Introduction

From the age of 60, the lifetime risk of fracture for women is 44% [1]. The lifetime risk of hip fracture for a white woman (1/6) is higher than the lifetime risk of developing breast cancer (1/9) [2]. A preexisting fracture confers a substantially increased risk of subsequent fracture [3, 4] and premature mortality [5], which in turn incurs significant heath-care costs [6, 7]. Currently, fewer than one-third of patients who sustain a fragility fracture are properly diagnosed and treated [8]. Therefore, one of the major priorities in osteoporosis research is to develop valid models for identifying individuals at high risk of fracture for allocating appropriate medical intervention [9].

Accumulated research evidence during the past three decades has suggested that advancing age, low bone mineral density (BMD) [10], low body mass index [11], smoking [12], corticosteroid use [13], and prior fractures [3, 4] are major independent predictors of fracture in women. Using these clinical risk factors, a number of prognostic models have recently been developed for predicting fracture risk [14–18]. The discriminatory value of these prognostic models is moderate to good, with areas under the receiver operating characteristic curve ranging 0.7–0.8 [17, 18]. However, there is still room for improving the predictive accuracy of these prognostic models by incorporating information from nonclinical risk factors.

Fracture liability is partially determined by genetic factors [19]. Twin studies suggest that approximately 48% of the liability to hip fracture is genetically related [20]. Among the genes that have been implicated in the regulation of fracture risk, the collagen I alpha 1 (COL1A1) gene has been consistently associated with fracture risk [21–28]. If the effect of COL1A1 on fracture risk is independent of clinical risk factors, then the incorporation of COL1A1 genotype could improve the prognosis of fracture.

A nomogram is a prognostic model that combines several risk factors to provide an accessible assessment tool to clinicians and patients. We have recently developed simple prognostic models for individualizing the risk of fracture [17] based on simple clinical risk factors. It is hypothesized that the incorporation of COL1A1 genotypes would improve the predictive accuracy of the nomogram. The present study was designed to test that hypothesis by developing a prognostic model for predicting fracture risk, taking into account COL1A1 genotype.

Materials and Methods

Study Design

The present study was part of the ongoing Dubbo Osteoporosis Epidemiology Study, which was designed as a prospective population-based investigation, in which women aged 60+ years in 1989 were invited to participate. The study details and protocol have been described elsewhere [1]. Briefly, Dubbo is a city located approximately 400 km from Sydney, with a population of 32,000 people. It was selected for the study due to its stable population and geographic isolation, which allow total ascertainment of fracture incidence. This study was approved by the St. Vincent’s Ethics Committee, and written informed consent was obtained from each participant.

Measurements and Fracture Ascertainment

The original study population included 888 men and 1,402 women progressively recruited from 1989 and 1994. However, in this study, clinical history and anthropometric data from 915 women, who had provided blood samples for genetic analysis, were analyzed. Each woman was interviewed by a nurse coordinator at initial and subsequent visits at approximately 2-year intervals. At each visit a structured questionnaire was administered to collect data, including lifestyle factors (i.e., smoking habits, dietary calcium intake) and anthropometric variables (including weight and height) were measured. In addition, any history of falls in the preceding 12 months and any history of fractures after the age of 50 were recorded.

BMD (g/cm2) was measured at the lumbar spine (LSBMD) and the femoral neck (FNBMD) by dual-energy X-ray absorptiometry with a LUNAR DPX-L densitometer (GE Lunar, Madison, WI). The precision of BMD measurements in normal subjects at our institution is 1.3% at the lumbar spine and 3.5% at the femoral neck [29]. In this study, FNBMD was used in the development of the prognostic model because it is less susceptible to age-related degenerative changes than LSBMD.

Fractures were recorded from radiology centers servicing the Dubbo area, and the circumstances surrounding the fractures were determined by personal interview. All fractures included in the study were low-trauma fractures caused by a fall from standing height or less. Vertebral fractures were clinically diagnosed. There was no systemic X-ray screening before the study to identify prevalent or asymptomatic vertebral fractures. Incidentally found, i.e., asymptomatic, vertebral fractures were included, provided that there was no known malignancy or metabolic bone disease. Morphometric vertebral fractures were not considered in the analysis.

Genotyping

Blood samples were collected at baseline and in each subsequent visit for genetic and biochemical analyses, and DNA was extracted from leukocytes. The polymorphism of G to T in the first intron of the COL1A1 gene was genotyped by the restriction fragment length polymorphism method with MscI (New England Biolabs, Beverly, MA) and an isoschizomer of BalI. Digestions were analyzed by 3% agarose gel electrophoresis. Genotypes for this restriction site were denoted as GG, GT, and TT, with T being the minor allele. To validate the accuracy of genotyping, 30 subjects were randomly regenotyped with 100% consistency.

Data Analysis

The association between clinical risk factors and COL1A1 genotype and fracture was analyzed by the Cox proportional hazards model. In this model, the time from entry to fracture was considered an end point. Based on estimated parameters of the optimal model, nomograms with the inclusion of COL1A1, age, FNBMD, prior fracture, and fall were constructed using the Design library within the R system [30, 31]. In order to assess the incremental prognostic value attributable to COL1A1 genotypes, a reclassification analysis [32] was performed. Two specific models were considered: model I, with age and BMD, prior fracture, and fall, and model II, with age, BMD, prior fracture, fall, and COL1A1 genotypes. The area under the receiver operating characteristic curve [33] was not used because it is too insensitive to change [34]. In this reclassification analysis, the 10-year probability of fracture was estimated by each model and then classified into three risk groups: <10%, 10–20%, and >20%. The proportion of women who would be reclassified into three risk groups between the model without COL1A1 genotypes and the model with COL1A1 genotypes was calculated, and this proportion is regarded as the net improvement in prediction.

Results

During the follow-up period (median 9 years, range 0.2–18), 332 women had sustained at least one fragility fracture, including 77 hip and 127 clinical vertebral fractures. Women with fracture were older and had lower body weight, shorter stature, and lower BMD than women without fracture (Table 1). The differences were more pronounced in patients with hip fracture, in which FNBMD was approximately 1 standard deviation (SD) lower than in women without fracture. However, there were no significant differences in smoking habit, alcohol consumption, and dietary calcium intake between the fracture and nonfracture groups.

Table 1 Baseline clinical and anthropometric characteristics of participants stratified by fracture statures
Full size table

COL1A1 Genotypes and Fracture Risk

The Sp1 binding genotypes of the COL1A1 gene were successfully determined in 894 women. The distribution of COL1A1 genotypes was consistent with the expected frequencies by the Hardy–Weinberg equilibrium law: GG (63.8%), GT (32.6%), and TT (3.6%). There were no statistically significant differences in age, weight, and BMD among the genotypes (Table 2). However, carriers of the TT genotype had significantly greater height than those with the GG genotype. Women homozygous for the T allele had a greater dietary calcium intake than those without the allele (P = 0.02).

Table 2 Baseline clinical and anthropometric characteristics of participants stratified by COL1A1 genotype
Full size table

Among the fracture group, 6.2% (n = 20) of fracture cases were carriers of the TT genotype; this proportion was significantly higher than that in the nonfracture group (2.3%, P = 0.03) (Table 1). In addition, the frequency of this homozygote was more prevalent in the hip fracture group (9.4%) and the vertebral fracture group (8.9%) compared with the nonfracture group. Furthermore, women with the TT genotype had a greater cumulative probability of fracture than those with the GG or GT genotype (Fig. 1)

Fig. 1
figure 1

Cumulative probability of any fracture (a), hip fracture (b), and clinical vertebral fracture (c), stratified by COL1A1 genotype

Full size image

Risk Factors for Fracture

Advancing age, weight loss, shorter stature, lower BMD at the femoral neck or lumbar spine, and a history of fractures and falls were associated with high risk of any fracture, hip or vertebral (Table 3). In hip fracture, the magnitude of association was more pronounced in which every 5 years of advancing age or 10 kg of lower weight had a double risk of hip fracture (Table 3). Similarly, women who were 5 cm shorter had a 50% risk of sustaining a hip fracture, but the magnitude of effect was reduced in non-hip fractures (Table 3).

Table 3 Association between clinical and genetic risk factors and fracture: univariate analysis
Full size table

BMD at both the femoral neck and lumbar spine was consistently associated with the risk of all types of fracture. Indeed, FNBMD was the most consistent predictor of hip fracture with a 3.35-fold (95% confidence interval [CI] 2.67–4.19) increase in hip fracture risk for every SD lower in BMD, while each SD lower in LSBMD was associated with a double risk of vertebral fracture (Table 3). The same relationship was observed for other fracture types, albeit with lesser effect size. There was no significant association between a history of falls and vertebral fracture risk.

Women carrying the TT genotype had an increased risk of any fracture (relative risk [RR] = 1.91, 95% CI 1.21–3.04), hip (RR = 4.21, 95% CI 1.88–9.45), and vertebral (RR = 3.57, 95% CI 1.89–6.77) fracture compared to women with the GG genotype. There were no statistically significant associations between the GT and GG genotypes for any type of fractures (Table 3). Therefore, in multivariable analysis, participants with GT and GG genotypes were combined into one group and compared with those with the TT genotype. This was in agreement with a previous study [25]. When age, FNBMD, and a history of fractures and falls were adjusted in a multivariable Cox proportional hazards model, the strength of association between TT genotype and risk of all fractures did not change or was even stronger (Table 4).

Table 4 Independent risk factors of fracture: multivariable analysis
Full size table

Reclassification Analysis

For any fracture, approximately 13% of women were reclassified into higher or lower risk categories in the model containing COL1A1 genotypes (Table 5). This reclassification was 30% for hip fracture and 33% for clinical vertebral fracture (Table 5). On further analysis by fracture status to account for the “correct” movement into risk categories (higher risk for women in whom fracture had occurred and lower risk for women free of fracture), the inclusion of COL1A1 genotypes mainly improved the classification of nonfracture rather than fracture cases. For example, in 561 women who had not sustained a fragility fracture, the model with COL1A1 genotypes reclassified 17 down and eight up, a net improvement of nine (or ~2%). On the other hand, among 317 women who sustained a fracture, the model with COL1A1 genotypes was tied with the model without COLIA1 genotypes. The overall net reclassification improvement was 4% for hip fracture and 5% for clinical vertebral fracture.

Table 5 Comparison of 10-year predicted risk of any fracture, hip fracture, and clinical vertebral fracture in models including age and BMD T score with and without COL1A1 genotype
Full size table

Multivariable models including the three independent predictive factors age, FNBMD and COL1A1 genotype were developed for the risk of any, hip, or vertebral fracture. The areas under the curve of these models were 0.76, 0.88, and 0.72 for predictive models of any, hip, and vertebral fracture, respectively. Based on parameters obtained from multiple Cox proportional hazards models, nomograms were developed for prediction of any fracture (Fig. 2a), hip fracture (Fig. 2b), and clinical vertebral fracture (Fig. 2c).

Fig. 2
figure 2

Prognostic nomograms for predicting the probability of sustaining any fragility fracture (a), hip fracture (b), and vertebral fracture (c) based on age, FNBMD T score, history of fractures and falls, and genotype. Instruction for usage: Draw a vertical line from “Age” to the “Total Points” axis to determine the score for age. Repeat the process for the other risk factors to obtain total points, and from the “Total Points” axis draw a vertical line to the 5-year and 10-year risk axes to estimate the probability of sustaining a fracture in the next 5 and 10 years, respectively

Full size image

Clinical Applications of the Nomogram

Patient 1

Patient 1 was a 75-year-old woman with an FNBMD T score of –2.5 (osteoporosis), one prior fracture, one fall in the last 12 month, and GG genotype. Her age (Fig. 2b) scored 18 points, T score 60 points, prior fracture 9 points, fall 3 points, and GG genotype 0 points. Therefore, the total points for the woman were 90 on the “Total Points” axis, giving an estimated 5-year average risk of hip fracture of ~10%.

Patient 2

Patient 2 was a 60-year-old woman with an FNBMD T score of –2 (osteopenia), one prior fracture, one fall in the last 12 months, and TT genotype. Her age (Fig. 2b) scored 5 points, T score 55 points, prior fracture 9 points, fall 3 points, and TT genotype 23 points. Therefore, the total points for the woman were 95 on the “Total Points” axis, giving an estimated 10-year average risk of hip fracture of ~12%.

Discussion

Although several candidate genes have been suggested to be associated with fracture, the replication of these findings has been poor [35], as is often seen in other fields of research [36]. Among the candidate genes of osteoporosis that have been studied, the association between COL1A1 and fracture has been independently replicated in several studies [21]. However, the extent to which the genetic information could improve fracture prognosis is largely unknown. The present result confirmed that women carrying the COLIA1 Sp1 TT genotype had an increased risk of fragility fracture, including hip and vertebral fracture, and that the use of this polymorphism could improve the predictive accuracy of fracture prognosis over and above that of age and FNBMD.

Our result is consistent with an earlier meta-analysis which found an additive association between the COL1A1 gene variant and fracture risk [26]. In that meta-analysis, the risk of any fracture among TT carriers was increased by 1.78-fold compared to those with the G allele [26]; however, in a large-scale study on 20,708 Caucasian participants, the TT genotype was associated with a 33% increased risk of incident vertebral fracture, and this association was independent of BMD [37]. Taken together, these data suggest that the COL1A1 gene is a suitable candidate gene for improving the predictive value of existing prognostic models of fracture. In this study, we have demonstrated a simple way to translate a genetic effect into such prognostic models for individualizing fracture risk.

Developing a genetic test for assessment of genetic predisposition to fracture is one of the goals in osteoporosis research. This means using genetic testing to identify individuals with higher risk of fracture, who could be encouraged either to change lifestyle factors or to take medications to reduce fracture risk. Prognosis based on our nomogram model could be helpful in fracture prediction by providing an absolute risk rather than placing individuals into risk-group strata usually expressed by relative risk. This absolute risk approach allows a better appreciation of fracture risk to an individual because it takes into account the time dimension of the risk.

There are some major advantages of using genetic markers as prognostic factors of fracture risk. First, since an individual’s genotype is time-invariant, it is easier to estimate its effect size and to incorporate its information in a prognostic model. Second, as the association between COL1A1 and fracture risk appears to be independent of clinical risk factors, the use of such a genetic marker can potentially improve the predictive value. Third, although there is no genetic therapy for individuals at high risk of fracture, the use of genetic markers could help segregate individuals at high risk from those with low risk of fracture and help manage the burden of osteoporosis in the community.

The potential usefulness of genetic testing for fracture risk prediction can be evaluated by comparing the discriminative accuracy of predictions based on models which do and do not include the genetic information. Traditionally, the increase in the area under the receiver operating characteristic curve was considered a measure of improvement; however, in recent years it has become clear that such a measure is not optimal because it is relatively insensitive, leading to omission of important prognostic factors [33]. In our study, for example, for hip fracture prediction, the area under the receiver operating characteristic curve for model with age, BMD T score, prior fracture, and fall was 0.86; when COL1A1 genotypes were added to the model, the value increased to 0.88—a modest improvement. However, using the reclassification analysis, it was shown that the improvement of prognosis due to COL1A1 genotypes was substantial, with ~5% reclassification for hip and vertebral fractures. Moreover, in a further analysis, the improvement was mainly in the specificity, not the sensitivity, of prognosis. This is perhaps not surprising because the relative frequency of the “risk genotype” (TT) is only 5% in the population and the magnitude of association between the genotype and fracture risk is modest.

The ultimate aim of developing a prognostic model is to provide clinicians and each individual with his or her risk estimate to guide clinical decisions. At present, individuals with low BMD (i.e., T scores less than −2.5) or with a history of low-trauma fracture are recommended for therapeutic intervention [38, 39]. This recommendation is logical and appropriate since these individuals have higher risk of fracture [40, 41] and treatment can reduce their risk of fracture [42–44]. However, because fracture is a multifactorial event, there is more than one way that an individual can attain the risk conferred by either low BMD or a prior fracture [45]. Indeed, as shown in this study, a 60-year-old woman with a T score of −2.3 could have the same risk of fracture as a 75-year-old woman with a T score of −1 if the two are carriers of genotype TT. These data support the informativeness of a multivariable prognostic model and the limitation of a risk stratification–based approach for risk assessment for an individual.

It has been suggested that treatment should be considered for postmenopausal women and men aged 50 years or more [46] (1) with a preexisting hip or clinical vertebral fracture or a morphometric vertebral fracture, (2) with FN- or LSBMD T score ≤−2.5 after excluding secondary cause of osteoporosis, and (3) with FN- or LSBMD T score between −1 and −2.5 and a 10-year risk of hip fracture ≥3% or a 10-year risk of major osteoporotic fracture ≥20%. The nomogram presented here and the FRAX model [15] in conjunction with the above guidelines can help select suitable individuals for intervention.

A number of considerations should be taken into account in extrapolating the present findings to other populations. The present finding was based on an association—not linkage—analysis and, as such, does not necessarily show that the COL1A1 gene is directly involved, rather than being a marker for nearby genes that are linked to hip fracture liability. The data were based on a sample of Caucasian women, whose lifestyles and environmental living conditions are relatively homogeneous; hence, they may not apply to women in other populations or to men. The number of fractures, particularly hip fracture, together with the low frequency of the TT genotype were modest, resulting in a rather wide confidence interval of the estimate of association between the gene and fracture. The sample size, as in most studies to date, was not sufficient to examine potential gene–environment interactions, which might improve fracture risk prediction. There are likely many other genes associated with fracture risk which were not considered in the present model.

Both prognosis and treatment decisions are concerned with an individual. Each individual is a unique case because there is no “average individual” in the population. The uniqueness of an individual can be defined in terms of the individual’s environmental and genetic factors. The knowledge of genetics, in combination with environmental factors, can shift our current risk-stratification approaches to a more individualized evaluation and treatment of osteoporosis. To this end, these data indicate that a risk genotype of the COL1A1 gene is associated with an increased risk of fracture and that the incorporation of this genetic information into a prognostic model could enhance its accuracy and predictive value for an individual. These data support the potential utility of genetic information in an individual’s absolute fracture risk prediction.