Finite element analysis informed variable selection for femoral fracture risk prediction

Logistic regression classification (LRC) is widely used to develop models to predict the risk of femoral fracture. LRC models based on areal bone mineral density (aBMD) alone are poor, with area under the receiver operator curve (AUROC) scores reported to be as low as 0.63. This has led to researchers investigating methods to extract further information from the image to increase performance. Recently, the use of active shape (ASM) and appearance models (AAM) have resulted in moderate improvements, but there is a risk that inclusion of too many modes will lead to overfitting. In addition, there are concerns that the effort required to extract the additional information does not justify the modest improvement in fracture risk prediction. This raises the question, are we reaching the limits of the information that can be extracted from an image? Finite element analysis was used in combination with active shape and appearance modelling to select variables to develop LRC models of fracture risk. Active shape and active appearance models were constructed based on a previously reported cohort of 94 post-menopausal Caucasian women (47 with and 47 without a fracture). T-tests were used to identify differences between the two groups for each mode of variation. Femur strength was predicted for two load cases, stance and a fall. Stepwise multi-variate linear regression was used to identify shape and appearance modes that were predictors of strength for the femurs in the training set. Femurs were also synthetically generated to explore the influence of the first 10 modes of the shape and appearance models. Identified modes of variation were then used to generate LRC models to predict fracture risk. Only 6 modes, 4 active appearance and 2 active shape modes, were identified that had a significant influence on predicted fracture strength. Of these, only two active appearance modes were needed to substantially improve the predictive mode performance (ΔAUROC = 0.080). The addition of 3 more modes (1 AAM and two ASM) further improved the performance of the classifier (ΔAUROC = 0.123). Further addition of modes did not result in any further substantial improvements. Based on these findings, it is suggested that we are reaching the limits of the information that can be extracted from an image to predict fracture risk.