Improving the Hip Fracture Risk Prediction with a Statistical Shape-and-Intensity Model of the Proximal Femur

Abstract

AbstractSevere predictions have been made regarding osteoporotic fracture incidence for the next years, with major economic and social impacts in a worldwide greying society. However, the performance of the currently adopted gold standard for fracture risk prediction, the areal Bone Mineral Density (aBMD), remains moderate. To overcome current limitations, the construction of statistical models of the proximal femur, based on three-dimensional shape and intensity (a hallmark of bone density), is here proposed for predicting hip fracture in a Caucasian postmenopausal cohort. Partial Least Square (PLS)-based statistical models of the shape, intensity and their combination were developed, and the corresponding modes and components were identified. Logistic regression models using the first two shape, intensity and shape-intensity PLS components were implemented and tested within a 10-fold cross-validation procedure as predictors of hip fracture. It emerged that (1) intensity components were superior to shape components in stratifying patients according to their fracture status, and that (2) a combination of intensity and shape improved patients risk stratification. The area under the ROC curve was 0.64, 0.85 and 0.92 for the models based on shape, intensity and shape-intensity combination respectively, against a 0.72 value for the aBMD standard approach. Based on these findings, the presented methodology turns out to be promising in tackling the need for an enhanced fracture risk assessment.