- Email: [email protected]
Fracture Risk Assessment: From Population to Individual
ARTICLE IN PRESS Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health, vol. ■, no. ■, 1–11, 2017 © 2017 The International Society for Clinical Densitometry. 1094-6950/■:1–11/$36.00 http://dx.doi.org/10.1016/j.jocd.2017.06.023
Original Article
Fracture Risk Assessment: From Population to Individual Tuan V. Nguyen,*,1,2,3 and John A. Eisman1,2,4 1 Bone Biology Division, Garvan Institute of Medical Research, Darlinghurst, NSW, Australia; 2St Vincent’s Clinical School, UNSW Medicine, UNSW, Australia; 3Centre for Health Technology, University of Technology, Sydney, Australia; and 4School of Medicine Sydney, University of Notre Dame Australia, Fremantle, Australia
Abstract Fracture caused by osteoporosis remains a major public health burden on contemporary populations because fracture is associated with a substantial increase in the risk of mortality. Early identification of high-risk individuals for prevention is a priority in osteoporosis research. Over the past decade, few risk prediction models, including the Garvan Fracture Risk Calculator (Garvan) and FRAX®, have been developed to provide absolute (individualized) risk of fracture. Recent validation studies suggested that the area under the receiver operating characteristic curve in fracture discrimination ranged from 0.61 to 0.83 for FRAX® and from 0.63 to 0.88 for Garvan, with hip fractures having a better discrimination than fragility fractures as a group. Although the prognostic performance of Garvan and FRAX® for fracture prediction is not perfect and there is room for further improvement, these predictive models can aid patients and doctors communicate about fracture risk in the medium term and to make rational decisions. However, the application of these predictive models in making decisions for an individual should take into account the individual’s perception of the importance of fracture relative to other diseases. Key Words: Garvan Fracture Risk Calculator; fracture; FRAX®; osteoporosis; risk assessment. fracture (1.8-fold) is significantly greater than that in women (1.4-fold) (4–6). The increased mortality risk was also observed in younger individuals with fracture (3). Moreover, up to 24% of women and 38% of men will die within the first 3 mo after experiencing a hip fracture (7,8). Those who survive a fracture usually develop 1 or more of chronic pain, increased dependence, and reduce quality of life (9–11).Taken together, recent data clearly underline that osteoporotic fracture is a common and serious skeletal disorder that is expected to increase in magnitude over the next few decades as populations are rapidly aging. It is the association between fracture and premature mortality that is of great concern. There are, however, highquality data suggesting that treatment of individuals with fracture could reduce mortality risk. Indeed, a large randomized controlled trial has shown that zoledronic acid treatment reduces the risk of post-hip-fracture mortality by 28%, when given within 3 mo post hip surgery (12). Interestingly, only a small part of the benefit of reducing death post fracture was attributable to preventing refracture. More recent studies (13–15) have also suggested that individuals on oral bisphosphonates have a lower risk of mortality. Despite these
“All models are wrong, but some are useful.” George E.P. Box (1919–2013)
Introduction Osteoporosis and its consequence of fracture remain a major burden on contemporary populations. From the age of 50, 1 out of 2 women and approximately 1 out of 3 men will sustain a fracture during their remaining lifetime (1). In women, the remaining lifetime risk of hip fracture is equivalent to or higher than the risk of invasive breast cancer (1,2), and in men, the risk of hip and clinical vertebral fractures (17%) is comparable to the risk of prostate cancer (2,3). More importantly, fracture is associated with an increased risk of mortality, and the risk is greater in men than in women. Indeed, numerous studies, including our own, have consistently shown that the relative risk of death in men with *Address correspondence to: Tuan V. Nguyen, PhD, DSc, Osteoporosis and Bone Biology Program, Garvan Institute of Medical Research, 384 Victoria Street, Sydney, NSW 2010 Australia. E-mail: [email protected]
1
ARTICLE IN PRESS 2 pieces of evidence (14,15), less than 30% of women and less than 10% of men who have already had an osteoporotic fracture receive treatment to reduce their risk of subsequent fractures (16). Thus, osteoporosis is an undertreated disease, and the undertreatment status could partly be responsible for the excess mortality associated with fracture (11). One way to improve treatment uptake is by identifying high-risk individuals for closer observation and, possibly, for selecting individuals for an appropriate treatment option. The identification of individuals at high risk of osteoporotic fracture remains a significant challenge; the susceptibility to fracture is highly variable among individuals within a population as well as between populations. This high variability in fracture risk is expected, because fracture itself is associated with multiple factors, some of which may be causal. Therefore, any single risk factor is unlikely to be helpful for risk stratification; a more logical and better approach is to tailor the risk of fracture for an individual based on the individual’s risk profile. A risk profile must consider multiple risk factors by taking into account the magnitude of association between a risk factor and a fracture. Research during the past 3 decades or so have identified several risk factors for fracture, including advancing age, low bone mineral density (BMD), a personal history of fracture, fall, low body mass index, and, more interestingly, genetic factors (17). Advancing age is clearly a major risk factor of fracture, as the incidence of fracture increases exponentially with advancing age in both men and women (18–21). Recent data have shown that the 10-yr probability of fracture at the forearm, humerus, spine, or hip increases between the ages of 45 and 85 by as much as 8-fold for women and 5-fold for men (22). Low BMD is the most robust predictor of fracture risk. Each standard deviation difference in BMD is associated with about a 2-fold change in the risk of fracture (23,24). However, BMD alone cannot reliably predict an individual who is (or is not) going to sustain a fracture. It has been estimated that less than 40% of fracture cases occur in people with BMD in the osteoporotic range (25). Among women aged 60 yr or older with low BMD (high-risk group), only 40% sustained an osteoporotic fracture within 13 years of followup (25). On the other hand, among those who sustained a fracture, almost 60% had BMDs above the osteoporotic cutpoint (T-score ≤−2.5). In other words, more than half of individuals with low BMD were “resistant to fracture.” The situation in elderly men is similar: 70% of men with low BMD did not sustain a fracture; and among fracture cases, 77% occurred in those with nonosteoporotic BMD levels (25). This finding is consistent with the point raised by Geoffrey Rose in his influential essay (26), where he postulates that the majority of chronic disease cases arise from the mass of the population with risk around the average. Clearly, BMD alone cannot identify most individuals at high risk of fracture. A rethinking is needed for the majority of individuals whose BMD values are near the osteoporotic threshold. A simple fact is that, at any age and level of BMD, fracture risk varies widely in relation to the burden of other
Nguyen and Eisman risk factors, including prior fracture and falls. A prior fragility fracture signals a substantially elevated risk of future fracture (6,27–29). The elevated risk is 1.5- to 9.5-fold, depending on the age at assessment, the number of prior fractures, and the site of the incident fracture. Pooling the results from all studies (women and men) and for all fracture sites, the risk of subsequent fracture among those with a prior fracture at any site is 2.2 times more than the risk among those without a prior fragility fracture (29). Fall is also an important risk factor for fracture. Indeed, in the elderly population, falls are associated with hip fractures, Colles’ fractures, pelvis fractures, and ankle fractures (30). Approximately 95% of hip fractures are results of falls (31). The prevalence of falls increases with advancing age (32) and is higher in women than in men (33). The genderrelated difference in fall prevalence could partly explain the difference in fracture risk between men and women. Thus, for any individual, the risk of fracture depends on a combination of factors (34). This means that 2 individuals, both with “osteoporosis,” can have different risks of fracture because they have different risk profiles. Similarly, an osteoporotic individual can have the same risk of fracture as a nonosteoporotic individual because of the difference in constellation of risk factors between the 2 individuals. In other words, the assessment of fracture risk can and should be based on a multivariable model.The advantage of a statistical multivariable model is that it yields a more accurate and consistent prediction than human experts’ prediction (35,36). A number of statistical multivariable models have been developed for individualized fracture risk assessment.Among the models, the Garvan Fracture Risk Calculator (37,38), FRAX® (39), and QFracture (40) are widely used. FRAX® uses 12 risk factors, including femoral neck BMD, anthropometric factors, lifestyle factors, and comorbidities. The Garvan Fracture Risk Calculator (Garvan) uses 5 risk factors, namely, age, gender, femoral neck BMD, prior fracture, and history of fall. Garvan considers the frequency of prior fractures and the number of falls during the past 12 mo, not simply binary (yes/no) variables. Whereas Garvan provides 5- and 10-yr risks of total fracture and hip fracture, FRAX® provides a 10-yr risk of hip fracture and major osteoporotic fractures (Table 1).
Predictive Performance The usefulness of a predictive model is usually quantified in terms of discrimination, calibration, and reclassification. Discrimination is the ability to separate individuals who will sustain a fracture along a continuum from those who will not. The primary metric of discrimination is the area under the receiver operating characteristic curve (AUC), which evaluates the compromise between sensitivity and specificity, and is thus a global estimate of prognostic accuracy (41).
Discrimination Several independent studies have been carried out to examine the prognostic performance of Garvan (42,43),
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment
3
Table 1 Risk Factors Included in FRAX® and Garvan Fracture Risk Calculator (Garvan)
Risk factors (inputs)
Output
Website
FRAX®
Garvan
Age Gender Femoral neck BMD Body weight Height History of prior fracture Parental history of hip fracture Current smoking Chronic glucocorticoid use Rheumatoid arthritis Secondary osteoporosis Alcohol (3 or more units/d) 10-yr Risk of hip fracture 10-yr Risk of major fractures
Age Gender Femoral neck BMD (or body weight) Number of prior fractures Number of falls during the past 12 mo
http://www.shef.ac.uk/FRAX
5-yr Risk of hip fracture 5-yr Risk of any fragility fracture 10-yr Risk of hip fracture 10-yr Risk of any fragility fracture www.fractureriskcalculator.com
Abbr: BMD, bone mineral density.
FRAX® (44–49), or both Garvan and FRAX® (42,50,51).The Garvan model has been validated in the Canadian Multicenter Osteoporosis Study (43), which followed up 4152 women and 1606 men for over 10 yr, and observed 123 hip fractures (93 women) and 672 fragility fractures (579 women). The application of Garvan models to predict fracture yielded good discrimination results, particularly for hip fracture (AUC 0.80 for women and 0.85 for men). In general, the discrimination in hip fracture was better than all fractures. In predicting hip fracture risk, the median AUC value for Garvan was 0.80, which was equivalent to that of FRAX® (AUC 0.78). In predicting major fracture risk, the median AUC value for Garvan and FRAX® was 0.76 and 0.69, respectively (Table 2). It appears that the discrimination of fracture in men was lower than that in women. In men, the AUC value for fracture discrimination was 0.76 by Garvan and 0.54 by FRAX® based on US data (42).
Calibration Results of discrimination analyses clearly showed that all predictive models are imperfect. However, in practice, the agreement between observed and predicted outcomes (i.e., calibration) is more important than discrimination. Several studies have indicated that the Garvan model had very good calibration. A validation study on 1422 postmenopausal women living in New Zealand observed 229 fracture cases and Garvan predicted 276 cases (99% agreement) (50). However, the study also found that Garvan overestimated the risk of fracture among individuals in the top quartile of fracture risk (50), which is also noted in the initial development study (37,38). In the Canadian Multicenter
Osteoporosis Study cohort, the Garvan model also shows a remarkable agreement between the predicted 10-yr probability of fracture and the observed 10-yr risk of fracture (43). Most validation studies suggest that the FRAX® underestimated the risk of fracture (Table 3). In the Canadian Multicenter Osteoporosis Study cohort, and results of validation suggest that FRAX® underestimated the risk of fracture in men: predicted 5.4% vs observed 6.4% (45). In a validation study in 5891 men of the MrOS cohort (49), the FRAX® 10-yr predicted probability of hip fracture was 1.4% whereas the actual risk was 3% (Table 3). For major osteoporotic fractures, the discrepancy between the 10-yr predicted risk (6.5%) and observed risk (6.9%) was not significant. A validation study in New Zealand found that FRAX® underestimated the risk of fracture by up to 50% for every risk level (50). In the FRIDEX cohort (48), FRAX® predicted only 41% of actual hip fracture cases and 46% of major fractures. In contrast, FRAX® tended to overestimate the risk of fracture in a Japanese cohort (47). Using the predicted 10-yr risk of 20% as a cutoff value to define high risk vs low risk, the sensitivity and specificity for predicting hip fracture of FRAX® were 39% and 79%, respectively (49). Another study by (53) observed that 50% of patients with fracture were not classified as “high risk” based on the FRAX®-predicted risk just before the fracture event occurred.
Reclassification It has been well known that AUC is a rather insensitive measure (54), such that a statistically significant risk
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 4
Nguyen and Eisman
Table 2 Area Under the Receiver Operating Characteristic Curve for FRAX® and Garvan Models in Predicting Hip Fracture and Major Osteoporotic Fractures Fracture (gender)
FRAX® (with BMD)
Garvan
Major fractures (women) Major fractures (men) Hip fracture (women) Major fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (men) Major fractures (men) Hip fracture (women) Major fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (women) Hip fracture (men) Major fractures (women) Major fractures (men) Hip fracture (women) Major fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (men and women) Major fractures (men and women) Hip fracture (men) Major fractures (men)
0.78 0.54 0.73 0.83
0.84 0.76 0.85 0.88 0.80 0.70 0.85 0.69 0.67 0.63 0.76 0.64
Study Sandhu et al (42) Pluskiewicz et al (51) Langsetmo et al (43)
Bolland et al (50) Sambrook et al (52) Leslie et al (44)
Ensrud et al (46) Tamaki et al (47) Azagra et al (48) Ettinger et al (49)
0.70 0.62 0.78 0.61 0.82 0.79 0.70 0.66 0.76 0.69 0.88 0.69 0.85 0.72 0.77 0.67
Abbr: BMD, bone mineral density.
Table 3 Comparison Between Predicted (FRAX® and Garvan) and Actual Fracture Risks in Some Validation Cohorts Study Bolland et al (50) Langsetmo et al (43)
Tamaki et al (47) Azagra et al (48) Ettinger et al (49)
Fracture (gender) Hip fracture (women) Major fractures (women) Hip fracture (men) Hip fracture (women) Osteoporotic fractures (men) Osteoporotic fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (men and women) Major fractures (men and women) Hip fracture (men) Major fractures (men)
FRAX® predicted/observed fractures
Garvan predicted/observed fractures
43/57 121/229
85/57 276/279 41/39 232/116 180/140 756/673
8/4 50/43 7/17 30/65 1.4/3.0 6.5/6.9
Note: Data are the actual number of fractures.
factor may not have a high AUC value, particularly in the presence of other variables in the model. Hence, a clinically meaningful difference in the prognostic value between 2 predictive models is not necessarily reflected by the
absolute difference in AUC. Another feature of usefulness of a predictive model is risk reclassification (55). For a given threshold of risk (e.g., 10-yr risk of 20%), an individual can be classified as “high risk” or “low risk.” With additional
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment risk factors, the individual may change the risk category from one to another. If a new risk factor or marker is materially useful, then the addition of the risk factor should result in more individuals who will subsequently fracture being classified in the high-risk group than in the low-risk group; conversely, among those who will not subsequently fracture, more would be classified in the low-risk group than in the high-risk group. The net difference between the 2 proportions of reclassification is referred to as net reclassification improvement (56). Thus, when treatment decisions are based on the risk threshold, the net reclassification improvement can be helpful for making a clinical decision concerning an individual. Whether including BMD into a predictive model can result in an improvement in the discrimination or not is a contentious issue. In the MrOS cohort, a FRAX® plus a BMD model produced a lower risk of fracture than a FRAX® without a BMD model (49). However, other studies (44,47,48) found that a FRAX® with a BMD model produced greater AUC values than the model without BMD. Nevertheless, a complicated model with multiple risk factors does not necessarily improve the prognostic performance of a predictive model. Indeed, a simple model with age and femoral neck BMD or a model with age and fracture history was as good as a more complicated model such as FRAX® in terms of fracture prediction (46,52,57).
Concordance Between Garvan and FRAX® The concordance in the predicted probabilities of fracture between the Garvan Fracture Risk Calculator and FRAX® was modest, with the coefficient of correlation being 0.67 (57). Another validation study in 2012 postmenopausal women of Polish background found that there was a considerable discrepancy in risk estimates between the Garvan and FRAX® models with the Garvan model predicting fracture more accurately than FRAX® (51). Despite the fact that there are differences in the predicted risk of fracture between Garvan and FRAX®, the majority of the differences do not seem to impact on the treatment recommendation (58). Discrepancy in predicted risks between Garvan and FRAX® (or any other models) is expected, because the models not only use different risk factors but also have different discriminations and calibrations. Moreover, an individual can have 2 different risks predicted by 2 valid models (59), because the individual has an indefinite number of risk profiles, which could assign the individual into different risk categories (Fig. 1). This paradox is known as the “reference class” problem (61), which, in practice, means that no individual is consistently assigned to the same risk group by different models even though all models are valid.
Threshold for Intervention One application of fracture prediction models is for selecting patients suitable for intervention. However, this application raises serious challenges because the predicted risk
5
Fig. 1. Equivalent but different risk stratification: 3 subjects with no event (1–3) and 3 subjects with events (4–6). Each row represents a risk stratification method for identifying high- and low-risk subpopulations. Adapted from Stern (60).
of fracture is a continuous probabilistic variable ranging from 0 to 1, and selecting a predicted probability to classify an individual as high risk or low risk requires a thorough research. Nevertheless, the National Osteoporosis Foundation guidelines recommend treatment in the following clinical situations in postmenopausal women and in men aged 50 yr or older (62) (1) with a hip or clinical vertebral fracture or a morphometric vertebral fracture, (2) with femoral neck or lumbar spine BMD T-scores being equal to or less than −2.5 after excluding the secondary cause of osteoporosis, and (3) with femoral neck or lumbar spine BMD T-scores between −1 and −2.5 and a 10-yr risk of hip fracture ≥3% or a 10-yr risk of major osteoporotic fracture ≥20%. The clinical benefit and cost-effectiveness of these recommendations should be subject to a more systematic research. The individualization of fracture risk assessment can be applied to optimize the number needed to treat (NNT). In several randomized clinical trials (63), the number of patients needed to be treated (NNT) to reduce 1 vertebral fracture compared to the untreated group ranged between 8 and 83. For hip fracture, the NNT ranged between 91 and 250 (63). The NNT varies inversely with the background risk, such that the treatment of high-risk individuals inherently yields a lower NNT (Fig. 2). The large variability in the NNTs among trials is assumed to be due to the variability in fracture rates among the study samples. However, the variability is expected given the multiple risk factors that affect the incidence of fractures. In the presence of such
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 6
Fig. 2. Relationship between background risk (rate of fracture in the placebo group) and the number of patients needed to treat to reduce 1 case of fracture. Each dot in the figure represents the result of a randomized controlled trial. Source of data: Delmas et al (63).
variability, selecting patients based on their absolute risk of fracture (rather than based on a BMD threshold value) may improve the consistency of the therapeutic efficacy and efficiency of trials. Trials specifically testing the efficacy of multivariable risk-based therapy have not been performed. However, such approaches could be expected to prove more cost-effective and yield a more consistent NNT, particularly if the duration, typically 3 yr, is standardized. Recent analyses of the correlation between FRAX®predicted fracture risks and antifracture efficacy yielded mixed results. One clinical trial (64) randomized 5212 women aged 75 yr and older into 2 groups, placebo receiving calcium and vitamin D with placebo or clodronate (800 mg daily po). Ten-year probability of fracture was computed for each woman using baseline clinical risk factors including body mass index, prior fracture, glucocorticoid use, parental hip fracture, smoking, alcohol, and secondary osteoporosis. In women in the top 25th percentile of fracture probability (average probability of 24%), treatment reduced the risk of fracture by 23% over 3 yr (hazard ratio 0.77, 95% confidence interval 0.63–0.95). Importantly, among those in the top 10% percentile of risk (average fracture probability of 30%), treatment reduced the fracture risk by 31% (hazard ratio 0.69, 0.53–0.90) (64). Thus, treatment of individuals at high risk or moderate risk could reasonably be expected to have a similar benefit in relative but greater effect in absolute risk reduction for the higher risk group. In a post hoc analysis of the Fracture Intervention Trial, the investigators used FRAX® to estimate each patient’s 10-yr risk of fracture, and then correlated the risk with antifracture efficacy. The investigators conclude that the
Nguyen and Eisman magnitude of effect of alendronate was constant across FRAX®-predicted risks (65). However, a close examination of the data reveals that the absolute risk reduction increased with absolute risk levels, and as a result, the NNT decreased as absolute risk increased. For example, among those in the first tertile of risk (4.8%–22.1%) the NNT was 76, but among those in the highest tertile (34.2%–85.4%), the NNT was 40. Results of post hoc analyses of clinical trials appear to suggest that, in relative risk terms, the magnitude of antifracture efficacy of denosumab (66) and bazedoxifene (67) was dependent on absolute risks of fracture, such that those at high risk had better relative efficacy than those at low risk of fracture. However, the antifracture of strontium ranelate (68) and raloxifene (69) seemed to be independent of the fracture risk assessed by FRAX®. However, it is not clear whether absolute risk reduction or the NNT was independent of patients’ absolute risk levels. Nevertheless, taken together, these results are consistent with the supposition that the antifracture effect size of pharmacological therapies is inversely associated with patients’ absolute risks, supporting the use of predictive models for selecting patients to include in future randomized controlled trials of osteoporosis.
Risk Communications: Individual vs Population At the heart of risk communication is the very meaning of “risk.” More specifically, what does a [say] 20% risk of fracture within the next 5 yr mean to an individual? This question is important, because only when an individual understands and appreciates the risk can the individual arrive at an informed decision. Given the way the risk is calculated, it is probably reasonable to say that “based on your age and risk profile, your risk of sustaining a fracture in the next 5 yr is 20%.” This information can, however, present a conflict between the individual and the population. In medical parlance, risk is commonly seen as the probability of getting a disease over a certain period of time, but epistemologically, risk is defined as the product of probability and the consequence of an adverse event (70). The probability component measures the uncertainty, whereas consequence measures the impact of an event. Virtually all predictive models in medicine, including Garvan, provide only the uncertainty, not the impact, of a health event. There is a conflict between the population and the individual uncertainty in the estimation of risk. The estimated probability of fracture is derived from data of a group of individuals, but the probability is used for an individual. In other words, a group-based estimate is used as an intrinsic property of an individual. However, an individual will either sustain or not sustain a fracture (i.e., a binary event), and risk in the form of continuous probability is thus not meaningful to an individual. Indeed, some argue that probabilistic risk is not meaningful to an individual (71) because, from a probabilistic viewpoint,
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment probability does not exist (72). Nevertheless, we would argue that risk is a meaningful measure for an individual because, from a Bayesian point of view, risk reflects an individual’s measure of belief that an event will occur. Therefore, individual risk is still a useful concept in fracture risk assessment. Because risk is a conditional probability, the predicted risk of fracture for an individual can vary, depending on the collection of risk factors considered. Moreover, statistical models of risk prediction do not necessarily produce a unique solution. It is important to appreciate that the predicted risk is actually an average, a typical value, with “true” values fluctuating below or above the typical value. Therefore, an individual does not necessarily have a unique risk value. This subtle fact also explains why different valid predictive models can yield substantially different results for an individual. Predicted risk can change with time because risk factors for fracture change with time. However, risk estimates are currently based on the assumption of constancy in risk factors. An individual’s absolute risk over a meaningful time period, arguably 5 rather than 10 yr (73), has to be considered alongside that individual’s concern about the outcome of interest. Some fractures will be of limited concern to some people but of great concern to others who have a more personal and perhaps more accurate view of their impact on the quality of life and even survival. Thus, whereas risk exists as an objective identity, its measure, probability, does not exist objectively.
Issues for Further Research Although the development and implementation of fracture prediction models represent a major advance in osteoporosis research, several issues need to be addressed in the future. A good predictive model should meet the “4R” criteria: reliability, reproducibility, relevance, and realworld value. A reliable model should have good sensitivity and specificity, and its predicted outcomes agree closely with observed outcomes. Reproducibility here means that risk factors included in a model should be highly reproducible and that the risk factors are found in various independent populations. Relevance refers to the clinical relevance of a predictive model, such that intervention on those with high risk predicted by the model could alter their outcome. A predictive model has “real-world” value if the model is easy to use and inexpensive. None of the existing predictive models in osteoporosis meet all the criteria. It is important to distinguish between prediction and association, a point that is not well recognized in clinical medicine (74). In the context of variable selection, the goal of association or explanation analysis is to identify risk factors that affect an outcome variable, but the goal of prediction or classification is about to find observed factors that predict new observation of the outcome variable. Therefore, statistically significant risk factors are not necessarily good predictors or classifiers (75,76), but a factor that is not
7 statistically significant can be a useful predictor (77). All existing models have been developed based on the idea of association, not prediction, using traditional statistical methods. The development of the Garvan model, for example, was based on the Bayesian model average method (78), which has been shown to be more reliable than traditional methods. In recent years, several new machine learning algorithms, such as the random forest and the retention-partition methods, can be useful for developing predictive models, but these algorithms have not been applied.
Bone Turnover Markers (BTMs) BTMs can be useful for predicting fracture risk independent of BMD. There is accumulating evidence that accelerated bone resorption with the associated imbalance in bone remodeling adversely affects skeletal microstructure and increases fracture risk. Several cross-sectional and longitudinal studies have found that fragility fractures occur not only because of low BMD but also as a result of rapid bone turnover that leads to architectural changes, such as disruption of trabecular networks and struts. Increased urinary levels of the pyridinium cross-link, deoxypyridinoline, was associated with a 2- to 3-fold increase in the risk of hip fracture (79). In the EPIdémiologie de l’OStéporose study, urinary cross-linked C-telepeptide (CTX) and free deoxypyridinoline levels above the premenopausal range were associated with a 2-fold increase in hip fracture risk after adjusting for BMD and physical mobility (80). In that study, 16% of the subjects with both a low BMD (T-score ≤−2.5) and an elevated urinary CTX (T-score ≤−2 standard deviation of the mean for premenopausal women) had a relative risk of 4.8 for hip fracture, compared with a relative risk of 2.7 for low BMD alone or a relative risk of 2.2 for high urinary CTX alone (80). Elevated bone resorption marker levels have been shown to be associated with increased fracture risk independent of BMD (81). In our recent study in men, increased bone resorption was associated with increased fracture risk in men (82). Taken together, these results strongly suggest that an incorporation of BTMs into the existing prognostic model could improve the prediction of absolute fracture risk.
Trabecular Bone Score (TBS) One of the promising factors for improving the accuracy of fracture prediction is the TBS (83–85), a graylevel texture measurement based on a variogram of the 2-dimensional BMD projection image (86). Thus, TBS is a texture parameter that reflects a pixel gray-level variation in dual-energy X-ray absorptiometry images. Previous studies have reported that TBS is not significantly correlated with BMD, but is significantly correlated with the trabecular number, trabecular separation, and the structure model index. Moreover, TBS was found to be associated with fracture risk in elderly women (87) and in diabetic patients (88) independent of BMD and classical clinical risk factors (84,85).
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 8 Genetics The variation in fracture susceptibility is partly determined by genetic factors. A twin study (89) has shown that almost 50% of the variance in the liability to fracture was attributable to genetic factors. Through large-scale collaborative studies, several genetic variants have been identified. To date, 12 genome-wide association studies have been conducted for BMD phenotypes (90), and 62 loci have been identified to be associated with BMD at the genomewide significance level.Among the 62 single-nucleotide polymorphisms (SNPs) identified, 8 SNPs were associated with fracture risk at the genome-wide significance level (90). However, these common variants have a modest association with fracture risk with odds ratios ranging between 1.1 and 1.4, suggesting that, individually, these variants have limited utility for fracture prediction. Nevertheless, a recent simulation study showed that a genetic profile of up to 50 genetic variants, with each having a modest effect size (odds ratio of 1.01–1.35), could improve the accuracy of fracture prediction by 10% points of AUC (91). Two recent studies in postmenopausal women of Korean background found that a genetic profiling of 39 SNPs in 30 human genomic loci could increase the precision of nonvertebral fracture prediction and help to define the risk threshold (92), whereas 35 risk alleles were significantly associated with the risk of vertebral fracture (92,93) in patients on bisphosphonate. Taken together, these results and the present finding suggest that genetic profiling is useful in the identification of high-risk individuals.
Fracture-Specific Prediction Current prognostic models have been developed based on the “one size-fits-all” approach. For example, each model uses a set of risk factors for predicting the 5- or 10-yr risk of all types of fracture, because it is assumed that the set of risk factors is associated with any type of fracture in any ethnic population in both men and women. This assumption is strong and likely untenable because a risk factor may be uniquely associated with a type of fracture. For instance, fall is a major risk factor for hip fracture and is not a risk factor for vertebral fracture.
Time-Variant Predictions An important weakness of current fracture prediction models is that they are based on a single measurement of risk factors, with the underlying but not stated assumption that the risk factors do not change with time. Obviously, this assumption is not true in many risk factors such as BMD and body weight, which are known to decline or change with time. Moreover, the rates of decline in BMD varied substantially among individuals. Similarly, the risk of a second fracture is higher and closer in time to the initial fracture with risks declining substantially after 10 yr. Therefore, 1 important aspect of future model development should take the time-varying nature of risk factors into account to achieve a better estimate of risk for an individual.
Nguyen and Eisman In conclusion, fracture caused by osteoporosis is a serious skeletal disorder because it is associated with considerable morbidity and increased risk of mortality. Despite the availability of many effective treatments, the treatment uptake is still low. In the era of evidence-based medicine, patients and doctors need to work together to make informed decisions concerning treatment decisions. Therefore, the individualized assessment of fracture risk is of growing interest to the general public, doctors, and researchers. The advance of predictive models for individualized fracture risk assessment represents a great success in the translational research of osteoporosis. Although these models remain to be improved, their availability helps patients and doctors make decisions that are more evidence based and better tailored to an individual’s specific needs. However, if predictive models are used to guide therapy, a higher standard of reliability, reproducibility, and relevance must be attained.
References 1. Nguyen ND, Ahlborg HG, Center JR, et al. 2007 Residual lifetime risk of fractures in women and men. J Bone Miner Res 22(6):781–788. 2. Cummings SR, Black DM, Rubin SM. 1989 Lifetime risks of hip, Colles’, or vertebral fracture and coronary heart disease among white postmenopausal women. Arch Intern Med 149(11):2445–2448. 3. Shortt NL, Robinson CM. 2005 Mortality after low-energy fractures in patients aged at least 45 years old. J Orthop Trauma 19(6):396–400. 4. Johnell O, Kanis JA, Oden A, et al. 2004 Fracture risk following an osteoporotic fracture. Osteoporos Int 15(3):175–179. 5. Bliuc D, Nguyen ND, Milch VE, et al. 2009 Mortality risk associated with low-trauma osteoporotic fracture and subsequent fracture in men and women. JAMA 301(5):513–521. 6. Center JR, Nguyen TV, Schneider D, et al. 1999 Mortality after all major types of osteoporotic fracture in men and women: an observational study. Lancet 353(9156):878–882. 7. Hindmarsh DM, Hayen A, Finch CF, Close JC. 2009 Relative survival after hospitalisation for hip fracture in older people in New South Wales, Australia. Osteoporos Int 20(2):221–229. 8. Haentjens P, Magaziner J, Colon-Emeric CS, et al. 2010 Metaanalysis: excess mortality after hip fracture among older women and men. Ann Intern Med 152(6):380–390. 9. Randell AG, Nguyen TV, Bhalerao N, et al. 2000 Deterioration in quality of life following hip fracture: a prospective study. Osteoporos Int 11(5):460–466. 10. Adachi JD, Adami S, Gehlbach S, et al. 2010 Impact of prevalent fractures on quality of life: baseline results from the global longitudinal study of osteoporosis in women. Mayo Clin Proc 85(9):806–813. 11. Frost SA, Nguyen ND, Center JR, et al. 2013 Excess mortality attributable to hip-fracture: a relative survival analysis. Bone 56(1):23–29. 12. Lyles KW, Colon-Emeric CS, Magaziner JS, et al. 2007 Zoledronic acid in reducing clinical fracture and mortality after hip fracture. N Engl J Med 357:1799–1809. 13. Beaupre LA, Morrish DW, Hanley DA, et al. 2011 Oral bisphosphonates are associated with reduced mortality after hip fracture. Osteoporos Int 22(3):983–991.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment 14. Center JR, Bliuc D, Nguyen ND, et al. 2011 Osteoporosis medication and reduced mortality risk in elderly women and men. J Clin Endocrinol Metab 96(4):1006–1014. 15. Bolland MJ, Grey AB, Gamble GD, Reid IR. 2010 Effect of osteoporosis treatment on mortality: a meta-analysis. J Clin Endocrinol Metab 95(3):1174–1181. 16. Eisman J, Clapham S, Kehoe L. 2004 Osteoporosis prevalence and levels of treatment in primary care: the Australian BoneCare Study. J Bone Miner Res 19(12):1969– 1975. 17. Ralston SH. 2002 Genetic control of susceptibility to osteoporosis. J Clin Endocrinol Metab 87(6):2460–2466. 18. Kanis JA, Johnell O, Oden A, et al. 2001 Ten year probabilities of osteoporotic fractures according to BMD and diagnostic thresholds. Osteoporos Int 12(12):989–995. 19. Nguyen TV, Eisman JA, Kelly PJ, Sambrook PN. 1996 Risk factors for osteoporotic fractures in elderly men. Am J Epidemiol 144(3):255–263. 20. Cummings SR, Nevitt MC, Browner WS, et al. 1995 Risk factors for hip fracture in white women. Study of Osteoporotic Fractures Research Group. N Engl J Med 332(12):767– 773. 21. Hui SL, Slemenda CW, Johnston CCJ. 1988 Age and bone mass as predictors of fracture in a prospective study. J Clin Invest 81(6):1804–1809. 22. Wasnich RD, Davis JW, Ross PD. 1994 Spine fracture risk is predicted by non-spine fractures. Osteoporos Int 4(1):1– 5. 23. Marshall D, Johnell O, Wedel H. 1996 Meta-analysis of how well measures of bone mineral density predict occurrence of osteoporotic fractures. BMJ 312(7041):1254–1259. 24. Nguyen T, Sambrook P, Kelly P, et al. 1993 Prediction of osteoporotic fractures by postural instability and bone density. BMJ 307(6912):1111–1115. 25. Nguyen ND, Eisman JA, Center JR, Nguyen TV. 2007 Risk factors for fracture in nonosteoporotic men and women. J Clin Endocrinol Metab 92(3):955–962. 26. Rose G. 2001 Sick individuals and sick populations. Int J Epidemiol 30(3):427–432, discussion 433-4. 27. Pongchaiyakul C, Nguyen ND, Jones G, et al. 2005 Asymptomatic vertebral deformity as a major risk factor for subsequent fractures and mortality: a long-term prospective study. J Bone Miner Res 20(8):1349–1355. 28. Ismail AA, Cockerill W, Cooper C, et al. 2001 Prevalent vertebral deformity predicts incident hip though not distal forearm fracture: results from the European Prospective Osteoporosis Study. Osteoporos Int 12(2):85–90. 29. Klotzbuecher CM, Ross PD, Landsman PB, et al. 2000 Patients with prior fractures have an increased risk of future fractures: a summary of the literature and statistical synthesis. J Bone Miner Res 15(4):721–739. 30. Cummings SR, Nevitt MC, Study of Osteoporotic Fractures Research Group. 1994 Non-skeletal determinants of fractures: the potential importance of the mechanics of falls. Osteoporos Int 4(Suppl 1):67–70. 31. Nyberg L, Gustafson Y, Berggren D, et al. 1996 Falls leading to femoral neck fractures in lucid older people. J Am Geriatr Soc 44(2):156–160. 32. Lord SR, Sambrook PN, Gilbert C, et al. 1994 Postural stability, falls and fractures in the elderly: results from the Dubbo Osteoporosis Epidemiology Study. Med J Aust 160(11):684– 685, 688-91. 33. Prudham D, Evans JG. 1981 Factors associated with falls in the elderly: a community study. Age Ageing 10(3):141– 146.
9 34. Nguyen ND, Pongchaiyakul C, Center JR, et al. 2005 Identification of high-risk individuals for hip fracture: a 14-year prospective study. J Bone Miner Res 20(11):1921–1928. 35. Grove WM, Meehl PE. 1996 Comparative efficiency of informal (subjective, impressionistic) and formal (mechanical, algorithmic) prediction procedures: the clinical–statistical controversy. Psychol Public Policy Law 2(293–323). 36. Oberije C, Nalbantov G, Dekker A, et al. 2014 A prospective study comparing the predictions of doctors versus models for treatment outcome of lung cancer patients: a step toward individualized care and shared decision making. Radiother Oncol 112(1):37–43. 37. Nguyen ND, Frost SA, Center JR, et al. 2007 Development of a nomogram for individualizing hip fracture risk in men and women. Osteoporos Int 18(8):1109–1117. 38. Nguyen ND, Frost SA, Center JR, et al. 2008 Development of prognostic nomograms for individualizing 5-year and 10-year fracture risks. Osteoporos Int 19:1431–1444. 39. Kanis JA, Johnell O, Oden A, et al. 2008 FRAX and the assessment of fracture probability in men and women from the UK. Osteoporos Int 19(4):385–397. 40. Collins GS, Mallett S, Altman DG. 2011 Predicting risk of osteoporotic and hip fracture in the United Kingdom: prospective independent and external validation of QFracture scores. BMJ 342:d3651. 41. Harrell FEJ, Califf RM, Pryor DB, et al. 1982 Evaluating the yield of medical tests. JAMA 247(18):2543–2546. 42. Sandhu SK, Nguyen ND, Center JR, et al. 2010 Prognosis of fracture: evaluation of predictive accuracy of the FRAX algorithm and Garvan nomogram. Osteoporos Int 21(5):863– 871. 43. Langsetmo L, Nguyen TV, Nguyen ND, et al. 2011 Independent external validation of nomograms for predicting risk of low-trauma fracture and hip fracture. CMAJ 183(2):E107– E114. 44. Leslie WD, Lix LM, Johansson H, et al. 2010 Independent clinical validation of a Canadian FRAX tool: fracture prediction and model calibration. J Bone Miner Res 25(11):2350– 2358. 45. Leslie WD, Lix LM, Langsetmo L, et al. 2011 Construction of a FRAX(R) model for the assessment of fracture probability in Canada and implications for treatment. Osteoporos Int 22(3):817–827. 46. Ensrud KE, Lui LY, Taylor BC, et al. 2009 A comparison of prediction models for fractures in older women: is more better? Arch Intern Med 169(22):2087–2094. 47. Tamaki J, Iki M, Kadowaki E, et al. 2011 Fracture risk prediction using FRAX(R): a 10-year follow-up survey of the Japanese Population-Based Osteoporosis (JPOS) Cohort Study. Osteoporos Int 22(12):3037–3045. 48. Azagra R, Roca G, Encabo G, et al. 2012 FRAX(R) tool, the WHO algorithm to predict osteoporotic fractures: the first analysis of its discriminative and predictive ability in the Spanish FRIDEX cohort. BMC Musculoskelet Disord 13:204. 49. Ettinger B, Ensrud KE, Blackwell T, et al. 2013 Performance of FRAX in a cohort of community-dwelling, ambulatory older men: the Osteoporotic Fractures in Men (MrOS) study. Osteoporos Int 24(4):1185–1193. 50. Bolland MJ, Siu AT, Mason BH, et al. 2011 Evaluation of the FRAX and Garvan fracture risk calculators in older women. J Bone Miner Res 26(2):420–427. 51. Pluskiewicz W, Adamczykb P, Franekc E, et al. 2009 Conformity between 10-year probability of any osteoporotic fracture assessed by FRAX and nomogram by Nguyen et al. Bone 44(Suppl 2):S229–S230.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 10 52. Sambrook PN, Flahive J, Hooven FH, et al. 2011 Predicting fractures in an international cohort using risk factor algorithms without BMD. J Bone Miner Res 26(11):2770–2777. 53. Aubry-Rozier B, Stoll D, Krieg MA, et al. 2013 What was your fracture risk evaluated by FRAX(R) the day before your osteoporotic fracture? Clin Rheumatol 32(2):219–223. 54. Huntjens KM, Kosar S, van Geel TA, et al. 2010 Risk of subsequent fracture and mortality within 5 years after a nonvertebral fracture. Osteoporos Int 21(12):2075–2082. 55. Pencina MJ, D’Agostino RB Sr, D’Agostino RB Jr, Vasan RS. 2008 Evaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond. Stat Med 27(2):157–172, discussion 207-12. 56. Cook NR. 2008 Statistical evaluation of prognostic versus diagnostic models: beyond the ROC curve. Clin Chem 54(1):17– 23. 57. van Geel TA, Nguyen ND, Geusens PP, et al. 2011 Development of a simple prognostic nomogram for individualising 5-year and 10-year absolute risks of fracture: a populationbased prospective study among postmenopausal women. Ann Rheum Dis 70(1):92–97. 58. Bolland MJ, Grey A, Gamble G, Reid IR. 2013 Comment on Kanis et al.: pitfalls in the external validation of FRAX. Osteoporos Int 24(1):389–390. 59. Lemeshow S, Klar J, Teres D. 1995 Outcome prediction for individual intensive care patients: useful, misused, or abused? Intensive Care Med 21(9):770–776. 60. Stern R. 2010 Discordance of individual risk estimates. J Am Coll Cardiol 56(9):743, author reply 743-4. 61. Hájek A. 2007 The reference class problem is your problem too. Synthese 156:563–585. 62. NOF. Clinician’s guide to prevention and treatment of osteoporosis. National Osteoporosis Foundation 2008;Washington DC. 63. Delmas PD, Rizzoli R, Cooper C, Reginster JY. 2005 Treatment of patients with postmenopausal osteoporosis is worthwhile. The position of the International Osteoporosis Foundation. Osteoporos Int 16(1):1–5. 64. McCloskey E, Johansson H, Oden A, et al. 2007 Efficacy of clodronate on fracture risk in women selected by 10-year fracture probability. J Bone Miner Res 22(10):S131. 65. Donaldson MG, Palermo L, Ensrud KE, et al. 2012 Effect of alendronate for reducing fracture by FRAX score and femoral neck bone mineral density: the Fracture Intervention Trial. J Bone Miner Res 27(8):1804–1810. 66. McCloskey EV, Johansson H, Oden A, et al. 2012 Denosumab reduces the risk of osteoporotic fractures in postmenopausal women, particularly in those with moderate to high fracture risk as assessed with FRAX. J Bone Miner Res 27(7):1480–1486. 67. Kanis JA, Johansson H, Oden A, McCloskey EV. 2009 Bazedoxifene reduces vertebral and clinical fractures in postmenopausal women at high risk assessed with FRAX. Bone 44(6):1049–1054. 68. Kanis JA, Johansson H, Oden A, McCloskey EV. 2011 A meta-analysis of the effect of strontium ranelate on the risk of vertebral and non-vertebral fracture in postmenopausal osteoporosis and the interaction with FRAX((R). Osteoporos Int 22(8):2347–2355. 69. Kanis JA, Johansson H, Oden A, McCloskey EV. 2010 A meta-analysis of the efficacy of raloxifene on all clinical and vertebral fractures and its dependency on FRAX. Bone 47(4):729–735. 70. Kaplan S, Garrick BJ. 1981 On the quantitative definition of risk. Risk Anal 1:11–27.
Nguyen and Eisman 71. Sniderman AD, D’Agostino RB Sr, Pencina MJ. 2015 The role of physicians in the era of predictive analytics. JAMA 314(1):25–26. 72. Nau RF. 2001 De Finetti was right: probability does not exist. Theory Decis 51:89–124. 73. Bolland MJ, Jackson R, Gamble GD, Grey A. 2013 Discrepancies in predicted fracture risk in elderly people. BMJ 346:e8669. 74. Shmueli G. 2010 To explain or to predict. Stat Sci 25(3):289– 310. 75. Lo A, Chernoff H, Zheng T, Lo SH. 2015 Why significant variables aren’t automatically good predictors. Proc Natl Acad Sci USA 112(45):13892–13897. 76. Pepe MS, Janes H, Longton G, et al. 2004 Limitations of the odds ratio in gauging the performance of a diagnostic, prognostic, or screening marker. Am J Epidemiol 159(9):882– 890. 77. Guyon I, Elisseeff A. 2003 An introduction to variable and feature selection. J Mach Learn 3:1157–1182. 78. Hoeting J. 1999 Bayesian model averaging: a tutorial. Stat Sci 14:382–417. 79. van Daele PL, Seibel MJ, Burger H, et al. 1996 Case-control analysis of bone resorption markers, disability, and hip fracture risk: the Rotterdam study. BMJ 312(7029):482– 483. 80. Garnero P, Hausherr E, Chapuy MC, et al. 1996 Markers of bone resorption predict hip fracture in elderly women: the EPIDOS Prospective Study. J Bone Miner Res 11(10):1531– 1538. 81. Akesson K, Ljunghall S, Jonsson B, et al. 1995 Assessment of biochemical markers of bone metabolism in relation to the occurrence of fracture: a retrospective and prospective population-based study of women. J Bone Miner Res 10(11):1823–1829. 82. Meier C, Nguyen TV, Center JR, et al. 2005 Bone resorption and osteoporotic fractures in elderly men: the Dubbo osteoporosis epidemiology study. J Bone Miner Res 20(4):579– 587. 83. Hans D, Barthe N, Boutroy S, et al. 2011 Correlations between trabecular bone score, measured using anteroposterior dualenergy X-ray absorptiometry acquisition, and 3-dimensional parameters of bone microarchitecture: an experimental study on human cadaver vertebrae. J Clin Densitom 14(3):302– 312. 84. Silva BC, Leslie WD, Resch H, et al. 2014 Trabecular bone score: a noninvasive analytical method based upon the DXA image. J Bone Miner Res 29(3):518–530. 85. Ulivieri FM, Silva BC, Sardanelli F, et al. 2014 Utility of the trabecular bone score (TBS) in secondary osteoporosis. Endocrine 47(2):435–448. 86. Pothuaud L, Barthe N, Krieg MA, et al. 2009 Evaluation of the potential use of trabecular bone score to complement bone mineral density in the diagnosis of osteoporosis: a preliminary spine BMD-matched, case-control study. J Clin Densitom 12(2):170–176. 87. Hans D, Goertzen AL, Krieg MA, Leslie WD. 2011 Bone microarchitecture assessed by TBS predicts osteoporotic fractures independent of bone density: the Manitoba study. J Bone Miner Res 26(11):2762–2769. 88. Leslie WD, Aubry-Rozier B, Lamy O, , et al. 2013 TBS (trabecular bone score) and diabetes-related fracture risk. J Clin Endocrinol Metab 98(2):602–609. 89. Michaelsson K, Melhus H, Ferm H, et al. 2005 Genetic liability to fractures in the elderly. Arch Intern Med 165(16):1825–1830.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment 90. Richards JB, Zheng HF, Spector TD. 2012 Genetics of osteoporosis from genome-wide association studies: advances and challenges. Nat Rev Genet 13(8):576– 588. 91. Tran BNH, Nguyen ND, Nguyen VX, et al. 2011 Genetic profiling and individualized prognosis of fracture. J Bone Miner Res 26(2):414–419.
11 92. Lee SH, Lee SW, Ahn SH, et al. 2013 Multiple gene polymorphisms can improve prediction of nonvertebral fracture in postmenopausal women. J Bone Miner Res 28(10):2156–2164. 93. Lee SH, Cho EH, Ahn SH, et al. 2016 Prediction of future osteoporotic fracture occurrence by genetic profiling: a 6-year follow-up observational study. J Clin Endocrinol Metab doi:10.1210/jc.2015-3972. jc20153972.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
Original Article
Fracture Risk Assessment: From Population to Individual Tuan V. Nguyen,*,1,2,3 and John A. Eisman1,2,4 1 Bone Biology Division, Garvan Institute of Medical Research, Darlinghurst, NSW, Australia; 2St Vincent’s Clinical School, UNSW Medicine, UNSW, Australia; 3Centre for Health Technology, University of Technology, Sydney, Australia; and 4School of Medicine Sydney, University of Notre Dame Australia, Fremantle, Australia
Abstract Fracture caused by osteoporosis remains a major public health burden on contemporary populations because fracture is associated with a substantial increase in the risk of mortality. Early identification of high-risk individuals for prevention is a priority in osteoporosis research. Over the past decade, few risk prediction models, including the Garvan Fracture Risk Calculator (Garvan) and FRAX®, have been developed to provide absolute (individualized) risk of fracture. Recent validation studies suggested that the area under the receiver operating characteristic curve in fracture discrimination ranged from 0.61 to 0.83 for FRAX® and from 0.63 to 0.88 for Garvan, with hip fractures having a better discrimination than fragility fractures as a group. Although the prognostic performance of Garvan and FRAX® for fracture prediction is not perfect and there is room for further improvement, these predictive models can aid patients and doctors communicate about fracture risk in the medium term and to make rational decisions. However, the application of these predictive models in making decisions for an individual should take into account the individual’s perception of the importance of fracture relative to other diseases. Key Words: Garvan Fracture Risk Calculator; fracture; FRAX®; osteoporosis; risk assessment. fracture (1.8-fold) is significantly greater than that in women (1.4-fold) (4–6). The increased mortality risk was also observed in younger individuals with fracture (3). Moreover, up to 24% of women and 38% of men will die within the first 3 mo after experiencing a hip fracture (7,8). Those who survive a fracture usually develop 1 or more of chronic pain, increased dependence, and reduce quality of life (9–11).Taken together, recent data clearly underline that osteoporotic fracture is a common and serious skeletal disorder that is expected to increase in magnitude over the next few decades as populations are rapidly aging. It is the association between fracture and premature mortality that is of great concern. There are, however, highquality data suggesting that treatment of individuals with fracture could reduce mortality risk. Indeed, a large randomized controlled trial has shown that zoledronic acid treatment reduces the risk of post-hip-fracture mortality by 28%, when given within 3 mo post hip surgery (12). Interestingly, only a small part of the benefit of reducing death post fracture was attributable to preventing refracture. More recent studies (13–15) have also suggested that individuals on oral bisphosphonates have a lower risk of mortality. Despite these
“All models are wrong, but some are useful.” George E.P. Box (1919–2013)
Introduction Osteoporosis and its consequence of fracture remain a major burden on contemporary populations. From the age of 50, 1 out of 2 women and approximately 1 out of 3 men will sustain a fracture during their remaining lifetime (1). In women, the remaining lifetime risk of hip fracture is equivalent to or higher than the risk of invasive breast cancer (1,2), and in men, the risk of hip and clinical vertebral fractures (17%) is comparable to the risk of prostate cancer (2,3). More importantly, fracture is associated with an increased risk of mortality, and the risk is greater in men than in women. Indeed, numerous studies, including our own, have consistently shown that the relative risk of death in men with *Address correspondence to: Tuan V. Nguyen, PhD, DSc, Osteoporosis and Bone Biology Program, Garvan Institute of Medical Research, 384 Victoria Street, Sydney, NSW 2010 Australia. E-mail: [email protected]
1
ARTICLE IN PRESS 2 pieces of evidence (14,15), less than 30% of women and less than 10% of men who have already had an osteoporotic fracture receive treatment to reduce their risk of subsequent fractures (16). Thus, osteoporosis is an undertreated disease, and the undertreatment status could partly be responsible for the excess mortality associated with fracture (11). One way to improve treatment uptake is by identifying high-risk individuals for closer observation and, possibly, for selecting individuals for an appropriate treatment option. The identification of individuals at high risk of osteoporotic fracture remains a significant challenge; the susceptibility to fracture is highly variable among individuals within a population as well as between populations. This high variability in fracture risk is expected, because fracture itself is associated with multiple factors, some of which may be causal. Therefore, any single risk factor is unlikely to be helpful for risk stratification; a more logical and better approach is to tailor the risk of fracture for an individual based on the individual’s risk profile. A risk profile must consider multiple risk factors by taking into account the magnitude of association between a risk factor and a fracture. Research during the past 3 decades or so have identified several risk factors for fracture, including advancing age, low bone mineral density (BMD), a personal history of fracture, fall, low body mass index, and, more interestingly, genetic factors (17). Advancing age is clearly a major risk factor of fracture, as the incidence of fracture increases exponentially with advancing age in both men and women (18–21). Recent data have shown that the 10-yr probability of fracture at the forearm, humerus, spine, or hip increases between the ages of 45 and 85 by as much as 8-fold for women and 5-fold for men (22). Low BMD is the most robust predictor of fracture risk. Each standard deviation difference in BMD is associated with about a 2-fold change in the risk of fracture (23,24). However, BMD alone cannot reliably predict an individual who is (or is not) going to sustain a fracture. It has been estimated that less than 40% of fracture cases occur in people with BMD in the osteoporotic range (25). Among women aged 60 yr or older with low BMD (high-risk group), only 40% sustained an osteoporotic fracture within 13 years of followup (25). On the other hand, among those who sustained a fracture, almost 60% had BMDs above the osteoporotic cutpoint (T-score ≤−2.5). In other words, more than half of individuals with low BMD were “resistant to fracture.” The situation in elderly men is similar: 70% of men with low BMD did not sustain a fracture; and among fracture cases, 77% occurred in those with nonosteoporotic BMD levels (25). This finding is consistent with the point raised by Geoffrey Rose in his influential essay (26), where he postulates that the majority of chronic disease cases arise from the mass of the population with risk around the average. Clearly, BMD alone cannot identify most individuals at high risk of fracture. A rethinking is needed for the majority of individuals whose BMD values are near the osteoporotic threshold. A simple fact is that, at any age and level of BMD, fracture risk varies widely in relation to the burden of other
Nguyen and Eisman risk factors, including prior fracture and falls. A prior fragility fracture signals a substantially elevated risk of future fracture (6,27–29). The elevated risk is 1.5- to 9.5-fold, depending on the age at assessment, the number of prior fractures, and the site of the incident fracture. Pooling the results from all studies (women and men) and for all fracture sites, the risk of subsequent fracture among those with a prior fracture at any site is 2.2 times more than the risk among those without a prior fragility fracture (29). Fall is also an important risk factor for fracture. Indeed, in the elderly population, falls are associated with hip fractures, Colles’ fractures, pelvis fractures, and ankle fractures (30). Approximately 95% of hip fractures are results of falls (31). The prevalence of falls increases with advancing age (32) and is higher in women than in men (33). The genderrelated difference in fall prevalence could partly explain the difference in fracture risk between men and women. Thus, for any individual, the risk of fracture depends on a combination of factors (34). This means that 2 individuals, both with “osteoporosis,” can have different risks of fracture because they have different risk profiles. Similarly, an osteoporotic individual can have the same risk of fracture as a nonosteoporotic individual because of the difference in constellation of risk factors between the 2 individuals. In other words, the assessment of fracture risk can and should be based on a multivariable model.The advantage of a statistical multivariable model is that it yields a more accurate and consistent prediction than human experts’ prediction (35,36). A number of statistical multivariable models have been developed for individualized fracture risk assessment.Among the models, the Garvan Fracture Risk Calculator (37,38), FRAX® (39), and QFracture (40) are widely used. FRAX® uses 12 risk factors, including femoral neck BMD, anthropometric factors, lifestyle factors, and comorbidities. The Garvan Fracture Risk Calculator (Garvan) uses 5 risk factors, namely, age, gender, femoral neck BMD, prior fracture, and history of fall. Garvan considers the frequency of prior fractures and the number of falls during the past 12 mo, not simply binary (yes/no) variables. Whereas Garvan provides 5- and 10-yr risks of total fracture and hip fracture, FRAX® provides a 10-yr risk of hip fracture and major osteoporotic fractures (Table 1).
Predictive Performance The usefulness of a predictive model is usually quantified in terms of discrimination, calibration, and reclassification. Discrimination is the ability to separate individuals who will sustain a fracture along a continuum from those who will not. The primary metric of discrimination is the area under the receiver operating characteristic curve (AUC), which evaluates the compromise between sensitivity and specificity, and is thus a global estimate of prognostic accuracy (41).
Discrimination Several independent studies have been carried out to examine the prognostic performance of Garvan (42,43),
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment
3
Table 1 Risk Factors Included in FRAX® and Garvan Fracture Risk Calculator (Garvan)
Risk factors (inputs)
Output
Website
FRAX®
Garvan
Age Gender Femoral neck BMD Body weight Height History of prior fracture Parental history of hip fracture Current smoking Chronic glucocorticoid use Rheumatoid arthritis Secondary osteoporosis Alcohol (3 or more units/d) 10-yr Risk of hip fracture 10-yr Risk of major fractures
Age Gender Femoral neck BMD (or body weight) Number of prior fractures Number of falls during the past 12 mo
http://www.shef.ac.uk/FRAX
5-yr Risk of hip fracture 5-yr Risk of any fragility fracture 10-yr Risk of hip fracture 10-yr Risk of any fragility fracture www.fractureriskcalculator.com
Abbr: BMD, bone mineral density.
FRAX® (44–49), or both Garvan and FRAX® (42,50,51).The Garvan model has been validated in the Canadian Multicenter Osteoporosis Study (43), which followed up 4152 women and 1606 men for over 10 yr, and observed 123 hip fractures (93 women) and 672 fragility fractures (579 women). The application of Garvan models to predict fracture yielded good discrimination results, particularly for hip fracture (AUC 0.80 for women and 0.85 for men). In general, the discrimination in hip fracture was better than all fractures. In predicting hip fracture risk, the median AUC value for Garvan was 0.80, which was equivalent to that of FRAX® (AUC 0.78). In predicting major fracture risk, the median AUC value for Garvan and FRAX® was 0.76 and 0.69, respectively (Table 2). It appears that the discrimination of fracture in men was lower than that in women. In men, the AUC value for fracture discrimination was 0.76 by Garvan and 0.54 by FRAX® based on US data (42).
Calibration Results of discrimination analyses clearly showed that all predictive models are imperfect. However, in practice, the agreement between observed and predicted outcomes (i.e., calibration) is more important than discrimination. Several studies have indicated that the Garvan model had very good calibration. A validation study on 1422 postmenopausal women living in New Zealand observed 229 fracture cases and Garvan predicted 276 cases (99% agreement) (50). However, the study also found that Garvan overestimated the risk of fracture among individuals in the top quartile of fracture risk (50), which is also noted in the initial development study (37,38). In the Canadian Multicenter
Osteoporosis Study cohort, the Garvan model also shows a remarkable agreement between the predicted 10-yr probability of fracture and the observed 10-yr risk of fracture (43). Most validation studies suggest that the FRAX® underestimated the risk of fracture (Table 3). In the Canadian Multicenter Osteoporosis Study cohort, and results of validation suggest that FRAX® underestimated the risk of fracture in men: predicted 5.4% vs observed 6.4% (45). In a validation study in 5891 men of the MrOS cohort (49), the FRAX® 10-yr predicted probability of hip fracture was 1.4% whereas the actual risk was 3% (Table 3). For major osteoporotic fractures, the discrepancy between the 10-yr predicted risk (6.5%) and observed risk (6.9%) was not significant. A validation study in New Zealand found that FRAX® underestimated the risk of fracture by up to 50% for every risk level (50). In the FRIDEX cohort (48), FRAX® predicted only 41% of actual hip fracture cases and 46% of major fractures. In contrast, FRAX® tended to overestimate the risk of fracture in a Japanese cohort (47). Using the predicted 10-yr risk of 20% as a cutoff value to define high risk vs low risk, the sensitivity and specificity for predicting hip fracture of FRAX® were 39% and 79%, respectively (49). Another study by (53) observed that 50% of patients with fracture were not classified as “high risk” based on the FRAX®-predicted risk just before the fracture event occurred.
Reclassification It has been well known that AUC is a rather insensitive measure (54), such that a statistically significant risk
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 4
Nguyen and Eisman
Table 2 Area Under the Receiver Operating Characteristic Curve for FRAX® and Garvan Models in Predicting Hip Fracture and Major Osteoporotic Fractures Fracture (gender)
FRAX® (with BMD)
Garvan
Major fractures (women) Major fractures (men) Hip fracture (women) Major fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (men) Major fractures (men) Hip fracture (women) Major fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (women) Hip fracture (men) Major fractures (women) Major fractures (men) Hip fracture (women) Major fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (men and women) Major fractures (men and women) Hip fracture (men) Major fractures (men)
0.78 0.54 0.73 0.83
0.84 0.76 0.85 0.88 0.80 0.70 0.85 0.69 0.67 0.63 0.76 0.64
Study Sandhu et al (42) Pluskiewicz et al (51) Langsetmo et al (43)
Bolland et al (50) Sambrook et al (52) Leslie et al (44)
Ensrud et al (46) Tamaki et al (47) Azagra et al (48) Ettinger et al (49)
0.70 0.62 0.78 0.61 0.82 0.79 0.70 0.66 0.76 0.69 0.88 0.69 0.85 0.72 0.77 0.67
Abbr: BMD, bone mineral density.
Table 3 Comparison Between Predicted (FRAX® and Garvan) and Actual Fracture Risks in Some Validation Cohorts Study Bolland et al (50) Langsetmo et al (43)
Tamaki et al (47) Azagra et al (48) Ettinger et al (49)
Fracture (gender) Hip fracture (women) Major fractures (women) Hip fracture (men) Hip fracture (women) Osteoporotic fractures (men) Osteoporotic fractures (women) Hip fracture (women) Major fractures (women) Hip fracture (men and women) Major fractures (men and women) Hip fracture (men) Major fractures (men)
FRAX® predicted/observed fractures
Garvan predicted/observed fractures
43/57 121/229
85/57 276/279 41/39 232/116 180/140 756/673
8/4 50/43 7/17 30/65 1.4/3.0 6.5/6.9
Note: Data are the actual number of fractures.
factor may not have a high AUC value, particularly in the presence of other variables in the model. Hence, a clinically meaningful difference in the prognostic value between 2 predictive models is not necessarily reflected by the
absolute difference in AUC. Another feature of usefulness of a predictive model is risk reclassification (55). For a given threshold of risk (e.g., 10-yr risk of 20%), an individual can be classified as “high risk” or “low risk.” With additional
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment risk factors, the individual may change the risk category from one to another. If a new risk factor or marker is materially useful, then the addition of the risk factor should result in more individuals who will subsequently fracture being classified in the high-risk group than in the low-risk group; conversely, among those who will not subsequently fracture, more would be classified in the low-risk group than in the high-risk group. The net difference between the 2 proportions of reclassification is referred to as net reclassification improvement (56). Thus, when treatment decisions are based on the risk threshold, the net reclassification improvement can be helpful for making a clinical decision concerning an individual. Whether including BMD into a predictive model can result in an improvement in the discrimination or not is a contentious issue. In the MrOS cohort, a FRAX® plus a BMD model produced a lower risk of fracture than a FRAX® without a BMD model (49). However, other studies (44,47,48) found that a FRAX® with a BMD model produced greater AUC values than the model without BMD. Nevertheless, a complicated model with multiple risk factors does not necessarily improve the prognostic performance of a predictive model. Indeed, a simple model with age and femoral neck BMD or a model with age and fracture history was as good as a more complicated model such as FRAX® in terms of fracture prediction (46,52,57).
Concordance Between Garvan and FRAX® The concordance in the predicted probabilities of fracture between the Garvan Fracture Risk Calculator and FRAX® was modest, with the coefficient of correlation being 0.67 (57). Another validation study in 2012 postmenopausal women of Polish background found that there was a considerable discrepancy in risk estimates between the Garvan and FRAX® models with the Garvan model predicting fracture more accurately than FRAX® (51). Despite the fact that there are differences in the predicted risk of fracture between Garvan and FRAX®, the majority of the differences do not seem to impact on the treatment recommendation (58). Discrepancy in predicted risks between Garvan and FRAX® (or any other models) is expected, because the models not only use different risk factors but also have different discriminations and calibrations. Moreover, an individual can have 2 different risks predicted by 2 valid models (59), because the individual has an indefinite number of risk profiles, which could assign the individual into different risk categories (Fig. 1). This paradox is known as the “reference class” problem (61), which, in practice, means that no individual is consistently assigned to the same risk group by different models even though all models are valid.
Threshold for Intervention One application of fracture prediction models is for selecting patients suitable for intervention. However, this application raises serious challenges because the predicted risk
5
Fig. 1. Equivalent but different risk stratification: 3 subjects with no event (1–3) and 3 subjects with events (4–6). Each row represents a risk stratification method for identifying high- and low-risk subpopulations. Adapted from Stern (60).
of fracture is a continuous probabilistic variable ranging from 0 to 1, and selecting a predicted probability to classify an individual as high risk or low risk requires a thorough research. Nevertheless, the National Osteoporosis Foundation guidelines recommend treatment in the following clinical situations in postmenopausal women and in men aged 50 yr or older (62) (1) with a hip or clinical vertebral fracture or a morphometric vertebral fracture, (2) with femoral neck or lumbar spine BMD T-scores being equal to or less than −2.5 after excluding the secondary cause of osteoporosis, and (3) with femoral neck or lumbar spine BMD T-scores between −1 and −2.5 and a 10-yr risk of hip fracture ≥3% or a 10-yr risk of major osteoporotic fracture ≥20%. The clinical benefit and cost-effectiveness of these recommendations should be subject to a more systematic research. The individualization of fracture risk assessment can be applied to optimize the number needed to treat (NNT). In several randomized clinical trials (63), the number of patients needed to be treated (NNT) to reduce 1 vertebral fracture compared to the untreated group ranged between 8 and 83. For hip fracture, the NNT ranged between 91 and 250 (63). The NNT varies inversely with the background risk, such that the treatment of high-risk individuals inherently yields a lower NNT (Fig. 2). The large variability in the NNTs among trials is assumed to be due to the variability in fracture rates among the study samples. However, the variability is expected given the multiple risk factors that affect the incidence of fractures. In the presence of such
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 6
Fig. 2. Relationship between background risk (rate of fracture in the placebo group) and the number of patients needed to treat to reduce 1 case of fracture. Each dot in the figure represents the result of a randomized controlled trial. Source of data: Delmas et al (63).
variability, selecting patients based on their absolute risk of fracture (rather than based on a BMD threshold value) may improve the consistency of the therapeutic efficacy and efficiency of trials. Trials specifically testing the efficacy of multivariable risk-based therapy have not been performed. However, such approaches could be expected to prove more cost-effective and yield a more consistent NNT, particularly if the duration, typically 3 yr, is standardized. Recent analyses of the correlation between FRAX®predicted fracture risks and antifracture efficacy yielded mixed results. One clinical trial (64) randomized 5212 women aged 75 yr and older into 2 groups, placebo receiving calcium and vitamin D with placebo or clodronate (800 mg daily po). Ten-year probability of fracture was computed for each woman using baseline clinical risk factors including body mass index, prior fracture, glucocorticoid use, parental hip fracture, smoking, alcohol, and secondary osteoporosis. In women in the top 25th percentile of fracture probability (average probability of 24%), treatment reduced the risk of fracture by 23% over 3 yr (hazard ratio 0.77, 95% confidence interval 0.63–0.95). Importantly, among those in the top 10% percentile of risk (average fracture probability of 30%), treatment reduced the fracture risk by 31% (hazard ratio 0.69, 0.53–0.90) (64). Thus, treatment of individuals at high risk or moderate risk could reasonably be expected to have a similar benefit in relative but greater effect in absolute risk reduction for the higher risk group. In a post hoc analysis of the Fracture Intervention Trial, the investigators used FRAX® to estimate each patient’s 10-yr risk of fracture, and then correlated the risk with antifracture efficacy. The investigators conclude that the
Nguyen and Eisman magnitude of effect of alendronate was constant across FRAX®-predicted risks (65). However, a close examination of the data reveals that the absolute risk reduction increased with absolute risk levels, and as a result, the NNT decreased as absolute risk increased. For example, among those in the first tertile of risk (4.8%–22.1%) the NNT was 76, but among those in the highest tertile (34.2%–85.4%), the NNT was 40. Results of post hoc analyses of clinical trials appear to suggest that, in relative risk terms, the magnitude of antifracture efficacy of denosumab (66) and bazedoxifene (67) was dependent on absolute risks of fracture, such that those at high risk had better relative efficacy than those at low risk of fracture. However, the antifracture of strontium ranelate (68) and raloxifene (69) seemed to be independent of the fracture risk assessed by FRAX®. However, it is not clear whether absolute risk reduction or the NNT was independent of patients’ absolute risk levels. Nevertheless, taken together, these results are consistent with the supposition that the antifracture effect size of pharmacological therapies is inversely associated with patients’ absolute risks, supporting the use of predictive models for selecting patients to include in future randomized controlled trials of osteoporosis.
Risk Communications: Individual vs Population At the heart of risk communication is the very meaning of “risk.” More specifically, what does a [say] 20% risk of fracture within the next 5 yr mean to an individual? This question is important, because only when an individual understands and appreciates the risk can the individual arrive at an informed decision. Given the way the risk is calculated, it is probably reasonable to say that “based on your age and risk profile, your risk of sustaining a fracture in the next 5 yr is 20%.” This information can, however, present a conflict between the individual and the population. In medical parlance, risk is commonly seen as the probability of getting a disease over a certain period of time, but epistemologically, risk is defined as the product of probability and the consequence of an adverse event (70). The probability component measures the uncertainty, whereas consequence measures the impact of an event. Virtually all predictive models in medicine, including Garvan, provide only the uncertainty, not the impact, of a health event. There is a conflict between the population and the individual uncertainty in the estimation of risk. The estimated probability of fracture is derived from data of a group of individuals, but the probability is used for an individual. In other words, a group-based estimate is used as an intrinsic property of an individual. However, an individual will either sustain or not sustain a fracture (i.e., a binary event), and risk in the form of continuous probability is thus not meaningful to an individual. Indeed, some argue that probabilistic risk is not meaningful to an individual (71) because, from a probabilistic viewpoint,
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment probability does not exist (72). Nevertheless, we would argue that risk is a meaningful measure for an individual because, from a Bayesian point of view, risk reflects an individual’s measure of belief that an event will occur. Therefore, individual risk is still a useful concept in fracture risk assessment. Because risk is a conditional probability, the predicted risk of fracture for an individual can vary, depending on the collection of risk factors considered. Moreover, statistical models of risk prediction do not necessarily produce a unique solution. It is important to appreciate that the predicted risk is actually an average, a typical value, with “true” values fluctuating below or above the typical value. Therefore, an individual does not necessarily have a unique risk value. This subtle fact also explains why different valid predictive models can yield substantially different results for an individual. Predicted risk can change with time because risk factors for fracture change with time. However, risk estimates are currently based on the assumption of constancy in risk factors. An individual’s absolute risk over a meaningful time period, arguably 5 rather than 10 yr (73), has to be considered alongside that individual’s concern about the outcome of interest. Some fractures will be of limited concern to some people but of great concern to others who have a more personal and perhaps more accurate view of their impact on the quality of life and even survival. Thus, whereas risk exists as an objective identity, its measure, probability, does not exist objectively.
Issues for Further Research Although the development and implementation of fracture prediction models represent a major advance in osteoporosis research, several issues need to be addressed in the future. A good predictive model should meet the “4R” criteria: reliability, reproducibility, relevance, and realworld value. A reliable model should have good sensitivity and specificity, and its predicted outcomes agree closely with observed outcomes. Reproducibility here means that risk factors included in a model should be highly reproducible and that the risk factors are found in various independent populations. Relevance refers to the clinical relevance of a predictive model, such that intervention on those with high risk predicted by the model could alter their outcome. A predictive model has “real-world” value if the model is easy to use and inexpensive. None of the existing predictive models in osteoporosis meet all the criteria. It is important to distinguish between prediction and association, a point that is not well recognized in clinical medicine (74). In the context of variable selection, the goal of association or explanation analysis is to identify risk factors that affect an outcome variable, but the goal of prediction or classification is about to find observed factors that predict new observation of the outcome variable. Therefore, statistically significant risk factors are not necessarily good predictors or classifiers (75,76), but a factor that is not
7 statistically significant can be a useful predictor (77). All existing models have been developed based on the idea of association, not prediction, using traditional statistical methods. The development of the Garvan model, for example, was based on the Bayesian model average method (78), which has been shown to be more reliable than traditional methods. In recent years, several new machine learning algorithms, such as the random forest and the retention-partition methods, can be useful for developing predictive models, but these algorithms have not been applied.
Bone Turnover Markers (BTMs) BTMs can be useful for predicting fracture risk independent of BMD. There is accumulating evidence that accelerated bone resorption with the associated imbalance in bone remodeling adversely affects skeletal microstructure and increases fracture risk. Several cross-sectional and longitudinal studies have found that fragility fractures occur not only because of low BMD but also as a result of rapid bone turnover that leads to architectural changes, such as disruption of trabecular networks and struts. Increased urinary levels of the pyridinium cross-link, deoxypyridinoline, was associated with a 2- to 3-fold increase in the risk of hip fracture (79). In the EPIdémiologie de l’OStéporose study, urinary cross-linked C-telepeptide (CTX) and free deoxypyridinoline levels above the premenopausal range were associated with a 2-fold increase in hip fracture risk after adjusting for BMD and physical mobility (80). In that study, 16% of the subjects with both a low BMD (T-score ≤−2.5) and an elevated urinary CTX (T-score ≤−2 standard deviation of the mean for premenopausal women) had a relative risk of 4.8 for hip fracture, compared with a relative risk of 2.7 for low BMD alone or a relative risk of 2.2 for high urinary CTX alone (80). Elevated bone resorption marker levels have been shown to be associated with increased fracture risk independent of BMD (81). In our recent study in men, increased bone resorption was associated with increased fracture risk in men (82). Taken together, these results strongly suggest that an incorporation of BTMs into the existing prognostic model could improve the prediction of absolute fracture risk.
Trabecular Bone Score (TBS) One of the promising factors for improving the accuracy of fracture prediction is the TBS (83–85), a graylevel texture measurement based on a variogram of the 2-dimensional BMD projection image (86). Thus, TBS is a texture parameter that reflects a pixel gray-level variation in dual-energy X-ray absorptiometry images. Previous studies have reported that TBS is not significantly correlated with BMD, but is significantly correlated with the trabecular number, trabecular separation, and the structure model index. Moreover, TBS was found to be associated with fracture risk in elderly women (87) and in diabetic patients (88) independent of BMD and classical clinical risk factors (84,85).
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 8 Genetics The variation in fracture susceptibility is partly determined by genetic factors. A twin study (89) has shown that almost 50% of the variance in the liability to fracture was attributable to genetic factors. Through large-scale collaborative studies, several genetic variants have been identified. To date, 12 genome-wide association studies have been conducted for BMD phenotypes (90), and 62 loci have been identified to be associated with BMD at the genomewide significance level.Among the 62 single-nucleotide polymorphisms (SNPs) identified, 8 SNPs were associated with fracture risk at the genome-wide significance level (90). However, these common variants have a modest association with fracture risk with odds ratios ranging between 1.1 and 1.4, suggesting that, individually, these variants have limited utility for fracture prediction. Nevertheless, a recent simulation study showed that a genetic profile of up to 50 genetic variants, with each having a modest effect size (odds ratio of 1.01–1.35), could improve the accuracy of fracture prediction by 10% points of AUC (91). Two recent studies in postmenopausal women of Korean background found that a genetic profiling of 39 SNPs in 30 human genomic loci could increase the precision of nonvertebral fracture prediction and help to define the risk threshold (92), whereas 35 risk alleles were significantly associated with the risk of vertebral fracture (92,93) in patients on bisphosphonate. Taken together, these results and the present finding suggest that genetic profiling is useful in the identification of high-risk individuals.
Fracture-Specific Prediction Current prognostic models have been developed based on the “one size-fits-all” approach. For example, each model uses a set of risk factors for predicting the 5- or 10-yr risk of all types of fracture, because it is assumed that the set of risk factors is associated with any type of fracture in any ethnic population in both men and women. This assumption is strong and likely untenable because a risk factor may be uniquely associated with a type of fracture. For instance, fall is a major risk factor for hip fracture and is not a risk factor for vertebral fracture.
Time-Variant Predictions An important weakness of current fracture prediction models is that they are based on a single measurement of risk factors, with the underlying but not stated assumption that the risk factors do not change with time. Obviously, this assumption is not true in many risk factors such as BMD and body weight, which are known to decline or change with time. Moreover, the rates of decline in BMD varied substantially among individuals. Similarly, the risk of a second fracture is higher and closer in time to the initial fracture with risks declining substantially after 10 yr. Therefore, 1 important aspect of future model development should take the time-varying nature of risk factors into account to achieve a better estimate of risk for an individual.
Nguyen and Eisman In conclusion, fracture caused by osteoporosis is a serious skeletal disorder because it is associated with considerable morbidity and increased risk of mortality. Despite the availability of many effective treatments, the treatment uptake is still low. In the era of evidence-based medicine, patients and doctors need to work together to make informed decisions concerning treatment decisions. Therefore, the individualized assessment of fracture risk is of growing interest to the general public, doctors, and researchers. The advance of predictive models for individualized fracture risk assessment represents a great success in the translational research of osteoporosis. Although these models remain to be improved, their availability helps patients and doctors make decisions that are more evidence based and better tailored to an individual’s specific needs. However, if predictive models are used to guide therapy, a higher standard of reliability, reproducibility, and relevance must be attained.
References 1. Nguyen ND, Ahlborg HG, Center JR, et al. 2007 Residual lifetime risk of fractures in women and men. J Bone Miner Res 22(6):781–788. 2. Cummings SR, Black DM, Rubin SM. 1989 Lifetime risks of hip, Colles’, or vertebral fracture and coronary heart disease among white postmenopausal women. Arch Intern Med 149(11):2445–2448. 3. Shortt NL, Robinson CM. 2005 Mortality after low-energy fractures in patients aged at least 45 years old. J Orthop Trauma 19(6):396–400. 4. Johnell O, Kanis JA, Oden A, et al. 2004 Fracture risk following an osteoporotic fracture. Osteoporos Int 15(3):175–179. 5. Bliuc D, Nguyen ND, Milch VE, et al. 2009 Mortality risk associated with low-trauma osteoporotic fracture and subsequent fracture in men and women. JAMA 301(5):513–521. 6. Center JR, Nguyen TV, Schneider D, et al. 1999 Mortality after all major types of osteoporotic fracture in men and women: an observational study. Lancet 353(9156):878–882. 7. Hindmarsh DM, Hayen A, Finch CF, Close JC. 2009 Relative survival after hospitalisation for hip fracture in older people in New South Wales, Australia. Osteoporos Int 20(2):221–229. 8. Haentjens P, Magaziner J, Colon-Emeric CS, et al. 2010 Metaanalysis: excess mortality after hip fracture among older women and men. Ann Intern Med 152(6):380–390. 9. Randell AG, Nguyen TV, Bhalerao N, et al. 2000 Deterioration in quality of life following hip fracture: a prospective study. Osteoporos Int 11(5):460–466. 10. Adachi JD, Adami S, Gehlbach S, et al. 2010 Impact of prevalent fractures on quality of life: baseline results from the global longitudinal study of osteoporosis in women. Mayo Clin Proc 85(9):806–813. 11. Frost SA, Nguyen ND, Center JR, et al. 2013 Excess mortality attributable to hip-fracture: a relative survival analysis. Bone 56(1):23–29. 12. Lyles KW, Colon-Emeric CS, Magaziner JS, et al. 2007 Zoledronic acid in reducing clinical fracture and mortality after hip fracture. N Engl J Med 357:1799–1809. 13. Beaupre LA, Morrish DW, Hanley DA, et al. 2011 Oral bisphosphonates are associated with reduced mortality after hip fracture. Osteoporos Int 22(3):983–991.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment 14. Center JR, Bliuc D, Nguyen ND, et al. 2011 Osteoporosis medication and reduced mortality risk in elderly women and men. J Clin Endocrinol Metab 96(4):1006–1014. 15. Bolland MJ, Grey AB, Gamble GD, Reid IR. 2010 Effect of osteoporosis treatment on mortality: a meta-analysis. J Clin Endocrinol Metab 95(3):1174–1181. 16. Eisman J, Clapham S, Kehoe L. 2004 Osteoporosis prevalence and levels of treatment in primary care: the Australian BoneCare Study. J Bone Miner Res 19(12):1969– 1975. 17. Ralston SH. 2002 Genetic control of susceptibility to osteoporosis. J Clin Endocrinol Metab 87(6):2460–2466. 18. Kanis JA, Johnell O, Oden A, et al. 2001 Ten year probabilities of osteoporotic fractures according to BMD and diagnostic thresholds. Osteoporos Int 12(12):989–995. 19. Nguyen TV, Eisman JA, Kelly PJ, Sambrook PN. 1996 Risk factors for osteoporotic fractures in elderly men. Am J Epidemiol 144(3):255–263. 20. Cummings SR, Nevitt MC, Browner WS, et al. 1995 Risk factors for hip fracture in white women. Study of Osteoporotic Fractures Research Group. N Engl J Med 332(12):767– 773. 21. Hui SL, Slemenda CW, Johnston CCJ. 1988 Age and bone mass as predictors of fracture in a prospective study. J Clin Invest 81(6):1804–1809. 22. Wasnich RD, Davis JW, Ross PD. 1994 Spine fracture risk is predicted by non-spine fractures. Osteoporos Int 4(1):1– 5. 23. Marshall D, Johnell O, Wedel H. 1996 Meta-analysis of how well measures of bone mineral density predict occurrence of osteoporotic fractures. BMJ 312(7041):1254–1259. 24. Nguyen T, Sambrook P, Kelly P, et al. 1993 Prediction of osteoporotic fractures by postural instability and bone density. BMJ 307(6912):1111–1115. 25. Nguyen ND, Eisman JA, Center JR, Nguyen TV. 2007 Risk factors for fracture in nonosteoporotic men and women. J Clin Endocrinol Metab 92(3):955–962. 26. Rose G. 2001 Sick individuals and sick populations. Int J Epidemiol 30(3):427–432, discussion 433-4. 27. Pongchaiyakul C, Nguyen ND, Jones G, et al. 2005 Asymptomatic vertebral deformity as a major risk factor for subsequent fractures and mortality: a long-term prospective study. J Bone Miner Res 20(8):1349–1355. 28. Ismail AA, Cockerill W, Cooper C, et al. 2001 Prevalent vertebral deformity predicts incident hip though not distal forearm fracture: results from the European Prospective Osteoporosis Study. Osteoporos Int 12(2):85–90. 29. Klotzbuecher CM, Ross PD, Landsman PB, et al. 2000 Patients with prior fractures have an increased risk of future fractures: a summary of the literature and statistical synthesis. J Bone Miner Res 15(4):721–739. 30. Cummings SR, Nevitt MC, Study of Osteoporotic Fractures Research Group. 1994 Non-skeletal determinants of fractures: the potential importance of the mechanics of falls. Osteoporos Int 4(Suppl 1):67–70. 31. Nyberg L, Gustafson Y, Berggren D, et al. 1996 Falls leading to femoral neck fractures in lucid older people. J Am Geriatr Soc 44(2):156–160. 32. Lord SR, Sambrook PN, Gilbert C, et al. 1994 Postural stability, falls and fractures in the elderly: results from the Dubbo Osteoporosis Epidemiology Study. Med J Aust 160(11):684– 685, 688-91. 33. Prudham D, Evans JG. 1981 Factors associated with falls in the elderly: a community study. Age Ageing 10(3):141– 146.
9 34. Nguyen ND, Pongchaiyakul C, Center JR, et al. 2005 Identification of high-risk individuals for hip fracture: a 14-year prospective study. J Bone Miner Res 20(11):1921–1928. 35. Grove WM, Meehl PE. 1996 Comparative efficiency of informal (subjective, impressionistic) and formal (mechanical, algorithmic) prediction procedures: the clinical–statistical controversy. Psychol Public Policy Law 2(293–323). 36. Oberije C, Nalbantov G, Dekker A, et al. 2014 A prospective study comparing the predictions of doctors versus models for treatment outcome of lung cancer patients: a step toward individualized care and shared decision making. Radiother Oncol 112(1):37–43. 37. Nguyen ND, Frost SA, Center JR, et al. 2007 Development of a nomogram for individualizing hip fracture risk in men and women. Osteoporos Int 18(8):1109–1117. 38. Nguyen ND, Frost SA, Center JR, et al. 2008 Development of prognostic nomograms for individualizing 5-year and 10-year fracture risks. Osteoporos Int 19:1431–1444. 39. Kanis JA, Johnell O, Oden A, et al. 2008 FRAX and the assessment of fracture probability in men and women from the UK. Osteoporos Int 19(4):385–397. 40. Collins GS, Mallett S, Altman DG. 2011 Predicting risk of osteoporotic and hip fracture in the United Kingdom: prospective independent and external validation of QFracture scores. BMJ 342:d3651. 41. Harrell FEJ, Califf RM, Pryor DB, et al. 1982 Evaluating the yield of medical tests. JAMA 247(18):2543–2546. 42. Sandhu SK, Nguyen ND, Center JR, et al. 2010 Prognosis of fracture: evaluation of predictive accuracy of the FRAX algorithm and Garvan nomogram. Osteoporos Int 21(5):863– 871. 43. Langsetmo L, Nguyen TV, Nguyen ND, et al. 2011 Independent external validation of nomograms for predicting risk of low-trauma fracture and hip fracture. CMAJ 183(2):E107– E114. 44. Leslie WD, Lix LM, Johansson H, et al. 2010 Independent clinical validation of a Canadian FRAX tool: fracture prediction and model calibration. J Bone Miner Res 25(11):2350– 2358. 45. Leslie WD, Lix LM, Langsetmo L, et al. 2011 Construction of a FRAX(R) model for the assessment of fracture probability in Canada and implications for treatment. Osteoporos Int 22(3):817–827. 46. Ensrud KE, Lui LY, Taylor BC, et al. 2009 A comparison of prediction models for fractures in older women: is more better? Arch Intern Med 169(22):2087–2094. 47. Tamaki J, Iki M, Kadowaki E, et al. 2011 Fracture risk prediction using FRAX(R): a 10-year follow-up survey of the Japanese Population-Based Osteoporosis (JPOS) Cohort Study. Osteoporos Int 22(12):3037–3045. 48. Azagra R, Roca G, Encabo G, et al. 2012 FRAX(R) tool, the WHO algorithm to predict osteoporotic fractures: the first analysis of its discriminative and predictive ability in the Spanish FRIDEX cohort. BMC Musculoskelet Disord 13:204. 49. Ettinger B, Ensrud KE, Blackwell T, et al. 2013 Performance of FRAX in a cohort of community-dwelling, ambulatory older men: the Osteoporotic Fractures in Men (MrOS) study. Osteoporos Int 24(4):1185–1193. 50. Bolland MJ, Siu AT, Mason BH, et al. 2011 Evaluation of the FRAX and Garvan fracture risk calculators in older women. J Bone Miner Res 26(2):420–427. 51. Pluskiewicz W, Adamczykb P, Franekc E, et al. 2009 Conformity between 10-year probability of any osteoporotic fracture assessed by FRAX and nomogram by Nguyen et al. Bone 44(Suppl 2):S229–S230.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS 10 52. Sambrook PN, Flahive J, Hooven FH, et al. 2011 Predicting fractures in an international cohort using risk factor algorithms without BMD. J Bone Miner Res 26(11):2770–2777. 53. Aubry-Rozier B, Stoll D, Krieg MA, et al. 2013 What was your fracture risk evaluated by FRAX(R) the day before your osteoporotic fracture? Clin Rheumatol 32(2):219–223. 54. Huntjens KM, Kosar S, van Geel TA, et al. 2010 Risk of subsequent fracture and mortality within 5 years after a nonvertebral fracture. Osteoporos Int 21(12):2075–2082. 55. Pencina MJ, D’Agostino RB Sr, D’Agostino RB Jr, Vasan RS. 2008 Evaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond. Stat Med 27(2):157–172, discussion 207-12. 56. Cook NR. 2008 Statistical evaluation of prognostic versus diagnostic models: beyond the ROC curve. Clin Chem 54(1):17– 23. 57. van Geel TA, Nguyen ND, Geusens PP, et al. 2011 Development of a simple prognostic nomogram for individualising 5-year and 10-year absolute risks of fracture: a populationbased prospective study among postmenopausal women. Ann Rheum Dis 70(1):92–97. 58. Bolland MJ, Grey A, Gamble G, Reid IR. 2013 Comment on Kanis et al.: pitfalls in the external validation of FRAX. Osteoporos Int 24(1):389–390. 59. Lemeshow S, Klar J, Teres D. 1995 Outcome prediction for individual intensive care patients: useful, misused, or abused? Intensive Care Med 21(9):770–776. 60. Stern R. 2010 Discordance of individual risk estimates. J Am Coll Cardiol 56(9):743, author reply 743-4. 61. Hájek A. 2007 The reference class problem is your problem too. Synthese 156:563–585. 62. NOF. Clinician’s guide to prevention and treatment of osteoporosis. National Osteoporosis Foundation 2008;Washington DC. 63. Delmas PD, Rizzoli R, Cooper C, Reginster JY. 2005 Treatment of patients with postmenopausal osteoporosis is worthwhile. The position of the International Osteoporosis Foundation. Osteoporos Int 16(1):1–5. 64. McCloskey E, Johansson H, Oden A, et al. 2007 Efficacy of clodronate on fracture risk in women selected by 10-year fracture probability. J Bone Miner Res 22(10):S131. 65. Donaldson MG, Palermo L, Ensrud KE, et al. 2012 Effect of alendronate for reducing fracture by FRAX score and femoral neck bone mineral density: the Fracture Intervention Trial. J Bone Miner Res 27(8):1804–1810. 66. McCloskey EV, Johansson H, Oden A, et al. 2012 Denosumab reduces the risk of osteoporotic fractures in postmenopausal women, particularly in those with moderate to high fracture risk as assessed with FRAX. J Bone Miner Res 27(7):1480–1486. 67. Kanis JA, Johansson H, Oden A, McCloskey EV. 2009 Bazedoxifene reduces vertebral and clinical fractures in postmenopausal women at high risk assessed with FRAX. Bone 44(6):1049–1054. 68. Kanis JA, Johansson H, Oden A, McCloskey EV. 2011 A meta-analysis of the effect of strontium ranelate on the risk of vertebral and non-vertebral fracture in postmenopausal osteoporosis and the interaction with FRAX((R). Osteoporos Int 22(8):2347–2355. 69. Kanis JA, Johansson H, Oden A, McCloskey EV. 2010 A meta-analysis of the efficacy of raloxifene on all clinical and vertebral fractures and its dependency on FRAX. Bone 47(4):729–735. 70. Kaplan S, Garrick BJ. 1981 On the quantitative definition of risk. Risk Anal 1:11–27.
Nguyen and Eisman 71. Sniderman AD, D’Agostino RB Sr, Pencina MJ. 2015 The role of physicians in the era of predictive analytics. JAMA 314(1):25–26. 72. Nau RF. 2001 De Finetti was right: probability does not exist. Theory Decis 51:89–124. 73. Bolland MJ, Jackson R, Gamble GD, Grey A. 2013 Discrepancies in predicted fracture risk in elderly people. BMJ 346:e8669. 74. Shmueli G. 2010 To explain or to predict. Stat Sci 25(3):289– 310. 75. Lo A, Chernoff H, Zheng T, Lo SH. 2015 Why significant variables aren’t automatically good predictors. Proc Natl Acad Sci USA 112(45):13892–13897. 76. Pepe MS, Janes H, Longton G, et al. 2004 Limitations of the odds ratio in gauging the performance of a diagnostic, prognostic, or screening marker. Am J Epidemiol 159(9):882– 890. 77. Guyon I, Elisseeff A. 2003 An introduction to variable and feature selection. J Mach Learn 3:1157–1182. 78. Hoeting J. 1999 Bayesian model averaging: a tutorial. Stat Sci 14:382–417. 79. van Daele PL, Seibel MJ, Burger H, et al. 1996 Case-control analysis of bone resorption markers, disability, and hip fracture risk: the Rotterdam study. BMJ 312(7029):482– 483. 80. Garnero P, Hausherr E, Chapuy MC, et al. 1996 Markers of bone resorption predict hip fracture in elderly women: the EPIDOS Prospective Study. J Bone Miner Res 11(10):1531– 1538. 81. Akesson K, Ljunghall S, Jonsson B, et al. 1995 Assessment of biochemical markers of bone metabolism in relation to the occurrence of fracture: a retrospective and prospective population-based study of women. J Bone Miner Res 10(11):1823–1829. 82. Meier C, Nguyen TV, Center JR, et al. 2005 Bone resorption and osteoporotic fractures in elderly men: the Dubbo osteoporosis epidemiology study. J Bone Miner Res 20(4):579– 587. 83. Hans D, Barthe N, Boutroy S, et al. 2011 Correlations between trabecular bone score, measured using anteroposterior dualenergy X-ray absorptiometry acquisition, and 3-dimensional parameters of bone microarchitecture: an experimental study on human cadaver vertebrae. J Clin Densitom 14(3):302– 312. 84. Silva BC, Leslie WD, Resch H, et al. 2014 Trabecular bone score: a noninvasive analytical method based upon the DXA image. J Bone Miner Res 29(3):518–530. 85. Ulivieri FM, Silva BC, Sardanelli F, et al. 2014 Utility of the trabecular bone score (TBS) in secondary osteoporosis. Endocrine 47(2):435–448. 86. Pothuaud L, Barthe N, Krieg MA, et al. 2009 Evaluation of the potential use of trabecular bone score to complement bone mineral density in the diagnosis of osteoporosis: a preliminary spine BMD-matched, case-control study. J Clin Densitom 12(2):170–176. 87. Hans D, Goertzen AL, Krieg MA, Leslie WD. 2011 Bone microarchitecture assessed by TBS predicts osteoporotic fractures independent of bone density: the Manitoba study. J Bone Miner Res 26(11):2762–2769. 88. Leslie WD, Aubry-Rozier B, Lamy O, , et al. 2013 TBS (trabecular bone score) and diabetes-related fracture risk. J Clin Endocrinol Metab 98(2):602–609. 89. Michaelsson K, Melhus H, Ferm H, et al. 2005 Genetic liability to fractures in the elderly. Arch Intern Med 165(16):1825–1830.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017
ARTICLE IN PRESS Fracture Risk Assessment 90. Richards JB, Zheng HF, Spector TD. 2012 Genetics of osteoporosis from genome-wide association studies: advances and challenges. Nat Rev Genet 13(8):576– 588. 91. Tran BNH, Nguyen ND, Nguyen VX, et al. 2011 Genetic profiling and individualized prognosis of fracture. J Bone Miner Res 26(2):414–419.
11 92. Lee SH, Lee SW, Ahn SH, et al. 2013 Multiple gene polymorphisms can improve prediction of nonvertebral fracture in postmenopausal women. J Bone Miner Res 28(10):2156–2164. 93. Lee SH, Cho EH, Ahn SH, et al. 2016 Prediction of future osteoporotic fracture occurrence by genetic profiling: a 6-year follow-up observational study. J Clin Endocrinol Metab doi:10.1210/jc.2015-3972. jc20153972.
Journal of Clinical Densitometry: Assessment & Management of Musculoskeletal Health
Volume ■, 2017