The National Health Insurance Database (NHID) was used in this study. The NHID and Cause of Death database were provided and authorized access by the Health and Welfare Data Science Center, Ministry of Health and Welfare, Taiwan. These databases provide person-level information through personal identification number linkage [13]. The NHID is populated from a universal single-payer healthcare program that covers 99% of Taiwan’s population. It is composed of a registry of beneficiaries, ambulatory care claims, inpatient claims, and prescription dispensing claims of pharmacies. Each claim contains a diagnosis code according to the International Classification of Diseases, Ninth Edition, Clinical Modification (ICD-9-CM, up to December 31, 2015) and Tenth Edition (ICD-10-CM, after January 1, 2016). In addition, medications, laboratory tests, and surgeries are recorded with National Health Insurance codes. For details of the NHID, refer to Hsieh et al. [13].

We conducted a nested case–control study [14] to investigate the association between CDI and duration of PPI or H2RA use in hospitalized patients. The study cohort was identified as patients with a first admission who began to use PPIs or H2RAs in the NHID from 2012 to 2018. Patients were excluded if they were aged under 20 years, had a history of PPI or H2RA prescription during a 180-day period prior to the admission date, had a diagnosis of peptic ulcer (ICD-10-CM codes K25–K28), had acquired immunodeficiency syndrome (ICD-10-CM codes B20–B22, B24), or were pregnant (O00-O99, O9A). A length of stay (LOS) longer than 90 days was considered excessive, and as these patients could have a greater illness severity or other critical issues they were excluded from our study cohort (Fig. 1).

Assembly of the study population

The duration of follow-up was based on PPI or H2RA exposure days in an “as treated” way. The exposure days to PPIs or H2RAs began on the date of the first prescription filling and was censored at the date of discontinuation, switching, death, occurrence of CDI, discharge from hospital, or the end of the study period after the first prescription, whichever came first. Discontinuation was defined as patients who stopped receiving PPIs or H2RAs within 7 days after the date of the last prescription. Baseline variables were constructed from the NHID during the 180-day pre-admission period. Comorbidities and co-medications were assessed according to the Charlson comorbidity index (CCI) [15], and CCI scores were calculated in this study. Information regarding the antibiotic use in our study, it was measured as a co-medication during the 2-week pre-admission period and was considered as a baseline variable in our analysis. The date of incident CDI was defined as the index date. Details of the baseline variables are presented in Table 1 and in Supplemental Table 1 in the supplementary data.

According to the nested case–control study, the source population was our study cohort of hospitalized patients who were prescribed a PPI or H2RA for treatment. We then followed these patients, and if a patient had CDI during the PPI or H2RA exposure period, it was considered a case. Patients with CDI in this study were considered to be those who received metronidazole, fidaxomicin, or vancomycin orally for treatment for at least 7 days [16]. If a patient becomes a case, we would randomly select a patient who did not have CDI at that time from the source population and matched by age, sex, and calendar year of entry into the cohort according to incidence density sampling as a control until the end of follow-up during the PPI or H2RA exposure period. The final case and control groups were matched to a case on a 1:1 basis for age, sex, and calendar year of cohort entry. The nested case–control study design in this study is shown in Supplemental Fig. 1.

To confirm that our results were not biased by misclassification of CDI, we performed two sensitivity analyses. First, we restricted our study cohort to between 2017 and 2018, and cases were defined as patients with records of CDI testing or diagnosis of CDI (ICD-10-CM codes 8.45, A04.7) in addition to metronidazole, fidaxomicin or vancomycin oral treatment for at least 7 days. The C. difficile GDH Ag rapid test and C. difficile toxin A/B rapid test have been reimbursed by the National Health Insurance program since 2017. We therefore used a strict definition to identify patients with CDI and performed the analysis again. Second, in order to avoid community-acquired infection, we excluded patients who had CDI within 3 days between admission date and index date and performed the analysis again.

Descriptive statistics were used to characterize case and control characteristics. Continuous variables were described as means with standard deviations (SD), and categorical variables by numbers and proportions. The standardized mean difference (SMD) was used to characterize the differences in baseline characteristics between the case and control groups. We used univariable conditional logistic regressions to estimate the sensitivity, specificity, and area under the receiver operating characteristic curve (AUC ROC) for CDI associated with duration of PPI or H2RA use, respectively. Youden’s J statistic was used to identify the optimal cutoff duration for the length of PPI or H2RA use. According to the results of Youden’s J statistic, we stratified the duration of PPI or H2RA use into shorter cutoff duration and longer cutoff duration groups, and the adjusted odds ratios (aORs) were estimated to compare the risk of CDI in patients using PPIs and H2RAs for shorter or longer cutoff durations by multivariable conditional logistic regression. An example of estimating odds ratios is shown in Supplemental Table 2.