Use of DXA-based statistical shape models of the femur for hip fracture risk prediction.

Osteoporosis is a systematic skeletal disorder characterized by low bone mass and microarchitectural deterioration of bone tissue, with a consequent increase in bone fragility and susceptibility to fracture, affecting about 30% of post-menopausal women. Nowadays, the main operational criterion to establish the presence of osteoporosis is based on the measurement of the areal Bone Mineral Density (aBMD) through Dual energy X-ray Absorptiometry (DXA). However, different studies have shown that nearly half of the subjects experiencing a low trauma hip fracture were classified as "low risk" according to the aBMD value. To overcome this problem and improve fracture prediction, different studies have been conducted, ranging from the analysis of the parameters of Hip Structural Analysis (HSA) and Trabecular Bone Score (TBS), up to the use of QCT images for the creation of Finite Element (FE) models to predict the bone load to failure or for the development of three-dimensional statistical shape models, in order to identify fracture-prone features. Nonetheless, the main problem with the aforementioned analyses is that QCT is not routinely performed in a clinical environment. In order to overcome this problem, in this work the possibility to build statistical models of DXA-derived proximal femur shapes is investigated, aiming at the hip fracture risk prediction in a post-menopausal Caucasian cohort. Fifty post-menopausal women, aged 55-90 years who had sustained a hip fracture were recruited as fractured cases, and for each case a post-menopausal woman matched with age, weight and height, was enrolled as control. The patient-specific geometry of the proximal femur was extracted by performing a semiautomatic segmentation of the DXA images in which the femoral head was simplified as a circle and the lesser trochanter was not considered because not present in all DXA images. Subsequently, the patient-specific 2D femur shapes were given as inputs to Deformetrica, who allowed the extraction of the template, i.e. the mean anatomical shape, and of the so called moment vectors, which gather the patient-specific anatomical features. Then, the moment vectors were used to build Statistical Shape Models (SSMs). Principal Component Analysis (PCA) and Partial Least Square (PLS) were adopted, leading to two distinct SSMs. While PCA maximized the variance found in the femurs anatomical features, PLS identified the modes maximizing the covariance between femurs anatomical features and the known patient-specific fracture status. Later, the identified modes were used for the implementation of logistic regression models for the prediction of the patients’ fracture status, which were tested using a 10-fold cross-validation procedure. The first five PCA modes and PLS modes were selected, which could explain at least 90% of the total shape variance. The predictive model with the first two, three, four and five PCA components used as predictors resulted in AUC values all settled between 0.59 and 0.62; instead, for the PLS components provided AUC values between 0.62 and 0.63. An AUC value of 0.73 was obtained using the gold standard aBMD. In conclusion, the use of purely SSM does not seem to outperform the current gold standard for hip fracture risk prediction. The inclusion of Statistical Intensity Models (SIMs), built starting from the local BMD values of DXA images, might allow an enhanced fracture risk assessment.