Study design and ethical approval

We conducted a retrospective cohort study utilising nationwide administrative health data, and reported it in line with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement. ^{Reference Von Elm, Altman, Egger, Pocock, Gøtzsche and Vandenbroucke13} Ethical approval was granted by the Institutional Review Board of Seoul National University Bundang Hospital (X-2309-852-901). The NHIS reviewed the study protocol (NHIS-2025-04-1-102) and permitted data access. Since the data-set was anonymised and originally collected for reimbursement and administrative use, the requirement for informed consent from individuals was waived, consistent with Korean regulations and the principles of the Declaration of Helsinki (1975, revised 2008).

Data source

The analysis relied on the NHIS database, which is South Korea’s nationwide compulsory health insurance programme covering nearly the entire population. This resource includes longitudinal, person-level information on demographics, diagnostic codes, healthcare utilisation, prescriptions, and mortality. Diagnoses are coded according to the ICD-10. Although most clinical care is delivered in private medical settings, mandatory claims reporting to the NHIS and centralised data management ensure high completeness and reliability. ^{Reference Lee, Lee, Park, Shin and Kim14}

Study population

Within the NHIS database, all physicians – including psychiatrists – are required to record ICD-10 diagnostic codes for reimbursement, which supports the accuracy of psychiatric diagnoses. We included adults aged ≥18 years who received at least one diagnosis of a major psychiatric disorder in 2021 (exposure period). The psychiatric conditions of interest were schizophrenia spectrum disorders (F20, F23, F23.2, F25), bipolar disorder (F30, F31), autism spectrum disorders (F84.0, F84.1, F84.5, F84.9), eating disorders (F50), alcohol use disorders (F10), attention-deficit hyperactivity disorder (F90), obsessive–compulsive disorder (F42), major depressive disorder (single and recurrent episodes; F32, F33), tic disorders (F95), post-traumatic stress disorder (F43.1) and anxiety disorders (F40, F41). A total of 331 020 individuals satisfied these inclusion criteria (psychiatric disorder group).

For the reference cohort, we identified 668 980 adults without any psychiatric diagnoses in 2021 (non-psychiatric disorder group). To ensure a rigorous comparison and minimise confounding, we performed 1:1 propensity score matching between both groups, based on all baseline covariates measured in 2021. The final matched cohort for the primary analysis consisted of 567 884 participants, with 283 942 in each group. Figure 1 illustrates the detailed flow of cohort construction.

Fig. 1

Flowchart of study population selection.

Fig. 1 Long description

The flowchart begins with two groups: the Psychiatric disorder group with 331,020 individuals and the Non-Psychiatric disorder group with 668,980 individuals. These groups converge into a Total adult cohort of 1,000,000 individuals. The next step involves 1:1 propensity score matching, resulting in two matched groups: the psychiatric disorder group with 283,942 individuals and the Non-psychiatric disorder group with 283,942 individuals.

Study outcomes

The primary end-point was to evaluate and compare the sustained use of oral glucocorticoids between the psychiatric disorder and non-psychiatric disorder cohorts during the follow-up period (1 January to 31 December 2022). To ensure a clear temporal sequence, only glucocorticoid prescriptions occurring after the established psychiatric diagnosis in 2021 were considered for the primary outcome analysis.

Sustained use was operationally defined as a cumulative supply of systemic glucocorticoids lasting at least 90 days within the follow-up year. Consistent with previous epidemiological studies, this was further refined as continuous treatment maintained for 90 or more days, permitting prescription gaps of up to 30 days to account for potential variations in medication adherence or refill timing. ^{Reference Benard-Laribiere, Pariente, Pambrun, Begaud, Fardet and Noize1,Reference Iki, Fujimori, Nakatoh, Tamaki, Ishii and Okimoto15} For participants whose initial glucocorticoid prescription began in the latter part of the follow-up period (after 1 September 2022), prescription records were tracked through 31 March 2023, to ensure a sufficient window for the appropriate classification of sustained exposure.

Covariates

Baseline characteristics included demographic and socioeconomic indicators. We accounted for age and gender, as well as employment status, residential setting and household income at the index date. In the NHIS system, the lowest-income beneficiaries exempt from insurance premiums are covered by the Medical Aid programme, and the remaining population is distributed into quartiles based on income (quartile 1 being the lowest and quartile 4 being the highest). Residential area was classified as metropolitan (Seoul and other large cities) versus non-metropolitan regions.

Overall morbidity burden was assessed using the Charlson Comorbidity Index (CCI), derived from ICD-10 claims data (Supplementary Table 1). Disability status was also incorporated. In Korea, disability registration requires physician certification and formal review, with severity ratings reflecting functional limitations in daily life. For this analysis, disability was grouped into mild-to-moderate or severe, with detailed categories presented in Supplementary Table 2.

We additionally adjusted for baseline musculoskeletal disorders, autoimmune conditions, and clinical indications and metabolic status that may affect glucocorticoid prescribing, including fibromyalgia, chronic low back pain, other chronic spinal pain, degenerative osteoarthritis, rheumatoid arthritis, headache, trigeminal neuralgia and myofascial pain. Autoimmune conditions were also considered, specifically ankylosing spondylitis, autoimmune hepatitis, Behçet’s disease, Crohn’s disease, polymyositis, polyarteritis, Sjögren syndrome, systemic lupus erythematosus, systemic sclerosis and ulcerative colitis. Moreover, clinical indications and metabolic status included chronic respiratory diseases (asthma and chronic obstructive pulmonary disease (COPD)) and metabolic factors (obesity, dyslipidaemia, and other metabolic disorders) (ICD-10 codes listed in Supplementary Table 3).

Statistical analysis

Continuous variables were expressed as means with standard deviations, whereas categorical variables were summarised as frequencies with percentages. Although the non-psychiatric disorder cohort was initially generated using age- and gender-based stratification, residual differences in other baseline variables remained. To mitigate these imbalances, additional 1:1 propensity score matching was applied. ^{Reference Rosenbaum and Rubin16} Propensity score values were estimated through logistic regression including all predefined covariates, and nearest-neighbour matching without replacement was used, with a caliper width set to 0.25 times the standard deviation of the logit of the propensity score. Covariate balance before and after matching was assessed using absolute standardised mean differences, with thresholds <0.1 considered acceptable.

Primary outcomes (sustained systemic glucocorticoid use) were evaluated using conditional logistic regression models for the matched cohort to account for the paired nature of the data. Results are presented as odds ratios with 95% CIs. We also fitted multivariable logistic regression models in the total population, adjusting for all covariates and incorporating specific psychiatric diagnoses – including major depressive disorder (single and recurrent episodes), anxiety disorder, post-traumatic stress disorder and eating disorders – as separate predictors. Multicollinearity was assessed using variance inflation factors (<2.0), and model calibration was verified using the Hosmer–Lemeshow test.

To evaluate the relationship between physical comorbidities and healthcare-seeking behaviour, we examined the correlation between the CCI and the total number of out-patient clinic visits by using Spearman’s rank correlation analysis. The analysis revealed a moderate correlation (rho = 0.48, P < 0.001), confirming that these variables capture distinct clinical constructs. Furthermore, to avoid potential overadjustment bias, our primary propensity score model excluded out-patient visit frequency, as healthcare utilisation was identified as a plausible mediator rather than a baseline confounder on the causal pathway between psychiatric status and subsequent prescribing. A sensitivity analysis including visit frequency was conducted to assess the robustness of the findings (Supplementary Tables 4 and 5). Additionally, E-values were calculated to quantify the strength of unmeasured confounding required to nullify our observed associations.

Subgroup analyses were conducted to evaluate the robustness of the findings according to gender, age (<65 v. ≥65 years), and the presence of underlying musculoskeletal diseases, autoimmune conditions or chronic respiratory diseases (COPD or asthma). All statistical analyses were carried out using R software version 4.5.1 for Windows (R Foundation for Statistical Computing, Vienna, Austria; see https://www.r-project.org/foundation/), and two-sided P-values < 0.05 were regarded as statistically significant.