Introduction
Several multi-factorial diseases (e.g., diabetes, hypertension) are defined on the basis of important risk factors for the clinical outcome. Hypertension, for example, is defined from the measurement of blood pressure that provides information on the likelihood of stroke. Hypercholesterolaemia and osteoporosis are also examples. In the case of osteoporosis, bone mineral density (BMD) is measured both to provide a diagnosis, and to yield information on fracture risk [
1].
Many well-controlled prospective studies with dual energy X-ray absorptiometry (DXA), particularly in elderly women, indicate that the risk of fracture about doubles for each SD reduction in BMD [
2,
3]. The measurement of a risk factor for diagnostic use, however, can only capture one aspect of the likelihood of the outcome when the disease is multifactorial, and in osteoporosis, assessment with BMD captures a minority of the fracture risk. For example, the annual incidence of hip fracture increases approximately 30-fold between the ages of 50 and 90 years; but, from the known relationship between BMD and fracture risk and the loss of bone with age, it is expected that hip fracture risk would rise only fourfold [
4‐
6]. Thus, the increase in risk with age is approximately sevenfold greater than can be explained on the basis of BMD alone.
The imperfect capture of risk with BMD alone poses several problems for the clinical assessment of fracture risk. In the context of population screening with BMD alone, the performance characteristics of the test are less than optimal in terms of the trade-off between sensitivity and specificity [
1,
7,
8]. Thus, osteoporotic fractures affect a substantial minority of the population, but intervention thresholds based on BMD alone lack sensitivity over most reasonable assumption, i.e., the detection rate is low. For example, at the age of 50 years the proportion of women with osteoporosis is approximately 5% [
4]. The proportion of these who will fracture in the next 10 years (i.e., positive predictive value) is about 20%. The detection rate for these fractures (sensitivity) is, however, low, and 96% of fragility fractures would arise in women without osteoporosis given a test like BMD where the fracture risk doubled for each SD decrease [
9]. Low sensitivity is one of the reasons why wide-spread population screening is not widely recommended in women at the menopause [
1]. Moreover, a normal BMD measurement is no guarantee that a fracture will not occur.
The use of risk factors that add information on fracture risk independently of BMD improves the sensitivity of the assessment for any specificity [
8,
9]. Over the past several years, we have undertaken a series of meta-analyses to identify clinical risk factors for fracture that provide independent information on fracture risk [
3,
10‐
16]. The analyses were based on the primary data from prospective population based studies. This permits the inter-dependence of each of the candidate risk factors to be examined so that they can be accurately combined for clinical use. In the case of the clinical risk factors, BMI is used as a continuous variable, so that its distribution is preserved by the addition of dichotomous variables [
8]. For hip fracture prediction, the gradient of risk (increase in fracture risk per standard deviation increase in risk score) for the CRFs is comparable to the use of DXA alone, and the gradient of risk is further enhanced by the addition of BMD to the CRFs [
17]. These inter-relationships, assessed from multiple populations and validated in independent cohorts, permit the more accurate identification of individuals who will fracture, so that the average risk in any given proportion of the population identified for treatment will be higher [
8].
These considerations indicate that assessment can be improved by the integration of clinical risk factors with or without BMD. In other words, treatment should be directed not only on the basis of T-score for BMD, but also on the independent contribution of other validated risk factors. However, the use of different metrics (the T-score, gradients of risk, risk ratios, etc) is confusing for clinicians and patients alike. For this reason, there has been interest in the development of algorithms that express absolute risk, or the probability of fracture within a given time period [
18‐
20], as has been done for cardiovascular and other diseases [
21‐
25]. The aim of the present study was to develop a model for the clinical assessment of fracture probability in men and women based on the epidemiology of the UK.
Discussion
The present study provides a model for the assessment of fracture probability in men and women. The model (FRAX™) uses data derived from nine cohorts from around the world, including centres from North America, Europe, Asia and Australia and has been validated in 11 independent cohorts with a similar geographic distribution [
17]. The use of primary (but anonymized) data for the model construct permits the interaction of each of the risk factors to be determined to improve the accuracy whereby fracture probability can be computed. The large sample permitted the examination of the general relationship of each risk factor by age, sex, duration of follow up and, for continuous variables (BMD and BMI), the relationship of risk with the variable itself in a manner hitherto not possible. The use of primary data also eliminates the risk of publication bias. The validity of the clinical risk factors identified are supported by the expected relationships between BMD and fracture risk [
3].
In the present study, the FRAX™ model has been calibrated to the epidemiology of the UK, but could be calibrated to any country where the epidemiology of fracture and death is known (see
Appendix). FRAX™ models for the UK and some other countries are available through the web (
http://www.shef.ac.uk/FRAX/index.htm). The approach uses easily obtained clinical risk factors to estimate risk. The estimate can be used alone or with BMD to enhance fracture risk prediction.
Several previous studies have developed models to predict fracture risk from the combination of clinical risk factors and BMD [
18,
65‐
78]. The risk factors used include activities of daily living, impaired cognition, liability to falls, poor overall health, history of stroke, seizure disorder and several different medications. A limitation of many of these studies is that, with the exception of the SOF study [
18], and one of the GPRD studies [
76], they have not been tested in other cohorts. The model described in this paper has been validated in 11 independent prospectively studied cohorts with in excess of one million patient years [
17].
The use of risk factors for case finding presupposes that the risk so identified is responsive to a therapeutic intervention. To test this hypothesis, it would be necessary to recruit patients selected on the basis of the risk factor(s) to a randomised controlled trial (RCT). The risk factor that is best evaluated in this way is BMD, and indeed the vast majority of therapeutic studies have recruited patients on the basis of low BMD as recommended by regulatory agencies in the US and Europe [
79,
80]. In recent years, other trials have recruited patients on the basis of age, gender, a prior vertebral fracture and current exposure to glucocorticoids irrespective of BMD, and have shown therapeutic effects similar to those noted in RCT’s based on BMD selection [
14,
81‐
83].
For other risk factors, comparable data are lacking. In the absence of empirical data, an alternative approach is to demonstrate that the presence (or absence) of a risk factor does not adversely influence therapeutic efficacy against fractures. Several studies have shown no significant interaction between response to treatment and the presence or absence of the risk factors used in the present study including age, height, family history of fracture, low body weight or BMI, smoking, alcohol intake or prior non-vertebral fracture [
84‐
88]. In contrast, some risk factors may be associated with less therapeutic efficacy. For example, patients selected on the basis of risk factors for falling may respond less completely to agents that preserve bone mass than patients selected on the basis of low BMD [
89]. This concern is greatest in models that omit BMD, because pharmacological agents may not be equally effective across the entire range of BMD [
90].
The present model has several unique features. FRAX™ uses Poisson regression to derive hazard functions of death and fracture. Such hazard functions are continuous as a function of time, unlike Cox’s regression for which the corresponding hazard functions are zero except at the time points of a fracture or death. There are also several advantages of the Poisson model over logistic regression analysis. Logistic regression does not take account of when a fracture occurred, nor whether individuals without a fracture died or when death occurred. Secondly, for the assessment of 10-year probabilities by logistic regression, the observation period should be for 10 years. Moreover, information longer than the 10-year period cannot be used for analysis. The cost of ignoring information when fractures occur and whether and when deaths occur is on the precision of the estimate. In simulation experiments, the Poisson model gives the same precision as logistic regression with fewer numbers of individuals. In our own simulations in the present context (data on file), the Poisson model gave the same precision as logistic regression with half the number of individuals. Finally, the Poisson model allows adjustments to be made for time trends. The ability to use several Poisson models permits the use of data from different sources to integrate fracture and death hazards, and to calibrate to different countries.
A further feature of the FRAX™ model is that it takes account of deaths from all causes. In several recent models of disease probability, this has not been accounted for [
23‐
25]. For example, the probability of stroke has been determined as a function of age, race, smoking, body mass index, atrial fibrillation, HbA1c, systolic blood pressure, ratio of total to HDL cholesterol and duration of diabetes [
25], but the risk of dying from other reasons was not taken into account. In the context of osteoporosis, fracture probabilities are markedly underestimated when no account is taken of the competing death hazard [
27,
35,
36]. For example, in a study of men followed after orchidectomy, the cumulative incidence of fractures was 19% after 15 years, but the figure was 40% when deaths were considered as a competing event [
36].
FRAX™ also takes account of the impact of risk factors on the death hazard. For example, smoking and low BMD are risk factors for fracture but also significant risk factors for death. Thus, at very low T-scores for BMD, hip fracture probabilities decrease with age (see Table
3), in part related to the higher mortality associated with the lower values for BMD.
There are several limitations that should be mentioned. As with nearly all randomly drawn populations, non-response bias may have occurred. The effect is likely to exclude sicker members of society, and may underestimate the absolute fracture risk for example by age. The analyses also have significant limitations that relate to the outcome variables and the characterisation of risk factors. The definition of what was considered to be an osteoporotic fracture was not the same in all cohorts, but the effect of this inconsistency is likely to weaken rather than strengthen the associations that were found. For the hip fracture outcome, the definition was similar in all cohorts, and may explain in part the higher risk ratios associated for this fracture rather than for osteoporotic fracture. Also, the analyses were confined to clinical fractures, and the results might differ from vertebral fractures diagnosed by morphometry or as an incidental radiographic finding.
There are also limitations with the risk factors themselves. In the case of BMI, this was chosen rather than weight as the measure for body composition. This has the advantage that there is less variability across countries and between sexes. A potential drawback is that BMI can be influenced by height loss associated with vertebral deformities. Therefore, in individuals with important loss of height, the risk conferred through BMI could be underestimated [
91]. The use of maximal attained height, rather than current height, might be a solution in the future, if it were shown that fracture risk prediction could be improved.
Further problems relate to the construct of the questions to elicit the presence or absence of risk factors, which varied between cohorts. These included questions on family history, prior fracture, smoking and glucocorticoid use. The effect of this heterogeneity is likely to weaken rather than strengthen the associations found. Recall is also subject to errors and was not validated in any of these cohorts. This is particularly problematic in the elderly. In addition, the validity of self-reported alcohol intake is notoriously unreliable [
92]. Indeed, alcohol consumption was significantly less in both men and women than that assessed in the UK [
93]. Given that these studies were prospective, however, much of this error (with the exception of alcohol intake) should be random, giving rise to non-directional misclassification. Thus, the associations may actually be stronger than reported here. Any underestimate may have limited consequences for case-finding, since the populations to be tested are similar to the populations interrogated. Biases that arise have more significance where causality is inferred.
A further limitation is that several of the clinical risk factors identified take no account of dose-response, but give risk ratios for an average dose or exposure. By contrast, there is good evidence that the risk associated with excess alcohol consumption and the use of glucocorticoids is dose-responsive [
14,
94]. In addition, the risk of fracture increases progressively with the number of prior fractures [
95]. These limitations are nearly all conservative.
It should be acknowledged that there are many other risk factors that might be considered for incorporation into assessment algorithms. These include BMD at other skeletal sites, ultrasonography, quantitative computed tomography and the biochemical indices of bone turnover. The available information was too sparse to provide a meta-analytic framework, but they should be incorporated into risk assessment algorithms when they are more adequately characterised. Notwithstanding, the present model provides a mechanism to enhance patient assessment by the integration of clinical risk factors alone and/or in combination with BMD.
The application of this methodology to clinical practice will demand a consideration of the fracture probability at which to intervene, both for treatment (an intervention threshold) and for BMD testing (assessment thresholds). These are currently being developed for the UK, based on cost-effectiveness analyses [
96]. Intervention thresholds developed for the United Kingdom may not be applicable to other countries. The 10-year probability of fracture varies markedly in different countries [
97]. For countries with low hip fracture rates, as found in developing countries, the relative risk at which intervention is cost-effective will be higher, though the absolute risk at which intervention is cost-effective would not change assuming comparable costs. Intervention thresholds would, however, change with differences in costs, particularly fracture costs, which vary markedly world wide. There is also the issue of affordability or willingness to pay for a strategy. The gross domestic product (GDP) per capita provides an index of affordability. The GDP varies markedly in different regions of the world. In the UK, the GDP per capita is estimated at US$ 25,300 in 2002, as compared with US$ 7,000 in Turkey. Thus, for the same fracture risk and the same costs, treatment will be less affordable (at least to health services) in Turkey than in the UK. Nevertheless, individuals in Turkey, rather than society as a whole, may be willing to pay “United Kingdom prices” for health care. There is also a marked heterogeneity in the proportion of GDP devoted to health care, and in the proportion of the population at risk from osteoporotic fracture (i.e., elderly people) [
98,
99]. For all these reasons, it is important to define intervention and assessment thresholds on a country by country basis that takes into account the setting for service provision and willingness to pay, as well as considerations of absolute costs.