Optimization of BMD Measurements to Identify High Risk Groups for Treatment—A Test Analysis

Authors


Abstract

The aim of this study was to optimize the use of BMD measurements in case finding strategies. The use of clinical risk factors with and without BMD was explored in a random sample of the Sheffield female population ≥75 years of age. The use of clinical risk factors alone could identify women well above or well below a threshold of fracture risk. BMD assessment can be confined to a minority of women (∼20%) in whom the measurement aids in prognostication of fracture.

Introduction: The aim of this study was to develop a methodology to optimize the role of BMD measurements in a case finding strategy. We studied 2113 women ≥75 years of age randomly selected from Sheffield, UK, and adjacent regions. Baseline assessment included hip BMD and clinical risk factors. Outcomes included death and fracture in women followed for 6723 person-years.

Materials and Methods: Poisson models were used to identify significant risk factors for all fractures and for death with and without BMD and the hazard functions were used to compute fracture probabilities. Women were categorized by fracture probability with and without a BMD assessment. A 10-year fracture probability threshold of 35% was taken as an intervention threshold. Discordance in categorization of risk (i.e., above or below the threshold probability) between assessment with and without BMD was examined by logistic regression as probabilities of re-classification. Age, prior fracture, use of corticosteroids, and low body mass index were identified as significant clinical risk factors.

Results: A total of 16.8% of women were classified as high risk based on these clinical risk factors. The average BMD in these patients was ∼1 SD lower than in low-risk women; 21.5% of women were designated to be at high risk with the addition of BMD. Fifteen percent of all women were reclassified after adding BMD to clinical risk factors, most of whom lay near the intervention threshold. When a high probability of reclassification was accepted (without a BMD test) for high risk to low risk (p1 ≤ 0.8) and a low probability accepted for low to high risk (p2 ≤ 0.2), BMD tests would be required in only 21% of the population.

Conclusion: We conclude that the use of clinical risk factors can identify elderly women at high fracture risk and that such patients have a low average BMD. BMD testing is required, however, in a minority of women—a fraction that depends on the probabilities accepted for classification and the thresholds of risk chosen. These findings need to be validated in other cohorts at different ages and from different regions of the world.

INTRODUCTION

IN THE ABSENCE OF adequate evidence for the use of population screening, case-finding strategies are widely accepted as a method of identifying individuals suitable for the treatment of osteoporosis.(1–4) The general principles are that individuals are identified by the presence of risk factors and subsequently undergo a BMD examination. When BMD is below a given threshold, intervention for osteoporosis is recommended. The threshold for BMD that is used varies in different guidelines. For example, the European Foundation for Osteoporosis (EFFO; now the International Osteoporosis Foundation [IOF]) suggests a threshold equivalent to a T-score of −2.5 SD, corresponding to the WHO diagnostic criterion for osteoporosis, whereas in the United States, the National Osteoporosis Foundation recommends that a T-score of −2.0 SD be used.(2,5) Also, the threshold may be fixed (IOF) or vary according to the presence of risk factors that contribute to risk independently of BMD.(5)

It is self-evident that intervention should be targeted to those at highest risk, always provided that effective treatments are available to reduce that risk. The use of threshold risks based on BMD alone are imprecise, because the same T-score has a different prognostic significance at different ages.(6) Thus, a T-score of −2.5SD at the femoral neck is associated with a 10-year fracture probability of fracture at the hip, spine, forearm, or proximal humerus of 11% in Swedish women at the age of 50 years but a probability of 26% at the age of 80 years.(7) This is because age contributes to risk independently of BMD. There are risk factors in addition to age and BMD that provide independent risks. Examples include prior fragility fractures,(8–11) high rates of bone turnover,(12–14) the presence of some diseases, and the use of drugs such as corticosteroids.(15,16)

The use of multiple risk factors that are at least in part mutually independent improves the detection rate (sensitivity) of assessment of risk in case-findings strategies without sacrificing specificity.(17) In other words, more individuals in a population can be detected in whom the risk of fracture is above a given threshold. In practice, three categories of individual may be identified based on clinical risk factors: (1) individuals much above a threshold risk; (2) individuals close to a threshold risk in whom BMD might be measured to more accurately categorize risk; and (3) individuals much below a threshold risk.

It may be assumed that patients in categories 1 or 3 are unlikely to have their categorization of risk changed by performing a BMD test. The assumption can be tested by determining the probability that a BMD test would change an individual's categorization from high to low risk and vice versa.

The aim of this study was to determine the performance characteristics of clinical risk assessment on the categorization of threshold risks and the impact of measuring BMD on this strategy.

MATERIALS AND METHODS

We studied women ≥75 years of age that were selected randomly from the population of Sheffield, UK, and surrounding districts between 1993 and 1999. Approximately 35,000 women identified from General Practitioner listings were contacted by letter and invited to attend for assessment of skeletal status. A total of 5873 women were willing to attend for the screening visit. Apart from age and sex, the only inclusion criteria were the ability to give informed consent and a willingness to take part in the 3-year study. Exclusion criteria comprise the taking of bone active agents, known malabsorption states, and lack of compliance because of a poor mental state or concurrent illnesses. Laboratory exclusion criteria were a serum creatinine >0.3 mM, leukopenia (white cell count, <2 × 109/liter), hyper- or hypocalcemia, and elevated transaminases (greater than twice the upper reference limit). On these criteria, 281 women were excluded, and the remainder was randomly allocated after informed consent to treatment with the bisphosphonate, clodronate, or an identical placebo. Randomization was done with the SAS/PLAN procedure for one site, two treatments, and a block size of 10. This study was conducted after enrollment of 2796 women into the placebo arm of the study, in whom follow-up data were available in 2175. Data that were made anonymous were provided to an independent statistician, and the investigators remained fully blinded. Women remained under the supervision of their general practitioner, including the treatment of intercurrent fractures. The study had the approval of the relevant Ethics Committees.

Baseline variables measured included age, height, weight, body mass index (BMI), personal and family history of fracture, smoking (never/ever), milk (never, occasional, and 1-2, 3-4, and >5 glasses/day), oral corticosteroids, and disorders associated with osteoporosis or fracture. These comprised self-reported rheumatoid arthritis, stroke (transient ischemic attacks or cerebrovascular accidents), diabetes, hyperparathyroidism, and osteoarthrosis. In addition, we documented age at menopause, prior use of hormone replacement treatment, hysterectomy, and oophorectomy. Bone assessment included BMD by DXA (Hologic 4500) at the total hip and its regions (neck, trochanter, intertrochanteric, Ward's triangle). T-scores for BMD at the femoral neck were computed using the Third National Health and Nutrition Examination Survey (NHANES III),(18) using the ages 20-29 years as a reference for young healthy women.(19)

Outcome variables collected were fractures according to site and death from any cause. Deaths were verified by death certificate. Fractures were asked about at six monthly home visits by study nurses. All self-reported fractures were independently verified from radiographic inspection, radiographic reports, or operation and hospital summaries. A comparison of self-reported events and those independently documented showed that >99% of fractures were captured by the self-reports. Of 2175 women followed up, a full baseline assessment was available in 2113 women (97%). Follow-up was for 6723 patient-years, during which time 208 deaths and 282 fractures were recorded. Fractures considered to be caused by osteoporosis (232 fractures; 82%)(20) were comprised of 53 at the hip, 26 clinical vertebral fractures, 36 axial nonvertebral fractures, and 117 appendicular fractures.

A Poisson model was used to identify significant risk factors for all fractures and for death with and without the inclusion of hip BMD. Current time (time since assessment) was an additional co-variate that describes the change in risk with time from entry into the study. The hazard functions for mortality and fracture were used to compute 10-year fracture probability by previously published methods.(17) A 10-year time frame was chosen, because in the context of assessment and treatment, this covers the likely duration of treatment and any continued benefits that might accrue when treatment is stopped.(21,22) For the purpose of this paper, we used a threshold probability of fracture of 35% over 10 years, above which an intervention threshold was attained. The threshold approximates the 10-year probability of any osteoporotic fracture in women at the same age with a T-score of −2.5 SD.(6) The effect of a less and more stringent threshold was also examined (27% and 37% 10-year probabilities to select 50% or 10% of the population for treatment). The 10-year fracture probabilities were computed in the absence of a BMD measurement (S1) and with the inclusion of BMD (S2).

The discordance between S1 and S2 was categorized according to the threshold probability (35%). Individuals with discordant results were those classified at high risk without BMD (S1 ≥ 0.35) that were subsequently classified at low risk with a BMD measurement (S2 < 0.35). Conversely, reclassification was documented in women characterized initially at low risk (S1 < 0.35) in whom BMD assessment subsequently showed them to be at high risk (S2 ≥ 0.35).

A logistic regression model was used to determine the probabilities that an individual at low risk without a BMD measurement would be reclassified to be at high risk with the addition of a BMD measurement (false negatives). Conversely, a separate model was applied to define probabilities that those with a high risk would, with a BMD examination, be classified at low risk (false positives). Threshold probabilities were used to determine the proportion of the population in whom BMD assessment would be required to optimize a case finding strategy. p1 was the probability of reclassifying a high-risk patient to low risk, and for the base case, was set at 0.8. Thus, if p1 is exceeded, a BMD measurement would be required. p2 was the probability of reclassifying a low-risk patient as high risk, and for the base case, was set at 0.2. Thus, if p2 was exceeded, a BMD measurement would be required.

RESULTS

The characteristics of the patients studied are shown in Table 1. Of the variables tested, low body weight, BMI, previous fracture, corticosteroid use, and osteoarthritis were significantly associated with fracture risk (see Table 1). Weight, however, was no longer a significant risk factor when BMI was used in the model, and osteoarthritis was no longer significant when a prior fracture was used. Thus, age, prior fracture, use of corticosteroids, and BMI were used in the model in the absence of BMD.

Table Table 1.. Clinical Characteristics of the Women Studied and Risk Ratio for Fracture
original image

The hazard functions for fracture and death with and without BMD are shown in Table 2. In the presence of hip BMD, BMI was no longer significantly associated with fracture. The risk of death was significantly associated with age, ever use of corticosteroids, and current time. The latter term indicates an increasing risk of death over and above that explained by age or the other risk factors. This is likely to be because of recruitment bias (exclusion of the sickest patients) and the trend with time for selected patients to approximate the characteristics of the normal population. The 10-year probability of fracture ranged from 11% to 55% (28 ± 7%).

Table Table 2.. Hazard Function of Fracture and Death With or Without Hip BMD
original image
Table Table 3.. Distribution of 10-Year Fracture Probabilities in Women Assessed With (S2) and Without (S1) BMD Measurements
original image

Three hundred fifty-four women (16.8%) were classified as high risk based on clinical risk factors, that is, a 10-year probability of fracture that exceeded 35% using the model without BMD (Fig. 1; Table 3). When BMD was additionally used, 455 women (21.5%) were identified at high risk. The avoidance of BMD measurement gave rise to errors of risk stratification. Thus, 210 women that were classified by clinical risk factors to be at low risk were found to be at high risk after the measurement of BMD (11.9% of low risk population), and 46.2% of the high-risk population was not detected by clinical risk factors alone (Fig. 1). Conversely, 109 patients classified at high risk were, based on BMD, found to be at low risk (30.8% of high-risk patients). Errors of classification were most frequent close to the threshold value chosen (see Table 3). Conversely, errors were infrequent the larger the difference between the calculated probability (S1) and the intervention threshold. For example, no women with a probability of fracture (S1) between 0% and 15% were incorrectly classified and none with a probability of more than 50% needed to be reclassified.

Figure FIG. 1..

The impact of stratification of fracture risk by the use of clinical risk factors alone and subsequent measurements of BMD.

In women characterized by risk factors without BMD, mean BMD values decreased with increasing 10-year probability (see Table 3). In women below the threshold, BMD was ∼1 SD higher than in women above the threshold (0.77 and 0.64 g/cm2, respectively). The corresponding T-scores were −1.5 and −2.5 SD.

Logistic regression models were used to identify the factors that determined the probability of reclassification. Only the estimated 10-year probability S1 was significantly related to the probability of reclassification either from high to low risk or low to high risk. The estimated probability from high to low risk was 1/{1 + exp[−(11.23 − 31.48.S1)]} and from low to high risk was 1/{1 + exp[−(11.08 + 31.66.S1)]}.

Table Table 4.. Percentage of 2113 Women ≥75 Years of Age in Whom BMD Tests Would Be Required to Classify a Fracture Risk According to the Probabilities of Misclassification Accepted
original image

The proportion of patients that require a BMD test is shown in Table 4 according to the probability of reclassification. If a very low probability of reclassification is accepted, say p1 and p2 = 0.0, 354 patients judged to be at high risk without BMD would require a BMD, and 1759 patients judged clinically to be at low risk would require a BMD measurement (i.e., the whole sample). At the other extreme, accepting a higher probability of classification (0.5), 59 individuals (2.8% of total) would require a BMD measurement. Under the assumptions given in the Materials and Methods section, a higher probability of classification was accepted for high to low risk (p1 = 0.8), whereas a low probability was accepted for low to high risk (p2 = 0.2). Under these assumptions, no individuals considered to be at high risk would require a BMD because the probability of reclassification was consistently <0.8. In contrast, 452 women classified initially at low risk would require a BMD test, representing 21% of the population (Fig. 2). This strategy implied that 13% (59 of 455) high-risk women were not detected, and the proportion of reclassified women of the whole population was 8% (59 + 109 of 2113). The requirements for BMD testing for other permutations of p1 and p2 are given in Table 4.

Figure FIG. 2..

Effect of the selective use of BMD tests in the cohort where the probability of reclassification with BMD from high risk to low risk is set at p1 = 0.8 and from low to high risk at p2 = 0.2. False positives and false negatives refer to individuals who would have been reassigned to low- or high-risk groups, respectively, had a BMD test been done.

The clinical characteristics of women allocated to high and low risk using the base case are shown in Table 5. As expected, there were highly significant differences in the risk indicators, as well as in the fracture outcomes. The T-score for BMD at the hip was −2.96 SD for the high-risk group and −1.79 SD for the low-risk group.

Table Table 5.. Clinical Characteristics of Women Identified at High or Low Risk Using Clinical Risk Factors and the Selective Use of BMD Tests (Base Case)
original image

Changing the intervention threshold such that 10% or 50% of the population would be selected would require that BMD tests be undertaken in 19% and 12% of the population, respectively (data not shown).

In the base case, the use of clinical risk factors alone gave a gradient of fracture risk of 1.42/SD change in risk score. For BMD alone, the gradient of risk was 1.56/SD, and for BMD in all patients combined with the risk score, the risk gradient was 1.64/SD.

If all patients are identified by risk factors and BMD assessed in the whole population (i.e., sensitivity = 100%), 54% of patients are identified by the use of risk factors (245/455; see Fig. 1) and 87% with the addition of the selective use of BMD tests [(2113 − 1549 − 109 − 59)/(2113 − 1549 − 109); see Fig. 2].

DISCUSSION

A primary objective of a case-finding strategy in osteoporosis is to identify individuals above or below a threshold fracture risk. In this study, we have used fracture probability as an index of risk and used a threshold risk of 35% 10-year fracture probability in the base case. This threshold identifies ∼21.5% of the population to be above the threshold risk. The threshold we use is, however, arbitrary and could be set at a number of different levels depending on health economic or public health criteria.

Irrespective of the threshold used, this study indicates that patients can be categorized by fracture probability according to clinical risk factors with or without the use of BMD. As expected,(17) the range of risk identified was greater with the addition of BMD. Using clinical risk factors alone, 10-year probabilities range from 11% to 55%, whereas, with the addition of BMD measurements, the range was 5-70%. It is evident, therefore, that there is not a perfect concordance between fracture probability assessed by clinical risk factors alone and that assessed by the additional use of BMD measurements. For this reason, clinical risk factors alone cannot be used to predict BMD accurately,(23) and the avoidance of BMD testing is confined to a minority.

In contrast, this study shows that a substantial majority of individuals can be accurately characterized to be above or below a threshold value of risk. The performance (sensitivity) of the case-finding algorithm with risk factors alone is 54%, but it increases to 87% with selective use of BMD. If errors are judged according to threshold values (i.e., patients characterized at high or low risk by risk factors alone who subsequently are characterized at low or high risk, respectively, with the addition of BMD), errors are confined to a minority (8%) of the total population. Of particular interest is that errors of misclassification were greatest around the intervention threshold chosen. For an intervention threshold of 35% probability, the vast majority of classification errors occurred in women in whom clinical risk factors suggested a fracture probability that lay between 30% and 40%. Indeed, individuals with probabilities between 30% and 40% accounted for 82.5% of classification errors (see Table 3). In a separate regression analysis that determined 10-year probability assessed with the use of BMD, the only risk factor of significance was the 10 year probability assessed without the use of BMD. Thus, classification errors were rare where the use of clinical risk factors alone identified very high or very low fracture probabilities. This indicates an important principle, namely that not all patients require a BMD test to assess fracture risk.

The question arises as to what probabilities of reclassification should be accepted by clinicians. In this paper, we have modeled all probabilities (see Table 4). We have, however, suggested that different probabilities might be taken for reclassification errors from high to low risk and vice versa. If a patient is classified at high risk by a clinical risk factor profile alone, but reclassified at low risk with a BMD measurement, we have suggested that a probability of reclassification could be set at a high value (p1 = 0.8). This is because, even after reclassification occurs, the fracture probability is still high (in practice, no patient had such a high probability of reclassification, and this never exceeded 0.6). In contrast, in women characterized at low risk with the use of clinical risk factors, we suggest that more confidence is required to be sure that an error has not occurred and accept a lower probability of reclassification (p2 = 0.2). Under this scenario, 21.4% of the population would need a BMD test. Even accepting both probabilities (p1 and p2) be set at 0.2, only one-third of women would require a BMD test. Thus, this study provides a frame-work for an approach to case finding that is highly economic for BMD testing. It is relevant that the model that we used was for the prediction of all fractures rather than for osteoporotic or hip fractures. Modeling for fractures related to osteoporosis might be expected to improve the efficiency of case finding. Indeed, when hip fracture alone was predicted, fewer individuals required a BMD test (12%; data not shown).

The purpose of this study was to enhance a case-finding strategy. Although most countries adopt a case-finding strategy in the management of osteoporosis,(1–5) population screening with BMD is recommended in the United States for women ≥65 years of age.(5,24,25) The same principles may be applied to screening and thereby improve the cost-effectiveness of screening by decreasing the burden of BMD tests required.

In selecting patients for treatment, however, it is important to be confident that intervention would be of value. The efficacy of inhibitors of bone resorption has been well characterized in individuals with low bone mass. Their efficacy in individuals with normal bone density is less secure, and it has been suggested that efficacy is less likely.(26) Many recent studies indicate, however, that pharmacologic interventions have efficacy in patients with osteopenia or in whom BMD was not assessed.(27–33) In this study, there is a clear decrease in BMD with increasing fracture probability that was not solely dependent on age (see Table 3). This is because the risk factors identified provide a surrogate to some extent for BMD. It is of interest that very few risk factors were required, and in this particular cohort, they were confined to age, BMI, prior fracture, and the use of corticosteroids. Hip BMD was ∼1 SD lower in individuals characterized to be at high risk by clinical risk factors compared with the low-risk group.

It should be acknowledged that physicians may be reluctant to give pharmacologic interventions without objective evidence of a diagnosis of osteoporosis or other BMD threshold. In practice, BMD testing might be considered in individuals characterized at high risk, in whom a test was not required for this purpose, to monitor treatment. In such cases, it would be appropriate to measure BMD at the lumbar spine, rather than the hip, because the spinal site is most responsive to the effects of pharmacologic interventions. In other words, a test is done to provide a baseline on which to judge the effects of treatment rather than to decide who to treat. This approach is more logical than BMD testing at the hip and may still have economies. Under the scenario that we tested, 907 women would require a BMD test (43% of all women), and all patients committed to treatment would have had a BMD test at the lumbar spine.

This study is confined to a randomly selected cohort of women ≥75 years of age from the United Kingdom. The prevalence of risk factors and their importance for fracture prediction will, however, differ by age and sex, and possibly also in different regions of the world. For this reason, our data should not be applied to other settings. Rather, more cohorts, preferably drawn randomly from the population, should be examined to determine which risk factors to use and what thresholds are appropriate. In developing approaches that are applicable internationally, it would also be important to identify those risk factors that have international validity for any given age or sex.

Within these limitations, we conclude that relatively few clinical risk factors can be used to characterize three groups of individuals. The first is comprised of individuals with low risk identified from clinical risk factors alone in whom a BMD measurement would not alter their classification. The second is comprised of individuals who, based on clinical risk factors alone, have such a high risk that, irrespective of their BMD test, their risk would exceed an intervention threshold. The third identifies a group of individuals in whom a BMD test is worthwhile to more accurately characterize risk as being above or below a given threshold. The latter group is comprised of a minority of individuals over a range of threshold assumptions, indicating that case-finding strategies can be economically developed.

Acknowledgements

We thank the Alliance for Better Bone Health, GE Lunar, Hologic, Lilly, Novartis, Pfizer, Roche, Wyeth, the IOF, and the International Society for Clinical Densitometry for unrestricted grants. We also thank the Medical Research Council and Leiras who co-sponsored the MRC Hip Fracture Trial from which the data sets were derived.

Ancillary