Introduction

There is mounting recognition that “diabetic bone disease” [1] and fractures are end-organ complications of diabetes mellitus (DM) [2,3,4,5]. Although DM is often associated with higher bone mineral density (BMD) [5], one critical feature of diabetic bone disease is diminished bone quality [1]. One bone quality parameter that is altered in DM is trabecular microarchitecture, which can be indirectly assessed using trabecular bone score (TBS). TBS is a textural parameter derived from DXA-based lumbar spine (LS) images and correlates with HR-qQCT measurements of trabecular microstructure [6]. Indeed, the pathophysiologic importance of impaired trabecular microarchitecture in diabetic bone disease is supported by studies showing that TBS is lower in diabetics versus non-diabetics [7,8,9,10,11,12,13], and in persons with DM, lower TBS predicts fracture risk independent of BMD [8, 14].

Although the negative relation between DM and trabecular microarchitecture is well-studied [7,8,9,10,11,12,13], whether prediabetes is similarly associated with lower TBS is less certain [10, 11, 15]. This knowledge is important because it could shed light on whether prediabetes has clinical implications beyond increased risk for developing DM. Clinicians disagree on how aggressively to treat prediabetes [16, 17], in part, because not every prediabetic develops diabetes, and prediabetes on its own has not been definitively linked to end-organ complications. To address this knowledge gap, we set out to determine if prediabetes is associated with lower TBS. Because a rise in insulin resistance precedes the onset of prediabetes and DM, our second objective was to examine whether greater insulin resistance is related to lower TBS, even in the absence of DM.

To accomplish our objectives, we conducted a cross-sectional analysis of participants in the Trabecular Bone Score (TBS) Study, which was done as part of the Study of Women’s Health Across the Nation (SWAN). SWAN is a cohort study of the menopause transition in a multi-racial/ethnic community sample. For this analysis, we used data from the SWAN TBS Study baseline visit, at which time participants were between 42 and 52 years.

Methods

SWAN is a multi-center, longitudinal study of 3,302 diverse, community-dwelling women. At study inception, participants were between 42 and 52 years, and in premenopause (no change from usual menstrual bleeding) or early perimenopause (less predictable menstrual bleeding at least once every three months). Potential volunteers were excluded if they did not have an intact uterus and at least one ovary or were using sex steroid hormones. A total of seven clinical sites recruited study participants: Boston, Chicago, Detroit, Pittsburgh, Los Angeles, Newark, and Oakland. The SWAN Bone Cohort was composed of 2,407 women from five sites (excludes Chicago and Newark, where BMD assessments were not performed). In turn, the SWAN TBS Study included 1,436 women from three Bone Cohort sites (Boston, Detroit, and Los Angeles). Pittsburgh and Oakland were not TBS Study sites because TBS could not be measured using scans acquired from their older DXA machines. Participants provided written informed consent, and each site obtained institutional review board approval.

Samples

Of the 1,436 women from the SWAN TBS Study, we excluded those who reported taking either bone-beneficial medications (hormone therapy, calcitonin, calcitriol, bisphosphonates, denosumab, or parathyroid hormone) or bone-detrimental medications (oral or injectable glucocorticoids, aromatase inhibitors, gonadotropin releasing hormone agonists, or anti-epileptic medications) at the time of the first TBS measurement (N = 61), women with DM (fasting glucose ≥ 126 mg/dL, or use of DM medications [metformin, sulfonylurea, meglitinide, thiazolidinedione, DPP-IV inhibitor, GLP-agonist, insulin]) (N = 70) and those for whom HOMA-IR could not be calculated (N = 58). This left us with a sample of 1,248 women.

Outcomes

The outcome in analyses was TBS, a textural parameter derived from lumbar spine (LS) BMD scans. SWAN TBS Study sites measured LS BMD using Hologic 4500A instruments. For the LS region of interest, vertebrae were excluded if local structural change or artifact was visualized. In addition, anatomically abnormal vertebrae were removed if there was a > 1.0 T-score difference between the vertebra in question and adjacent vertebrae. A standard BMD quality-control program, conducted in collaboration with Synarc, Inc. (Newark, CA), included daily phantom measurements, SWAN site cross-calibration with a circulating anthropomorphic spine standard, local site review of all scans, and central review of scans that met problem-flagging criteria.

The SWAN TBS Study baseline data consists of TBS values calculated from the first available LS BMD from SWAN baseline through follow-up visit 3. The study used thickness-corrected TBS (Med-Imaps, Pessac, France), which corrects for errors due to overlying soft tissue, by using direct measures of thickness by DXA. Earlier versions of Med-Imaps software (iNsight v3.0 and older) used BMI as a surrogate for soft tissue thickness to estimate the soft-tissue correction. This led to a residual negative correlation between BMI and earlier TBS estimates on scans acquired using Hologic densitometers [18]. One validation study of TBS corrected for directly measured soft tissue thickness reported no correlation between BMI and TBS [19]. The SWAN TBS study confirmed that there was no correlation between BMI and TBS for BMI in the middle range, 24 to 31 kg/m2. However, there was a positive correlation for BMI < 24 kg/m2 and negative correlation when BMI was greater than 31 kg/m2 [20]. We therefore accounted for BMI in analyses using splines (see the “Data analysis” section).

Primary exposures

For our first analysis, the primary exposure was glycemic status (prediabetes vs. normal fasting glucose), determined from blood glucose measured from a fasting morning draw, and medication use determined by an inventory of pill bottles brought by the participant to the study visit. Serum glucose was measured using a hexokinase-coupled reaction (Roche Molecular Biochemicals Diagnostics, Indianapolis, IN). Prediabetes was defined as a fasting glucose ≥ 100 and ≤ 125 mg/dL without use of DM medications. For our analyses, we divided prediabetics into 2 groups, low prediabetes (those with fasting glucose 100–109 mg/dL), and high prediabetes (those with fasting glucose 110–125 mg/dL). We made this distinction based on prior data that fasting glucose levels ≥ 110 mg/dL, specifically, are associated with adverse health markers [21, 22]. Participants that did not fall into either low or high prediabetes categories were defined as normoglycemic.

The primary exposure for the second analysis was insulin resistance, as assessed by HOMA-IR, calculated as fasting blood glucose (mg/dL) times fasting serum insulin (U/mL) divided by the constant 405. Insulin was measured from the same fasting blood sample as glucose. Serum insulin was measured using a radioimmunoassay (Coat-a-Count; Diagnostic Products Corp., Los Angeles, CA). The quality control program for serum insulin in SWAN has been previously described [23].

Covariates

Factors that could impact the TBS outcome were included as covariates in analyses. These included self-reported age (years), race/ethnicity, menopause transition (MT) stage (premenopause [no change in menstrual bleeding] vs. perimenopause [less predictable menstrual bleeding at least once every 12 months]), supplemental vitamin D use (yes/no), supplemental calcium use (yes/no), cigarette use (yes/no), and BMI calculated from measured height and weight.

Our fully adjusted models also included lumbar spine (LS) BMD as a covariate. DXA estimates LS BMD by quantifying the average gray-level over all pixels within the LS image region of interest, and TBS indexes the variation in gray-level between pixels in the same region [6]. In SWAN, there was a positive, biphasic relation between LS BMD and TBS with an inflection point at a BMD value of 1.080 g/cm2 (more positive before, less positive after). We, therefore, accounted for LS BMD in analyses using splines (see the “Data analysis” section). Conceptually, controlling for BMD permits us to quantify the associations of glycemic status or insulin resistance with variation in pixel gray-level (TBS) for a given amount of average gray-level (BMD). Prior studies confirm that adjusting for BMD strengthens the association between DM and TBS [8].

Data analysis

In our first analysis, we examined the association of glycemic status (high prediabetes, low prediabetes, normoglycemia) with TBS, using multivariable linear regression with TBS as outcome and glycemic status as categorical primary predictor. Our initial model included the following covariates: age, race/ethnicity, BMI, MT stage (pre- vs. perimenopause), supplemental vitamin D use (yes/no), supplemental calcium use (yes/no), cigarette use (yes/no), and study site. The final, fully adjusted model also controlled for LS BMD. We adjusted for BMI and LS BMD using splines because the relations of these variables with TBS were non-linear [24]. In the case of BMI and TBS, the non-linear relation was tri-phasic (positive for BMI < 24 kg/m2, flat for BMI from 24 to 31 kg/m2, and negative for BMI > 31 kg/m2) [20]. Thus, to adjust for BMI, we used a 3-piece linear spline, with knots (inflection points where the slope changes) at 24 kg/m2 and 31 kg/m2 to model this relationship as piece-wise linear. Specifically, this 3-piece spline allows for different slopes between BMI and TBS in the three BMI categories [24]. Analogously, the relation between BMD and TBS was also non-linear (bi-phasic: positive when LS BMD < 1.080 g/cm2, less positive for LS BMD ≥ 1.080 g/cm2). Therefore, to account for BMD in the model, we used a 2-piece linear spline (with knot at 1.080 g/cm2).

Our second analysis assessed whether greater insulin resistance (assessed by HOMA-IR) is related to lower TBS. We used multivariable linear regression with TBS as the outcome, and HOMA-IR (base 2 log transformed) as continuous primary predictor. We first explored the shape of the relationship between log2HOMA-IR and TBS using LOESS (Fig. 1) and found a biphasic relation with an inflection point at log2HOMA-IR level of 0.7 (corresponding to a raw HOMA-IR value of 1.62). We therefore modeled log2(HOMA-IR) using a 2-piece linear spline with a single knot at 0.7. The decision to assess for a non-linear relation between HOMA-IR and TBS was made a priori, based on experimental data that insulin can have anabolic properties on bone [25, 26], but in insulin resistant states, insulin signaling leads to expansion of bone marrow adipose tissue and decreased trabecular BMD [27]. Covariates were again handled in a two-step fashion as in the first analysis.

Fig. 1
figure 1

LOESS plot of TBS with HOMA-IR (log base 2 transformed). Bottom and top 5% of HOMA-IR distribution dropped to reduce the effect of outliers. Smoothing parameter = 0.8

Full size image

Results

Sample characteristics

Table 1 presents participant characteristics. Mean age, log2HOMA-IR, TBS, and LS BMD were 46 years, 0.97, 1.443, and 1.073 g/cm2, respectively. Thirty-two percent of women were Black, 21% Japanese, and the remaining White. Six and 14% were categorized as having high and low prediabetes, respectively. Two hundred three women had BMI values ≥ 35 kg/m2.

Table 1 Participant characteristicsa for analytic samples
Full size table

Glycemic status and trabecular bone score

Adjusted for age, race/ethnicity, MT stage, BMI, cigarette use, calcium and vitamin D supplementation, and study site, in multivariable regression, compared to TBS in the normoglycemic state, TBS was significantly lower in high prediabetes (p = 0.01), but not in low prediabetes (p = 0.4) (Table 2). After accounting for lumbar spine BMD, TBS was 0.21 SD lower in women with high prediabetes (p = 0.007) vs. those with normal blood glucose (Table 2).

Table 2 Adjusted associationsa of glycemic status and lumbar spine bone mineral density (BMD) with trabecular bone score (TBS)
Full size table

Insulin resistance and trabecular bone score

On visual inspection of the LOESS plot, the relation between HOMA-IR and TBS was non-linear (Fig. 1), with an inflection point at log2HOMA-IR value of 0.7 (corresponding to raw HOMA-IR level of 1.62). Forty-four percent of participants had HOMA-IR measurements < 1.62; median [IQR] HOMA-IR in these women was 1.20 [0.99, 1.39]. The remaining 56% of women had HOMA-IR levels ≥ 1.62 (median [IQR] = 2.61 [2.05, 3.81]).

In multivariable linear regression adjusting for age, race/ethnicity, MT stage, smoking, calcium and vitamin D supplementation, and study site, TBS was not associated with log2(HOMA-IR) when HOMA-IR was < 1.62 (p = 0.3) but was negatively associated with log2(HOMA-IR) when HOMA-IR was ≥ 1.62 (p = 0.02) (Table 3). After additionally accounting for lumbar spine BMD, each doubling of HOMA-IR was associated with 0.15 SD greater TBS when HOMA-IR was < 1.62 (p = 0.01) and 0.15 SD lower TBS when HOMA-IR was ≥ 1.62 (p < 0.0001) (Table 3).

Table 3 Adjusted associationsa of HOMA-IR and lumbar spine bone mineral density (BMD) with trabecular bone score (TBS)
Full size table

Discussion

The overarching goal of this cross-sectional analysis was to determine whether, in women without diabetes, prediabetes and insulin resistance are associated with lower TBS. We report that women with high prediabetes (fasting glucose 110–125 mg/dL) have lower TBS than women with normal blood sugars. In addition, more insulin resistance was associated with lower TBS when HOMA-IR was ≥ 1.62.

A negative relation between DM and TBS has been reported in numerous studies [7,8,9,10,11,12,13]. However, few studies to date have examined the potential associations of prediabetes or insulin resistance (in the absence of DM) with TBS [10, 11, 15], and the results of these studies are inconsistent. One possible explanation for these discrepancies is that hyperglycemia and insulin resistance may not adversely impact trabecular microarchitecture until they exceed certain thresholds. In this analysis, TBS was lower in prediabetes, but only among those with fasting glucose ≥ 110 mg/dL. Similarly, when HOMA-IR was < 1.62, there was a positive association between insulin resistance and TBS, suggesting an anabolic effect of insulin at lower levels [25]. However, once HOMA-IR was ≥ 1.62, greater insulin resistance was related to lower TBS. This is consistent with in vivo models demonstrating that in insulin resistant states, insulin signaling leads to expansion of bone marrow adipose tissue and decreased trabecular bone mineral density (BMD) [27]. Osteoblasts may also become resistant to insulin signaling in insulin resistant states [28].

Knowing that high prediabetes (fasting glucose 110–125 mg/dL) may lead to skeletal complications would be important for public health and clinical reasons. Nearly 20% of US adults have prediabetes with fasting glucose in this range [29], but clinicians remain unsure about how aggressively to treat it. Our results suggest that impaired trabecular bone quality (one mechanism of skeletal fragility in DM) may also affect those with fasting glucose ≥ 110 mg/dL. If high prediabetes also leads to fractures, treating it to prevent adverse bone outcomes could be warranted. Several studies testing for a potential relation between prediabetes and fracture have led to inconsistent results [30,31,32,33], but none specifically focused on individuals with fasting blood sugars ≥ 110 mg/dL or with HOMA-IR levels ≥ 1.62. Future studies should ascertain whether prediabetes is indeed a risk factor for incident fracture and, if it is, whether treating it improves trabecular microarchitecture and prevents fractures. Although a prior SWAN analysis reported that TBS did not predict fracture, independent of BMD in midlife women [34], identifying risk factors for lower TBS in this cohort is nonetheless important. This is because TBS does predict fracture, even when accounting for BMD, in older adults [7, 8], and women with lower TBS in midlife are more likely to have lower TBS in later life [20]. This is similar to how women with lower peak bone mass will have lower BMD in older age [35, 36].

This study has limitations that warrant mention. First, is the cross-sectional study design, which limits our ability to draw causal inference. Our results, nonetheless, lay the foundation for future work to examine the associations of prediabetes (with fasting glucose ≥ 110 mg/dL) and HOMA-IR ≥ 1.62 µU/mL as predictors of change in TBS. The second limitation is that we did not include bone quality parameters other than TBS as outcomes in our analysis. Beyond altered trabecular microarchitecture, increased cortical porosity [37], accumulation of advanced glycation end products [38], and decreased bone material strength [39] are features of the diabetic bone phenotype. Unfortunately, cortical microarchitectural assessments were not available in the full SWAN cohort [40], and measuring bone material strength is not feasible in large cohorts. Third, TBS measures are artifactually affected by the thickness of soft tissue (which increases with BMI). Although we used soft-tissue-thickness-corrected TBS, there remained a relationship between TBS and BMI at the low and high ends (below 24 and above 31 kg/m2). To account for this, we modeled BMI using a 3-segment linear spline. Lastly, we had a relatively small number of participants who had fasting glucose ≥ 110 mg/dL.

In conclusion, we report that high prediabetes (fasting glucose 110–125 mg/dL), and greater HOMA-IR (when HOMA-IR is ≥ 1.62) were associated with lower TBS. Future studies should examine the longitudinal associations of prediabetes with higher fasting glucose and HOMA-IR levels above 1.62 µU/mL with longitudinal declines over time in measures of bone strength, including BMD, TBS, and composite strength indices, as well as with incident fracture.