC) based on HR-pQCT measurement results and machine learning methodology using the random survival forest method, which assembles decision trees and provides an intuitive classification resembling clinical decision-making. In a large dataset from several cohorts of older men and women (n = 6 802), the association between incident fractures (609 incident fractures during the average 4.7-year follow-up) and HR-pQCT variables were investigated using random survival forest models. They found that in evaluations of sensitivity and specificity in predicting any incident fracture, using ROC curves, the developed μFRAC tool performed better than other models or the calculated FRAX scores for the 10-year probability of MOF. These results suggest that incorporating HR-pQCT or other methods to measure bone microstructure could improve fracture prediction in a clinical setting. The study has considerable strengths. It is one of the largest yet performed and uses a novel machine learning methodology to identify the most accurate fracture prediction model incorporating HR-pQCT data. Both full models, using all standard parameters and minimal models, performed reasonably well for MOF prediction. The use of standard HR-pQCT outcome measures as input to the models makes this approach relatively straightforward to translate to external datasets for validation and practical to apply to both historical and newly collected data for investigative purposes. However, some issues and limitations with the analysis also need to be acknowledged. First, there were few (n = 50) hip fractures in the analysis, even though the gradient of risk was similar for hip fractures. The comparator model used, FRAX, is known to performmost optimally for hip fractures, indicating that comparing μFRAC performance for the prediction of other fractures has its limitations. In addition, Whittier et al. used calculated FRAX scores rather than the individual clinical risk factors. It is well known that using cohort-specific risk factors provides better fracture risk prediction than generic FRAX scores designed to work across a large number of cohorts with very different characteristics. This may also be an important factor in explaining the differences in model performance found. Furthermore, the μFRAC prediction model was developed and tested in the same combined cohort, which indicates that model performance should also be evaluated externally, as has been done for the FRAX tool. Because HR-pQCT-derived failure load is an excellent proxy for bone strength, inherently reflects bone mass and architecture, and was previously found to be an independent predictor of fracture risk, it is perplexing why this bone trait alone was not used as a comparator for the μFRAC model to investigate whether the combination of bone microstructural measurements had an added and intrinsic value that could improve the prediction of fractures beyond failure load. Although the μFRAC model could be clinically useful in patients with available bone microstructure measurements, much work remains to be done to demonstrate the practicality and generalizability of this approach. The μFRAC method