Selection of Bone fragility-Related Features Obtained with Bi-Directional Axial Transmission, Through a Machine Learning Strategy

Osteoporosis is a widespread public health problem worldwide, characterized by low bone mass, which compromises strength and increases the risk of fracture. Currently, the gold standard for assessing fracture risk is measurement of areal bone mineral density with dual-energy X-ray absorptiometry (DXA). Several ultrasound techniques, such as Bi-Directional Axial Transmission (BDAT) have been presented as alternatives. For the first studies, classification between fractured and non fractured patients was based on classical ultrasonic parameters, such as velocities or cortical thickness and porosity, obtained from an inverse problem. Recently, novel parameters obtained from structural analysis guided wave spectrum images (GWSI) have been introduced. The aim of this study is to merge both points of view and explore which parameters are the most important to obtain a robust classification using a machine learning approach. This study uses the same set of patients used in previous studies with 195 patients associated with 8 ultrasonic parameters and 3 clinical factors (age BMI and cortisone intake). In addition, each patient corresponds to 10 GWSI, from which 32 parameters of structural analysis are extracted per image, leading to a total of 43 features per image. The dataset was divided into 70% of patients (n = 136) as training and 30% as testing (n = 59). The distribution of patients was adjusted for age and target class. The accuracy was calculated for an increased number of features, which ranking was obtained using Recursive Feature Elimination (RFE). The highest accuracy of 71% is obtained with the optimized parameters and a combination between 22 and 25 features. These result, comparable to femoral DXA (AUC = 0.71, adjusted linear regression), opens perspective towards robust detection of patients at risk of fracture with ultrasound.