The aim of our study was to validate a hip fracture risk function, composed of age and femoral neck bone mineral density (BMD). This estimate of the 1-year cumulative risk was previously developed on the basis of Dutch hip fracture incidence data and BMD in men and women. A cohort of 7046 persons (2778 men) aged 55 years and over was followed for an average of 3.8 years. The 1-year hip fracture risk estimate was calculated for each participant according to the risk function and categorized as low (<0.1%), moderate (0.1 to <1%), or high (≥1%). Observed first hip fracture incidence was then analyzed for each of these risk categories by age and gender. Additionally, we calculated the relative risk per standard deviation (SD) decrease in femoral neck BMD in this population. At baseline, 2360 individuals were categorized as low risk, 2567 as moderate risk, and 378 as high risk. During follow-up, 110 first hip fractures were observed corresponding to an incidence rate of 4.1/1000 person-years (pyrs) (95% confidence interval 3.4–5.0). The observed incidence rate in the low risk group was 0.2/1000 pyrs (0.1–0.9), 2.7/1000 pyrs (1.8–3.9) in the moderate risk group, and 18.4/1000 pyrs (12.4–27.2) in the high risk group. Below the age of 70 years, incidence was low in all categories, and very few individuals were considered at high risk. Above the age of 70 years, the observed incidence was high in the high risk group, while in the low and moderate risk groups, the incidence remained low even over 80 years of age. In women, the age-adjusted relative risk for hip fractures was 2.5 per SD decrease in femoral neck BMD (1.8–3.6), while in men this relative risk was 3.0 per SD (1.7–5.4). In conclusion, we observed a similar relation of hip fracture with femoral neck BMD in men and women and were able to predict accurately hip fracture rates over a period of almost 4 years.
IN WESTERN SOCIETIES, hip fractures cause major morbidity and mortality in the elderly.1 Consequently, they generate substantial costs due to acute hospital treatment and subsequent rehabilitation.2,3 Improved life expectancy and the demographic evolution will cause the number of hip fractures worldwide to increase from about 1.7 million in 1990 to over 6 million in 2050.1 To target prevention at those with the highest risk, it is important to be able to predict hip fractures.
Although the immediate cause of a hip fracture is mostly a fall, their occurrence is closely associated with osteoporosis.4 Osteoporosis is defined as a condition characterized by low bone mass and microarchitectural deterioration of bone tissue, with a consequent increase in bone fragility and susceptibility to fracture.5 The conventional method of estimating bone mass is by bone densitometry. The World Health Organization defines osteoporosis as a bone mineral density (BMD) of 2.5 standard deviation (SD) below the mean for young adults.6
It is, however, important to distinguish between BMD as a tool for defining osteoporosis and BMD measurement as a tool for fracture risk prediction.7 The association of BMD with subsequent hip fracture has been demonstrated in several studies with an estimated relative risk, for women, of 2.6 per SD decrease in femoral neck BMD.4,8 With this relative risk estimate, the registered hip fracture incidence in The Netherlands, and the BMD distribution from a Dutch population-based sample, we derived a theoretical 1-year hip fracture risk function by age and femoral neck BMD for men and women.9 In the present study, we validated this tool for predicting hip fracture incidence rates in a prospective follow-up study. Additionally, we calculated the risk of hip fracture per SD decrease in femoral neck BMD for both men and women.
MATERIALS AND METHODS
We previously estimated the 1-year cumulative risk of hip fracture by age and BMD, based on Dutch hip fracture incidence data, the distribution of BMD in a sample of 5814 men and women aged 55 years and over, and on data from the literature.9 In the present study, we validated this risk estimate in the Rotterdam Study, a population-based prospective cohort study of the occurrence and determinants of disease and disability in the elderly. The Rotterdam Study focuses on neurologic, cardiovascular, locomotor, and ophthalmologic diseases. The study started in 1990 and all 10,275 men and women aged 55 and over, living in Ommoord, a district of Rotterdam, were invited to participate. The study was approved by the Medical Ethics Committee of Erasmus University Medical School, and participants provided written informed consent. From those eligible for participation, 7983 did participate, bringing the overall response rate of this study to 78%. The design of this study has been described previously.10
The baseline survey included an initial home interview followed by two visits to the research center for a series of clinical examinations and laboratory assessments. Baseline assessments included dual-energy X-ray absorptiometry scans of the femoral neck in the independently living participants, using a Lunar DPX-L densitometer (Lunar Corp., Madison, WI, U.S.A.). Measurement procedures have been reported previously.11
For each participant, we calculated the 1-year cumulative risk according to the risk function, as an individual risk estimate. The equations used for this purpose are given in the appendix. Participants with an incident hip fracture prior to BMD measurement were excluded from the risk classification. Based on the same risk functions, a nomogram indicating the 1-year hip fracture risk by age and BMD was constructed for men and women. To avoid that these nomograms would only be useful to users of Lunar machines, we also converted those risk functions to other brands of densitometers. This conversion was done using the conversion algorithms provided by Genant et al.,12 and the resulting equations can also be found in the appendix.
Hip fracture follow-up
Follow-up for hip fractures was achieved through a link with the computer systems of the general practitioners of the district and on hospital admission data, covering about 80% of the study population. For all participants not covered by this system, annual checks were performed on the complete medical records of their general practitioners. Reported fractures were verified by retrieval and review of the appropriate discharge reports from the patient record. Follow-up started either at January 1, 1991 or at the time of the baseline interview when this was later. Follow-up ended either at death, at the time of the first hip fracture, or February 29, 1996.
To assess the performance of the risk function, we divided the population with a valid BMD measurement into categories by gender and risk. The individual 1-year risk estimate was categorized as either low (<0.1%), moderate (0.1 to <1%) or high (≥1%). These cut-off levels were arbitrary but based on clinical common sense, a 1-year risk of 0.1% approximately corresponds to the risk for an average women at the age of 60,9 while a 1% risk approximately corresponds to the risk for an average 80-year-old women. Incidence rates of first hip fracture were then calculated for each of these categories. To account for the events occurring in participants without a BMD measurement, hip fracture incidence was calculated separately for those living in institutions for residential care and for community-dwelling persons without BMD measurement.
To assess the effect of age, we subdivided these risk categories into 10-year age groups. To avoid categories without outcome events, we combined for this analysis by age the low and moderate risk categories into a single category. For categories where follow-up time was <100 person-years (pyrs), we did not to calculate incidence rates since those results were considered too unreliable.
Finally, we calculated the age-adjusted relative risk per SD decrease in BMD using Cox's proportional hazards model. All point estimates are presented with their 95% confidence intervals.
Baseline risk classification
Figures 1 and 2 represent, as a nomogram, the 1-year hip fracture risk by age and BMD, calculated from the risk functions, for men and women, respectively. The shaded areas depict our Dutch reference BMD distribution, while the curves connect points of equal risk at all ages. The curves are very similar in men and women, as we described before,9 while the average BMD is higher in men than in women at every given age, and the decrease of BMD with age is slower in men.
Complete follow-up was achieved for 7046 persons (2778 men, 4268 women) with an average follow-up time of 3.8 years. In 5305 of those (2227 men, 3078 women), we also had a valid BMD measurement at baseline, and the BMD distribution by age is shown in Table 1. Based on bone density, age, and gender, we classified everybody at baseline in risk categories according to the risk functions: 2360 individuals were categorized as low risk, 2567 as moderate risk, and 378 as high risk. For 1741 individuals, no baseline BMD measurement was available: 614 lived in residential care and 1127 lived independently. From the independently living, 487 did not come to the study center, the others had no BMD measurement because of technical reasons, mostly caused by machine downtime.
Table TABLE 1. AVERAGE BMD AT BASELINE (SD) BY AGE AND GENDER
Observed hip fracture incidence
Follow-up totaled 26,771 pyrs (10,333 for men, 16438 for women), and 110 first hip fractures were observed, 23 in men and 87 in women. The observed hip fracture incidence rate for the whole population was 4.1/1000 pyrs (3.4–5.0). The observed incidence in the low risk group (two fractures) was 0.22/1000 pyrs (0.05–0.87). In the moderate risk group (27 fractures), the incidence was 2.7/1000 pyrs (1.8–3.9), and in the high risk group (25 fractures), the observed incidence was 18.4/1000 pyrs (12.4–27.2). In the group without BMD measurement, incidence in residential care (36 fractures) was 19.7/1000 pyrs (14.2–27.4). In the independently living participants without BMD measurement (20 fractures), the overall incidence was 4.5/1000 pyrs (2.9–7.0). Figure 3 presents these observed hip fracture incidences for men and women separately.
Table 2 gives an overview of the baseline classification according to the risk function, by gender and 10-year age groups, and the number of hip fractures that occurred during follow-up. As expected from the nomograms, the high risk group is more prominent at older ages and more important in women than in men. At ages over 80 years, almost nobody is still in the low risk group, but an important proportion of the independently living population can be considered, a priori, at only moderate risk, especially men. Table 3 lists the observed incidence rates for all categories, and the point estimates confirm the a priori risk classification. However, in this analysis by age, the confidence intervals are much wider, reflecting the smaller number of events and follow-up time per category.
Table TABLE 2. RISK CLASSIFICATION AT BASELINE BY AGE AND GENDER (NUMBER OF OBSERVED HIP FRACTURES)
Table TABLE 3. OBSERVED INCIDENCE RATES/1000 PYRS BY AGE AND GENDER (95% CI)
The observed age-adjusted relative risk for hip fractures was similar in men and women. The relative risk was 2.5 for each SD decrease in femoral neck BMD (1.8–3.6) in women. In men, this relative risk was 3.0 (1.7–5.4). These relative risks were not statistically different (p = 0.65).
In the community-dwelling individuals, the high risk group consisted mainly of individuals aged 70 years and older, predominantly women. The observed incidence rates and their precision demonstrate that the risk function accurately predicted hip fracture incidence in the various risk subgroups in men and women. The low and moderate risk groups taken together identified a large proportion of the study group with low hip fracture incidence, even at ages over 80 years. However, a smaller group of individuals with high hip fracture incidence, starting at age 70, could be identified.
Participants living in residential care institutions had a hip fracture risk similar to the highest risk category. This is in agreement with previous findings of high hip fracture incidence rates in nursing homes and institutions for residential care.13,14 Both in residential care and in the high risk category, the observed incidence was slightly lower in men. This was probably due to the different age distribution, with men being, on average, younger within each group.
The relative risk for hip fracture per SD decrease in BMD observed in this study confirms previous estimates. While for women this estimate was based on large studies,4,8 the estimate for men was only based on a small sample.15 For the derivation of the 1-year cumulative risk function, we assumed that the relative risk was equal in men and women. Our present findings confirm the validity of this assumption.
The risk evaluation in this study was based on a 1-year risk estimate and, although follow-up amounted to almost 4 years, prediction on a longer term is important for therapeutic decisions. Prediction over a longer time period is, however, dependent on assumptions about the BMD evolution. If we assume linear decline, as we found in our cross-sectional study, we can estimate long-term risk by extrapolation of the current BMD level on the nomogram. There are, however, indications for a more rapid BMD decline at older ages,16,17 and, if so, future risk would be underestimated. Additionally, it is unclear how well an individual BMD measurement correlates with bone density in the future, and long-term risk prediction is dependent on assuming a high correlation.18 Follow-up studies over a longer period are needed to answer this question more precisely.
The risk function was developed based on Dutch data, and its applicability in other populations might be limited if either the hip fracture incidence or the bone density distribution were different. But, comparing Dutch hip fracture incidence data with recent international data,19,20 they appeared remarkably similar to incidences in Sweden, Scotland, and Switzerland, and slightly lower than in U.S. Caucasians. There have also been indications that BMD distributions might be different in several European countries, but those estimates were based on small samples.21,22 In this large sample, however, the bone density distribution was almost identical to that in U.S. women,23 after correction for scanner type.12 This suggests that the application of the risk function is not limited to a Dutch population, although additional validation in other populations remains desirable.
Even while the number of hip fractures in men was small (23 fractures), we believe this is an important first attempt to validate risk estimates in a male population. We are not aware of any comparable study where an a priori risk was validated in a follow-up design in males, but especially here, additional validation remains necessary.
To facilitate the usability by users of other brands of densitometers, we have provided converted equations in the appendix, while the converted nomograms are available from the authors. These conversions, obviously, do not change the risk classification itself, but only affect the absolute levels of BMD, and the associated risk curves.
Although population based, there was self-selection in the participation to the study. Even with the high response rates, this might cause selection bias toward more healthy individuals. There might also be concern about potential misclassification. Since the cases that were reported were well documented, this misclassification would most likely result in an underreporting of fractures. Both phenomena would lead to a lower hip fracture incidence than expected from Dutch overall hip fracture incidence rates. However, a comparison of our results with national incidence rates only indicates a slightly lower incidence than expected: 110 hip fractures versus 132 expected. This is equivalent to 83% (69–100%) of expected hip fractures, suggesting that those effects were probably small. It is, moreover, unlikely that this selection would influence the validity of the risk function.
BMD was not measured in some independently living participants because they did not come to the study center, and this could, in theory, again lead to selection bias toward more healthy individuals. When BMD was not measured for technical reasons, selection bias seems unlikely, and this applied to the majority of participants. In any case, the observed hip fracture rate in this group was very similar to the overall incidence for the same gender.
The validation was done in broad risk and age categories, which was necessary to acquire enough follow-up time to obtain stable results. In Figs. 1 and 2, we have indicated 1-year risk levels up to 10%, corresponding to the clinically relevant observations. In this study, there were only three women exceeding this 10% risk, and two of them suffered a hip fracture during follow-up. Although this would correspond to an incidence rate of 203/1000 pyrs (51–813), we believe this result should not be overinterpreted. We analyzed the effect of age by 10-year age categories, and although we are aware that the baseline risk is very different within those categories, using smaller age categories would have frustrated any validation effort.
For the original derivation of the risk functions, we used the baseline BMD data from the same reference group as where the risk function was validated afterward. Since the hip fractures in these studies were observed afterward and prospectively, we do not perceive this as a problem. An alternative would have been to use reference data from the manufacturer, but since those data were not based on Dutch data, we preferred to use our own.
Finally, the risk function generated a 1-year cumulative risk, while the outcome measure of this study was incidence rates. With the expected and observed incidence rates, the difference between those two measures is negligible.
We conclude that hip fracture rates can be predicted accurately, from age and BMD, in both men and women. In prevention programs, we need a tool for risk stratification of the population. This follow-up study showed that our risk function is a valid instrument for that purpose. The majority of hip fractures (61 out of 110) occurred either in the high risk group or in the residential care group, even though these groups accounted for barely 14% of the study population. In addition, we found a similar relation of femoral neck BMD with hip fracture in men and women.
We are very grateful to the participants of the Rotterdam Study, the general practitioners, and the many field workers in the research center in Ommoord. We also thank the dual-energy X-ray absorptiometer technicians, L. Buist and H.W.M. Mathot. This study is part of the research program of the Erasmus Center for Research on Aging of the Erasmus University, Rotterdam and the University Hospital, Rotterdam, Dijkzigt, The Netherlands.
Although both the derivation of the risk function by age and femoral neck bone density and the reference data were described previously,9 we felt it necessary to include them here. Additionally, and to expand the usability, we converted the risk functions to other brands of bone density equipment.
In the Rotterdam Study, we observed a linear decline of bone density with age in men and women. The average bone density at all ages, measured with a Lunar DPX-L densitometer, was given by:
BMD was normally distributed at all ages with an overall SD of 0.134 in women and 0.135 in men.
Using the conversion algorithms of Genant et al.12 we converted these regression formulas to:
Table TABLE 4. RISK FUNCTION VALUES FOR LUNAR DPX-L, HOLOGIC QDR-2000 AND NORLAND XR26MARK II DENSITOMETERS
Using this conversion on the BMD data from women in our cohort, the relation of BMD with age became almost identical to the NHANES III data.23 Unfortunately, no such widely accepted reference data are available for men.
For the derivation of the risk function, we used Dutch nationwide hip fracture registration to estimate the hip fracture incidence function by age and gender, and additionally, we used the assumption of a 2.6 relative risk per SD lower femoral neck bone density. Under those assumptions,9 the 1-year cumulative risk is given by: