Prognosis of fracture: evaluation of predictive accuracy of the FRAX™ algorithm and Garvan nomogram

Summary We evaluated the prognostic accuracy of fracture risk assessment tool (FRAX™) and Garvan algorithms in an independent Australian cohort. The results suggest comparable performance in women but relatively poor fracture risk discrimination in men by FRAX™. These data emphasize the importance of external validation before widespread clinical implementation of prognostic tools in different cohorts. Introduction Absolute risk assessment is now recognized as a preferred approach to guide treatment decision. The present study sought to evaluate accuracy of the FRAX™ and Garvan algorithms for predicting absolute risk of osteoporotic fracture (hip, spine, humerus, or wrist), defined as major in FRAX™, in a clinical setting in Australia. Methods A retrospective validation study was conducted in 144 women (69 fractures and 75 controls) and 56 men (31 fractures and 25 controls) aged between 60 and 90 years. Relevant clinical data prior to fracture event were ascertained. Based on these variables, predicted 10-year probabilities of major fracture were calculated from the Garvan and FRAX™ algorithms, using US (FRAX-US) and UK databases (FRAX-UK). Area under the receiver operating characteristic curves (AUC) was computed for each model. Results In women, the average 10-year probability of major fracture was consistently higher in the fracture than in the nonfracture group: Garvan (0.33 vs. 0.15), FRAX-US (0.30 vs. 0.19), and FRAX-UK (0.17 vs. 0.10). In men, although the Garvan model yielded higher average probability of major fracture in the fracture group (0.32 vs. 0.14), the FRAX™ algorithm did not: FRAX-US (0.17 vs. 0.19) and FRAX-UK (0.09 vs. 0.12). In women, AUC for the Garvan, FRAX-US, and FRAX-UK algorithms were 0.84, 0.77, and 0.78, respectively, vs. 0.76, 0.54, and 0.57, respectively, in men. Conclusion In this analysis, although both approaches were reasonably accurate in women, FRAX™ discriminated fracture risk poorly in men. These data support the concept that all algorithms need external validation before clinical implementation.