Prediction of osteoporotic hip fracture in postmenopausal women through patient-specific FE analyses and machine learning

•Supervised machine learning (ML) is applied in conjunction with clinical and computationally driven mechanical attributes in order to obtain a predictive model of the hip fracture risk.•A semi-automatic and patient-specific DXA-based FE model was used to generate the mechanical response of the bone after a sideways-fall.•Support Vector Machine (SVM) with radial basis function (RBF), Logistic Regression, Shallow Neural Networks and Random Forest were tested through a comprehensive validation procedure to compare their predictive performance.•SVM generated the best-learned algorithm for both experimental setups, including 19 attributes and only clinical attributes, outperforming BMD by 14pp for the first case. A great challenge in osteoporosis clinical assessment is identifying patients at higher risk of hip fracture. Bone Mineral Density (BMD) measured by Dual-Energy X-Ray Absorptiometry (DXA) is the current gold-standard, but its classification accuracy is limited to 65%. DXA-based Finite Element (FE) models have been developed to predict the mechanical failure of the bone. Yet, their contribution has been modest. In this study, supervised machine learning (ML) is applied in conjunction with clinical and computationally driven mechanical attributes. Through this multi-technique approach, we aimed to obtain a predictive model that outperforms BMD and other clinical data alone, as well as to identify the best-learned ML classifier within a group of suitable algorithms. A total number of 137 postmenopausal women (81.4 ± 6.95 years) were included in the study and separated into a fracture group (n = 89) and a control group (n = 48). A semi-automatic and patient-specific DXA-based FE model was used to generate mechanical attributes, describing the geometry, the impact force, bone structure and mechanical response of the bone after a sideways-fall. After preprocessing the whole dataset, 19 attributes were selected as predictors. Support Vector Machine (SVM) with radial basis function (RBF), Logistic Regression, Shallow Neural Networks and Random Forest were tested through a comprehensive validation procedure to compare their predictive performance. Clinical attributes were used alone in another experimental setup for the sake of comparison. SVM was confirmed to generate the best-learned algorithm for both experimental setups, including 19 attributes and only clinical attributes. The first, generated the best-learned model and outperformed BMD by 14pp. The resul