Database and study design

This cohort study used data from the TriNetX [13] database, which collects deidentified, patient-level data from electronic health records. Information in the TriNetX database comes from health care organizations (HCOs), typically academic health care centers, that collect data from their main and satellite hospitals and outpatient clinics. Available data include demographics, diagnoses (based on International Classification of Diseases, Tenth Revision, Clinical Modification codes), procedures (classified by International Classification of Diseases, Tenth Revision Procedure Coding System or Current Procedural Terminology), medications (Veterans Affairs Drug Classification System and RxNorm codes), laboratory tests (organized using Logical Observation Identifiers Names and Codes), and health care utilization records. We used the US Collaborative Network in TriNetX, which includes data from more than 100 million patients from 67 HCOs in the database. This study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki and was reviewed and approved by the Institutional Review Board of Chi Mei Medical Center (IRB No. 11402-E02). The requirement for informed consent was waived because this study used de-identified data.

Patient selection

This study utilized the TriNetX Research Network database to identify the study population between January 1, 2017, and December 31, 2019.

Adult patients (aged ≥18 years) with T2DM who were using either SGLT2 inhibitors or GLP-1 RAs were identified using the 10th edition of the International Classification of Diseases (ICD-10) and RxNorm Concept Unique Identifiers. The diagnosis of T2DM was based on the ICD-10 code E11.

Patients with a prior diagnosis of osteoarthritis (ICD-10: M15–M19) were excluded from the study. To differentiate the two cohorts, one group was defined by the use of SGLT2 inhibitors and the other by the use of GLP-1 RAs. Patients who had a history of using the comparator medication were excluded from each respective cohort (e.g., individuals in the SGLT2 inhibitor cohort with prior GLP-1 RA use were excluded).

Only adults newly prescribed either an SGLT2 inhibitor or a GLP-1 RA were included. New users were defined as individuals with no prescriptions for the index medication or its comparator within the 12 months preceding the index date (defined as the date of the first prescription).

SGLT2 inhibitor use was identified based on prescriptions for canagliflozin (RxNorm: 1373458), dapagliflozin (1488564), empagliflozin (1545653), ertugliflozin (1992672), bexagliflozin (1617044), or sotagliflozin (2638675). GLP-1 RA use was identified through prescriptions for semaglutide (1991302), albiglutide (1534763), dulaglutide (1551291), liraglutide (4755968), lixisenatide (1440051), or exenatide (60548).

Patients were classified and followed according to their initially prescribed index medication, regardless of subsequent treatment discontinuation, switching, or add-on use of the comparator medication. Post-index treatment changes were not used to reclassify exposure status in the primary analysis.

Covariates

We extracted covariates from each patient for the propensity score models, selecting variables that were likely to confound the relationship between the choice of SGLT2 inhibitors versus GLP-1 RAs and the prespecified outcomes [14]. These covariates included demographic factors, hemoglobin A1c levels, and comorbid conditions such as essential hypertension, ischemic heart disease, chronic kidney disease, cerebrovascular disease, and peripheral vascular disease, obesity, and concurrent medications (including metformin and insulin). A comprehensive list of covariates is provided in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Baseline characteristics before and after propensity-score matching. Continuous variables are presented as mean ± standard deviation and categorical variables as number (%). SMD, standardized mean difference. Covariate balance was primarily assessed using SMDs, with an SMD < 0.1 considered acceptable.

https://doi.org/10.1371/journal.pone.0353956.t001

Outcome

Over the five-year follow-up period, the primary outcome was the incidence of OA, defined as the presence of ICD-10 codes M15–M19 recorded on at least two separate clinical encounters. This approach was used to improve diagnostic specificity and reduce potential misclassification, consistent with methods commonly adopted in claims-based epidemiological research. The incidence of OA was subsequently compared between patients prescribed SGLT2 inhibitors and those prescribed GLP-1 RAs.

To further investigate specific manifestations of OA and related interventions, several secondary outcomes were evaluated. These included the incidence of hip OA (ICD-10: M16), knee OA (ICD-10: M17), total hip arthroplasty (THA; CPT: 27130), total knee arthroplasty (TKA; CPT: 27447), and major joint injections (CPT: 20610, 20611). Hip OA and knee OA were defined as the corresponding ICD-10 diagnosis recorded on at least two separate clinical encounters during follow-up. Major joint injections were defined as at least two recorded injection events during the follow-up period to enhance diagnostic specificity and better reflect clinically relevant treatment exposure.

The rates of these outcomes were calculated and compared across the matched cohorts of patients receiving either SGLT2 inhibitors or GLP-1 RAs. Follow-up continued until the earliest of the following events: occurrence of the outcome, death, or December 31, 2024. The diagnostic, medication, exclusion, outcome, demographic, and laboratory data sources used to define the study cohorts, exposures, outcomes, and baseline covariates are summarized in S1 Table.

Statistical analysis

Continuous variables were expressed as mean ± standard deviation (SD), while categorical variables were presented as counts and percentages. To minimize the impact of confounding factors, propensity scores were estimated using a logistic regression model including age at index, sex, race, HbA1c, obesity, essential hypertension, ischemic heart disease, chronic kidney disease, cerebrovascular disease, peripheral vascular disease, metformin use, and insulin use. Patients were matched in a 1:1 ratio using the TriNetX built-in greedy nearest-neighbor algorithm without replacement, with a caliper width of 0.2 standard deviations of the logit of the propensity score. Missing or unavailable covariate information was handled within the TriNetX built-in propensity-score matching workflow, and no additional imputation was performed by the investigators. Covariate balance was assessed using standardized mean differences, with an SMD < 0.1 considered acceptable. Hazard ratios (HRs) with corresponding 95% confidence intervals (CIs) were calculated for both primary and secondary outcomes using Cox proportional hazards models. Additionally, Kaplan–Meier analysis was used to evaluate time-to-event and frequency outcomes between groups. All statistical analyses were conducted using the TriNetX platform. Statistical significance was defined as P < 0.05.

Patient selection

A flowchart of the cohort construction is provided in Fig 1. This study analyzed data from the TriNetX network (as of December 30, 2024), which includes 67 HCOs and 104,785,291 individuals. We selected patients aged 18 or older who had at least two HCO visits between 2017 and 2019, resulting in 17,480,946 eligible individuals. Among them, 6,436,562 patients had used either SGLT2 inhibitors or GLP-1 RAs. After applying exclusion criteria, the SGLT2 inhibitor cohort included 586,195 users, while the GLP-1 RAs cohort comprised 545,583. Using 1:1 PSM, we established a final cohort of 452,445 patients in each group.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Flowchart of cohort construction.

Patients with type 2 diabetes mellitus who initiated either sodium–glucose cotransporter 2 (SGLT2) inhibitors or glucagon-like peptide-1 receptor agonists (GLP-1 RAs) between January 1, 2017, and December 31, 2019, were identified from the TriNetX US Collaborative Network. Patients with pre-existing osteoarthritis or prior use of the comparator medication were excluded. The remaining patients were further restricted to new users of the index medication, defined as having no recorded use of the assigned drug class or comparator medication within the 12 months before the index date. After 1:1 propensity-score matching using baseline covariates, 452,445 patients were included in each group. T2DM, type 2 diabetes mellitus; SGLT2, sodium–glucose cotransporter 2; GLP-1 RAs, glucagon-like peptide-1 receptor agonists.

https://doi.org/10.1371/journal.pone.0353956.g001

Demographic features of included patients

Table 1 presents a comparison between the SGLT2 inhibitor and GLP-1 RAs groups before and after PSM. Before matching, the SGLT2 inhibitor group was older on average than the GLP-1 RAs (60.9 ± 13.3 vs. 55.4 ± 14.1, P < 0.001). There were also significant differences in gender distribution, with a higher proportion of males in the SGLT2 inhibitor group (57.9% vs. 46.1%, P < 0.001). Racial composition differed significantly between groups, except for the proportion of White individuals (P = 0.096).

Regarding laboratory data, the SGLT2 inhibitor group had a slightly higher HbA1c level than the GLP-1 RAs group (7.49 ± 2.03 vs. 7.33 ± 2.03, P < 0.001). In terms of comorbidities, the SGLT2 inhibitor group had a lower prevalence of essential hypertension, ischemic heart disease, chronic kidney disease, cerebrovascular disease, and peripheral vascular disease than the GLP-1 RAs group (P < 0.001 for all).

After PSM, 452,445 matched patients were included in each group, leading to a more balanced distribution of baseline characteristics. All standardized mean differences after matching were below 0.1, indicating acceptable covariate balance. Although chronic kidney disease (P = 0.004), cerebrovascular disease (P = 0.001), peripheral vascular disease (P = 0.001), and HbA1c levels (P < 0.001) remained statistically different between groups, the magnitude of these differences was small according to SMD values.

Clinical and research implications

These findings may be considered hypothesis-generating and may inform future studies evaluating whether antidiabetic drug selection is associated with musculoskeletal outcomes in patients with T2DM. They also underscore the need for individualized treatment strategies, as different therapeutic agents may be associated with different musculoskeletal outcome profiles depending on patient-specific risk factors and comorbidities.

From a research perspective, these findings warrant further investigation into the mechanisms that may explain the observed association between SGLT2 inhibitor use and lower OA-related outcomes. Future studies could explore the effects of these agents on cartilage tissue, inflammatory cytokine profiles, oxidative stress markers, and other pathways relevant to OA development. Additionally, long-term prospective studies and randomized controlled trials are needed to confirm the observed associations and to determine whether SGLT2 inhibitors influence OA onset or progression in broader patient populations.

Limitations and strengths

This study has several limitations. First, its retrospective observational design precludes the establishment of causality. Second, although propensity-score matching was used to improve baseline comparability, residual confounding cannot be excluded. Unmeasured or incompletely captured factors, including changes in body weight, duration and severity of diabetes, physical activity, healthcare utilization intensity, medication adherence, and post-index treatment changes, may have influenced the observed associations. Patient-level post-index medication trajectories could not be reliably extracted within the aggregate TriNetX analytic workflow available for this study. Therefore, the proportion of patients who subsequently discontinued, switched to, or combined the comparator medication during follow-up could not be directly ascertained, and a sensitivity analysis using an as-treated framework was not feasible. As patients were classified according to their initially prescribed index medication in the primary analysis, potential exposure misclassification related to post-index treatment discontinuation, switching, or add-on use cannot be fully excluded. Third, OA-related outcomes were identified using ICD-10 diagnosis codes and procedure codes rather than radiographic findings, clinical severity grading, or patient-reported symptoms. Therefore, outcome misclassification may have occurred, and this study could not distinguish different stages or severities of OA.

This study also has several strengths. The use of a large multi-institutional database allowed the inclusion of a substantial number of patients with T2DM and enabled evaluation of multiple OA-related outcomes. In addition, the active-comparator, new-user design may have reduced confounding related to treatment indication and prior medication exposure. Propensity-score matching further improved measured baseline comparability between the SGLT2 inhibitor and GLP-1 RAs groups.