The prediction of the stress-state and fracture risk induced in bones by various loading conditions in individual patients using subject-specific finite element models still represents a challenge in orthopaedic biomechanics. The accuracy of the strain predictions reported in the literature is variable and generally not satisfactory. The aim of the present study was to evaluate if a proper choice of the density–elasticity relationship can lead to accurate strain predictions in the frame of an automatic subject-specific model generation strategy. To this aim, a combined numerical–experimental study was performed comparing finite element predicted strains with strain-gauges measurements obtained on eight cadaver proximal femurs, each instrumented with 15 rosettes mostly concentrated in the bone metaphyses, tested non-destructively in vitro under six different loading scenarios. Three different density–elasticity power relationships were selected from the literature and implemented in the finite element models derived from computed tomography data. The results of the present study confirm the great influence of the density–elasticity relationship used on the accuracy of numerical predictions. One of the tested constitutive laws provided a very good agreement (R²=0.91, RMSE lower than 10% of the maximum measured value) between numerical calculations and experimental measurements. The presented results show, in addition, that the adoption of a single density–elasticity relationship over the whole bone density range is adequate to obtain an accuracy that is already suitable for many applications.