HR-pQCT skeletal phenotype clustering associated with muscle function in older men and women: the Study of Muscle, Mobility and Aging (SOMMA)

Abstract

Age-related changes to BMD, morphometry, and microarchitecture do not occur uniformly across the population and the common skeletal phenotypes beyond BMD are not well defined. Additionally, the associations between bone and muscle are critical to understanding fall and fracture risk. We hypothesized that unsupervised clustering of HR-pQCT measures at the distal tibia (DT) and radius (DR), separately, would reveal unique skeletal phenotypes; and certain phenotypes would be associated with worse muscle function. In the Study of Muscle, Mobility and Aging (SOMMA; first annual follow-up visit), a cohort of community-dwelling older women and men (61% women; 87% White), HR-pQCT parameters acquired at the DT (N=321; 76.3 ± 4.6 yr) and DR (N=295; 76.1 ± 4.5 yr) were standardized within-sex then combined to form clusters. This resulted in 3 phenotypic clusters, (C1) high total BMD (Tt.BMD) and cortical area (Ct.Ar); (C2) medium Tt.BMD, Ct.Ar, and low trabecular BMD (Tb.BMD); and (C3) low Tt.BMD, and Ct.Ar. DT C2 and C3 exhibited lower micro-finite element analysis failure loads, with the cortical load fraction higher in C2 and lower in C3. C2 and C3 both had a similar proportion of osteoporotic and osteopenic/low bone density individuals, highlighting the novel granularity of HR-pQCT clusters vs aBMD clinical cutoffs. In linear regression models for women, DT C3 was associated with lower leg power (p <.05). For men, DT C3 was associated with lower stair climb and leg power (p <.05). No significant difference was found in grip strength between DT clusters. For DR, no significant difference or association was found between muscle function and clusters for women and men. These findings suggest the concept of bone phenotypic-specific associations with lower but not upper extremity muscle function and have possible implications for the interaction between skeletal phenotypes and muscle function as potential contributory factors to fracture risk.

Lay Summary

Current measures of bone health do not capture the individual variation of combined measures of bone density, shape, and structure. Also, how these bone characteristics relate to muscle function is unclear. In older women and men, 3 clusters were identified based on the bone characteristics. One cluster, with low bone density and thickness, was associated with lower leg power for women and men, and with lower stair climb power for men, vs other clusters. Therefore, this cluster has poorer bone characteristics and lower muscle function. These findings could suggest targets for muscle function to improve fall and fracture risk.

Graphical Abstract

Introduction

Current tools to assess fracture risk, such as DXA, do not identify the majority of individuals who sustain a fragility fracture.¹ HR-pQCT measures have been independently associated with incident fracture² even after controlling for standard DXA measures, such as FN areal BMD (aBMD) or the fracture risk assessment tool (FRAX) score.³ Specifically, the failure load estimated by micro-finite element analysis (μFEA) was most predictive of incident fracture, and the majority of the participants with fractures did not have an aBMD below the cutoff for osteoporosis, highlighting the added value of HR-pQCT measures over DXA.⁴ Failure load is a composite measure accounting for density, morphometry and architecture—all of which vary across the population. This variability necessitates a move away from a “one-size-fits-all” approach. Combining HR-pQCT with clustering techniques has revealed different combinations of individual HR-pQCT parameters or phenotypes that exist^5–7 and their distinct fracture risks.^8,9 These phenotypes have different morphometry and architecture thus different loading schemas which may also affect bone–muscle interactions.

How bone–muscle interactions influence bone health and fracture risk has gained greater emphasis in recent years.^10,11 Sarcopenia has been associated with osteoporosis,¹² and lower muscle function has been associated with fracture.^10,11,13 Lower grip strength, leg strength, balance, and walking ability and speed have been shown to increase fracture risk.^14–16 In relation to HR-pQCT, lower grip strength was associated with lower cortical thickness, cortical area, and trabecular microarchitecture for men,¹⁷ cortical thickness and total BMD for women,¹⁸ and lower cross-sectional area, and μFEA estimated failure load¹⁹ for men and women. Whether combinations of features or specific skeletal phenotypes are associated with lower muscle function is unknown and is the focus of the current investigation.

In the Study of Muscle, Mobility and Aging (SOMMA),²⁰ a subset of older women and men underwent HR-pQCT and DXA scanning and muscle function testing (stair climb power, leg power, leg strength, and grip strength). We aimed to identify (1) skeletal phenotypic clusters for both tibial and radial HR-pQCT measures standardized within sex and (2) which phenotypic combination of low bone density and diminished cortical or trabecular morphometry is associated with lower muscle function.

Materials and methods

Study design, setting, and participants

Our study was a cross-sectional analysis with assessments collected at first annual follow-up visit after baseline from an ongoing cohort study of mobility in community-dwelling older men and women from the SOMMA (https://sommaonline.ucsf.edu/) and SOMMA Bone Ancillary study. Participants for SOMMA were enrolled between April 2019 and December 2021 from 2 clinical sites: University of Pittsburgh (Pittsburgh, PA; n =439) and Wake Forest University School of Medicine (Forsyth County, NC; n =440). To be eligible, participants had to be aged ≥70 yr at baseline, able to complete a 400 m walk within 15 min, able to walk ≥0.6 m/s (4 m), and free from contraindications to MRI and muscle biopsy. Participants were excluded if they (1) reported inability to walk ¼ mile or climb a flight of stairs; (2) had active cancer or an advanced chronic disease (ie, heart failure, renal failure on dialysis, Parkinson’s disease, or dementia); or (3) had BMI>40 kg/m². A detailed SOMMA study protocol and recruitment was previously published.²⁰

At the first annual follow-up visit, participants at the University of Pittsburgh site (n =385) were invited to participate in the SOMMA Bone Ancillary study, which aimed to determine the association between muscle and bone using HR-pQCT and DXA to assess bone parameters. Among those who completed the first annual follow-up 329 (85%) enrolled in SOMMA Bone (n =48 refused participation, n =8 non-response). Western IRB-Copernicus Group (WCG-IRB; study number 20180764) approved the protocol. All participants provided written informed consent.

HR-pQCT and DXA

HR-pQCT

Of 329 participants enrolled in SOMMA Bone, 327 had a vailable HR-pQCT data (n =1 examiner safety concern, n =1 poor scan quality and re-scan refusal). Scans were eligible for inclusion unless participants had bilateral wrist, ankle, or hip contraindications (ie, prior fracture, metal artifacts, non-removable hardware or jewelry, significant weakness, or recent unloading for ≥6 wk for tibia scans). Of 327 with scans, 321 had available scans at the distal tibia (DT; excluded: n =1 poor image quality, n =2 prior bilateral fracture and/or hardware, n =1 bone deformity, and n =1 data transfer error) and 295 had scans at the distal radius (DR; excluded: n =22 poor image quality, n =7 prior bilateral fracture and/or hardware, n =1 image artifact, n =1 data transfer error, and n =1 unable to remove jewelry).

Participants were scanned at the DR and the DT, unilaterally, using HR-pQCT (Xtreme CT II; Scanco Medical AG) with a nominal isotropic voxel size of 61 μm. Laterality for HR-pQCT scans was matched to lower-extremity muscle function tests. These scans were done on the right leg and ipsilateral forearm except in the case of contraindications, in which case the contralateral side was scanned. At the time of the scan, limb length was measured using flexible anthropometric tape. Tibia length (Le, mm) was measured with participants in a seated position as the distance from the tibial plateau to the distal edge of the medial malleolus. Ulna length (mm) was measured as the distance between the olecranon process and the distal edge of the styloid process. A carbon cast was used to immobilize the limb during scan acquisition to reduce motion artifact. Quality control calibration was performed daily and weekly by a trained technician using the standard quality control phantom provided by the manufacturer. Scans (168 slices) at the distal sites were used for analyses and obtained using a fixed offset at 9.5 mm proximal to the DR endplate and 22.5 mm proximal to the tibial plafond at the DT. Raw images of the scans were evaluated by technicians for motion artifacts and rated on a scale of 1–5, with 1 representing no motion and 5 representing extreme motion artifact. If scans had a rating of 3–5, up to an additional 2 repeat scans per site were collected in an attempt to reduce motion artifact.²¹ Scans with a grade of 1–3 were included in the analysis. A trained technician then performed the region of interest contouring for 2- and 3-dimensional evaluation following best practice guidelines.²² HR-pQCT parameter values that were ± 2 SDs outside of the mean values for the study and/or flagged by the technician were reviewed for accuracy.

The following variables were quantified from the HR-pQCT images: total volumetric BMD (Tt.BMD, mg HA/cm³), cortical volumetric BMD (Ct.BMD, mg HA/cm³), trabecular volumetric BMD (Tb.BMD, mg HA/cm³), cortical area (Ct.Ar, mm²), trabecular area (Tb.Ar, mm²), cortical thickness (Ct.Th, mm), trabecular thickness (Tb.Th, mm), trabecular number (Tb.N, mm^–1), and trabecular separation (Tb.Sp, mm). The 2D analyses performed included Tt.Ar, Ct.Ar, and Tb.Ar as the mean of each respective area in all image slices, consistent with recommended guidelines.²² MicroFEA was performed using a homogeneous elastic modulus of 10 GPa and a Poisson’s ratio of 0.3 (Linux V8 4–2). Failure load (FL, kN) was estimated by calculation of the reaction force at which 2% of the elements exceeded 0.7% strain.²³ Stiffness (kN/mm), apparent modulus (N/mm²), and fraction of the load carried by the cortical compartment (Ct.LF, %) were also computed for each model.

DXA

Unilateral hip scans were performed to assess total and regional aBMD of the proximal femur using Lunar iDXA (GE Healthcare Lunar). Scans were acquired on the same side as the tibial HR-pQCT scan, except in the case of contraindications (ie, prior hip fracture or hip replacement surgery, metal artifacts), in which the contralateral side was scanned. Participants with bilateral contraindications were excluded from DXA scans. Certified Bone Densitometry Technologists (International Society for Clinical Densitometry, ISCD-CBDT) acquired the scans following the manufacturer’s guidelines for participant positioning along with daily calibration using the GE Lunar calibration phantom.²⁴ Scans were analyzed using enCORE Software version 15 (GE Healthcare Lunar). Values that were ± 2 SDs outside of the mean values for the study and/or flagged by the technician were reviewed for accuracy. Scans with severe motion artifact, attenuation, or discontinuity were excluded. DXA aBMD (g/cm²) at the FN was used in analyses. Individuals were assigned to clinical classification based upon WHO cut-offs of T-score of normal (T≥−1.0), low bone density/osteopenia (−1.0> T > −2.5), and osteoporosis (T≤ −2.5) using a reference database. FRAX score was calculated for major osteoporotic fracture including aBMD (fraxplus.org).

Clustering protocol

Choice of parameters was based upon those that were significantly associated with fracture,^2,4 and those previously shown to identify skeletal phenotypic clusters from the Bone Microar-chitecture International Consortium (BoMIC).⁸ These parameters included Tt.BMD, Ct.BMD, Tb.BMD, Tt.Ar, Ct.Ar, Tb.Ar, Ct.Th, Tb.Th, Tb.N, and Tb.Sp and either tibia length or ulnar length for the DT and DR sites, respectively.

The clustering protocol followed a similar previous approach using HR-pQCT parameters⁸; however, our analysis standardized the HR-pQCT parameters within sex (ie, separated men and women to calculate sex-specific z-scores, then recombined men and women), and performed the clustering for the DT and DR site separately. This within-sex standardization accounts for the differential distribution of HR-pQCT parameters between sexes.²⁶ Initially, using within sex standardized parameters, for each site (DT and DR), agglomerative clustering was conducted. Agglomerative clustering conducts pairwise similarity observations by a distance or dissimilarity function (like Euclidean, Manhattan), then a linkage function (Ward’s, average, and complete) extends this to pairs of clusters.²⁷ As a result, this agglomerative or hierarchical technique grouped individuals based upon similarity, sequentially collapsing more and more similar clusters of individuals into larger clusters. Several distance functions (Euclidean, Manhattan), and linkage functions (Ward, average, and complete) were also used to determine which combination of functions best fit the dataset. From this initial approach and upon visual inspection, 3 clusters fit the data with Euclidean and Ward as distance metric and linkage method, respectively. Using Euclidean distance, Fuzzy C-means clustering was conducted. The Fuzzy C-means clustering assigns a membership probability coefficient to each individual, and individuals were then assigned to clusters with their highest membership coefficient (ie, K-means). To confirm whether 3 clusters was appropriate, the Fuzzy Silhouette Index (FSI) and the total within sum of squares error were calculated for 1–10 clusters. The FSI (value of −1 to 1) is a measure of the homogeneity of each cluster with higher FSI values being more favorable. Also, using the total within sum of squares error, the elbow method (ie, determining the number of clusters after which increasing the number of clusters leads to only slight reductions in within sum of squares error) was applied to determine the ideal cluster number where adding more yields diminishing returns in error reduction. Both techniques confirmed that 3 clusters were appropriate for the DT and DR sites. (DT: FSI=0.356; DR: FSI=0.341). As a final confirmation of cluster stability, the Jaccard similarity index (JSI) was calculated by randomly resampling 80% of the individuals repeatedly (100 times) and comparing their cluster assignment as a measure of cluster stability.²⁸ Values closer to 1 indicate higher stability. All clustering analyses were performed in R v4.3.0 with packages for agglomerative (cluster v2.1.6), Fuzzy C-means (ppclust v1.1.01) clustering, and FSI calculation (fclust v.2.1.1.1).

Muscle function assessment

Leg power and strength were assessed on a Keiser pneumatic resistance device (A420 model; Keiser Sports Health Equipment).²⁹ Participants were excluded if they had difficulty bending or straightening either knee fully due to pain, arthritis, injury, or other condition; significant weakness; unable to move through range of motion with proper form; and self-reported pain during or after movement. Participants were seated and positioned using 1 leg at a starting position of 90°. The right leg was prioritized unless the participant had significant weakness; difficulty bending or straightening the knee due to pain; or injury within the last 12 wk that affected the right side. After 2 warm-up repetitions, resistance was gradually increased starting from 40 lbs until achieving a 1-repetition maximum (1-RM). After a subsequent 30-min rest, participants completed 2 trials each at 40%, 50%, 60%, and 70% of 1-RM, with 30 s of rest between trials and 1 min between increases in resistance, and the peak leg power (Watts) was obtained.

Participants performed 3 stair climb laps on a set of 4 standard stairs and the peak stair climb power was calculated as (Power, W)=[(Weight, kg)×(9.8 m/s²)×(Stair height, m)]/(Time to complete ascend lap, s) for the lap with the highest power.³⁰ Participants were excluded if they were unable to complete the task without using an assistive device or excluded for other health or safety concerns. Grip strength was obtained using a Jamar handheld dynamometer (Sammons Preston Rolyan)³¹ and maximum grip strength (kg) from 2 trials on the left and right hand was reported. Participants were excluded from grip strength testing if they had a recent surgery on their hand or wrist in the past 3 mo. Of the individuals with HR-pQCT scans at the DT, there were 4 exclusions for stair climb peak power, 29 for leg press peak power and strength, and 2 for grip strength. At the DR site, there were 2 exclusions for stair climb peak power, 24 for leg press peak power and strength, and 2 for grip strength.

Covariates, first annual follow-up visit

Demographics collected included self-reported age, sex, race (based on current census categories), and ethnicity classified as non-Hispanic White (y/n). Height (m) was obtained using a wall-mounted stadiometer, weight (kg) was obtained with a balance beam or digital scale, and BMI was calculated (kg/m²). Self-reported smoking (current or non-/past smoker), alcohol use (≥1 drink per week or <1 drink per week), and history of falls in the past year, fractures since age 50, and arthritis (y/n) were also collected.

Multimorbidity was classified using a modified Rochester Epidemiology Multimorbidity Scale (0–11).³² The multimorbidity index was comprised of 11 age-related conditions, with 1 point added to the score for each condition, including self-reported physician diagnosis of cancer (except nonmelanoma skin cancer), cardiac arrhythmia, CKD, chronic obstructive pulmonary disease, coronary artery disease, congestive heart failure, dementia, diabetes, stroke, aortic stenosis, and depressive symptoms based on the 10-item Centers for Epidemiologic Studies Depression scale (CESD-10) score.³³ Hypertension was defined as a systolic blood pressure ≥140 mmHg.

Total activity (counts/min) was collected using the wrist-worn accelerometry (ActiGraph GT9X; valid wear ≥3 d of ≥17 h) with a sampling rate of 80 Hz, processed in one-minute epochs (activity counts/min), and then reprocessed without the low frequency extension (LFE) filter. Accelerometry data were scored using Acti-Graph LLC ActiLife Software. For the minutes of non-wear, activity data without the LFE filter was imputed.

Statistical analyses

Descriptive statistics using pairwise 2-sided t-tests (or Wilcoxon signed-rank test, if appropriate), and pairwise Fisher’s exact test were conducted to compare characteristics between each cluster for men and women separately. Age-adjusted ANCOVAs were performed followed by Tukey’s post-hoc tests to identify the differences in sex-standardized HR-pQCT parameters among clusters. Fleiss kappa was conducted to compare DT vs DR cluster assignments, and to compare DXA T-score clinical classifications of normal, low bone density/osteopenia (−1.0 > T-score > −2.5), and osteoporosis to cluster assignment.

Multivariable linear regression models were used to assess the association between cluster assignment (referenced against cluster 1), and muscle function measures, for men and women separately with standardized β coefficient and p-value reported. Muscle function measures included power (leg power, stair climb power) and strength (leg strength, grip strength). Model covariates were added sequentially as groups. Model 1: univariate; Model 2: Model 1+age, race; Model 3: Model 2+height, weight; Model 4: Model 3+smoking status, alcohol use, total activity counts, multimorbidity, arthritis, and hypertension; Model 5: age, race, height, and weight forced into the model, but covariates (smoking status, alcohol use, activity counts, multimorbidity, arthritis, and hypertension) were removed one by one if nonsignificant (alpha level > 0.10). Covariates were harmonized across final models for consistency. No additional covariates besides age, race, height, and weight were retained for the final models of leg power, leg strength, or grip strength. Final models for stair climb power included activity, alcohol use, and smoking. A sensitivity analysis was conducted including CKD or diabetes added as individual covariates. SOMMA dataset release from May 2024 was utilized for all analyses.

Results

Descriptive characteristics of clusters

Three clusters were identified at the DT site. The number of individuals in cluster 1 (C1) was 121 (60% women), cluster 2 (C2) was 86 (63% women), and cluster 3 (C3) was 114 (61% women) with no significant difference in the proportion of women across clusters. Women in C2 were significantly younger vs C3 (Table 1). For men, no significant difference in age was observed among clusters. Anthropometrically, women in C3 were taller than C1 and C2, and C3 had a lower BMI vs C1 and C2. Men in C3 were also taller than C1 and C2 and men in C3 also weighed more than C2. More women in C3 reported alcohol use (≥1 drink per week) vs C1. No significant difference between clusters was observed for the percentage of individuals identifying as non-Hispanic White, smoking status, total activity counts, self-reported falls, fractures, hip fractures, multimorbidity counts, arthritis, or hypertension. Descriptive results for clusters formed using the DR site were similar; however, there were fewer significant differences in age and anthropometry between clusters (Table S1). Comparing the DT and DR clusters, there was 58.8% agreement in an individual’s cluster assignment, an overall Fleiss Kappa value of 0.38 indicating some agreement, and individual Kappa values of 0.49, 0.27, and 0.38 for C1, C2, and C3, respectively, with higher values indicating more agreement (Figure S1). The JSI for the clusters was 0.962 for the DT and 0.955 for the DR indicating high stability of the clustering.

Table 1.

Descriptive characteristics of skeletal phenotypic distal tibia (DT) clusters by sex.

	Women			p-values			Men			p-values
	C1	C2	C3	C1 to C2	C1 to C3	C2 to C3	C1	C2	C3	C1 to C2	C1 to C3	C2 to C3
N	72	54	69				49	32	45
Demographics
Age at baseline, yr	75.8 (4.4)	75.4 (5.0)	77.3 (4.8)#	.24	.06	.018	75.8 (3.7)	76.5 (4.8)	76.7 (5.0)	.72	.66	.98
White race, (%)	81	83	87	.82	.37	.61	88	94	96	.47	.27	1.00
Anthropometry
Height, m	1.58 (0.06)	1.59 (0.06)	1.63 (0.06)‡#	.40	<.001	.003	1.72 (0.07)	1.70 (0.07)	1.77 (0.07)‡#	.19	<.001	<.001
Weight, kg	72.1 (13.0)	70.1 (14.0)	68.5 (11.2)	.43	.08	.48	84.0 (14.6)	79.7 (13.8)	85.6 (14.2)#	.11	.43	.032
BMI, kg/m2	28.9 (5.1)	27.8 (5.1)	25.9 (4.3)‡#	.27	<.001	.041	28.3 (4.5)	27.6 (4.9)	27.2 (4.1)	.33	.28	.90
Lifestyle characteristics
Current smoker, (%)	6.9	0	4.3	.07	.72	.26	0	3.1	2.2	.40	.48	1.00
≥1 alcohol drink/wk, (%)	38.0	55.8	60.3‡	.07	.011	.71	69.4	68.8	71.1	1.00	1.00	1.00
Tot. activity, counts/min 105	20.36 (5.74)	20.46 (5.99)	20.54 (6.89)	.98	.93	.97	17.92 (5.40)	16.99 (4.70)	18.13 (4.63)	.44	.84	.32
History of falls & fracture
Fall in past year, (%)	31.9	27.8	20.3	.70	.13	0.39	20.4	40.6	22.2	.08	1.00	.13
Fracture (any) age >50, (%)	25	27.8	29	.84	.71	1.00	12.2	18.8	11.1	.53	1.00	.51
Fracture (hip) age > 50 (%)	1.4	3.7	1.4	.58	1.00	.58	0	3.1	0	.40	1.00	.42
Health
Multimorbidity, (0–11)	0.8 (1.0)	0.6 (0.6)	0.6 (0.9)	.28	.22	.87	1.0 (0.9)	0.8 (0.8)	0.8 (0.9)	.29	.17	.76
Arthritis, (%)	72.2	64.8	63.8	.44	.37	1.00	46.9	40.6	44.4	.65	.84	.82
Hypertension, (%)	25.0	35.2	24.6	.24	1.00	.23	20.4	28.1	28.9	.44	.47	1.00
Muscle function
Stair climb power, W	154.3 (38.0)	145.9 (34.9)	140.5 (35.2)‡	.20	.030	.39	194.1 (43.4)	181.7 (39.2)	180.0 (42.5)	.19	.12	.86
Leg power, W	293.1 (79.2)	289.2 (89.3)	271.6 (84.9)	.81	.15	.30	524.7 (141.3)	464.7 (106.4)†	446.4 (166.4)‡	.036	.021	.58
Leg strength 1-RM, kg	145.4 (35.1)	146.4 (43.9)	138.2 (38.8)	.83	.28	.23	216.9 (52.4)	200.7 (56.2)	192.6 (53.4)‡	.16	.012	.44
Grip strength, kg	22.87 (5.42)	24.13 (5.25)	23.59 (5.52)	.29	.70	.48	36.69 (7.88)	35.44 (8.50)	36.80 (8.17)	.63	.88	.50
DXA measures
Femoral neck aBMD, g/cm2	0.86 (0.11)	0.79 (0.10)†	0.79 (0.10)‡	.001	.001	.91	1.04 (0.17)	0.88 (0.11)†	0.91 (0.11)‡	<.001	<.001	.32
T-score	−1.30 (0.78)	−1.78 (0.76)†	−1.80 (0.74)‡	.001	.001	.91	−0.25 (1.30)	−1.45 (0.88)†	−1.21 (0.88)‡	<.001	<.001	.33
FRAX score	14.72 (6.21)	17.44 (7.44)†	18.41 (9.76)‡	.031	.010	.81	6.01 (2.90)	8.77 (3.54)†	7.93 (3.44)‡	<.001	.001	.23

Symbols: † 1 vs 2,‡ 1 vs 3, and # 2 vs 3 significant differences (p-values in columns). Abbreviations: 1-RM, 1-repetition maximum. Non-significant (p-value < .05) reported to 2 decimal places.

HR-pQCT phenotypes of clusters

The within-sex standardized HR-pQCT measures used to form the DT clusters were compared to understand the skeletal phenotype of each cluster (Figure 1). C1 exhibited high-Tt.BMD, Tb.BMD, Ct.Ar, and Ct.Th, while C2 had intermediate Tt.BMD, Ct.Ar, with low Tb.BMD, and Tb.N, and C3 demonstrated low-Tt.BMD, Ct .Ar with high Tt.Ar, Tb.Ar, and Le. Figure 2 presents representative images of each DT cluster for men and women. For clusters formed using the DR site, DR C2 did not exhibit similar deficiencies in Tb.Sp or Tb.N (Figures S2 and S3).

Figure 1.

Box plot of HR-pQCT within-sex standardized z-scores compared by distal tibia (DT) cluster, adjusted for age. Total BMD (Tt.BMD), cortical BMD (Ct.BMD), trabecular BMD (Tb.BMD), total area (Tt.Ar), cortical area (Ct.Ar), trabecular area (Tb.Ar), cortical thickness (Ct.Th), trabecular thick-ness (Tb.Th), trabecular number (Tb.N), trabecular separation (Tb.Sp), and limb length (Le). Key: †p <.05; ‡p ≤.01; #p ≤ .001.

Figure 2.

Representative HR-pQCT cross-sections and 3D renderings of each distal tibia (DT) cluster by sex. Each column is a single representative individual in each cluster.

Examining the μFEA parameters for the DT clusters, for both women and men, C1 exhibited higher estimated failure load vs C2 and C3 (Figure 3A) with no difference between C2 and C3. For women and men, the estimated load carried by the cortical compartment (Ct.LF) was lower in C3 vs C1 and C2 (Figure 3B), the estimated stiffness was highest in C1 vs C2 and C3 (Figure 3C), with the apparent modulus highest in C1, followed by C2, and C3 as the lowest (Figure 3D). For DR clusters, there were significant differences in estimated failure load among clusters with DR C1 as the highest, followed by C2, and then C3, with no difference in Ct.LF observed between C2 and C3 (Figure S4).

Figure 3.

Boxplots of micro-finite element analysis (μFEA) measures between distal tibia (DT) clusters by sex for (A) estimated failure load, (B) cortical load fraction (Ct.LF), (C) stiffness, and (D) modulus. Key: †p <.05; ‡p ≤0.01; #p ≤.001.

Summarizing the skeletal phenotype of each cluster, DT C1 exhibited the highest density with favorable cortical morphometry and trabecular microarchitecture, and the highest estimated failure load. C2 and C3 had comparably low estimated failure loads. C2 demonstrated medium density with deficits in trabecular density and microarchitecture and greater reliance on the cortical compartment. C3 had the lowest density, with high cross-sectional area and deficits in cortical area and thickness.

Cluster assignment compared to clinical cutoffs

To contextualize the clustering results against current clinical standards, the DXA T-score at the FN was compared between each DT cluster (Table 1). For both women and men, the FN aBMD, T-score and FRAX score were all lower in C2 and C3 vs C1 with no significant difference between C2 and C3. A mosaic plot of the classifications of normal, low bone density/osteopenia, and osteoporosis vs DT clusters found that C1 represented the majority of normal DXA T-scores, and the smallest proportion of osteoporotic individuals vs C2 and C3 (Figure 4). However, C2 and C3 each captured a similar proportion of individuals with normal, low bone density/osteopenia, or osteoporosis T-scores. Comparing the clusters with clinical classifications, these indicated no agreement, and a minimal level of disagreement (Fleiss Kappa score = −0.24). For the DR site, C1 captured fewer individuals with a normal T-score vs DT C1, and indicated no agreement, with a minimal level of disagreement (Figure S5; Fleiss Kappa score = −0.238).

Figure 4.

Mosaic plot of distal tibia (DT) cluster and DXA T-score clinical cut-offs of normal, low bone density/osteopenia, and osteoporosis. The area of each box is proportional to the number of individuals. No corresponding DXA scan for C1 (n= 2), C2 (n=1), and C3 (n=7).

Muscle function comparisons and associations with cluster phenotype

For the DT clusters, stair climb power in women was higher in C1 vs C3 (Table 1). For men, C1 exhibited higher leg power than both C2 and C3, while leg strength was higher in C1 vs C3. No differences were found for any muscle function measure in women and men between DR clusters (Table S1). Muscle function measures were compared between clusters in multivariable linear regressions referenced against C1. For the DT clusters in women, C3 was associated with lower leg power (Table 2), while C2 was not. In men, C3 was associated with lower stair climb power and lower leg power, while C2 was not. For the sensitivity analysis, including CKD and diabetes as individual covariates vs as components of multimorbidity, no change in significance of any model (p >.05) or direction of effect size was found. For DR clusters, no significant association existed between C2 or C3 with any measure of muscle function (Table S2).

Table 2.

Adjusted multivariable linear regression, reported as standardized β regression coefficients of categorical distal tibia (DT) cluster assignment (ref. C1) as exposures of sex-specific muscle function z-score.

	Women					Men
		β		p-value			β		p-value
	Mean ± SD	C2	C3	C2	C3	Mean ± SD	C2	C3	C2	C3
Power
Stair climb power (W)	147.1 ± 36.2	−0.21	−0.20	.192	.228	185.1 ± 42.1	−0.19	−0.43	.363	.039
Leg power (W)	285.5 ± 84.0	−0.05	−0.39	.782	.042	481.3 ± 144.5	−0.31	−0.55	.157	.010
Strength
Leg strength (kg)	143.6 ± 38.9	0.02	−0.28	.909	.163	204.8 ± 53.9	−0.28	−0.32	.234	.152
Grip strength (kg)	23.5 ± 5.4	0.18	−0.10	.281	.563	36.5 ± 8.0	−0.10	−0.12	.659	.601

Mean ± SD reported for each muscle function measure by sex. The final model for all measures of muscle function included: age, race, height, and weight. Additionally, final models for stair climb power also included activity, alcohol use, and smoking while the final model for leg power included arthritis.

Discussion

This study identified novel clusters of HR-pQCT skeletal phenotypes and determined whether these phenotypes differentially associate with muscle function. We demonstrated that C3 had not only an unfavorable skeletal phenotype (high Tt.Ar, low Tt.BMD, and Ct.Th) but was associated with lower stair climb power and leg power in men, and lower leg power in women. Prior studies have demonstrated associations of individual bone parameters to muscle function, but a person’s skeletal fragility is not determined by one individual parameter. Thus, a major strength of our study is that it is the first study to demonstrate that a specific skeletal phenotype is associated with poor muscle function.

The use of clustering techniques applied to HR- pQCT is being established^5–8 to move away from a “one-size-fits-all” approach and the use of only DXA aBMD to characterize bone health. Using the BoMIC dataset, 3 clusters were found with highest fracture risk in the low density cluster followed by the low volume, and healthyclusters.⁸ Using this same technique, a cohort of recent hip fracture patients were more likely to have a low density phenotype for both men and women.⁹ The healthy, low-volume, and low-density phenotypes may roughly correspond to the DT C1, C2, and C3, respectively, in our study. In a follow-up of the Hertfordshire Cohort Study, clustering analysis found for men, one cluster with high Tt.Ar, low Ct.Th, and a second with low Tb.BMD both of which had higher fractures. For women, one cluster with low Ct.Ar, and Ct.Th and a second with low Tb.BMD had higher fractures.⁵ Each of these clusters may align with the DT C3 and C2 found in our study, respectively. In the Global Longitudinal study of Osteoporosis in Women, clustering analysis found a tibial cluster with high-Tt.Ar and a radius cluster with alterations to trabecular microarchitecture were both associated with higher fractures.⁶ Ideally, all these clustering techniques would converge on a similar set of phenotypic clusters, and they do appear to have revealed a consistent low-density, high-Tt.Ar phenotype similar to DT C3. Our results of low-density, higher Tt.Ar (C3) phenotype associated with low muscle function also suggests that the higher fractures observed in prior clustering analyses in this similar C3 phenotype may reflect a contribution of muscle function to fracture risk beyond skeletal properties alone.

The past studies used varied methods and populations. Unlike our study, the BoMIC approach did not use within-sex standardization and combined DT and DR parameters to form each cluster rather than examining them separately by site. As result, they found significant differences in the sex-composition of each of their phenotypes with women dominating a low-volume pheno-type, and men restricted to healthy and low-density phenotypes. Other approaches have clustered men and women entirely separately.^5–7 With our approach using within-sex standardizing, number of men and women did not significantly vary across clusters. Sex-based differences in bone aging has been well established. Over 3 yr, post-menopausal women, and older men (age > 50) both lost predominantly cortical bone by HR-pQCT, but women had greater magnitude of losses.³⁴ Over 5-yr by hip QCT, women began with lower BMD but both sexes lost BMD and Ct.Th with higher magnitude of losses in women.³⁵ Within-sex standardizing accounts for the differing initial BMD between sexes, and the higher magnitude of losses associated with aging in women.³⁵ By not within-sex standardizing, men potentially would not be identified as having a low-volume phenotype or the number of women who fall into a low-density phenotype could be overestimated. We formed clusters separately by skeletal site, as we observed significant differences in cluster assignment between the weight bearing DT and non-weight bearing DR which suggested different phenotypes may exist between anatomical sites. Therefore, combining DT and DR to form the clusters could overgeneralize an individual’s skeletal phenotype. Our results indicate that future studies should consider cluster formation by within-sex standardization and by separating skeletal site.

Other techniques to characterize skeletal phenotype include statistical shape or shape-density modeling which identifies principal components from 3-dimensional imaging data to explain the underlying variation in a population.^36–38 This technique does not classify individuals into a phenotype, rather it reduces variable dimensionality. However, in this report, we were not attempting to reduce the number of bone parameters, but instead we were interested identifying subgroups of individuals with common patterns of bone measurements. Thus, centroid-based clustering algorithms, such as fuzzy C-means or K-means were a more optimal statistical approach since they group individuals upon similarity.

Given our aim of characterizing the phenotype, we formed clusters using density, morphometry, and microarchitecture. Micro-FEA mechanical measures were not included in the clustering parameters but were instea d compared between clusters after formation. While the μFEA estimated failure load has been demon-strated to be the most predictive measure of fracture among HR-pQCT measures,⁴ it is critical to understand how skeletal phenotypes contribute to estimated failure load, which is dictated by the bone microarchitecture and morphometry.²² We found 2 clusters with similarly low estimated failure loads, but distinctly different phenotypes emphasizing future interpretations of μFEA outcomes should incorporate skeletal phenotype to understand the mechanism of each individual’s fragility. Inclusion of functional outcomes like estimated failure load into clustering algorithms may obfuscate the relative contributions of morphometry or microarchitecture to a skeletal phenotype and should be carefully considered when determining which parameters to input into clustering algorithms. Also, μFEA estimated failure load and identification of phenotypic clusters requires HR-pQCT, of which a limited number of systems exist and only in research settings.³⁹ Other FEA models have been applied to more widely available imaging modalities, such as DXA⁴⁰ or CT,⁴¹ which do not capture microarchitecture, but in our study, there were prominent differences between clusters for parameters, such as Tt.BMD, Tt.Ar, and bone length, which do not require HR-pQCT resolution suggesting it may be possible to identify skeletal phenotypes with other imaging modalities. Clustering techniques, machine learning, and individualized approaches to identifying individuals at highest risk of fracture have yet to be fully applied into these clinical imaging modalities. There are several important factors to consider in order to translate these findings and approaches, including whether consistent phenotypes could be identified between HR-pQCT and CT or DXA, whether μFEA estimated failure load is comparable to other FEA models at lower imaging resolutions, and whether CT or DXA-based FEA models are able to capture compartment-level(cortical vs trabecular) differences in contribution to failure load as we identified herein. Additionally, identification of clusters and phenotypes will need to be further replicated, and a consensus methodology agreed upon in order to incorporate these approaches into clinical decision making.

We found disagreement by Fleiss Kappa for both DT and DR clusters to clinical DXA cut-points and no difference in T-score between C2 and C3 for men and women. This disagreement and lack of difference in T-score highlights the unique potential clinical importance of HR-pQCT cluster identification vs DXA. Areal BMD alone also cannot explain sex-based differences in bone strength. For example, in past cadaveric studies, men exhibited higher proximal femoral bone strength than women at the same aBMD.^42,43 The disagreement between cluster phenotype and DXA clinical cut-points, as well as the inability of DXA to explain sex-based differences in strength, support the additional contribution of HR-pQCT parameters in understanding skeletal phenotype beyond aBMD to improve fracture risk prediction. Future studies could evaluate whether examining the individual components of BMD (BMC and area) or cortical and trabecular compartment could improve the field’s understanding of sex-based differences, bone strength, and fracture risk prediction.

The descriptive characteristics of the DT clusters were very similar, with notable differences being C3 being taller in men and women likely due to the inclusion of bone length in the parameters to identify clusters. Additionally, for women DT, C2 was slightly younger with lower BMI in C3 again mainly driven by height differences, but in men C3 weighed more but being taller there was no difference in BMI. As C2 and C3 capture a similar proportion of individuals with osteoporosis, we would not expect to capture differences in BMI, with osteoporosis typically being associated with lower weight and BMI. Establishing how representative our sample is, in NHANES the mean FN aBMD is 0.623–0.729 for older women, and 0.733–0.813 for older men,⁴⁴ our study had a slightly higher FN aBMD of 0.814 for women and 0.952 for men. About 25% of women and 6% of men aged 65 yr or older are classified as having osteoporosis⁴⁵ with our study having 10% of the women and 6% of the men qualifying as having osteoporosis. The entry criteria for the study also include mobility as participants were required to walk 400 m in order to be enrolled. It is possible that those excluded for this study would have poorer bone density and may contribute to the slightly higher FN aBMD and lower proportion of women with osteoporosis in our study. But we aim to improve our identification of individuals missed by current clinical cutoffs and suggest that inclusion of other bone parameters or consideration of phenotypes may improve fracture risk identification. Additionally, C2 and C3 captured a similar proportion of individuals with low bone density and osteoporotic density, and both clusters have higher FRAX scores than C1 highlighting the added novelty of the phenotypic cluster identification within the aBMD range of our study compared to current clinical cutoffs.

For associations between specific HR-pQCT bone parameters and muscle function, lower Ct.Th, Ct.Ar, and trabecular bone volume were associated with lower muscle mass or grip strength.^17,18,46 Lower Ct.Ar, trabecular bone volume fraction, and estimated failure load were associated with lower stair climb power.⁴⁷ Combining these individual parameter associations with our finding of DT C3 associations with lower muscle function, and that C2 and C3 had similarly low aBMD and Tt.BMD, this suggests that low aBMD or Tt.BMD may only be associated with poor muscle function in a phenotype with concomitant low Ct.Ar, Ct.Th, and high Tt.Ar such as C3.

Traditionally, bone and muscle were thought to be related insofar as they both rely on mechanical stimulation for tissue health, but more studies are exploring the complex interplay in paracrine and endocrine signaling between muscle and bone.¹⁰ In close anatomical proximity, periosteal expansion occurs with bone aging as an attempt to maintain mechanical strength.⁴⁸ However, our results showed higher Tt.Ar in C3 coincided with lower Ct.Th, and Ct.Ar, suggesting that lower muscle function is associated with increased bone resorption or an inability to maintain bone at the endosteal surface. While our study is unable to attribute causality, it is possible that decreased loading leads to both the C3 phenotype and lower muscle power and strength. Beyond mechanical stimulation, myokines and osteokines are also implicated in bone–muscle cross-talk.¹³ It remains unclear how myokines, isolated from concurrent mechanical stimulation, affect bone morphometry or trabecular bone microarchitecture. Whether specific myokines, fiber types, motor neurons, or loading profiles may explain these differences in muscle power vs strength with bone phenotypes or sex differences, is also unknown. Leg power is more dependent on fast-twitch fibers, meanwhile, fewer fast-twitch fibers is associated with lower hip aBMD,⁴⁹ and the proportion of fast twitch fibers decreases with age.⁵⁰ Future studies should evaluate how histological measures of fiber type vary with HR-pQCT measures of bone, and what specific myokines or signaling pathways could explain associations. Additionally, sex-based differences in muscle function and fracture risk have been previously observed with declines in muscle strength associated with increased fracture risk in women, but not men.¹⁵

Strengths

This study included both older women and men, at 2 skeletal sites (DT and DR), with a thorough characterization by HR-pQCT and DXA. We had multiple measures of muscle function, which is unique for a study of HR-pQCT parameters. We also standardized within sex which accounted for the differential distribution of the HR-pQCT parameters between sexes. There were also minimal differences in descriptive characteristics between clusters and no difference in proportion of women between clusters. Fuzzy c-means clustering was employed, in a rigorous statistical approach, to identify key HR-pQCT phenotypes.

Limitations

Our sample size was limited compared to prior clustering approaches, which may have impacted statistical power for comparisons between clusters. Additionally, we did not adjust our final models for multiple comparisons as the associations of bone phenotypic cluster membership with muscle function had not been previously reported. Our study discretized the individuals into the cluster by the highest clustering coefficient homogenizing their cluster identity, but a more detailed evaluation of cluster coefficients may be done in future studies. We also relied on μFEA measures of bone mechanics, rather than ex vivo experimental mechanics, or endpoints of fracture. Whether inclusion of skeletal phenotypes and muscle function may better identify individuals at increased risk of fracture without DXA-defined osteoporosis would be best studied with fracture as an endpoint, though many studies such as ours are not statistically powered for fracture endpoints. We also limited our DXA-based measures to FN aBMD and FRAX scores. As a cross-sectional study, we were unable to investigate the causality of these bone–muscle associations, how these phenotypes developed over time or the underlying bone remodeling during growth and adulthood.

Conclusions

In older women and men, clustering of within-sex standardized HR-pQCT measures of bone density, morphometry, and microarchitecture yielded 3 unique phenotypic clusters. For the DT, C2 and C3 exhibited lower μFEA estimated failure loads. The Ct.LF was higher in C2, indicating reliance on cortical bone and while the Ct.LF in C3 was lower, indicating reliance on the trabecular bone to achieve similarly lower μFEA failure loads in C2 and C3. Muscle function measures differed between clusters, and membership in DT C3 was associated with lower leg power. This study highlights the importance of HR-pQCT to distinguish skeletal phenotypes which have unique associations with lower power, with possible implications for bone–muscle interaction. These findings emphasize the need for use of other metrics to characterize bone phenotypes beyond aBMD.

Supplementary Material

Supplementary material is available at Journal of Bone and Mineral Research online.

Data availability

The analysis dataset for this specific manuscript is available on request from the corresponding author. The SOMMA dataset will be available via the study’s public data release site (https://sommaonline.ucsf.edu/). The SOMMA dataset release from 2024 was utilized for all analyses.