Introduction The socioeconomic burden imposed by osteoporosis is growing exponentially in every economically developed country [1]. The ability to predict subject at high risk of bone fracture is necessary to modulate the therapies but also to motivate necessary restrictions to the life style. Such risk is usually predicted with epidemiology-based models, the most popular being the FRAX system developed at the University of Sheffield. However, these models show in various studies limited predictive accuracy (expressed as area under the Receiver Operating Characteristic curve (AUC)), typically around 0.65 [2]. The primary reason for this moderate accuracy is that while the main cause of osteoporotic bone fractures is the loss of mechanical strength due to the increased porosity of bone tissue, another important cause is the tendency for elders to fall and overload the skeleton more frequently than normal healthy, a factor that can be modelled only as a stochastic process. In the VPHOP project we developed a multiscale model of the neuromusculoskeletal system that can be individualised using patient data from clinical, imaging, and instrumental observation, capable of predicting the risk of bone fracture in a patient as a function of her degree of osteoporosis, of her propensity to fall, and of the expected progression of the disease over time without or with pharmacological treatment. In this work we present such multiscale model, describe its current limitations, and show how a subset of it was validated over multiple restrospective cohorts, and confirmed with population studies. Materials and methods A complete body-organ-tissue-cell multiscale model of the skeletal biomechanics for the prediction of the hip and spine fractures have been realised by the VPHOP consortium, along the lines proposed in [3]. The multiscale model describes the skeletal loading as a stochastic process identified over a large database of musculo-articular forces during physiological and pathological activities, as predicted by an individualised whole body musculoskeletal dynamics model. Such loading spectra are applied, within a Monte Carlo scheme, to a individualised finite element model of the hip and of the spine, which predicts for each load instance if the bone would fracture or not in that condition; since those skeletal sites under those loading conditions fail both in fragile and ductile mode, a full anisotropic elastic-viscoplastic constitute equation was used. The bone constitutive equation is defined at the tissue scale, and homogenised at the continuum level. A bone-remodelling algorithm predicts the changes over time of the tissue morphology due to the cellular activity, which change the constitutive equation, and thus the strength of the bone. The full multiscale model required 62,055 core hours to be solved, which makes it unpractical for large scale clinical use at this point in time. While we expect that in five years the computational cost will be reduced on 20 times only with the improvement of the computational resources, at the present it was decided to conduct the clinical validation using a subset of the multiscale model. This reduced model retrains the body and organ models, uses a linear-elastic failure criterion (accurate for the fragile fractures observed at the hip), and replace the tissue-cell models, with a phenomenological model that predicts the changes of the bone density at the continuum level as a function of time, removing the need to solve repeatedly tissue-scale finite element models involving over 20 million degrees of freedom. Results Various retrospective cohorts were investigated to assess the predictive accuracy of the reduced multiscale model. Here we present the results on a cohort originally investigated at the University of Sheffield, formed by 50 post-menopause women with a femoral neck fracture, and 50 non-fractured post-menopause women age, height, and weight matched. On this cohort the FRAX algorithm showed an AUC of 0.64, whereas the VPHOP model showed an AUC of 0.75, an improvement usually considered sufficient enough to justify the adoption of a new technology in the clinical practice. Comparable results were obtained in other studies conducted by our consortium, over two additional retrospective cohorts in Reykjavik and in Bologna. In order to obtain a confirmation, the same multiscale model was used to conduct three public health studies where entire populations were modelled, using statistical distributions of input values derived from the literature for the specific population being modelled. In all three cases good agreement was found [4-6].