Semaglutide's links to bone health in people with obesity, with and without type 2 diabetes

This retrospective study used the TriNetX US Collaborative Network, a federated database of de‐identified records from approximately 120 million patients across 65 healthcare organisations in the United States, containing diagnoses (ICD‐10‐CM), procedures, prescriptions and laboratory data.

The Institutional Review Board at Chung Shan Medical University Hospital approved this study (CS2‐24004, CS2‐24100). TriNetX provides only de‐identified aggregate data; accordingly, the IRB waived consent under 45 CFR 46.104 (d) (4). Reporting followed STROBE guidelines.

We emulated two parallel target trials in TriNetX to compare skeletal outcomes after initiation of semaglutide in people living with obesity, stratified by type 2 diabetes (T2D) status. Eligibility windows were May 13, 2022 to Feb 28, 2026 for obesity with T2D; Nov 8, 2023 to Feb 28, 2026 for obesity without T2D. Three comparator categories are used throughout the manuscript. Active comparators denote individual drugs compared separately with semaglutide: empagliflozin, sitagliptin and glipizide in the T2D cohort; naltrexone–bupropion, phentermine and phentermine–topiramate in the non‐T2D cohort. Usual care denotes patients who did not initiate any index or comparator medication and were treated with metformin in both cohorts. In both cohorts, usual care was defined as metformin use without initiation of semaglutide or any active comparator medication. Pooled active comparators, reported only in Supporting Information, combine all individual active comparators within each cohort into a single reference group. These windows were aligned with the respective FDA approval dates (Ozempic, December 2017; Wegovy, June 2021). An approval‐aligned sensitivity analysis restricted each comparison to the later drug's approval date. The first qualifying prescription defined the index date (T₀). A new‐user design required no prescription for the index drug or same‐class medication during the 6 months before T₀. A directed acyclic graph (Figure S1) guided covariate selection, separating confounders from mediators. Post‐baseline mediators (weight change, glycemic trajectories) were not adjusted for to avoid blocking the pathway of interest.

For semaglutide, active comparators were selected based on prescribing frequency within each stratum. In the T2D cohort, empagliflozin (SGLT2 inhibitor), sitagliptin (DPP‐4 inhibitor) and glipizide (sulfonylurea) were the three most commonly started second‐line glucose‐lowering agents during the study period; usual care was defined as stable metformin monotherapy without initiation of any index or comparator medication. In the non‐T2D cohort, naltrexone–bupropion, phentermine and phentermine–topiramate were selected as FDA‐approved anti‐obesity medications with the largest user base; usual care was defined as metformin treatment without initiation of semaglutide or any active comparator medication. Each comparator carries a distinct skeletal risk profile. Early CANVAS trial data raised concern about fractures associated with canagliflozin, though subsequent studies of empagliflozin and dapagliflozin did not replicate this signal [14, 15]. Phentermine is generally prescribed for shorter durations to younger patients with fewer comorbidities [16]. These differences in patient profiles were addressed through 1:1 propensity score matching within each comparison. We conducted parallel active‐comparator analyses, including head‐to‐head contrasts with semaglutide and additional comparisons with selected glucose‐lowering or weight‐lowering agents. We also evaluated pooled comparator strategies within each cohort, defined as initiation of any prespecified active comparator in that stratum, to summarise associations against a broader treatment mix. All primary analyses report individual drug comparisons.

Following target trial emulation guidance [17, 18, 19], we prespecified eligibility, treatment strategies, assignment procedures, follow‐up, outcomes, contrasts and the analysis plan (Table S1). All analyses were interpreted as comparative associations under real‐world treatment initiation and follow‐up patterns, rather than as causal effects.

In each cohort, exposure groups were defined by initiation of semaglutide identified from prescription records. Active comparators were defined by initiation of the corresponding comparator medication. For pooled comparator analyses, the comparator group comprised initiators of any prespecified active comparator within that cohort (Figure 1). The index date was the first prescription within the observation window. Within the usual care groups, metformin use was part of the operational definition. Baseline use of these drugs was captured from the supplemental medication records and included as indicator variables in the propensity score.

Exclusions included other/unspecified diabetes (E08, E09, E13), type 1 diabetes (E10), prior bariatric surgery or organ transplantation, HIV and end‐stage renal disease. The primary analysis followed an intention‐to‐treat approach. Bariatric surgery was not a censoring event because it is a post‐baseline intervention potentially influenced by the index treatment; censoring at surgery would introduce informative censoring bias if surgical probability differs between groups. Under the primary ITT estimand, follow‐up started at T0 and continued regardless of treatment discontinuation, switching or augmentation, and ended at the first occurrence of the outcome, death, loss to follow‐up or administrative end of follow‐up. Prescription records capture orders not confirmed dispensing; resultant misclassification would attenuate ITT estimates toward the null. The per‐protocol analysis partially addresses this.

The primary endpoint was MOF, defined as a composite of hip (S72), clinical vertebral (S32.0–S32.2), distal radius/ulna (S52.5–S52.6), proximal humerus (S42.2–S42.3) and fragility fractures (M80). Secondary fracture outcomes included individual MOF components. The secondary outcome was incident osteoporosis (M80–M81). Exploratory outcomes were knee osteoarthritis (M17), hip osteoarthritis (M16) and gout (M10), disorders of bone density and structure (M85.8–M85.9), bone‐active medication use (alendronate, zoledronic acid, denosumab), healthcare utilisation (outpatient, emergency, inpatient) and described longitudinal BMI and HbA1c trajectories under the ITT principle, with measurements attributed to the original treatment group regardless of subsequent treatment changes. These trajectories reflect the combined influence of the index drug, switching, dose adjustment and concomitant medications during follow‐up. Knee osteoarthritis, hip osteoarthritis and gout were exploratory outcomes. A validated hip fracture endpoint, defined by hip fracture diagnosis combined with inpatient surgical procedure codes, was analysed as a sensitivity analysis. To probe residual confounding, we included negative control outcomes with no expected relation to exposure: dog bites, ganglion cysts, adhesive capsulitis of shoulder, blepharitis, hernias and skin cancer (). Supporting Information

We adjusted for 215 covariates to balance treatment groups. These included demographics (age, sex, race, socioeconomic status), lifestyle factors (nicotine dependence, alcohol‐related disorders), comorbidities (hypertension, ischemic heart disease, heart failure, hyperlipidemia, acute and chronic kidney disease), concomitant therapies (insulin, metformin, SGLT2 inhibitors, thiazolidinediones, sulfonylureas, DPP‐4 inhibitors; ACE inhibitors, ARBs, beta‐blockers, calcium‐channel blockers, statins; bisphosphonates), laboratory measures (BMI, HbA1c, eGFR, proteinuria, lipid profile), systemic glucocorticoid prescriptions within 12 months before index, DXA utilisation, healthcare utilisation intensity (emergency, ambulatory and inpatient visits during 12 months before index), fall history and prior traumatic fracture. Comorbidities were ascertained via ICD‐10 codes, medications from prescription records and labs from structured EHR fields. These variables entered the propensity‐score matching to improve balance (). Supporting Information

Propensity score matching was used because exact matching on 215 covariates (including continuous measures such as BMI, HbA1c, eGFR and lipid values) was not feasible. Each contrast was analysed in an independently matched cohort using 1:1 nearest‐neighbour matching with a calliper of 0.2 standard deviations of the logit propensity score; balance was confirmed by standardised mean differences below 0.10. Cox proportional hazards models estimated hazard ratios with 95% confidence intervals. The Benjamini–Hochberg procedure controlled the false discovery rate across all comparisons; both raw and adjusted p‐values are reported. E‐values were calculated for selected significant associations. The primary prespecified hypothesis was that semaglutide initiation would be associated with lower MOF hazard compared with active comparators in the T2D cohort. Osteoporosis was prespecified as the secondary outcome. Non‐T2D comparisons and exploratory outcomes (knee OA, hip OA, gout) were interpreted as exploratory. Subgroup analyses were hypothesis‐generating.

Per‐protocol (PP) analyses were performed by restricting to patients persistent on their assigned treatment across cumulative 75‐day refill intervals. Grace periods were tested: 75 days (30‐day prescription plus 45‐day permissible gap) [20]. The proportion of participants remaining on index treatment at each grace period threshold was reported for all treatment groups. The ITT analysis was retained as the primary approach because it reflects the clinical question of treatment initiation and avoids informative censoring that may arise when discontinuation is related to the outcome. Osteoporosis medication use (alendronate, zoledronic acid, denosumab) and healthcare utilisation were analysed as time‐to‐event outcomes using the same Cox proportional hazards approach in each matched cohort. Longitudinal changes in BMI and HbA1c were tracked from the index date across the follow‐up period using the TriNetX analytics module.

Pre‐specified subgroups included eGFR (≥ 45 vs. < 45 mL/min/1.73 m²), HbA1c (≥ 7% vs. < 7%), BMI categories (< 30, 30–34.9, 35–39.9, ≥ 40 kg/m²), cardiovascular comorbidities (hypertension, heart failure, ischemic heart disease), proteinuria and background therapy (insulin, metformin, renin–angiotensin system inhibitors). Sensitivity analyses evaluated a class‐based sensitivity analysis by replacing semaglutide with any GLP‐1 RA (semaglutide, liraglutide, dulaglutide, exenatide) under the same specification. Missing data were handled without imputation. The proportions of missing data for key continuous covariates before propensity score matching are reported in Table S2. To assess the robustness of fracture findings to outcome misclassification, two additional analyses were performed. First, the fracture endpoint was restricted to hip fracture confirmed by surgical procedure codes (CPT 27235, 27 244, 27 245; ICD‐10‐PCS 0QS6 and 0QS7), a validated endpoint with a reported positive predictive value exceeding 95% in administrative databases [21]. Second, site‐specific fracture outcomes (hip, clinical vertebral, distal radius/ulna, proximal humerus) were analysed individually to evaluate whether the composite MOF association was consistent across fracture types. To address the potential influence of differential mortality as a competing event, we examined a composite endpoint of MOF plus all‐cause mortality across all comparisons. Complete‐case analysis was used; sensitivity sets with alternative completeness rules were also examined.

Analyses were run on the TriNetX platform (Java 11.0.16), with statistics in R 4.0.2 and visualisation in Python 3.7 and Prism 10.0. Cell counts < 10 were rounded to the nearest 10 to protect privacy.

Our study included two populations with obesity: those with T2D, compared against glucose‐lowering agents or usual care over 3 years, and those without T2D, compared against anti‐obesity medications or usual care over 2 years (Figure 1; Tables S3–S10). Across the nine T2D comparisons, matched cohorts ranged from 19 824 to 93 519 pairs per arm. After matching, mean age ranged from 55.1 to 62.5 years, and the proportion of female participants varied from 43.0% to 59.2%, reflecting the different sex distributions of each comparator population. Mean BMI fell between 34.1 and 38.2 kg/m², mean HbA1c between 7.1% and 7.5%, and all post‐matching SMDs were below 0.10.

We examined composite MOF across all treatment comparisons (Table 1 and Figure 2). In the T2D cohort over 3 years, semaglutide was associated with lower MOF hazards versus empagliflozin (HR 0.69, 95% CI: 0.61–0.77), glipizide (HR 0.72, 95% CI: 0.63–0.83) and usual care (HR 0.84, 95% CI: 0.76–0.93), all of which remained significant after FDR correction (Table 1). The semaglutide versus sitagliptin comparison showed a similar direction (HR 0.84, 95% CI: 0.70–0.99) and nominal significance before adjustment (raw p = 0.045), but it did not remain significant after FDR correction (adjusted p = 0.113). In the non‐T2D cohort over 2 years, none of the semaglutide comparisons reached statistical significance for MOF (Table 2).

Secondary skeletal outcomes are summarised in Table 1 and Figure 2. In the T2D cohort, semaglutide was not associated with osteoporosis risk versus any individual comparator, though the comparison with empagliflozin showed a trend (HR 0.90, 95% CI: 0.81–1.01; p = 0.071). In the non‐T2D cohort, osteoporosis associations were similarly non‐significant across all comparisons. Among exploratory outcomes, semaglutide was associated with modestly higher knee osteoarthritis hazards in the T2D cohort versus empagliflozin (HR 1.14, 95% CI: 1.06–1.23; p < 0.001) and sitagliptin (HR 1.16, 95% CI: 1.04–1.29; p = 0.006), and in the non‐T2D cohort versus usual care (HR 1.20, 95% CI: 1.08–1.32; p < 0.001; Table 2). Hip osteoarthritis was higher versus usual care in the non‐T2D cohort (HR 1.26, 95% CI: 1.07–1.48; p = 0.006) but not in other comparisons. Gout showed no significant associations across either cohort.

In the PP analysis, the associations between semaglutide and skeletal outcomes were consistent with the ITT results across all persistence thresholds (e.g., semaglutide vs. empagliflozin for MOF: HR 0.59, 95% CI: 0.49–0.71, p < 0.001 at day 75; Table S11). Semaglutide consistently showed higher persistence than their respective comparators in both cohorts (Table S12).

Subgroup analyses are detailed in Tables S13–20. In the T2D cohort, semaglutide showed a stronger association with reduced osteoporosis risk among patients with HbA1c < 7% (HR 0.46, 95% CI 0.24–0.85; p for interaction = 0.03) compared with empagliflozin. Most other subgroup comparisons lacked statistical significance.

In the T2D cohort, site‐specific fracture HRs for semaglutide versus empagliflozin ranged from 0.52 (hip) to 0.77 (humerus), all in the same direction (Tables S21 and S22). In the surgical code–confirmed hip fracture analysis (Figure S2), semaglutide was associated with a lower hazard versus empagliflozin (HR 0.45, 95% CI 0.30–0.67) and glipizide (HR 0.59, 0.35–1.00); comparisons versus sitagliptin and usual care did not reach significance, consistent with smaller event counts in this more restrictive definition. When MOF and all‐cause mortality were combined (Figure S3), the HR was 0.38 (0.36–0.41) versus 0.69 (0.61–0.77) for MOF alone; this gap was largely attributable to the mortality component. In the GLP‐1 RA class‐level sensitivity analysis, the associations with MOF were consistent with the individual semaglutide results, with lower hazards observed versus empagliflozin (HR 0.74, 95% CI: 0.66–0.83, p < 0.001), glipizide and usual care in the obesity with T2D cohort (Figure S4). Semaglutide was also associated with a lower hazard of MOF versus the pooled active comparator in the obesity with T2D cohort (HR 0.72, 95% CI: 0.65–0.80, p < 0.001), consistent with the individual comparisons (Figure S5). Results from the global network and three‐month landmark sensitivity analyses were consistent with the main findings (Figures S6 and S7).

Exploratory analyses of bone‐active medication patterns, healthcare utilisation and longitudinal BMI and HbA1c trajectories are reported in Figures. S8–S10