A genome-wide genomic score added to standard recommended stratification tools does not improve the identification of patients with very low bone mineral density

This cross-sectional sub-study of the multicenter Danish Study of Non-Invasive Diagnostic Testing in Coronary Artery Disease (Dan-NICAD) included participants (n=1675) referred to a cardiac CT between September 2014 and March 2016 at two centers. All participants had an indication for a coronary CT angiography, a low to intermediate risk profile of coronary artery disease (CAD), and did not have a history of CAD. Inclusion and exclusion criteria have previously been reported [19]. All participants gave written, informed consent before enrolment.

The Central Denmark Regional Committee on Health Research Ethics (date: 2014-07-30, ref.: 1-10-72-190-14) and the Danish Data Protection Agency (ref.: 1-16-02-345-14) approved this study. The trial was registered at ClinicalTrials.​gov (NCT02264717). The study followed the principles of the World Medical Association Declaration of Helsinki – Ethical Principles for Medical Research Involving Human Subjects.

All participants were referred for routine cardiac CT on the clinical indication of CAD. Using the non-enhanced cardiac CT, a mean BMD was analyzed by an experienced reader (Josephine Therkildsen (JT)) blinded to participant data for every participant based on three consecutive thoracic vertebrae beginning from the left main coronary artery. The selected thoracic vertebrae have previously been demonstrated to include Th6 (11%), Th7 (48%), Th8 (37%), and Th9 (4%) [20]. The Mindways QCT Pro software (Mindways Software Inc, TX) was used to measure BMD together with the calibration phantom, Mindways Solid 3 (Mindways Software Inc, TX) as previously described [5]. The mean BMD was grouped as very low (<80 mg/cm³), low (80–120 mg/cm³), and normal BMD (>120 mg/cm³) as described by the American College of Radiology (ACR) for lumbar spine QCT [21].

During interviews at the baseline visit and from electronic patient records, trained study nurses obtained information on self-reported osteoporosis risk factors including age, sex, smoking status, alcohol history, and family predisposition for osteoporosis (1-degree relatives having osteoporosis or a history of hip fractures). Anthropometric measurements, such as weight and height, were also obtained.

Danish Patient Registries were used to collect history of previous fractures, both any type of fracture and osteoporosis-related fractures (hip, spine, forearm, and proximal humerus fracture), use of glucocorticoids and anti-osteoporotic treatment, and known diseases to calculate Charlson comorbidity index as previously reported [4].

Using the Bone Health and Osteoporosis Foundation (BHOF), formerly known as the National Osteoporosis Foundation, recommendations for BMD testing, participants were divided into an “indication for BMD testing” and a “no indication for BMD testing” group. The categorization was based on [1] age + sex (age cutoff in women ≥ 65 years and men ≥ 70 years irrespective of osteoporosis risk factors) and [2] age + sex + clinical risk factors (postmenopausal women and men older than age 50 years with clinical risk factors which here includes body mass index<20 kg/m², previous osteoporosis-related fracture, family history of osteoporosis, active smoking, glucocorticoid treatment within the previous 6 months, and secondary osteoporosis (here defined as diabetes mellitus type 1, early menopause < 45 years of age or Charlson comorbidity index ≥2)) [22].

The FRAX tool estimates the 10-year probability of fracture (hip and major osteoporotic fracture) (https://​www.​sheffield.​ac.​uk/​FRAX/​tool.​aspx?​country=​29). A previous osteoporosis-related fracture was defined as a previous fracture at the hip, spine, proximal humerus, or forearm registered in the National Danish Patient registry using the 10th edition of the International Classification of Diseases [23] as previously reported [4]. Participants were asked if they were predisposed to osteoporosis, including 1-degree relatives having known osteoporosis or a history of hip fractures. Current systemic glucocorticoid use was defined as a redeemed prescription within 6 months prior to the baseline visit. All participants with self-reported or known rheumatoid arthritis, Crohn’s disease, colitis ulcerosa, osteogenesis imperfecta, severe scoliosis, Scheuermann’s disease, lymphoma, leukemia, or myelomatosis were excluded from BMD analyses; thus, these risk factors were considered as not present for the entire cohort. Secondary osteoporosis was defined as diseases strongly associated with osteoporosis (type 1 diabetes mellitus, early menopause (<45 years), or Charlson comorbidity index ≥ 2). Consumption of alcohol in units/day was not recorded; however, participants with chronic alcoholism were excluded from BMD analyses, and so the risk of alcohol consumption ≥ 3 units/day was considered as not present for the entire cohort. In general, there were a few missing values for osteoporosis risk factors (Table 1 legends). If a participant had a missing risk factor, the FRAX score was recorded as missing (n=32). Seven participants had a weight above 125 kg, and as FRAX only allow a maximum weight of 125 kg to be included, these participants were regarded as having a weight of 125 kg.

All participants had blood samples collected (4 mL EDTA whole blood) at baseline for genotyping. DNA was extracted, and only high-quality DNA was used for genotyping using the Illumina Global Screening Array (Illumina, Inc., 5200 Illumina Way, San Diego, CA) at deCODE genetics (deCODE genetics, Inc., Reykjavik, Iceland). Quality control was performed with Plink 1.9 [24] removing samples and SNPs with missingness <0.02, sex discrepancy, minor allele frequency <0.01, and deviation from Hardy-Weinberg equilibrium (p<1 × 10⁻⁶), which resulted in a total of 310,531 autosomal SNPs. Furthermore, first and second-degree relatives were identified based on identity by descent (\(\hat{\pi}\)>0.2), and ancestral outliers were identified by multidimensional scaling, which in total removed 27 samples. Genotypes were imputed using the Michigan Imputation Server [25] (reference panel HRC: r1.1.2016 GRCh37). SNPs with a minor allele frequency below 0.01, missing genotype rate >5%, deviation from Hardy-Weinberg equilibrium (p<1 × 10⁻⁶), imputation quality below 0.9, and SNPs located within the major histocompatibility complex (MHC; 6p22.1–21.3: chr6:28,477,797-33,448,354, GRCh37) were removed, resulting in a total of 6,565,335 autosomal genetic variants, with positions annotated to genome build GRCh37.

The GS was computed based on a GWAS meta-analysis assessing total body (TB) BMD (GS_{TB_BMD}) as measured by dual-energy X-ray absorptiometry (DXA) in 66,628 individuals compromising of 86 % European ancestry, 2% African American, and 14% from admixed background [17].

The GS_{TB_BMD} was predefined as the primary GS based on an available full GWAS summary statistic.

Two other GSs (GS_{FN_BMD} and GS_{HQ_SOS}) were constructed and compared with the primary GS_{TB_BMD} (Supplementary Table S1). The summary statistics for the GS_{FN_BMD} were based on a GWAS meta-analysis assessing femoral neck (FN) BMD measured by DXA in 32,961 individuals with populations from both Europe, America, East Asia, and Australia [14]. Only 54 SNPs associated with BMD were publicly available. Finally, the GS_{HQ_SOS} was based on 11,709 SNPs from a study assessing heel quantitative ultrasound speed of sound (HQ_SOS) in 341,449 individuals mainly of white British ancestry [15].

The GSs were computed based on genetic markers pruned for linkage disequilibrium (r²<0.9) as \(GS=\sum_{i=1}^m{X}_i{\hat{\beta}}_i\), where X_i is the ith genotype encoded as 0, 1, or 2 counting the number of the first allele in the bim-file, and \({\hat{\beta}}_i\) is the marker effects of the ith marker. Each GS was computed using different SNP P value thresholds (0.001, 0.01, 0.1, 0.5, 0.7, 0.9) to find the optimal GS performance. Processing of the genetic data and computing the GS were performed with the R package qgg [26, 27].

Participants were divided into three groups based on their GS_{TB_BMD}; low (0–20% percentile), average (20–80% percentile), and high (80–100% percentile) comparable to previous reported results [28].

All GS scores were standardized to a mean of zero and a standard deviation (SD) of one. Demographics at baseline were compared by GS groups (low, average, and high) using a one-way analysis of variance, Kruskal-Wallis test, or Fischer’s exact test. Data are presented as a number with a percentage for categorical variables, as a mean with SD for normally distributed continuous variables, or as a median with an interquartile range for not normally distributed continuous variables.

The correlations between the GS and mean BMD, together with the GS and osteoporosis risk factors, were estimated using Pearson correlation, Spearman correlation, or logistic regression coefficients.

Four different logistic regression models were used to estimate odds ratio (OR) (and 95% confidence intervals) for having very low BMD (y_BMD, defined as mean BMD <80 mg/cm³ as described by ACR [21]):

Model (2 and 3) was based on a combined model for clinical risk factors (CRF) which included: age, sex, BMI, active smoker, osteoporosis disposition, previous osteoporosis-related fracture, glucocorticoid treatment, and secondary osteoporosis. The GS and BMI were multiplied with −1, as an inverse association with mean BMD was present and statistical tests assume higher values indicating higher risk. The ORs presented in Table 2 are per 1 SD in GS standardized to the current dataset using z scores.

Using logistic regression, the CRF and CRF-GS models were compared using the likelihood ratio test. We evaluated each model’s ability to discriminate participants having very low BMD using area under the receiver operating characteristic curve (AUC). Furthermore, we evaluated the ability of reclassifying participants using the CRF model vs. the CRF+GS model using pre-test probability cutoff values of 6%, 12%, and 24% [29]. These cutoff values were based on the prevalence (p) of having very low BMD (12%) in the total cohort, by calculating p/2, p, and 2p, as recommended when clinical cutoff values do not exist [30]. The net reclassification index (NRI) and integrated discrimination improvement (IDI) were calculated. In order to internally validate the CRF and CRF+GS model, we compared AUCs with the FRAX and FRAX+GS model (not fitted to the current study [31]) and tested AUCs using 5-fold cross-validation [32]. A P value<0.05 was considered statistically significant and in multiple comparisons, P values were adjusted according to the Bonferroni adjustment (P value/n_tests) when relevant. Sample size calculation has previously been published [19]. All analyses were performed using STATA/IC 17.0 (StataCorp, College Station, Tex).

In total, 1421 participants (672 men, 749 women) had BMD measurements and GS available. The participant flowchart is displayed in Fig. 1. The mean age for the total population was 57 ± 9 years (range 40–79 years). Baseline characteristics stratified by GS groups are shown in Table 1. Baseline characteristics stratified by clinical indication for BMD testing are presented in Supplementary Table S2.

Participants having low GS (0–20%) were more likely to have a family history of osteoporosis (P<0.01), known osteoporosis (P<0.001), to take anti-osteoporotic medication (P<0.05), have very low BMD (P<0.01), and a higher median FRAX_MOF score (P<0.05) compared to the average (20–80%) and high (80–100%) GS group (Table 1). Furthermore, subdividing GS showed that extremes affected mean BMD and prevalence of very low BMD (Supplementary Figure S1 and S2).

Assessing three different genomic scores, GS_{TB_BMD} and GS_{HQ_SOS} showed significant associations to mean BMD but the GS_{FN_BMD} did not (Supplementary Table S1). Overall, a greater GS_{TB_BMD} was associated with higher BMD in both men and women (Fig. 2) with a 3 % of the variance in the mean BMD explained by the GS. The GS was not statistically significantly correlated with osteoporosis risk factors, except for family predisposition in women likely capturing the inherited risk (Supplementary Table S3). GS and mean BMD were overall correlated across age and BMD groups, suggesting its predictive value irrespectively of age and BMD groups (Supplementary Table S4).

The GS showed a similar ability in identifying participants having very low BMD compared to standard clinical risk factors as displayed by the AUC in the total cohort and in participants with indication for BMD testing (Fig. 3A and Supplementary Figure S3). In the total cohort, the AUC for the GS was 0.57 (95% CI 0.52–0.61) not statistically different from the AUCs for individual risk factors (P=0.39). However, in participants without indication for BMD testing (n=841), the AUC for the GS alone (0.60 [95% CI 0.52–0.69]) performed above some individual clinical risk factors (sex, active smoking, and secondary osteoporosis, all P<0.0063) and equally for previous osteoporosis-related fracture, family predisposition, BMI, and glucocorticoid treatment (all P>0.0063); Bonferroni correction (0.05/8=P<0.0063) (Fig. 3B).

Per one SD decrease in GS, the odds of having very low BMD were increased by 34% (Table 2). Subgroup analyses including women vs men showed similar results (Table 2). In participants without indication for BMD testing, OR for very low BMD per SD decrease was 1.56 (95% CI 1.18–2.07, P<0.01) compared to 1.24 (95% CI 1.02–1.50, P<0.05) in participants with an indication for BMD testing.

No improvements in net reclassification nor the integrated discrimination were detected when adding the GS to the CRF model using predefined probability cutoff values of 6%, 12%, and 24% in participants without an indication for BMD testing (Table 3). We observed a few individuals without very low BMD being misclassified using the CRF+GS model (n=3) (Table 3). Furthermore, no improvements were detected in participants with an indication for BMD testing or for the total cohort (Supplementary Table S5 and S6).

To assess the possible benefits of using GS for risk stratification, we assessed the effect of including the GS in the stratification algorithm as recommended by BHOF (Supplementary Figure S4). In total, 841 individuals did not have an indication for BMD testing based on standard stratification tools; however, 54 cases with very low BMD (6%) fell into this category. A number of cases that would be screened if incorporating the low GS in risk stratification are n=150, with a possibility to identify 17 more cases with very low BMD.

The likelihood-ratio test showed a significant improvement of adding the GS to clinical risk factors (P<0.001). Subsequently, a CRF model including clinical risk factors (age, sex, BMI, active smoking, family history of osteoporosis, previous osteoporosis-related fracture, glucocorticoid treatment, and secondary osteoporosis) was developed to predict very low BMD using all available participants. Analysis of the model’s ability to discriminate very low BMD showed an AUC of 0.78 (95% CI 0.74–0.81). Next, a CRF+GS model was developed adding the GS to the clinical risk factor (CRF+GS) model resulting in an AUC of 0.78 (95% CI 0.75–0.81). Directly comparing AUCs did not show statistically significant improvement from the AUC for the CRF model alone (Supplementary Figure S3).

In participants without indication for BMD testing, adding GS to a clinical risk factor (CRF) model resulted in an AUC of 0.80 (95% CI 0.75–0.85) not statistically significant from the CRF model alone. AUCs stratified by indication for BMD testing are shown in Fig. 3.

Internal validation of the CRF and CRF+GS model is shown in Supplementary Figure S5. The models overall showed good calibration, although the CRF+GS model showed a tendency of better calibration at higher risks compared to the CRF model alone.

We computed the AUCs for the FRAX_MOF and a FRAX_MOF+GS model (0.73 [95% CI 0.69–0.77] and 0.72 [95% CI 0.68–0.76]), including the same risk factors as the CRF model, but without being fitted to the current study population (Supplementary Figure S3).

A five-fold cross-validation revealed AUCs for the CRF and CRF+GS model of 0.77 (95% CI 0.73–0.81) and 0.77 (95% CI 0.73–0.81), suggesting a small over-fitting but without affecting the difference in AUCs between the models.

The strengths of this study include the consecutive inclusion of a homogenous study population of both men and women at two centers. Participants had free access to medical attention and were referred to cardiac CT by experienced cardiologists, limiting possible selection bias. All participants had relevant and extensive osteoporosis risk factors collected at their baseline visit and a trained reader (JT) performed BMD analyses and calculated FRAX for all participants. BHOF releases comprehensive recommendations to guide clinicians to prevent and treat osteoporosis worldwide, including guidelines for BMD testing [22], and despite country-specific guidelines, we used these recommendations for a stratification algorithm to assess the GS in a clinical setting.

Limitations include the following: [1] Only participants with European ancestry referred to cardiac CT were included; thus, the results from this study may not be extrapolated to other cohorts. [2] FRAX includes previous fractures (defined as spontaneous or low-energy fractures in adult life). However, we defined a previous osteoporosis-related fracture based on diagnostic codes from Danish registries recorded at Danish hospitals since 1995. These registries do not differentiate low- from high-energy fractures, nor could we include non-clinical fractures, such as vertebral fractures only detected on imaging or vertebral fractures without imaging. Participants were asked if they were predisposed to osteoporosis, including 1-degree relatives having osteoporosis or a history of hip fractures, and this was entered into the FRAX tool. Thus, a history alone describing parental hip fracture was not recorded despite being used in the FRAX tool. Current systemic glucocorticoid use was defined as a redeemed prescription recorded 6 months prior to the baseline visit; however, the FRAX tool defines use as either currently using or exposed for 3 months of minimum 5 mg daily. Secondary osteoporosis was here defined as type 1 diabetes mellitus, early menopause (<45 years), or a Charlson comorbidity index ≥ 2. Although the FRAX tool does not include a Charlson comorbidity index in its users’ guide, this was chosen to include diseases with a known increased fracture risk [41] comparable to the clinical risk factor model. [3] Lumbar and thoracic spine BMD values show high correlations [20], but currently, only lumbar QCT BMD is recommended [43]. CT-measured thoracic BMD is a surrogate marker of fracture risk, currently not included in clinical guidelines to diagnose osteoporosis [20, 43]. Despite BMD being the strongest individual predictor of fracture risk, the developed models’ ability to predict future fractures needs to be assessed. In the lack of validated thoracic BMD cutoff values, we chose ACR’s recommended lumbar spine BMD cutoff values for very low and low BMD. This likely underestimates the very low and low BMD group due to physiological differences in BMD values between the thoracic and lumbar spine [44]. [4] In the lack of relevant clinical cutoff values, predefined cutoff values of 6%, 12%, and 24% were chosen according to the calculated p/2, p, and 2p as described by Cook et al. (2011) [30]. The results of the reclassification analyses were dependent on these cutoff values.

We performed cross-validation, as model synthesis tested in the original dataset tends to optimistic estimate the predictive performance [45]. Our analysis showed a more realistic estimate with the tendency of a slight over-fitting, however not affecting the overall results. As our models were developed and assessed in the same study population, there is a need to externally validate the results of this study in an independent cohort to fully assess the generalizability of the results.

In conclusion, adding a GS to a clinical risk factor model did not improve the overall prediction of very low BMD. However, we noticed that the GS performed equally or above individual risk factors in discrimination. Clinical risk factors combined showed superior discrimination to individual risk factors and the GS, underlining the value of combined CRFs in routine clinics as a stratification tool.