Predicting Hip Fracture Type With Cortical Bone Mapping (CBM) in the Osteoporotic Fractures in Men (MrOS) Study

Hip fracture risk is known to be related to material properties of the proximal femur, but fracture prediction studies adding richer quantitative computed tomography (QCT) measures to dual-energy X-ray (DXA)-based methods have shown limited improvement. Fracture types have distinct relationships to predictors, but few studies have subdivided fracture into types, because this necessitates regional measurements and more fracture cases. This work makes use of cortical bone mapping (CBM) to accurately assess, with no prior anatomical presumptions, the distribution of properties related to fracture type. CBM uses QCT data to measure the cortical and trabecular properties, accurate even for thin cortices below the imaging resolution. The Osteoporotic Fractures in Men (MrOS) study is a predictive case-cohort study of men over 65 years old: we analyze 99 fracture cases (44 trochanteric and 55 femoral neck) compared to a cohort of 308, randomly selected from 5994. To our knowledge, this is the largest QCT-based predictive hip fracture study to date, and the first to incorporate CBM analysis into fracture prediction. We show that both cortical mass surface density and endocortical trabecular BMD are significantly different in fracture cases versus cohort, in regions appropriate to fracture type. We incorporate these regions into predictive models using Cox proportional hazards regression to estimate hazard ratios, and logistic regression to estimate area under the receiver operating characteristic curve (AUC). Adding CBM to DXA-based BMD leads to a small but significant (p < 0.005) improvement in model prediction for any fracture, with AUC increasing from 0.78 to 0.79, assessed using leave-one-out cross-validation. For specific fracture types, the improvement is more significant (p < 0.0001), with AUC increasing from 0.71 to 0.77 for trochanteric fractures and 0.76 to 0.82 for femoral neck fractures. In contrast, adding DXA-based BMD to a CBM-based predictive model does not result in any significant improvement.