Data used in this nationwide cohort study were obtained from the National Health Insurance Research Database (NHIRD) in Taiwan since January 1, 2000, to December 31, 2015. The NHIRD was set up in 1997 from the National Health Insurance (NHI), which is a mandatory national health insurance program that covers more than 99% of the Taiwanese population . The database contains comprehensive information, including demographic characteristics, disease diagnoses, and details of prescriptions. Diseases were recorded using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM). The NHIRD has been widely used in epidemiological studies and is one of the most complete health care service databases in the world . The quality of the NHIRD, in terms of the diagnostic accuracy of major diseases and medications, has been well validated [26, 27]. Data confidentiality is assured by the regulations of the NHIRD Administration, Ministry of Health and Welfare, Taiwan. This study was performed in accordance with the Helsinki Declaration and was approved by the National Health Research Institutes and the Institutional Review Board of Taipei Veterans General Hospital (IRB No.: 2019-01-007AC).
The patient selection process is shown in Figure 1. Diabetic patients were evaluated between January 1, 2000, and December 31, 2015. Individuals were defined as having diabetes if there were at least three diagnosis codes (ICD-9-CM: 250) in their outpatient clinic records or one in their admission record, and if they had at least 3- months of anti-diabetic medications. To avoid confounding and to investigate the impact of DPP4i on BP, we excluded diabetic patients with cancers and neurological diseases, including cerebrovascular diseases, dementia, Alzheimer’s disease, Parkinson’s disease, epilepsy, multiple sclerosis, and psychiatric diseases, which are well-known risk factors for BP [28, 29].
Sitagliptin was the first DPP4i reimbursement by NHI in Taiwan in 2009, followed by saxagliptin and vidagliptin in 2011, and linagliptin in 2012. They were used alone or in combination with metformin as the second line treatment for diabetes. Since no data is available for the duration of DPP4i use to the risk of BP, a selection criterion of 3 months was used for enrolling DPP4i users. Patients who had used DPP4i ≥ 28 days per month for at least 3 months were identified as the DPP4i-treated group. An index date was assigned as the date after 3 months of DPP4i medication. The duration of diabetes was calculated from the date of diabetes diagnosis to the index date. We defined those who had used metformin at least 3 months over 1 year as metformin user. To further identify severely diabetic patients who required insulin treatment, insulin users were defined as diabetic patients who had used insulin for at least 3 months over 1 year.
Patients who used other antidiabetic medications, with the exception of DPP4is, for at least 3 months were assigned to the non-DPP4i-treated cohort. Because DPP4is were used as a second line anti-diabetic medication, those who had not taken DPP4is may have been new-onset or mild diabetic patients. Because diabetes has been reported as a risk factor for BP , we selected patients with the same duration of diabetes to include a comparable control. We assigned a pseudo-index date for each non-DPP4i patient and included those with the same duration of diabetes with the DPP4i-treated group. Patients who met the exclusion criteria before the pseudo-index date of non-DPP4i group were further excluded. A total of 124,619 patients who received DPP4i therapy were randomly matched 1:1 with those who had never received DPP4i by means of age (± 0.5 year), sex, duration of diabetes (± 0.5 year), insulin usage, and propensity scores (± 0.01) of comorbidities.
The occurrence of BP was the measured outcome. Patients who were diagnosed with BP were identified by ICD-9-CM 694.5. Individuals were defined as having BP if there were at least three diagnosis codes in their outpatient clinic records or one in admission record. Patients who developed BP before the index dates were excluded. Study participants were followed up until the date of BP diagnosis, death, or the end of the study period.
Assessment of accuracy of BP identification algorithm
To evaluate the accuracy of BP diagnosis by our defined algorithm, we retrieved BP data from the electronic medical records of the Taipei Veterans General Hospital (a 2,802-bed teaching hospital in Taipei) between January 1, 2011 and December 31, 2015. We reviewed patients’ clinical information, pathology results, direct and indirect immunoflourescence data, and medical records, as reference standards to estimate the positive predictive value of the BP identification algorithm. According to the BP identification algorithm, a total of 247 patients with BP were identified from the electronic medical records of the Taipei Veterans General Hospital. Among these algorithm-identified patients, 242 were confirmed through medical record reviews, suggesting that our algorithm had a good positive prediction value of 98.0% (95% confidence interval [CI]: 96.3-99.7%)
Major coexisting diseases
Several major coexisting diseases that might be associated with the risk of BP were identified by ICD-9-CM codes and included coronary artery disease, hypertension, hyperlipidemia, renal disease, chronic liver disease and cirrhosis, chronic obstructive pulmonary disease, and connective tissue disease. The diseases were included in the calculation of propensity score.
Continuous data were summarized with mean, median and interquartile range (IQR) whereas categorical data with number and proportion, unless otherwise specified. Cumulative incidence rates of BP during the follow-up were calculated. After adjusting for competing mortality, cumulative incidence rates were calculated and compared using modified Kaplan-Meier and Gray methods . A modified log-rank test was used to compare differences in the full time-to-event distribution between the DPP4i-treated and untreated groups.
Multivariable analyses using hazard ratios (HRs) were performed with modified Cox hazards models in the presence of competing risk events, and adjusted for covariate factors, including use of DPP4i, age, sex, duration of diabetes, all the major coexisting diseases, use of insulin, use of metformin, and the interaction term of DPP4i and insulin. Further stratified multivariable analyses were performed. Two-sided P values < .05 were considered statistically significant. All data were managed using SAS software, version 9.3 (SAS Institute, Gary, NC). The Cox proportional hazard regression models in the competing risk analysis were carried out using the “cmprsk” package of R (http://cran.r-project.org/web/packages/cmprsk/index.html).