This study assessed the additional utility of adding HRpQCT, BMD, microarchitectural and FEA parameters to enhance fracture discrimination in a cohort of elderly men and women. Whilst individual measures of bone microarchitecture and FEA discriminated fracture cases versus non-fracture groups, and were selected in cluster analyses, little benefit of adding the FEA parameters was found over and above femoral neck aBMD and bone microarchitecture in terms of fracture discrimination.
In terms of individual bone microarchitectural parameters, higher total and trabecular area and lower cortical thickness were associated with increased fracture risk in unadjusted and adjusted analyses among men; corresponding parameters among women were lower cortical area and porosity, trabecular density and thickness. The cortical porosity result in women seems a little counterintuitive though is consistent with previous work in this cohort [12], and in GLOW [13]. The observation may be due to the way the cortical porosity analysis script defines a pore [26] which is based on how many neighbouring voxels have a similar low attenuation value. If an individual has less cortical bone and thinner cortices (as shown in our current analyses), it may be that fewer pores meet these criteria, which would be reflected in a lower % porosity. Secondly, the bone of participants with fracture may have a lower turnover, and thus repair rate, which may also result in fewer pores and an increase in fracture risk. In men, no FEA parameters were associated with fracture risk, but in women, lower bone stiffness, failure load, Young modulus and trabecular stress and strain were all associated with higher fracture risk in both unadjusted and adjusted analyses.
It appears that men at higher risk of fracture had larger bones with a thinner cortex and deterioration of trabecular bone, which was not translated into reduced bone strength as measured by linear FEA. Linear FEA has been criticised as it only evaluates stresses and strains placed on the bone in one direction, representing a linear compressive force on the bone, whereas non-linear FEA may provide additional information on fracture risk [27]. Perhaps this phenotype in men may produce vulnerability to forces placed on bones from other directions, including bending, which would not be identified by linear FEA. If this were the case, it may be expected that these men may be at higher risk of hip or radial fractures, frequently sustained on falling, compared with vertebral, classically compressive, fractures. During ageing, men also compensate for bone loss with greater periosteal formation to a greater extent than women [28], which may protect against compressive forces better in males. Conversely, the impaired linear bone strength parameters seen in women may suggest these women are more prone to compressive fractures. Vertebral fractures were the most common fracture among both males and females, although due to the small number of fractures in our cohort, comparison of fracture types in different clusters was not possible. It would be interesting to investigate this further, using non-linear FEA, in a larger sample with larger numbers of fractures.
The ROC curve analyses showed the benefit of adding relevant bone microarchitectural parameters into DXA-derived femoral neck aBMD assessment of fracture risk. This benefit was not statistically significant, but this may be influenced by our small sample size. Despite the association of FEA parameters with fracture risk in females, ROC curve analysis revealed FEA had little additional benefit in predicting fracture over models using bone microarchitectural parameters and aBMD. FEA is computationally intensive, and therefore time consuming and expensive. Our analyses suggest the additional information may not justify its use in predicting fractures in clinical practice. These findings are in contrast with previous studies which have found FEA to be a valuable predictor of prevalent and incident fracture, more important than other HRpQCT measures [5,6,7, 29]. One study found a machine learning model incorporating HRpQCT measures could predict fractures better than aBMD [30]. Another study reported the additional benefit of a combination of a trabecular and cortical parameter or FEA failure load to femoral neck aBMD or FRAX-BMD in predicting fracture in postmenopausal women. However, failure load could be replaced with aBMD of the ultra-distal radius with no significant reduction in predictive worth [24]. This is similar to our findings, with the addition of FEA providing little additional benefit over bone microarchitectural parameters.
Our cluster analysis revealed similar clusters to previous findings [12] with Cluster 1 showing a predominantly ‘cortical deficiency’ phenotype based on bone microarchitectural parameters. In males, trabecular area was greater, perhaps reflecting a greater proportion of trabecular bone due to a reduction in cortical bone. In females, trabecular density and thickness were also lower, suggestive of more generalised deterioration of bone structure. FEA in this cluster showed greater percentage of load on trabecular bone, lower Young modulus and lower cortical stresses in both males and females and lower bone stiffness and bone failure load in females. Compared to the lowest risk cluster, this cluster also had lower mean aBMD, and was the only cluster with significantly greater fracture risk in females.
The trabecular phenotype was less definitive than in our previous analysis but was indicated in Cluster 2 where females tended towards a ‘trabecular deficiency’ phenotype, and tended towards higher fracture risk, although this did not reach statistical significance. This is likely due to the additional FEA parameters included in the cluster analysis compared to the previous analysis which only included bone microarchitectural parameters. Bone microarchitectural and FEA parameters were similar to the wider sample, even though there was a significantly lower aBMD in both males and females compared to the lowest risk cluster.
These findings demonstrate the importance of HRpQCT parameters and identify cortical deterioration as a key contributor to fracture risk. As with our findings, previous studies have found both cortical and trabecular microarchitectural deterioration important for fracture risk, some finding cortical [31, 32], and some finding trabecular [9, 19, 33] changes more important. One study found both lower trabecular and cortical volumetric density were independently associated with fracture incidence [24]. Our use of cluster analysis helps to elucidate different phenotypes within the population at risk of fractures, with different contributions of both trabecular and cortical parameters.
The strengths of our study include basing our analyses on the HCS, a well characterised cohort where data were rigorously collected by an experienced multidisciplinary team. Furthermore, our analyses used both bone microarchitectural and FEA parameters to provide a comprehensive illustration of bone phenotypes.
This study has some limitations. Firstly, a healthy responder bias has been observed in HCS and examining participant characteristics according to inclusion status has revealed healthier lifestyles at baseline for participants included in the analysis sample compared to those who were not. However, our analyses were internal, so bias would only arise if the associations of interest differed systematically between those who were included in the analysis sample and those who were not; this seems unlikely. Secondly, temporal causation cannot be inferred as our study has a cross-sectional design. It may be that the differences in bone microstructure seen are secondary to remodelling in response to fracture, rather than properties of the bone which predispose to fracture, especially as we have only collected information about previous fractures. Thirdly, fracture status was missing for some participants, although this information was available for the vast majority (91.9%) of the analysis sample. Finally, the low numbers of reported fractures and a relatively small sample size, along with the lack of stability regarding cluster analysis algorithms in general, may limit the generalisability of findings. However, the similarity of the clusters observed to those in other analyses and their biological plausibility suggests that they are robust.
In conclusion, microarchitectural deterioration, bone geometry and, in women, FEA-derived bone strength contributed to an increased risk of previous fracture. Cluster analysis revealed a cortical and a trabecular deficiency phenotype, which both showed lower aBMD in men and women. Only women with the cortical deficiency phenotype had significantly increased risk of previous fractures. In this cohort, adding bone microarchitectural parameters to aBMD could better predict previous fracture, but further addition of FEA conferred little benefit.