Predictive capacity of fracture risk assessment tools: overview of systematic reviews

Conduct an overview of systematic reviews of the current fracture risk prediction tools in use.

We included systematic reviews (SRs) that assessed the predictive ability of any tool, score, algorithm, or other instrument for fracture risk. The primary outcome measure was the area under the curve (AUC) representing predicted fracture risk within a specified timeframe obtained from receiver operating characteristic (ROC) analysis. We included SRs that studied both men and women with fractures in the general adult population.

The review identified 26 different tools currently in use to predict fracture risk. Within these tools a total, 21,717 different prediction variables were found. Among the different tools, a different number of factors were used ranging from the BWC model that used a single predictor variable to the GSOS tool that incorporated 21,717 predictor variables in its model (including many individual SNPs). Regarding the performance of the tools, AUC ranging from 0.58 to 0.90. None of the models had a prediction capacity greater than 90%. Most of the models are within the range of 0.7 and 0.75, but it cannot be said that any specific one stands out over the others. Rather, a fluctuating behavior is observed in all models within the different studies. The discrimination of the two most frequently validated models, including FRAX with and without BMD, varied among the studies with AUC/C index ranging from 0.58 to 0.90, respectively. Other commonly validated model, including the Garvan Model showed AUC between 0.57 to 0.84.

The vast majority of the models performance is within the range of 0.7 and 0.75. To compare the performance of different tools when predicting fracture, it is very important to consider the differences between prediction tools, the number of risk factors considered, as well as the nature of the variables as they will have an important impact on the feasibility of its use in clinical practice. Likewise, differences in the prediction results may depend on sex, age, types of fractures, as well as the temporal intervals of the prediction and could affect the use of the tools in the daily clinical routine.