INTRODUCTION
Osteoporosis is a silent chronic disease that is manifested by low bone mass and reduced structural integrity at the microarchitectural level, resulting in increased bone fragility and fracture risk with age.[1] By 2050, more than 50% of hip fractures (HFs) are expected to occur in Asia.[2] HFs are associated with high costs and mortality.[3,4] In Asia alone, the number of HFs is estimated to rise from 1.12 million to 2.56 million between 2018 and 2050, leading to an increase in the annual direct medical cost of treating HFs from USD 9.5 billion to USD 15 billion.[ 2] These fractures have a combined lifetime risk in terms of needing clinical attention equivalent to that of cardiovascular disease.[5] Osteoporotic fractures result in pain and disability, causing individuals to often lose their capacity for leading active, productive, and independent lives.[6] However, there remains an insufficient emphasis on managing the burden of osteoporotic fractures in Asia, leading to underdiagnosis and consequently undertreatment of this disease in the region.[6]
The main goal in managing osteoporosis is to prevent fractures, which necessitates identifying individuals at elevated risk. Bone mineral density (BMD) obtained from dual energy X-ray absorptiometry (DXA) imaging is the current gold standard for doing so. However, BMD alone has been deemed to lack sensitivity in terms of predicting future fractures.[7] Furthermore, DXA imaging is not widely accessible in many Asian countries.[6] This is especially critical in countries like China, India, Indonesia, Pakistan, Sri Lanka, Thailand, and Vietnam, where majority of the population live in rural areas.[6] This has led to increased interest in the use of the fracture risk assessment tool (FRAX), developed by the University of Sheffield, and launched in 2008. FRAX can estimate the 10-year fracture probability of a major osteoporotic fracture (MOF) and HF (https://frax.shef.ac.uk/FRAX/) based on clinical risk factors, with or without BMD. The clinical risk factors include age, gender, weight, height, fracture history, parent fracture history, smoking status, glucocorticoid intake, rheumatoid arthritis, secondary osteoporosis, alcohol intake, and BMD.[8] Some systematic reviews have evaluated the discriminatory ability of FRAX in predicting osteoporotic fractures and found that it performed moderately well in predicting the 10-year probability of MOF and HF with or without BMD.[9-11] Several studies have also demonstrated that FRAX, without BMD, performs comparably to BMD in identifying a densitometric diagnosis osteoporosis, making FRAX a convenient tool for identifying poor bone health in Asian populations, especially where access to medical imaging may be limited. [12-16]
While FRAX enables country-specific calculations of fracture probability, the development of FRAX was unfortunately primarily dependent on data from Caucasian populations.[ 9-11,17] There is a need to understand the performance of FRAX within Asian populations specifically since it is burdened with a rapidly increasing rate of osteoporotic fractures. Therefore, the aim of this systematic review and meta-analysis was to determine if FRAX has discriminative power to identify individuals at risk of fractures in Asia and to evaluate the limitations of FRAX with respect to Asia specifically. A secondary aim was to explore the discriminative power of FRAX to identify a densitometric diagnosis of osteoporosis in Asian populations.
METHODS
1. Search strategy
Electronic databases EMBASE, Cochrane, and PubMed were searched systematically from the earliest indexing date to January 21, 2024. ‘FRAX,’ ‘fracture risk assessment tool,’ and ‘fracture’ constituted the search terms and were formatted for different electronic databases. The full search term strategy can be found in Supplementary Appendix 1. We followed the preferred reporting items for systematic reviews and meta-analyses (PRISMA) guidelines for the review.[18] Rayyan (http://rayyan.qcri.org),[19] a free online application, was used to resolve duplicates and apply study eligibility criteria.
2. Study selection criteria
The following inclusion criteria was applied to all studies: (1) the study used FRAX to identify a densitometric diagnosis of osteoporosis or future osteoporotic fractures; (2) the study population was from Asia; (3) the study was in English and on human subjects; (4) the study was a journal article; and (5) reported area under the curve (AUC) values. Studies were excluded if found to be irrelevant during full-text review, or the studies were identified as statements, conference papers, case reports, or reviews. Authors DJ and ADP carried out independent screening of the entire collection based on the study selection criteria. The screening was two phased, where articles were first identified based on titles and abstracts and the second phase of screening involved selection based on full-text review. Additionally, manual search through references in the included studies was carried out for identifying studies not captured by the search strategy. Disagreements between the two reviewers were resolved by discussion.
3. Data extraction
The following data was extracted from the studies selected after full-text review by DJ and ADP independently: Country, study design, follow-up period in years if reported, sample size at baseline, subject demographics in terms of mean age, percentage of women, type of fracture recorded, fracture ascertainment method, BMD measurement site, and performance measures reported (AUC, sensitivity, specificity, and any other outcomes). The primary outcome considered in this review was the AUC from receiver operating characteristic curve. A formula (Eq. 1) was used to approximate the 95% confidence interval (CI) values for the studies that did not explicitly report it.[20]
where, zcrit=two-tailed critical value of the standard normal distribution
where n1=Number of cases, n2=number of non-cases; q0, q1, q3=Eq. 3a-3c
4. Quality assessment
The quality assessment was independently carried out by authors DJ and ADP using the quality assessment of diagnostic accuracy studies 2 (QUADAS-2) checklist.[21] The checklist was updated to carry out a quality assessment within the purview of this review (Supplementary Table 1). Disagreements were resolved by discussion. A potential selection bias was considered if fracture ascertainment was not verified and if more than 10% of eligible study participants were excluded from the analysis.
5. Statistical analysis
The included studies either evaluated the discriminatory power of FRAX in terms of predicting osteoporotic fractures or presence of osteoporosis. MOFs were defined as hip, vertebral, wrist, and humerus fractures. We analysed subgroups of the included studies that reported AUC estimates based on standard FRAX probabilities with and without BMD, adjusted FRAX probabilities, and BMD alone, in the prediction of either osteoporotic fractures or a densitometric diagnosis of osteoporosis. Subgroup analyses were performed separately for FRAX-MOF and FRAX-HF, except for the prediction using BMD alone due to the limited number of studies reporting AUC values. Adjusted FRAX refers to the modification of the standard FRAX by one or more clinical risk factors (i.e., increased age, reduced BMD, T-Score measured at different sites, falls, trabecular bone score [TBS], and sarcopenia). Separate meta-analyses were conducted on studies to evaluate the potential impact of class imbalance, as AUC values can be affected by unequal distribution between fracture and non-fracture cases.[22] We assumed that there was a class imbalance when less than 10% of the sample population experienced fractures. These studies reported AUC estimates for standard FRAX with and without BMD. A three-level random effects model was employed using the metafor package in the R programming environment (version 3.4.5; The R Foundation for Statistical Computing, Vienna, Austria; https://www.rproject.org/) [23] to account for multiple AUC values presented by most of the studies. This model was utilized to address the interdependence between reported values. Forest plots were used to compare AUC estimate values along with their corresponding 95% CIs. Heterogeneity was evaluated using the Q-test and the multi-level I2 statistic, allowing separate estimation of ‘within-study’ (I2Level 2) and ‘between-studies’ (I2Level 3) variance.[24] As individual effect sizes within a cluster or a study, are derived from the same sample, their sampling errors are presumed to be correlated.[25] To accommodate this within-study dependence, cluster-robust variance estimation was implemented by assuming a correlation factor of ρ=0.60. Statistical significance was defined as a P-value less than 0.05.
RESULTS
1. Study selection
The initial search strategy yielded 10,622 articles. After removing duplicates (N=2,894) and screening titles and abstracts, 53 articles were considered for full-text reading (Fig. 1). This led to further exclusion of fourteen articles. Reasons for exclusion were irrelevance for the purpose of the review (N=8), or articles found to be conference abstracts (N=6). Three additional articles were included after the manual search through references of studies that passed the inclusion criteria. In total, 42 articles were included in this review.
2. Study characteristics
The summary of study characteristics is provided in Table 1 for the use of FRAX for predicting osteoporotic fractures (N=32), and Table 2 for a densitometric diagnosis of osteoporosis (N=11). Studies were from the following eleven countries: China,[12,26-37] Taiwan,[38-42] Hong Kong,[7, 43-45] Japan,[46-49] Korea,[50-53] India,[54-56] Singapore,[57-59] Thailand,[60,61] Sri Lanka,[14,62] Iran,[63] and Palestine.[64] Studies that adjusted FRAX did so with increased age,[28] reduced BMD [28], T-score measured at different sites,[52] falls,[45] TBS,[32,44,46,48,51,55,62] and sarcopenia.[34] Nine studies reported discriminative ability of BMD alone for fracture prediction, in addition to FRAX outcomes.[26,27,29,30,33,47,50,53,60] There were seven cross-sectional studies for the prediction of osteoporotic fractures based on prevalent fracture history,[26,33,40,52, 53,55,61] and five cross-sectional studies for the prediction of osteoporosis.[12,36,37,59,64] Out of 16 prospective studies that used FRAX to predict fractures, the follow-up period ranged from 2.6 to 11 years.[7,28,31,32,34,38,39,41-48,51] Some studies focused on specific populations at risk of developing secondary osteoporosis, such as patients with differentiated thyroid carcinoma after postoperative thyroid-stimulating hormone suppression therapy,[35] breast cancer,[31] diabetes,[28,32] systemic sclerosis,[53] and rheumatoid arthritis.[41,53]
3. Assessment of methodological quality
Evaluation of study quality is presented in Supplementary Tables and Figures. Out of the 43 studies reviewed, 9 were assessed to have a low risk of bias for participant selection, 4 for index test, 38 for reference standard, and 34 for flow and timing (Supplementary Table 2). Additionally, 37 studies for participant selection and 41 studies each for index test and reference standard were assessed to have minimal concerns regarding applicability. Twenty-two studies did not report whether patient enrolment was randomized (Supplementary Fig. 1). Out of the studies analyzed, 12 used a case-control design, indicating a higher risk of bias. Most (36 studies) did not specify if investigators were blinded to fracture status. Fractures were confirmed using established protocols in 38 studies. Out of the studies, 23 had fewer than 100 events of interest, and only 16 had sample sizes exceeding 1,000. Risk factors were primarily collected through clinical interviews (31 studies), with 4 relying on self-reports and 8 not clearly stating the information source for clinical risk factors.
4. Meta-analysis
Twenty studies reported 49 AUC values for predicting osteoporotic fractures using standard FRAX without BMD (Fig. 2). The overall pooled AUC of FRAX for the prediction of any osteoporotic fracture was 0.72 (95% CI, 0.67-0.77; Q=1,008.97; df=48; P<0.01), with I2Level 2=24.79% and I2Level 3=71.29. Subgroup analysis for the prediction using FRAX-MOF without BMD revealed an AUC of 0.72 (95% CI, 0.66-0.77; Q=782.57; df=29; P<0.01) and was similar to the prediction using FRAX-HF with an AUC of 0.72 (95% CI, 0.65-0.80; Q=200.66; df=18; P<0.01). In the additional meta-analyses conducted only for studies, where more than 10% of the sample population experienced fractures, the exclusion resulted in a lower overall pooled AUC of 0.70 (95% CI, 0.64-0.77; Q=838.24; df=30; P<0.01) using FRAX without BMD (Supplementary Fig. 2), compared to an AUC of 0.72 without the exclusion (Fig. 2).
Twenty-two studies reported 62 AUC values for predicting osteoporotic fractures using standard FRAX with BMD (Fig. 3). The overall pooled AUC of FRAX for the prediction of any osteoporotic fracture was 0.74 (95% CI, 0.71-0.77; Q=615.67; df=61; P<0.01), with variances of I2Level 2= 35.57% and I2Level 3=53.95%. Subgroup analysis for the prediction using FRAX-MOF with BMD revealed an AUC of 0.73 (95% CI, 0.70-0.77; Q=334.03; df=40; P<0.01) and an AUC of 0.77 (95% CI, 0.71-0.83; Q=150.51; df=20; P< 0.01) for the prediction using FRAX-HF. Considering only the studies that had more than 10% of fracture events, the exclusion resulted in a lower AUC of 0.73 (95% CI, 0.68- 0.77; Q=388.29; df=43; P<0.01) using FRAX with BMD (Supplementary Fig. 3), compared to an AUC of 0.74 without the exclusion (Fig. 3).
Twelve studies reported 36 AUC values for predicting osteoporotic fractures using adjusted FRAX (Fig. 4) with or without BMD. The overall pooled AUC of FRAX for the prediction of any osteoporotic fracture was 0.71 (95% CI, 0.66-0.76; Q=547.69; df=35; P<0.01), with an I2Level 2 of 43.97% and I2Level 3 of 49.55%. Subgroup analysis for the prediction using adjusted FRAX-MOF and FRAX-HF revealed an AUC of 0.71 (95% CI, 0.65-0.77; Q=244.34; df=25; P<0.01) and an AUC of 0.75 (95% CI, 0.63-0.86; Q=56.32; df=9; P<0.01) respectively.
Nine studies reported 24 AUC values for predicting osteoporotic fractures using BMD (Fig. 5). The overall pooled AUC of BMD for the prediction of osteoporotic fractures was 0.68 (95% CI, 0.62-0.73; Q=298.95; df=23; P<0.01), with variances of I2Level 2=25.76% and I2Level 3=69.05%.
Eleven studies reported 52 AUC values for identifying a densitometric diagnosis of osteoporosis using standard FRAX (Fig. 6). The overall pooled AUC using FRAX for the densitometric diagnosis of osteoporosis was 0.78 (95% CI, 0.71-0.85; Q=2,217.58; df=50; P<0.01), with I2Level 2=21.84% and I2Level 3=75.89%. Subgroup analysis for the prediction using FRAX-MOF revealed an AUC of 0.77 (95% CI, 0.70- 0.84; Q=1,068.96; df=31; P<0.01) and an AUC of 0.76 (95% CI, 0.67-0.86; Q=1,127.19; df=18; P<0.01) using FRAX-HF. The AUC values obtained from each meta-analysis are shown in the forest plots (Fig. 2-6, Supplementary Fig. 2, 3) and were also tabulated according to the predictive outcomes and tools used (Table 3).
DISCUSSION
In this review, our primary objective was to assess the discriminatory power of FRAX in terms of predicting future fractures in Asian populations. The findings from our systematic review and meta-analysis suggest that FRAX may not perform as well in Asian populations, as it has been observed to do in meta-analyses conducted in Caucasian populations.[9,10] A systematic review encompassing 47 Caucasian and 6 Asian populations found that the discriminative ability of FRAX-MOF without BMD was 0.77 (95% CI, 0.73-0.80) and with BMD was 0.78 (95% CI, 0.75-0.81).[9] FRAX-HF, on the other hand, had an AUC of 0.75 (95% CI, 0.72-0.79) without BMD and 0.79 (95% CI, 0.77-0.81) with BMD.[9] These findings suggest that the AUC values were comparatively lower for both FRAX-MOF (AUC of 0.72 without BMD and 0.73 with BMD) and FRAX-HF (AUC of 0.72 without BMD and 0.77 with BMD) in Asian populations. The scarcity of FRAX calibration data for Asian populations highlights potential limitations in its predictive ability across diverse Asian ethnic groups.[65]
The reduction in discriminative power was more pronounced when only considering studies with over 10% of fracture cases in their sample populations. While there was no change in AUC for FRAX-MOF with BMD, subgroup analysis showed a 1-point decrease in AUC for FRAX-MOF without BMD. Additionally, the AUC for FRAX-HF showed a 4-point decrease without BMD and a 2-point decrease with BMD. Highly imbalanced data, like the occurrences of fracture events, represents a true clinical challenge. Several studies included in our review reported fracture events below 10% among their study populations.[7,28,31,32,38,39, 44-48,51,54,61] Therefore, this demonstrates the necessity for cautious interpretation of findings from such studies when establishing the reliability of FRAX in predicting fracture risk.
This review also found that FRAX demonstrated superior discriminatory performance compared to BMD alone in identifying individuals at risk of fractures. In fact, our review suggests that it may exhibit stronger, albeit moderate, predictive capabilities for identifying a densitometric diagnosis of osteoporosis in Asian populations. This could be attributed to the limitations of BMD as a two-dimensional measure of bone health,[66] whereas FRAX incorporates clinical risk factors that provide a more comprehensive assessment. Caution should be exercised when using FRAX to identify a densitometric diagnosis of osteoporosis, as it was originally developed for assessing fracture probability. Nonetheless, FRAX may be useful in guiding treatment decisions in resource-constrained settings with limited availability of DXA machines, particularly in some Asian countries, offering a practical alternative for identifying high-risk individuals.[14,56]
Our review indicates that FRAX with BMD tends to perform more reliably in predicting HF (AUC=0.77) compared to MOF (AUC=0.73). FRAX without BMD had the same AUC of 0.72 for both HF and MOF with a wider 95% CI for FRAX-HF indicating greater uncertainty of FRAX in predicting HFs in the absence of BMD. It is worth noting that in FRAX calculations where BMD is integrated, the femoral neck is typically utilized as the measurement site. Lumbar spine BMD could offer a modest improvement in predicting vertebral and potentially also MOF, though more research is needed to confirm these findings.[67]
FRAX was initially validated using 11 Caucasian and 1 Japanese cohort studies, with participants aged between 50 and 65 years.[17] This demographic skew might contribute to the underestimation of fracture risk in individuals above 65 years of age. For instance, a study conducted in Hong Kong revealed that FRAX failed to predict fracture risk accurately in the oldest age group (>80 years).[43] The National Health and Nutrition Examination Survey (NHANES) proposes fixed intervention cut-offs of 20% for MOF and 3% for HFs prediction. However, studies indicate that FRAX might underestimate fracture risk in certain Asian sub-groups due to the utilization of a single cut-off point. This approach could classify almost all individuals above 70 years as high risk solely based on age, which is a significant risk factor.[54] To establish appropriate and accurate cutoffs, various factors such as the cost of DXA, consultation fees, medication costs, and cost-effectiveness need careful consideration.[58] While the majority of studies used the fixed intervention cut-offs proposed by NHANES,[7,14,28, 31,32,34,38,40,41,43,46-48,50-52,57,63] some studies have evaluated the cut-offs based on the Youden’s Index and used these optimal cut-offs to determine the accuracy of FRAX.[12,26,27,29,30,33,35-37,39-42,44,45,49,54-56,58-62,64] Table 4 provides the recommended FRAX cut-off values for various countries to assess fracture risk. Studies reviewed for Japan and Korea utilized NHANES cut-offs for predicting fracture risk. In China, the recommended FRAX-MOF cut-off values ranged between 2.9% and 3.6%, with one study suggesting a cut-off of 9.25%.[30] This high cutoff of 9.25% could have resulted from the high risk of bias in patient selection and unclear applicability of the FRAX. [30] In India, the recommended cut-offs ranged between 9% and 10.5% for FRAX-MOF and between 2.5% and 3.5% for FRAX-HF.[54,55] There was only one study from Sri Lanka that recommended a cut-off of 9% for FRAX-MOF and 2.7% for FRAX-HF.[62] In Thailand, the recommended cutoffs for FRAX-MOF ranged from 8.9% to 10%, while for FRAX-HF they ranged from 3% to 4.9%.[60,61] In Hong Kong, two studies recommended a cut-off for FRAX-MOF of 12.5% for men and 15% for women.[44,45] The recommended cut-offs from Taiwan had the widest variability and ranged between 9.5% and 22.5% for FRAX-MOF and between 3.14% and 12.5% for FRAX-HF,[39-42] which could be attributed to the difference in follow-up times and sample population characteristics. Additionally, two studies from Taiwan compared the AUC for FRAX-MOF using NHANES cut-offs and recommended cut-offs.[40,41] The first study showed an increase in AUC with the recommended cut-off compared to the NHANES cut-off,[41] while the second study reported improved sensitivity and specificity for FRAX-MOF at the recommended cut-off, whilst maintaining the same AUC as the NHANES cut-off. [40] This variability in recommended cut-offs demonstrates the need for more standardized studies for better comparability. Also, the diverse recommendations across these studies emphasize the need to tailor FRAX cut-offs to the specific demographic characteristics and prevalent risk factors of each country. Interestingly, a recent study explored individual intervention thresholds (IIT) by comparing fracture risk with a comparator score derived solely from the history of osteoporotic fracture.[41] If the actual risk was higher than the comparator score, the fracture risk was deemed to be higher than or equal to the IIT score. However, the study also found that intervention thresholds determined by Youden’s Index remained more accurate than both IIT and fixed intervention thresholds suggested by NHANES.[41] Although this study focused on rheumatoid arthritis patients, it highlights the need for further investigation to establish optimal cut-offs for fracture risk assessment.
While FRAX incorporates multiple risk factors beyond BMD, such as age, gender, previous fracture history, and other clinical risk factors, these risk factors may not universally apply to all populations. Therefore, its widespread adoption could be limited, affecting its suitability for large-scale screening and community-level applications.[26] Many studies emphasize the importance of considering ethnic-specific risk factors and population characteristics when using FRAX. The effectiveness of FRAX in predicting fracture risk in Asian populations may vary due to differences in lifestyle factors, genetic predispositions, and bone structure. Firstly, alcohol use is limited in Asian countries and may be less relevant as a risk factor.[47,68] Incorporating handgrip strength, weight, age, postmenopausal status, and chronic joint pain yielded a more accurate model (AUC of 0.84) compared to the standard FRAX tool for identifying a densitometric diagnosis of osteoporosis.[59] Secondly, integrating genetic and physiological differences, such as fracture susceptibility due to varying physical activity patterns,[69,70] along with incorporating biomarkers like TBS,[71,72] using local epidemiological data for calibration, and conducting long-term cohort studies in specific ethnic groups, is crucial for refining FRAX’s predictive capabilities.
Some studies have also demonstrated improved fracture prediction accuracy using machine learning approaches in combination with FRAX.[31,51] This advancement often leverages data from wearable health technologies or utilizes advanced imaging techniques and post-processing methods to refine predictions. Structural biomarkers computed from finite element models based on computed tomography (CT) imaging have been shown to be more effective at predicting osteoporotic fractures, compared to BMD. [73,74] Advances in automatic CT scan segmentation using machine learning algorithms have also made the evaluation of these structural biomarkers computationally efficient and clinically feasible.[75] A recent study on a Swedish cohort has demonstrated an improved AUC of bone strength computed from finite element modeling based on DXA images over BMD for predicting HFs.[76] Future research with larger samples could help validate the inclusion of structural biomarkers such as bone strength in the FRAX algorithm to enhance fracture prediction accuracy.
This review has certain limitations that warrant consideration. First, while we compared the AUC values of FRAX, assessing the sensitivity and specificity of FRAX would have provided valuable insights. However, reported sensitivity and specificity values were limited, making meta-analyses infeasible. Additionally, studies were either cross-sectional or longitudinal in nature. However, it is crucial to acknowledge that fracture risk estimated by FRAX can only be accurately evaluated in studies with long-term follow-up. Correlations drawn from cross-sectional studies or studies with less than 10 years of follow-up may not fully capture the predictive ability of FRAX, especially considering that FRAX is designed to estimate 10-year fracture probability. Therefore, incorporating weightage based on study design and follow-up duration could potentially provide a more accurate estimation of the predictive ability of FRAX.
In conclusion, despite its limitations, FRAX remains a valuable tool, particularly in regions where guidelines for managing osteoporotic fractures are lacking. However, given that this review shows FRAX performs less effectively in Asian populations, its use requires careful attention to ethnic-specific factors, validation studies, and potential modifications to enhance its accuracy and usefulness in clinical settings. While FRAX exhibits better discrimination ability for osteoporosis, its effectiveness in predicting osteoporotic fractures in Asia remains uncertain. Future research endeavours should prioritise refining fracture risk assessment tools to enhance prediction accuracy so that ultimately, the burden of osteoporotic fractures in Asian communities can be alleviated.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement1
Print
