We assessed the validity of QUS as a screening method for osteoporosis in a Taiwanese population. We found a significant correlation between QUS and DXA T-scores in both the hip and spine (p < 0.05). The correlation coefficient between QUS and DXA-hip was 0.174, higher than that between QUS and DXA-spine (0.133). This may originate from the fact that the calcaneus and femoral neck belong to the lower limbs, sharing similar bony architectures. However, the DXA-spine remains the first-choice diagnostic method for osteoporosis in clinical practice. Therefore, once osteoporosis is suspected based on calcaneus QUS, further DXA examination is required to confirm the diagnosis of osteoporosis.
The correlations between QUS and DXA also differed within each sex. For example, in females, it was 0.298 for DXA-hip but 0.237 for DXA-spine; in males, the values were 0.216 and 0.255, respectively. This difference arises from several reasons, one of them is skin contact surface with the ultrasound transducer over the calcaneus region. The subcutaneous fat component is more abundant in females, providing a more consistent medium for conducting the ultrasound wave. In contrast, subcutaneous fat levels are lower in males, and more ultrasound energy is dissipated at various tissue junctions, resulting in inconsistent QUS measurements. In this study, about 31 outlier cases were identified, all with abnormally high SOS values. Males made up 23 of these outliers (about 67.5%). This phenomenon also can be explained by skin contact surface heterogeneity of the calcaneus region between both sexes.
The physical basis of QUS is measured by two ultrasound parameters on the calcaneus region, i.e., the SOS and BUA. The SOS and BUA are combined to calculate the T-score of QUS, as predefined by the manufacturer. The histogram data for QUS, SOS, and BUA are shown in Figure 1. The data are similar. However, the histogram data for DXA-hip and DXA-spine show a right-skewed (log-normal) distribution (Fig. 2 )[18]. This may have caused the discrepancies between QUS and DXA.
Many biological variables obey the law of normal distribution. On mathematical basis, a normal distribution is the limit resulting from a large binomial process. However, not all clinically relevant data are described by the normal distribution. For example, blood pressure data are reported as lognormal distribution. The right-skewed distribution is often transformed by natural logarithm function for subsequent processing. In this study, the lognormal distribution nature of bone mineral density is interesting and worth further studies.
Figure 3 shows the results of regression analyses, including scatter plots of QUS and DXA-Hip and DXA-spine data. The plots are concentrated around the -2.5 region, with fewer cases at QUS > -1.0. The reason for this is due to the inclusion criteria during the annual health examination; only subjects with QUS -2.0 were chosen for further DXA examination. This results in a highly concentrated data distribution; therefore, we adopted some cases with higher QUS values to study the regression line.
Figure 4 shows the sensitivity and specificity of using only QUS as a predictive variable for osteoporosis, according to ROC analyses. The area under curve (AUC) was 0.55 not far from the reference line which is AUC of 0.5. To increase discriminative power, we ran a multivariate logistic regression. The predictive variables included age, sex, body weight, height, BMI, BUS, and SOS. The logistic regression model had a sensitivity of 67.2% and a specificity of 64.9%, with an overall accuracy of 66.2%. Female had a 4.4-fold higher odds ratio than males with osteoporosis. Another significant predictive variable was QUS (where a lower value = osteoporosis more likely). Due to the improved performance of this more sophisticated logistic regression, the probability predicted by this model can be used as a new test variable for ROC analyses. These results suggest that using a more sophisticated model will increase the discrimination power for osteoporosis. The weighting of individual factors can also be examined based on the results of the logistic model described in Table 2. However, the use of a multivariate model requires a scoring system such as the Glasgow Coma Scale, and will take more time to perform.
The WHO defined the status of osteoporosis as a DXA T-score < -2.5. In this study, the ground truth of the examined osteoporotic subjects was established using the DXA T-score of either the spine or hip. The lumbar vertebrae T-score was recorded as the average value for the L1–L5 lumbar segments. This process automatically omits segments that are abnormal, such as those with a compression fracture. On the other hand, the DXA-hip value adopts only the femoral neck area as the sampling region. This is compatible with criteria used in clinical situations.
To determine the optimal cutoff value for QUS data, Youden’s J statistic was adopted. The Youden Index seeks the maximum value in an ROC curve (specificity + sensitivity – 1). As shown in Figure 6, the maximum value corresponded to a QUS of 2.725. Therefore, when the QUS is used alone as a predictive variable of osteoporosis, -2.725 is automatically the optimal threshold for defining disease status. This value is similar to the criteria established by the WHO when judging DXA results
There were two main limitations of this study. First, during the annual health examination for elderly patients, we did not collect data on subjects younger than 65 years old. Second, for cases that did not meet the inclusion criteria (age ≥ 65 years old and calcaneal QUS ≤ -2.0), the DXA scan was performed only for a small portion of patients. Therefore, we lack data on patients with higher QUS values.