Abstract
The Microarchitecture Fracture Risk Assessment Calculator (
FRAC) was developed to predict osteoporotic fracture risk using direct measures of bone microarchitecture. However, its performance has not yet been externally validated in women. This study aims to evaluate the performance of the 
FRAC model in predicting osteoporotic fracture risk in a cohort of older women. The performance of 
FRAC was assessed in a population of 2307 postmenopausal women aged 75-80 yr in Gothenburg, Sweden. All participants underwent HR-pQCT scanning of the distal radius and distal tibia (82 
m). Incident fracture information was collected from regional X-ray archives up until March 2023. The 
FRAC 5- and 10-yr risk of major osteoporotic fracture (MOF) and any osteoporotic fracture were calculated for all participants. The model discrimination was assessed using receiver operator characteristic (ROCs) curves and area under the curve (AUCs). FRAX and femoral neck areal bone mineral density (FN aBMD) were used as reference models for comparison. During the follow-up period (mean 6.82 
2.48 yr), 584 participants (25.3%) sustained an MOF. The AUCs for MOF models for 5- and 10-yr predictions were similar for 
FRAC (AUC = 0.601-0.634) compared with reference models of FRAX (AUC = 0.615-0.618) and FN aBMD (AUC = 0.596). Overall, 
FRAC showed good generalizability (regression slope 0.82-1.00) to the external cohort. Fracture risk prediction by 
FRAC was demonstrated to perform consistently with FRAX and FN aBMD in this external cohort of elderly females, suggesting it is a viable alternative fracture risk prediction tool. A benefit of 
FRAC is that it is based on bone microarchitecture rather than binary clinical risk factors, so it has potential to be sensitive to changes in bone fragility as a function of age-related deterioration or anti-osteoporosis treatment.
Keywords: fracture risk assessment, μFRAC, HR-pQCT, osteoporosis, FRAX
2 Introduction
Clinically, femoral neck areal bone mineral density (FN aBMD) is commonly used as a surrogate measure of bone strength.1 However, its predictive value is limited; approximately half of all fragility fractures occur in individuals who do not meet the World Health Organization’s definition of osteoporosis (T-score 
).2 To improve clinical management of osteoporosis, over 48 fracture risk assessment tools have been developed, 20 of which have undergone external validation.3,4 The majority of these tools combine FN aBMD with clinical risk factors and demographic data to estimate the 5- or 10-yr risk of fracture.5,6 However, reliance on risk factors has limitations because it is based on indirect measures of bone quality and fall risk; it frequently depends on patient recall or accurate medical history, and often involves binary (yes/no) indicators that may not capture complex dose-dependent cases, requiring additional physician discretion.
The development of 3-dimensional HR-pQCT provides a direct approach to assessing bone health, capturing key elements contributing to bone strength, including compartment-specific volumetric BMD, macrostructure, and microarchitecture in vivo. The recently developed Microarchitecture Fracture Risk Assessment Calculator (
FRAC) directly incorporates bone microarchitecture from HR-pQCT imaging to estimate osteoporotic fracture risk.7 The model was trained on data from the Bone Microarchitecture International Consortium (BoMIC), which encompassed 7 independent cohorts in North America and Europe, and included male and female participants with prospective fracture information.8 However, 
FRAC has not yet been validated in an independent female population.
Validation of the model in external cohorts is essential to confirm generalizability and robustness of 
FRAC before clinical application. Specifically, validation is particularly important for older women, who are at a high risk of fragility fractures and associated adverse outcomes. The most common fracture prediction tool, FRAX, generally performs well in this demographic and so comparison with 
FRAC is of particular interest. This study aims to validate 
FRAC using the Sahlgrenska University Hospital Prospective Evaluation of Risk of Bone Fractures (SUPERB) cohort of older women and to compare its performance with reference fracture prediction tools.
3 Materials and methods
3.1 Participants
The SUPERB study cohort is comprised of 3028 postmenopausal women aged 75-80 yr at baseline living in the Gothenburg area, Sweden. Participants were selected randomly from the Swedish national population registry between March 2013 and May 2016. Women were excluded if non-ambulant or unable to understand Swedish. Detailed cohort recruitment, inclusion/exclusion criteria, and study design have previously been described.9–13 All participants provided written and informed consent prior to study enrollment, and the study was approved by the regional Ethics Review Board in Gothenburg, Sweden. A subset of 2307 participants who had complete HR-pQCT and clinical data for the 
FRAC risk calculations were selected for the current study (Figure 1).
Figure 1.
Cohort selection workflow.
3.2 Medical imaging
Participants were scanned at the nondominant distal radius and ipsilateral distal tibia using HR-pQCT (XtremectCT, 82 
m, Scanco Medical AG, Brütisellen, Switzerland). In the case of a prevalent fracture the contralateral side was scanned. Scans were acquired at a fixed offset of 9.5 and 22.5 mm proximal to the articular surface of the distal radius and distal tibia, respectively. The scans were graded for motion artifacts on a scale of 1 (superior image quality) to 5 (poor image quality), with motion scores of 4 or higher excluded from analysis.
The scans were analyzed following the manufacturer’s (Scanco Medical AG) standard patient protocol.14 Contours were automatically placed in the periosteal and endosteal surfaces of the cortical bone, and manually corrected by the operator if needed.15 The following parameters were used in this study: total bone mineral density (TtBMD), cortical BMD (CtBMD), trabecular BMD (TbBMD), total area (TtAr), cortical area (CtAr), trabecular area (TbAr), trabecular number (TbN), trabecular inhomogeneity (Tb.1/N.SD), and cortical thickness (CtTh) at the radius and tibia.
Finite element (FE) analysis was performed using the manufacturer’s software (Scanco FE Software v1.15, Scanco Medical) on the segmented HR-pQCT images to obtain the estimated failure load. A tissue modulus of 10 GPa and a Poisson ratio of 0.3 was applied to the segmented bone tissue.16 A simulated uniaxial compression was applied, with the yield parameters used to estimate failure defined as at least 2% of elements surpassing 7000 microstrain.16 Estimated failure load was standardized to be consistent with the original 
FRAC model training using calibration equations derived from internal data (r2 = 0.99).17
The FN aBMD was measured using DXA (Discovery A; Hologic). Participants were scanned at the left femur, except in the case of a prior fracture, implant, or metal at the scan site, in which case the contralateral side was scanned. To maintain consistency with the original 
FRAC model training, FN aBMD was harmonized across DXA systems using previously reported equations.18
FRAX 10-yr probability of major osteoporotic fractures (MOFs) was calculated using FN aBMD and clinical risk factors including previous fracture, parental hip fracture, smoking, alcohol consumption, corticosteroids, rheumatoid arthritis, and secondary osteoporosis. Medical history and clinical risk factor information was obtained through validated study questionnaires. In the case of missing data, participants were excluded from the analysis. Previous fractures were self-reported and included any fractures sustained after the age of 50 at any location except the skull or face. FRAX was selected as the reference model because it is widely used internationally, and has been calibrated for use in more than 80 countries.19 Additionally, FN aBMD was selected as it is used in guidelines of several countries worldwide for osteoporosis screening, diagnosis, and fracture risk assessment.20–23
3.3 
FRAC model
The 
FRAC model was previously developed using the random survival forests algorithm,7 trained and internally validated using the BoMIC population cohort (
= 6836).8 The primary outputs of the model include the 5- and 10-yr risk of MOF and the secondary outputs are the 5- and 10-yr risk of any osteoporotic fracture (AOF). MOFs were defined as low-trauma hip, forearm, proximal humerus, and clinical lumbar spine fractures. AOF was defined as any fracture, regardless of trauma type, excluding fractures of the skull, fingers, and toes. There are 3 variants of the model that include varying levels of information as the input: (1) the complete model (
FRAC) that takes inputs of HR-pQCT bone parameters, 
FE-estimated failure load, FN aBMD, age, height, weight, sex (female/male), and prior fracture (yes/no); (2) a simplified model (
FRAC
) that does not include FN aBMD and any clinical risk factors except age; and (3) the simplest model (
FRAC
) that removes 
FE-estimated failure load and only uses microarchitecture and age.
3.4 Fracture outcome data
All incident fractures, regardless of trauma type, were identified from regional X-ray archives for the Västra Götaland region. All radiology reports were reviewed by 1 of 5 research nurses from baseline (March 2013 to April 2016) to March 2023. In the case of a missing report, an experienced orthopedic surgeon was consulted to determine the existence of a fracture. Incident vertebral fractures were only included when identified on examinations with a fracture inquiry. Mortality data were obtained using the regional database in Västfolket, Sweden.
3.5 Statistical analysis
Descriptive statistics of the cohort split into individuals with and without an incident MOF are summarized as the mean 
SD (minimum, maximum). Differences between groups were assessed using Chi-squared (for categorical data) and Mann-Whitney U test (for continuous data).
The performance of 
FRAC in predicting risk of AOF and MOF was assessed using time-dependent receiver operator characteristic (ROC) curves and area under the curve (AUC) for each model variant (
FRAC, 
FRAC
, 
FRAC
) for 5- and 10-yr risk estimates. ROCs and AUCs were also generated for FRAX and FN aBMD to provide a reference of model performance in the cohort. In the AUC analysis, FN aBMD was used as a stand-alone predictor variable without incorporating clinical risk factors (CRFs) such as age and sex. Due to the small age range and uniform sex distribution of the cohort, adding CRFs to the model reduced its predictive performance compared with FN aBMD alone. The mean and 95% CI of the time-specific AUCs were estimated from bootstrap sampling results with replacement over 100 iterations. AUC values are commonly interpreted as follows: 0.6-0.7 indicates sub-optimal discrimination, 0.7-0.8 acceptable, 0.8-0.9 good, and values above 0.9 are considered excellent. A one-way ANOVA was used to assess for significant differences between the AUCs of all models, and a Tukey test was performed for post-hoc comparisons.24
The categorical net reclassification index (NRI) was implemented to assess the potential of 
FRAC to shift clinical decision making based on clinical intervention thresholds. The NRI quantifies how well 
FRAC correctly categorizes individuals based on whether they fractured or not.25 The NRI was calculated with FRAX as the reference model and a risk cutoff of 26% was used for both FRAX and 
FRAC, representing the risk in a 70-yr-old Swedish woman with a previous fracture—a threshold derivation recommended by the National Osteoporosis Guideline Group (NOGG) in the United Kingdom.26 The 2 components of the NRI are reported: net reclassification of events (NRIe) and the net reclassification of nonevents (NRIne).
Tool calibration was assessed by fitting a Fine Gray competing risk regression model27 to obtain the observed 5- and 10-yr rates of osteoporotic fracture with 95% CI. The Fine Gray model accounts for event censoring and the competing risk of death. The cohort was stratified based on the 5- and 10-yr fracture risk estimates and divided into risk quintiles for the 3 variants of the 
FRAC model. Within each quintile, the average predicted fracture risk was compared with the observed rates of osteoporotic fracture. Analyses was undertaken on both MOF and AOF models. All analyses were implemented in R (v 4.3.3).
4 Results
4.1 Population
The cohort consisted of 2307 females (Figure 1) with an average age at baseline of 77.7 
1.6 yr, followed for a mean duration of 6.82 
2.48 yr (max 9.91 yr). During the follow-up time, 766 (33.2%) participants sustained fractures, of those 584 occurred at MOF sites, of which 176 were hip fractures and 144 were forearm fractures. A total of 247 participants (10.7%) died prior to any incident fracture. Descriptive characteristics of individuals with and without and incident MOF are presented in Table 1. Significant group-wise differences between those with and without an MOF were found across all parameters except height and weight. Individual 5-yr 
FRAC MOF risk estimates ranged between 0.9% and 27.3% and the 10-yr 
FRAC risk estimates ranged from 3.5% to 50.2%.
Table 1.
Descriptive statistics for participants at baseline with and without major osteoporotic fracture. Continuous parameters are reported as mean 
SDs and nominal parameters are reported as frequency and percentage.
| Parameter | Incident fracture | No incident fracture |
-value |
|---|---|---|---|
( = 584) |
( = 1723) |
||
| Age (yr) | 77.7  1.6 |
77.9  1.6 |
<0.01 |
| Any previous adulthood fracture | 264 (45.2%) | 569 (33.0%) | <0.01 |
| Follow-up time (yr) | 3.9  2.4 |
7.4  2.1 |
<0.01 |
| Anthropometrics | |||
| Height (cm) | 161.9  6.3 |
161.9  5.7 |
0.88 |
| Weight (kg) | 68.6  12.3 |
68.8  12.0 |
0.78 |
| DXA and risk estimates | |||
| Femoral neck areal BMD (g/cm2) | 0.71  0.10 |
0.75  0.12 |
<0.01 |
| FRAX 10-yr MOF risk (%) | 25.7  12.3 |
21.5  11.0 |
<0.01 |
 FRAC 10-yr MOF risk (%)
|
26.6  9.6 |
22.1  9.3 |
<0.01 |
| HR-pQCT - radius | |||
| Total BMD (mgHA/cm3) | 224.3  60.2 |
245.7  61.7 |
<0.01 |
| Cortical BMD (mgHA/cm3) | 757.2  78.8 |
773.5  77.6 |
<0.01 |
| Trabecular BMD (mgHA/cm3) | 110.4  39.7 |
124.8  40.4 |
<0.01 |
| Trabecular number (1/mm) | 1.62  0.44 |
1.75  0.42 |
<0.01 |
| Trabecular inhomogeneity (mm) | 0.37  0.27 |
0.31  0.24 |
<0.01 |
| Cortical thickness (mm) | 0.51  0.18 |
0.55  0.18 |
<0.01 |
| Total area (mm2) | 274.8  45.4 |
269.3  44.6 |
<0.01 |
| Cortical area (mm2) | 35.4  11.2 |
38.1  11.5 |
<0.01 |
| Trabecular area (mm2) | 228.6  45.8 |
221.4  45.4 |
<0.01 |
| Failure load (N) | 2309.4  463.0 |
2452.9  469.5 |
<0.01 |
| HR-pQCT - tibia | |||
| Total BMD (mgHA/cm3) | 215.2  45.6 |
232.7  47.5 |
<0.01 |
| Cortical BMD (mgHA/cm3) | 727.3  70.7 |
746.2  66.5 |
<0.01 |
| Trabecular BMD (mgHA/cm3) | 140.4  34.7 |
150.1  35.0 |
<0.01 |
| Trabecular number (1/mm) | 1.73  0.36 |
1.80  0.35 |
<0.01 |
| Trabecular inhomogeneity (mm) | 0.28  0.19 |
0.25  0.16 |
<0.01 |
| Cortical thickness (mm) | 0.70  0.24 |
0.78  0.24 |
<0.01 |
| Total area (mm2) | 741.5  108.0 |
724.9  103.7 |
<0.01 |
| Cortical area (mm2) | 73.4  22.4 |
81.2  22.8 |
<0.01 |
| Trabecular area (mm2) | 651.7  114.4 |
629.2  110.9 |
<0.01 |
| Failure load (N) | 6528.2  1051.6 |
6870.7  1058.4 |
<0.01 |
Abbreviations: FRAX, fracture risk assessment tool; 
FRAC,microarchitecture fracture risk assessment calculator; MOF, major osteoporotic fracture.
4.2 ROC analyses
Overall, the 
FRAC tool outperformed FN aBMD for prediction of MOF and AOF over 5- and 10-yr, and provided either improved or equivalent performance to FRAX across these outcomes performance measures (Figure 2). Specifically, the 5- and 10-yr MOF AUCs for the complete 
FRAC model were greater than FRAX by 0.016 and 0.011 (
FRAC AUCs: 0.601-0.634), respectively. For the AOF estimates, 
FRAC maintained an improved performance over FRAX and FN aBMD for all model variants (
FRAC AUCs: 0.608-0.631, Figure 2). A sub-analysis removed incident vertebral fractures from the MOF and AOF outcome measures. With the removal of incident vertebral fractures, there was an increase in predictive performance for all models (including FRAX and FN aBMD), but particularly in the 
FRAC
(AUCs: 0.619-0.648) and 
FRAC
(AUC: 0.618-0.639) models resulting in improved performance relative to FRAX (Table 2). Notably, this adjustment resulted in better consistency among the 
FRAC variants and was most apparent in the 5-yr estimates, where 160 (42.4%) vertebral fractures occurred compared with the 214 (36.6%) vertebral fractures that occurred over the 10-yr follow-up period.
Figure 2.
Time-dependent receiver operator characteristic (ROC) curves for prediction of major osteoporotic fracture (MOF), any osteoporotic fracture (AOF), hip fractures, and radius fractures by all 
FRAC variants, FRAX, and FN aBMD. The dashed line represents an area under the curve (AUC) of 0.50. In the figure, 
denotes a significant difference from the FRAX (+BMD) model, 
from the reference FN aBMD model,. from the 
FRAC model, and * from the 
FRAC
model in the current validation cohort.
Table 2.
Area under the curve (AUC) of all 
FRAC model variant for major osteoporotic fractures (MOF) and any osteoporotic fracture (AOF) with and without incident vertebral fractures.
| Model | AUC with incident vertebral fracture | AUC without incident vertebral fracture | |
|---|---|---|---|
| 5-yr MOF |
FRAC |
0.634 (0.631, 0.637) | 0.644 (0.641, 0.648) |
FRAC
|
0.607 (0.604, 0.610) | 0.648 (0.644, 0.651) | |
FRAC
|
0.601 (0.598, 0.605) | 0.634 (0.630, 0.638) | |
| FRAX | 0.618 (0.615, 0.622) | 0.632 (0.629, 0.635) | |
| FN aBMD | 0.596 (0.593, 0.599) | 0.606 (0.603, 0.610) | |
| 10-yr MOF |
FRAC |
0.626 (0.622, 0.629) | 0.637 (0.634, 0.640) |
FRAC
|
0.602 (0.599, 0.606) | 0.619 (0.615, 0.623) | |
FRAC
|
0.605 (0.601, 0.609) | 0.618 (0.614, 0.622) | |
| FRAX | 0.615 (0.612, 0.619) | 0.628 (0.625, 0.632) | |
| FN aBMD | 0.596 (0.592, 0.599) | 0.604 (0.600, 0.607) | |
| 5-yr AOF |
FRAC |
0.631 (0.628, 0.634) | 0.639 (0.636, 0.642) |
FRAC
|
0.615 (0.612, 0.618) | 0.619 (0.616, 0.622) | |
FRAC
|
0.608 (0.605, 0.612) | 0.639 (0.636, 0.642) | |
| FRAX | 0.604 (0.601, 0.607) | 0.606 (0.603, 0.609) | |
| FN aBMD | 0.580 (0.577, 0.582) | 0.580 (0.577, 0.583) | |
| 10-yr AOF |
FRAC |
0.624 (0.621, 0.627) | 0.626 (0.622, 0.630) |
FRAC
|
0.618 (0.614, 0.621) | 0.619 (0.616, 0.622) | |
FRAC
|
0.610 (0.606, 0.614) | 0.626 (0.622, 0.630) | |
| FRAX | 0.594 (0.590, 0.598) | 0.604 (0.600, 0.608) | |
| FN aBMD | 0.570 (0.567, 0.574) | 0.576 (0.572, 0.579) |
Sub-analyses of fracture types were further explored, including hip and radius fractures. For hip fractures, the complete 
FRAC model showed the most consistent performance with FRAX over 5- and 10-yr (Figures 2). The 5-yr AUCs (0.613-0.683) for all models, including FRAX and FN aBMD, were more robust in comparison with the 10-yr AUCs (0.587-0.628), emphasizing that there is better short-term fracture prediction in this older cohort. Analysis of radius fractures indicated improved performance of 
FRAC in estimating fracture risk compared with FRAX, particular for the 5-yr estimates (
FRAC AUCs: 0.637-0.655, Figure 2). The 
FRAC model variants that do not include FN aBMD (
FRAC
and 
FRAC
) exhibited the best model performance, highlighting the importance of peripheral measures in predicting radius fractures. Using 
FRAC
as an example, the improvement in AUC compared with FRAX was 0.067 (
) over 5-yr and 0.018 (
) over 10-yr, equating to an improvement in performance of 11.4% and 3.1%, respectively.
4.3 Net reclassification index
Using a 26% high-risk cutoff for both 
FRAC and FRAX, the 10-yr NRI revealed that 
FRAC was consistent or showed an improvement compared with FRAX in differentiating high-risk individuals (Table 3, NRI = −0.1-+8.3). The 
FRAC 5-yr prediction of fracture risk tended to classify those who did not go on to fracture as being in the lower risk category, with positive NRIne values (NRIne = 0.3-13.5) suggesting an improvement by 
FRAC in stratifying low-risk individuals over the short term. For the 10-yr estimates, 
FRAC tended to reclassify individuals who went on to fracture as being in the higher risk category, again with positive NRIe values (NRIe = 13.4-35.4), indicating improved stratification of higher risk individuals. The net NRI was positive for all models except 
FRAC
5-yr MOF (NRI = −7.5), 10-yr MOF (NRI = −0.1), and 5-yr AOF (NRI = −1.7). A positive NRI indicates that 
FRAC is improving risk classification by shifting individuals into categories that reflect their fracture outcomes.
Table 3.
Net reclassification index (NRI) of events, non-events, and net NRI for all model variants across both 5- and 10-yr fracture risk estimates for major osteoporotic fractures (MOF) and any osteoporotic fractures (AOF).
| Model | NRIe | NRIne | NRI | |
|---|---|---|---|---|
| 5-yr MOF |
FRAC |
−7.7 | 8.5 | 0.8 |
FRAC
|
−21.0 | 13.5 | −7.5 | |
FRAC
|
−0.3 | 0.6 | 0.3 | |
| 10-yr MOF |
FRAC |
13.4 | −9.8 | 3.6 |
FRAC
|
14.9 | −11.3 | 3.6 | |
FRAC
|
34.6 | −34.7 | −0.1 | |
| 5-yr AOF |
FRAC |
−0.6 | 4.0 | 3.4 |
FRAC
|
−8.5 | 6.8 | −1.7 | |
FRAC
|
1.4 | 0.3 | 1.7 | |
| 10-yr AOF |
FRAC |
28.9 | −21.2 | 7.7 |
FRAC
|
21.0 | −12.7 | 8.3 | |
FRAC
|
35.4 | −29.9 | 5.5 |
4.4 Calibration
Figure 3 shows the predicted versus observed fracture probability of 
FRAC in predicting MOF and AOF, stratified by probability quintile. For MOF, 
FRAC tended to underpredict fracture risk. However, removing incident vertebral fractures resulted in model probability predictions that were within the 95%CIs for the observed probability in most quintiles. The regression slope for the MOF predictions without vertebral fractures was 1.00 and 0.82 for the 5- and 10-yr estimates, respectively. For AOF, the models tended to show better calibration to the 10-yr risk estimates, particularly with the removal of incident vertebral fractures (Figure 3). The regression slope for AOF without vertebral fractures was 1.55 and 1.01 for the 5- and 10-yr estimates, respectively.
Figure 3.
Predicted 5- and 10-yr risk of major osteoporotic fracture (MOF) and any osteoporotic fracture (AOF) from 
FRAC versus the observed Fine-Gray fracture rates stratified by risk quintile. *indicates calibration without incident vertebral fractures. The dashed line depicts the line of identity. 95% error bars are shown.
5 Discussion
This study validates that 
FRAC performs consistently with, and in some cases slightly better than, the current standard FRAX in a cohort of older women (75-80 yr), as demonstrated by AUC and NRI analyses. This finding is of particular relevance as fracture prediction in this demographic is challenging, as evidenced by the limited performance of FRAX and FN aBMD. Given that 
FRAC uses extremity microarchitecture as a major input and achieves comparable results is notable. HR-pQCT-based fracture risk prediction models have consistently shown the potential of microarchitecture measures taken at various skeletal sites in predicting osteoporotic fracture risk.28–30 In this study, we further validate the robustness of 
FRAC models in an external cohort. Sub-analyses demonstrated that 
FRAC showed a marked improvement in predicting radius fracture risk, which is not surprising considering it includes distal radius bone microarchitecture. Although distal radius fractures are less life-threatening than other osteoporotic fractures, their prediction holds significant clinical value because they are often an early indicator of bone fragility and increase future fracture risk.31–33 It was also notable that hip fracture prediction performance was consistent despite 
FRAC not having microarchitectural information from that skeletal site. Overall, using microarchitecture as a primary model input, 
FRAC performs at least as well as traditional fracture prediction models we tested.
The longer mean follow-up of the SUPERB cohort (
7 yr) compared with the original 
FRAC training BoMIC cohort (
5 yr) was an opportunity to determine to what extent 
FRAC was suitable for 5- versus 10-yr predictions. In the original BoMIC cohort, a drop in the AUCs from the 5 to 10–yr risk estimates was observed;7 however, it was encouraging that in the SUPERB cohort that 
FRAC performed equally well at both the 5- and 10-yr risk estimates for MOF and AOF. It is notable that 5-yr fracture prediction in an older cohort may be more relevant clinically.
Vertebral fractures were removed as part of some sub-analyses because of their difficulty in being reliably clinically diagnosed, which can introduce variability and affect model performance. Following preliminary analyses, it was observed that model calibration and discrimination, particularly for 
FRAC, were adversely impacted by the inclusion of vertebral fractures. Upon this finding we performed an additional sub-analysis with these fractures excluded that resulted in improved performance across all models tested, including FRAX and FN aBMD. For 
FRAC, this improvement was most pronounced in the 
FRAC
and 
FRAC
models (both models exclude FN aBMD inputs) and overall there was greater consistency between the 
FRAC model variants. Standard methodologies for identifying fracture events in model development and validation typically rely on self-reporting and the use of International Classification of Diseases codes.7,34,35 However, these approaches can face challenges, including misremembered information, inaccurate reporting, and the complexity of certain fracture cases. The diagnosis of vertebral fractures in the SUPERB cohort was stringent, including pulling and reviewing radiology reports from regional X-ray archives, resulting in a higher incidence of findings. The original BoMIC training cohort, comprised of 7 individual cohorts, included more variability in approach to tracking vertebral fractures that may be more reflective of a real-world setting. As a result, the 
FRAC model may not be calibrated to the higher prevalence of vertebral fractures observed in the SUPERB cohort. Additionally, previous work has shown that accurate adjudication of incident clinical vertebral fractures requires examination of the X-rays, not just the radiographic reports, and comparison with baseline pre-fracture imaging.36 Thus, vertebral fractures represents a universal challenge for fracture prediction models, which was observed for both 
FRAC and FRAX performance when applied to the SUPERB cohort.
The potential for using microarchitecture measured at extremities to predict a fracture at the hip, which is an axial skeletal site, has not previously been fully explored.37 Testing 
FRAC during its development with the BoMIC cohort was limited due to a lack of hip fracture cases, but the hazard ratios reported demonstrated potential.7 More recently, a newly trained hip fracture prediction model showed promise reporting an AUC of 0.75.28 The model used 2 different HR-pQCT measurement sites on the same tibia. In our current validation tests using the radius and tibia HR-pQCT scans we found that the complete 
FRAC model demonstrated reasonable performance in predicting hip fractures (5-yr AUC = 0.683). The 
FRAC model performed similarly to FRAX in this cohort, with both 
FRAC and FRAX performing better than FN aBMD alone. Since both 
FRAC and FRAX use FN aBMD information, this suggests that for predicting hip fractures, the microarchitecture information included in 
FRAC is at least as useful information as the numerous CRFs used in FRAX. But, an advantage of using microarchitecture compared with CRFs is that it is not confined to a binary outcome.
The performance of 
FRAC was comparable with FRAX, but a potential advantage of 
FRAC is its minimal reliance on clinical risk factors, using only prior fracture history, or none at all, for prediction. This may be particularly beneficial in cases where static, binary risk factors fail to capture the nuances of an individual’s medical history, such as someone who smoked for 20 yr but quit 6 mo ago.38 Unlike FRAX, 
FRAC does not depend on patient recall or the availability of accurate and up-to-date medical records, which may be incomplete or inconsistent. Additionally, it may help reduce the burden on busy clinicians by eliminating the need to assess multiple risk factors during a patient visit. However, widespread implementation of 
FRAC is currently limited by the clinical availability of HR-pQCT. Despite this, HR-pQCT offers several clinical advantages. It demonstrates strong measurement precision, particularly for density-based parameters,39 which supports monitoring over shorter time intervals than DXA.40 Moreover, the reproducibility of the 
FRAC model is excellent (%
RMS 0.25%-0.40%41), further highlighting its potential for routine, repeated fracture risk assessment in high-risk individuals.
Evaluating the impact of fracture prediction tools on clinical decision-making can be challenging, but the NRI offers interesting insight. Despite the known limitations of the NRI measure,42,43 we found it to be a useful complement to the findings based on the ROC/AUC analysis. The net NRI revealed good concordance between the risk categories defined between FRAX and 
FRAC, consistent with what was observed in the AUC analysis. Over the 5-yr risk estimates, 
FRAC was reclassifying individuals who did not go on to fracture as being low risk, whereas over the 10-yr estimates 
FRAC was reclassifying individuals who went on to fracture as being high risk. This shows the potential for 
FRAC to be used to identify individuals who could benefit from short-term monitoring and avoid unnecessary treatment, while further discriminating high-risk individuals over the long-term. It is important to note that the 26% high fracture risk cutoff used for the NRI analysis has not been validated for 
FRAC, and was used here to demonstrate the potential clinical implications of the tool. Additionally, since there are no guidelines for interpreting the shorter 5-yr fracture risk, the cutoff was set to 13%. This assumes that the increase in fracture risk is linear; however, a limitation of this assumption is that it is known that fracture risk increases exponentially with age, particularly in older adults.44 Future work should investigate the optimal fracture risk cutoffs for the 
FRAC models for both 5- and 10-yr fracture risk estimates.
Without specific calibration to the cohort, 
FRAC displayed good generalizability to the older female cohort, particularly with the removal of incident vertebral fractures. When stratified by quintile, the majority of the error bars passed through the line of identity, which would indicate the model is well calibrated. Of the quintile predictions that fell outside the confidence interval, most exceeded the high-risk cutoff, meaning changes in risk categorization will not be affected by the calibration. Based on the older age of this cohort, the method used to assess tool calibration accounted for the competing risk of death. This consideration is especially important when evaluating long-term fracture risk, such as 10-yr predictions. However, the 
FRAC tool itself does not explicitly incorporate mortality risk when predicting fractures. In older individuals with a higher incidence of comorbidities, providing a fracture risk may be beneficial as it can allow physicians to consider the patient’s overall health status when determining treatment pathways.
A limitation of this study is the relatively small age range of participants. Given that fracture risk increases as a function of age, this led to a narrow range of fracture risk estimates in the population, which could have led to the overall lower AUC values of both 
FRAC and the reference models (FRAX and FN aBMD). Additionally, the advanced age of participants in this cohort resulted in most individuals being classified as high risk of fracture. Typically designing a new diagnostic tool to identify high-risk individuals is less of a concern than those with moderate risk where diagnoses is not as clear. Future studies should investigate the performance of 
FRAC in a broader range of fracture risk groups, particularly to demonstrate its value at clinical decision making for those who are not clearly high- or low-risk. Another factor to consider is race, the data collection for this study came from a single study center in Sweden with the majority of individuals being White. Therefore, future validation needs to be undertaken in multi-race cohorts in other geographic locations. Other researchers can explore this by using 
FRAC through the openly available online platform (www.normative.ca), providing the opportunity for independent external validation. Finally, although participants were followed for up to 10-yr, the majority of individuals who experience an osteoporotic fracture occurred within the first 5 yr after baseline imaging. Therefore, future research should validate the tool in a younger cohort, as they are more likely to sustain fractures over a longer term. However, in this older cohort there is a potential benefit of short-term fracture risk in older individuals.
In summary, 
FRAC demonstrated consistent performance with FRAX and FN aBMD at predicting MOF and AOF, and good generalizability to an external female cohort. The validation revealed that 
FRAC showed similar performance at predicting 5- and 10-yr risk of MOF for all model variants, particularly with the removal of incident vertebral fracture. An interesting finding was that 
FRAC showed a marked improvement in predicting 5-yr risk of radius fracture, which could be useful for identifying early signs of bone fragility. Overall, we envision 
FRAC not to be used stand-alone, but as a secondary screening tool for osteoporosis fracture risk, particularly in the case where standard, static binary risk indicators may not adequately capture complex medical history. This validation of 
FRAC serves as an important step toward the potential clinical translation of 
FRAC for fracture risk prediction.
Contributor Information
Annabel R Bugbird, McCaig Institute for Bone and Joint Health, Cumming School of Medicine, University of Calgary, Calgary AB, T2N 4Z6, Canada; Department of Biomedical Engineering, Schulich School of Engineering, University of Calgary, Calgary AB, T2N 1N4, Canada.
Raju Jaiswal, Sahlgrenska Osteoporosis Centre, Department of Internal Medicine and Clinical Nutrition, Institute of Medicine, University of Gothenburg, SE-413 45 Gothenburg, Sweden.
Danielle E Whittier, McCaig Institute for Bone and Joint Health, Cumming School of Medicine, University of Calgary, Calgary AB, T2N 4Z6, Canada; Department of Cell Biology and Anatomy, Cumming School of Medicine, University of Calgary, Calgary AB, T2N 4N1, Canada.
Steven K Boyd, McCaig Institute for Bone and Joint Health, Cumming School of Medicine, University of Calgary, Calgary AB, T2N 4Z6, Canada; Department of Biomedical Engineering, Schulich School of Engineering, University of Calgary, Calgary AB, T2N 1N4, Canada; Department of Radiology, Cumming School of Medicine, University of Calgary, Calgary AB, T2N 2T9, Canada.
Mattias Lorentzon, Sahlgrenska Osteoporosis Centre, Department of Internal Medicine and Clinical Nutrition, Institute of Medicine, University of Gothenburg, SE-413 45 Gothenburg, Sweden; Region Västra Götaland, Department of Geriatric Medicine, Sahlgrenska University Hospital, SE-431 80, Mölndal, Sweden.
Author contributions
Annabel R. Bugbird (Conceptualization, Formal analysis, Methodology, Writing—original draft), Raju Jaiswal (Data curation, Formal analysis, Writing—review & editing), Danielle Elizabeth Whittier (Conceptualization, Supervision, Writing—review & editing), Steven Kyle Boyd (Conceptualization, Supervision, Writing—review & editing), and Mattias Lorentzon (Data curation, Funding acquisition, Writing—review & editing)
Funding
The SUPERB study was funded by the Swedish Research Council Grant 2017-02229, the Swedish state under the agreement between the Swedish government and the county councils, the ALF-agreement (ALFGBG-716051), and by a grant from IngaBritt och Arne Lundbergs Forskningsstiftelse. This study was funded by the Canadian Institutes of Health Research (CIHR) PJT 178098 and the Bob and Nola Rintoul Chair in Bone and Joint Research. The Bone Microarchitecture International Consortium (BoMIC) was supported by an international team of researchers who contributed data from the Framingham Study and Mayo Clinic in the United States, QUALYOR, STRAMBO, and OFELY in France, GERICO in Switzerland and CaMos in Canada. Funding support was from the National Institute of Arthritis Musculoskeletal and Skin Diseases of the National Institutes of Health (R01AR061445 and AR027065), the National Heart, Lung, and Blood Institute Framingham Heart Study (N01-HC-25195, HHSN268201500001I), and research grants from the Investigator Initiated Studies Program of Merck Sharp & Dohme as well as Amgen.
Conflicts of interest
D.E.W. has received speaker honoraria from Amgen. M.L. has received lecture or consulting fees from Astellas, Amgen, UCB Pharma, Medison Pharma, Jansen-Cilag, Viatris, Medac, and Parexel International, Pharmacosmos, Sandoz, Alexion, all outside the submitted work. A.R.B., R.J., and S.K.B. declare no conflicts of interest.
Data availability
Data cannot be made publicly available for ethical and legal reasons. Such information is subject to legal restrictions according to national legislation. Specifically, in Sweden confidentiality regarding personal information in studies is regulated in the Public Access to Information and Secrecy Act (SFS 2009:400). The data underlying the results of this study might be made available upon reasonable request, after an assessment of confidentiality.
References
- 1. Srinivasan B, Kopperdahl DL, Amin S, et al. Relationship of femoral neck areal bone mineral density to volumetric bone mineral density, bone size, and femoral strength in men and women. Osteoporos Int. 2012;23(1):155–162. 10.1007/s00198-011-1822-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Fuggle NR, Curtis EM, Ward KA, Harvey NC, Dennison EM, Cooper C. Fracture prediction, imaging and screening in osteoporosis. Nat Rev Endocrinol. 2019;15(9):535–547. 10.1038/s41574-019-0220-8 [DOI] [PubMed] [Google Scholar]
- 3. Fuleihan GE-H, Chakhtoura M, Cauley JA, Chamoun N. Worldwide fracture prediction. J Clin Densitom. 2017;20(3):397–424. 10.1016/j.jocd.2017.06.008 [DOI] [PubMed] [Google Scholar]
- 4. Rubin KH, Friis-Holmberg T, Hermann AP, Abrahamsen B, Brixen K. Risk assessment tools to identify women with increased risk of osteoporotic fracture: complexity or simplicity? A systematic review. J Bone Miner Res. 2013;28(8):1701–1717. [DOI] [PubMed] [Google Scholar]
-
5.
Kanis JA, Oden A, Johansson H, Borgström F, Ström O, McCloskey E. Frax
and its applications to clinical practice. Bone. 2009;44(5):734–743. 10.1016/j.bone.2009.01.373
[DOI] [PubMed] [Google Scholar] - 6. Agarwal A, Leslie W, Nguyen T, Morin S, Lix L, Eisman J. Predictive performance of the Garvan Fracture Risk Calculator: a registry-based cohort study. Osteoporos Int. 2022;1–8 [DOI] [PubMed] [Google Scholar]
- 7. Whittier DE, Samelson EJ, Hannan MT, et al. A fracture risk assessment tool for high resolution peripheral quantitative computed tomography. J Bone Miner Res. 2023;38(9):1234–1244. 10.1002/jbmr.4808 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Samelson EJ, Broe KE, Xu H, et al. Cortical and trabecular bone microarchitecture as an independent predictor of incident fracture risk in older women and men in the bone microarchitecture international consortium (BoMIC): a prospective study. Lancet Diabetes Endocrinol. 2019;7(1):34–43. 10.1016/S2213-8587(18)30308-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Johansson L, Johansson H, Axelsson KF, et al. Improved fracture risk prediction by adding VFA-identified vertebral fracture data to BMD by DXA and clinical risk factors used in FRAX. Osteoporos Int. 2022;33(8):1725–1738. 10.1007/s00198-022-06387-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Larsson BA, Johansson L, Johansson H, et al. The timed up and go test predicts fracture risk in older women independently of clinical risk factors and bone mineral density. Osteoporos Int. 2021;32(1):75–84. 10.1007/s00198-020-05681-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Larsson BA, Johansson L, Mellström D, et al. One leg standing time predicts fracture risk in older women independent of clinical risk factors and BMD. Osteoporos Int. 2022;1–10. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Sundh D, Nilsson A, Nilsson M, Johansson L, Mellström D, Lorentzon M. Increased cortical porosity in women with hip fracture. J Intern Med. 2017;281(5):496–506. 10.1111/joim.12587 [DOI] [PubMed] [Google Scholar]
- 13. Johansson L, Sundh D, Magnusson P, et al. Grade 1 vertebral fractures identified by densitometric lateral spine imaging predict incident major osteoporotic fracture independently of clinical risk factors and bone mineral density in older women. J Bone Miner Res. 2020;35(10):1942–1951. 10.1002/jbmr.4108 [DOI] [PubMed] [Google Scholar]
- 14. Whittier DE, Boyd SK, Burghardt AJ, et al. Guidelines for the assessment of bone density and microarchitecture in vivo using high-resolution peripheral quantitative computed tomography. Osteoporos Int. 2020;31(9):1607–1627. 10.1007/s00198-020-05438-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Whittier D, Mudryk A, Vandergaag I, Burt L, Boyd S. Optimizing HR-pQCT workflow: a comparison of bias and precision error for quantitative bone analysis. Osteoporos Int. 2020;31(3):567–576. 10.1007/s00198-019-05214-0 [DOI] [PubMed] [Google Scholar]
- 16. Pistoia W, Van Rietbergen B, Lochmüller E-M, Lill C, Eckstein F, Rüegsegger P. Estimation of distal radius failure load with micro-finite element analysis models based on three-dimensional peripheral quantitative computed tomography images. Bone. 2002;30(6):842–848. 10.1016/S8756-3282(02)00736-6 [DOI] [PubMed] [Google Scholar]
- 17. Whittier DE, Manske SL, Kiel DP, Bouxsein M, Boyd SK. Harmonizing finite element modelling for non-invasive strength estimation by high-resolution peripheral quantitative computed tomography. J Biomech. 2018;80:63–71. 10.1016/j.jbiomech.2018.08.030 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Fan B, Lu Y, Genant H, Fuerst T, Shepherd J. Does standardized BMD still remove differences between Hologic and GE-Lunar state-of-the-art DXA systems? Osteoporos Int. 2010;21(7):1227–1236. 10.1007/s00198-009-1062-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Schini M, Johansson H, Harvey N, Lorentzon M, Kanis J, McCloskey E. An overview of the use of the fracture risk assessment tool (FRAX) in osteoporosis. J Endocrinol Investig. 2024;47(3):501–511. 10.1007/s40618-023-02219-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Cosman F, de Beur SJ, LeBoff M, et al. Clinician’s guide to prevention and treatment of osteoporosis. Osteoporos Int. 2014;25(10):2359–2381. 10.1007/s00198-014-2794-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Morin SN, Feldman S, Funnell L, et al. Clinical practice guideline for management of osteoporosis and fracture prevention in Canada: 2023 update. CMAJ. 2023;195(39):E1333-E1348. 10.1503/cmaj.221647 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Gregson CL, Armstrong DJ, Bowden J, et al. Uk clinical guideline for the prevention and treatment of osteoporosis. Arch Osteoporos. 2022;17(1):58. 10.1007/s11657-022-01061-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Kanis JA, McCloskey EV, Johansson H, et al. European guidance for the diagnosis and management of osteoporosis in postmenopausal women. Osteoporos Int. 2013;24(1):23–57. 10.1007/s00198-012-2074-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Tukey JW. Comparing individual means in the analysis of variance. Biometrics. 1949;5(2):99–114. 10.2307/3001913 [DOI] [PubMed] [Google Scholar]
- 25. Janes H, Pepe MS, Gu W. Assessing the value of risk predictions by using risk stratification tables. Ann Intern Med. 2008;149(10):751–760. 10.7326/0003-4819-149-10-200811180-00009 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Gregson CL, Armstrong DJ, Bowden J, et al. UK clinical guideline for the prevention and treatment of osteoporosis. Arch Osteoporos. 2022;17(1):58. 10.1007/s11657-022-01061-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Austin PC, Fine JP. Practical recommendations for reporting Fine-Gray model analyses for competing risk data. Stat Med. 2017;36(27):4391–4400. 10.1002/sim.7501 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Jaiswal R, Pivodic A, Zoulakis M, et al. Prediction of hip fracture by high-resolution peripheral quantitative computed tomography in older Swedish women. J Bone Miner Res. 2025;zjaf020. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Chapurlat R, Ferrari S, Li X, et al. Deep learning using high-resolution images of forearm predicts fracture. eLife. 2023;12:RP87990. 10.7554/eLife.87990.1 [DOI] [Google Scholar]
- 30. Lu S, Fuggle NR, Westbury LD, et al. Machine learning applied to HR-pQCT images improves fracture discrimination provided by DXA and clinical risk factors. Bone. 2023;168:116653. 10.1016/j.bone.2022.116653 [DOI] [PubMed] [Google Scholar]
- 31. Daruwalla ZJ, Huq SS, Wong KL, et al. Hip fractures, preceding distal radius fractures and screening for osteoporosis: should we be screening earlier? A minimum 10-year retrospective cohort study at a single Centre. Osteoporos Int. 2016;27(1):361–366. 10.1007/s00198-015-3375-8 [DOI] [PubMed] [Google Scholar]
- 32. Chen C-W, Huang T-L, Su L-T, et al. Incidence of subsequent hip fractures is significantly increased within the first month after distal radius fracture in patients older than 60 years. J Trauma Acute Care Surg. 2013;74(1):317–321. 10.1097/TA.0b013e31824bb325 [DOI] [PubMed] [Google Scholar]
- 33. Choi H-G, Kim D-S, Lee B, Youk H, Lee J-W. High risk of hip and spinal fractures after distal radius fracture: a longitudinal follow-up study using a national sample cohort. Int J Environ Res Public Health. 2021;18(14):7391. 10.3390/ijerph18147391 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Horton TG, Richardson TLJr, Hackstadt AJ, et al. Validation of an algorithm to identify fractures among patients within the veterans health administration. Pharmacoepidemiol Drug Saf. 2023;32(11):1290–1298. 10.1002/pds.5662 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Kristensen SB, Clausen A, Skjødt MK, et al. An enhanced version of FREM (fracture risk evaluation model) using national administrative health data: analysis protocol for development and validation of a multivariable prediction model. Diagn Progn Res. 2023;7(1):19. 10.1186/s41512-023-00158-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Schousboe JT, Langsetmo L, Szulc P, et al. Joint associations of prevalent radiographic vertebral fracture and abdominal aortic calcification with incident hip, major osteoporotic, and clinical vertebral fractures. J Bone Miner Res. 2020;36(5):892–900. 10.1002/jbmr.4257 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37. Lorentzon M, Burghardt AJ. The added value of high-resolution peripheral quantitative computed tomography in fracture risk prediction. J Bone Miner Res. 2023;38(9):1225–1226. 10.1002/jbmr.4909 [DOI] [PubMed] [Google Scholar]
- 38. Zarzour F, Leslie WD. The effect of smoking cessation versus current smoking on fracture risk: the Manitoba BMD registry. J Clin Densitom. 2024;27(4):101523. 10.1016/j.jocd.2024.101523 [DOI] [PubMed] [Google Scholar]
- 39. Mesinovic J, Breasail MO, Burt LA, et al. Bone imaging modality precision and agreement between DXA, pQCT, and HR-pQCT. JBMR plus. 2025;9(2):ziae158. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Kemp T, de Bakker C, Gabel L, et al. Longitudinal bone microarchitectural changes are best detected using image registration. Osteoporos Int. 2020;31(10):1995–2005. 10.1007/s00198-020-05449-2 [DOI] [PubMed] [Google Scholar]
-
41.
Bugbird AR, Whittier DE, Boyd SK. Transferability of bone phenotyping and fracture risk assessment by
FRAC from first-generation high-resolution peripheral quantitative computed tomography to second-generation scan data. J Bone Miner Res. 2024;39(5):571–579. 10.1093/jbmr/zjae039
[DOI] [PMC free article] [PubMed] [Google Scholar] - 42. Pepe MS, Fan J, Feng Z, Gerds T, Hilden J. The net reclassification index (NRI): a misleading measure of prediction improvement even with independent test data sets. Stat Biosci. 2015;7(2):282–295. 10.1007/s12561-014-9118-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43. Kerr KF. Net reclassification index statistics do not help assess new risk models. Radiology. 2022;306(3):e222343. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44. Liang W, Chikritzhs T. The effect of age on fracture risk: a population-based cohort study. J Aging Res. 2016;2016(1):5071438. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
Data cannot be made publicly available for ethical and legal reasons. Such information is subject to legal restrictions according to national legislation. Specifically, in Sweden confidentiality regarding personal information in studies is regulated in the Public Access to Information and Secrecy Act (SFS 2009:400). The data underlying the results of this study might be made available upon reasonable request, after an assessment of confidentiality.