How Important Are BMD Accuracy Errors for the Clinical Interpretation of DXA Scans?

How serious are accuracy errors in measurements of BMD for the clinical interpretation of DXA scans? Two recently published articles give radically different points of view on this important question.1, 2 In a discussion of the effect of BMD precision errors on the interpretation of T-score values, Kiebzak et al.1 stated their belief that given a properly calibrated DXA system, other types of accuracy error can be ignored. In contrast, a paper by Bolotin2 claimed that the errors caused by inhomogeneities in soft tissue composition are so large that they render all DXA scan findings totally unreliable. In this review, we discuss the size of the BMD accuracy errors and set out a middle way between these two opposite points of view. Most clinicians who interpret DXA scans are aware that BMD measurements are affected by both precision3, 4 and accuracy errors.5, 6 Precision errors measure the reproducibility of BMD results in individual patients and can be shown by performing repeated scans on a representative group of subjects.7 However, the clinical interpretation of DXA scans is also affected by accuracy errors in BMD measurements, of which the most important are the random errors that result from the variable lean and fat content of soft tissue.8-21 The cause of these errors is the inhomogeneous distribution of adipose tissue in the human body that affects both bone marrow and soft tissue external to bone in the path of the X-ray beam. Soft tissue accuracy errors are subtler than precision errors and their clinical effect are less easily appreciated, not least because the conditions necessary for a carefully conducted precision study also ensure that any scan-to-scan variation in the BMD accuracy error is minimized when repeated scans are performed on the same individual. Although the causes of soft tissue accuracy errors have been discussed for at least 40 yr,22 it is clear from the articles by Kiebzak et al.1 and Bolotin2 that their significance for scan interpretation remains a source of confusion. DXA scanners work by measuring the transmission of X-rays through the human body at two different photon energies.23 Because there are two attenuation processes (Compton scattering and the photoelectric effect) with different dependences on photon energy and atomic number, the two transmission measurements can be used to compute the areal densities (units: g/cm2) of two representative materials. In practice, the two materials chosen are bone mineral (hydroxyapatite) and soft tissue. However, the computed densities will only accurately reflect the true densities if the human body is made up solely of the two chosen materials. In practice, the body is composed of bone, lean tissue, and adipose tissue. The thickness of the human body through the lumbar spine is typically 15–20 cm for individuals of normal body weight and 25 cm or greater in obese subjects. Because of the large thickness of soft tissue, its areal density is much greater than that of bone mineral (in the range of 15–25 g/cm2 compared with a typical BMD figure of 1 g/cm2). For this reason, even the relatively small differences in X-ray attenuation between lean and adipose tissue13, 24 can create clinically significant measurement errors in the BMD results calculated by the DXA computer system. In practice, the soft tissue errors are minimized by comparing measurements over the bone region of interest (ROI) with those in an adjacent reference area of soft tissue. However, errors still arise because of differences in the percentage of adipose tissue between the bone ROI and the soft tissue reference area5, 16, 25 and variations in the composition of bone marrow.13, 26 These errors consist of a constant term (an average error obtained by taking the mean over many subjects) and a random term that varies from individual to individual. The average error has no effect on scan interpretation because its presence is allowed for through the use of an appropriate reference range. However, the random component of the error is important because it causes the apparent T-score value to misrepresent the patient's true bone status. Although the random error varies from person to person, in any individual, it always has the same value provided the same bone and soft tissue ROIs are used and there is no other relevant factor such as a change in body weight. It is the random component of the soft tissue accuracy errors that concerns us here. In situ measurements obtained by scanning the cadaver before the bones have been excised. In situ measurements are equivalent to clinical scans in terms of the effect of soft tissue accuracy errors on the BMD results. In vitro measurements obtained by scanning excised bones without defatting in a water bath, usually with a water depth of 16 cm. By scanning the bones in water, the errors caused by inhomogeneities in soft tissue composition external to bone are eliminated, but errors caused by bone marrow composition remain. True values determined either by bone ashing or by cleaning and defatting of excised bone to remove all marrow. When interpreting cadaver studies, care is necessary in distinguishing which types of measurements are being compared. The most relevant studies for assessing the overall effect of soft tissue accuracy errors on the clinical interpretation of DXA results are those that compare in situ measurements with true values (Fig. 1A).5, 9, 10 Other studies that present data on the comparison between in situ and in vitro measurements (Fig. 1B) give information on the effect of the inhomogeneities in soft tissue composition external to bone on the overall error.5, 8-10 Finally, studies that present data on the comparison between in vitro and true values (Fig. 1C) give information on the effect of the variations in bone marrow composition on the overall error.5, 6, 8-12 Summary of soft tissue accuracy errors in BMD (filled symbols) and BMC (open symbols) measurements reported in cadaver studies. The published measurements have been separated into errors between (A) in situ and true values; (B) in situ and in vitro values; (C) in vitro and true values. Error bars shown the 95% CI estimated from the number of independent measurements (degrees of freedom) using the χ2 distribution. The cadaver study providing the most comprehensive and relevant data for the clinical interpretation of spine and hip DXA scans is a study published by Svendsen et al.5 The authors performed both in situ and in vitro measurements, and true values were determined after cleaning and defatting excised bone to remove all marrow. The SE of the estimate (SEE) for the comparison between in situ and true values was 5.3% for posteroanterior (PA) spine BMD (L2–L4), 6.5% for total hip BMD, and 6.7% for femoral neck BMD (Table 1). With only 14 subjects, the statistical errors in the study of Svendsen et al. are quite large (for the 5.3% error for in situ versus true spine BMD, the 95% CI is 3.8–8.7%). It is therefore important to examine other studies for supporting evidence. These generally report accuracy errors in the range 5–10% (Fig. 1). Unfortunately, none of the other studies performed using DXA systems included in situ measurements and instead reported on the accuracy error between in vitro and true BMC values. Sabin et al.12 measured spine BMC (L2–L4) in 12 subjects and found an error of 4.9%. Ho et al.11 studied individual lumbar vertebrae in 11 subjects and reported an error of 8.9%. There is even less data for the hip. However, Kuiper et al.6 compared in vitro and true values of femoral neck BMC in 19 subjects and found an error of 5.2%. After the DXA studies, the next most relevant study is a study of accuracy errors in DPA measurements reported by Gotfredsen et al.10 In a study of 13 cadavers, they found BMD accuracy errors somewhat larger than those reported by Svendsen et al. (10.3% versus 5.3% for in situ versus true PA spine BMD), although the BMC errors for the two studies were quite similar. In part the larger BMD errors found by Gotfredsen et al. may reflect the poorer imaging properties of DPA systems compared with DXA. Schematic plot of the NHANES III total hip reference range35 shown as a plot of mean T-score ± 2 SD against age. (A) Cross showing a patient's result as plotted on a typical DXA scan report. (B) The same T-score result drawn with an error bar indicating the 95% CI of ±0.2 in the T-score assessment arising from BMD precision errors.1 (C) The same result drawn with an error bar indicating the 95% CI of ±1.0 in the T-score assessment arising when accuracy errors5 and precision errors1 are combined. Many clinicians will find the appearance of Fig. 2 surprising and uncomfortable. Although the effect of soft tissue accuracy errors on BMD measurements has been discussed for nearly 40 yr,22 many people are unaware of their importance for the interpretation of DXA results. In part this is because the soft tissue effects are so difficult to detect and measure. As explained above, they do not affect precision studies because the latter are based on the careful reproduction of both the soft tissue and bone ROIs, so the accuracy error in the BMD results for a particular individual at a given measurement site is unchanged on repeated scans. Even when results from different manufacturers' equipment are compared, similar regions of interest are studied and the accuracy errors on different systems are strongly correlated. Soft tissue errors may have a number of other observable effects on BMD measurements. Accuracy errors are one of a number of factors that contribute to the relatively poor correlation between measurements made at different skeletal sites, for example, between spine, hip, forearm, and heel BMD.28 This effect was discussed by Hassager et al.,29 who estimated that the observed correlation coefficient of r = 0.6 between spine and forearm measurements would improve to r = 0.8–0.9 if soft tissue errors were eliminated. However, this estimate was based on a spine BMD error of 10%,10 which is considerably larger than the figure in Table 1, suggesting that the effect of accuracy errors on the correlation coefficient might have been overestimated. Other factors that contribute to the poor correlation between different BMD sites include anatomical variability in bone size and shape and the effect of disease processes such as osteoarthritis of the spine. However, the relative importance of the different factors remains unclear. Soft tissue errors will also cause the precision error to be worse in situations where the accuracy error in an individual changes when the scan is repeated. Examples include the effect of changes in the distribution of adipose tissue during weight change. In addition, in individuals who are seriously overweight the variation in the disposition of adipose tissue when patients are repositioned on the scanning table can be expected to affect precision. There are few studies of these effects, but there is some evidence that both weight change and being overweight lead to poorer precision.30 At first sight, the suggestion that DXA T-score measurements have a 95% CI of ±1 unit is disturbing. It means that many patients found to have osteopenia could actually be osteoporotic or normal. However, we suggest that in practice these uncertainties have less of an impact on the clinical use of DXA scanning than one might at first think. The basic rationale of BMD examinations is to discriminate between patients at high risk of fracture and those at low risk. Data from fracture studies are fitted using a gradient-of-risk model in which the fracture risk P varies exponentially with Z-score with gradient β [P = P0 exp(−βZ)].31, 32 Results are usually expressed in terms of the relative risk (RR), the increased risk of fracture for each 1 SD decrease in BMD. The value of RR is found from β using the exponential function [RR = exp(β)]. Alternatively, the gradient β is found by taking the natural logarithm of RR [β = ln(RR)]. The larger the value of RR, the wider the separation of the BMD distribution curve representing the fracture population from the curve representing the study population at baseline (Fig. 3A). To understand the reason for this, we can assume that the BMD distribution of women referred for a DXA examination follows a bell-shaped curve (Fig. 3A, curve labeled normalized baseline distribution). The distribution of BMD values for the group of patients who will at some future date experience an osteoporotic fracture is found by multiplying the distribution curve of the baseline population by the gradient-of risk curve. When this is done, the BMD distribution curve representing the fracture population is found to be a second bell-shaped curve similar to the first but with its peak offset to lower BMD values (Fig. 3A, curve labeled normalized fracture distribution). When BMD values are expressed in Z-score units the two curves are offset by an amount equal to the gradient-of-risk β [ΔZ = β = ln(RR)] (Fig. 3A).33, 34 The larger the value of RR, the wider the separation of the two distribution curves (Fig. 3A, inset), and the more successful clinicians are at distinguishing those women at high risk of fracture from those at low risk. The clinical effectiveness of DXA examinations can be shown graphically by using the two BMD distributions in Fig. 3A to plot a receiver operating characteristic (ROC) curve in which the percentage of true-positive cases is plotted against the percentage of false-positive cases as the BMD threshold is varied (Fig. 3B).33, 34 (A) Schematic plot showing the bell-shaped BMD distributions (solid curves) of a population of patients referred for a DXA examination (baseline distribution) and the population of subsequent fracture cases (fracture distribution). For clarity, both curves have been normalized to the same peak value. The larger the gradient-of-risk β, the wider the separation of the two curves, and the more reliably clinicians can distinguish patients at high risk of fracture. To a first approximation, the separation of the two curves in Z-score units is given by the value of β.33, 34 The inset table lists values of relative risk (RR) and β. The dashed curves show the narrower BMD distribution curves that would be found if soft tissue accuracy errors were eliminated. (B) The curves of the fracture and baseline populations in A can be used to draw the ROC curve of true positive cases (patients thought to be at risk who go on to have a fracture) against false-positive cases (patients thought to be at risk who do not have a fracture). If soft tissue accuracy errors could be eliminated, the area under the ROC curve would increase (dashed curve). The effect of soft tissue accuracy errors on the clinical use of DXA scanning is best understood in terms of their effect on the ROC curve shown in Fig. 3B. If accuracy errors could be eliminated, the width of the BMD distribution curves would be narrower (Fig. 3A, dashed curves), resulting in more effective discrimination of fracture cases and a more favourable ROC curve (Fig. 3B). The data in Table 1 can be used to estimate the effect of soft tissue errors on RR values. The SD of the true biological BMD distribution is obtained by taking away the square of the soft tissue error from the square of the observed population SD and taking the square root. It follows given the 95% CI of ±1 T-score unit shown in Table 1 that the true population SD is √(1–0.52) = 87% of the observed SD. The broadening of the SD by measurement errors means that the observed value of the gradient-of-risk β is 87% of the value that would be found if accuracy errors were eliminated. It follows, for example, that an RR figure of 2.0 (equivalent to β = 0.69) found in a fracture study would have a value of RR = 2.2 (equivalent to β = 0.69/0.87 = 0.79) if the accuracy errors were eliminated. The above considerations suggest that the BMD measurement errors generated by inhomogeneities in soft tissue composition have less of an impact on the clinical use of DXA scanning than Fig. 2 might suggest. If these errors did not exist, BMD distribution curves would be narrower and fracture data would be fitted by gradient-of-risk models with higher values of RR. Although the existence of soft tissue errors causes a quantitative deterioration in the clinical performance of DXA examinations as measured by the area under the ROC curve, qualitatively there is remarkably little effect on the clinical use of DXA scanning. Clinicians treating patients for osteoporosis are always working with probabilities, because when fracture studies are plotted as ROC curves, it is clear that no bone densitometry technique gives perfect discrimination between fracture and nonfracture cases. Even if soft tissue errors could be eliminated, ROC curves would still fall short of the ideal case of 100% identification of true positive cases with no false positives (Fig. 3B). The practical effect of BMD measurement errors on scan interpretation is to make the clinical performance of DXA systems for predicting fracture risk marginally worse than would otherwise be the case.