Use of real-world evidence data to evaluate the comparative effectiveness of second-line type 2 diabetes medications on chronic kidney disease

Study population

The Northwestern Medicine Electronic Data Warehouse (NMEDW) is the primary data repository for all the medical records of patients who receive care within the Northwestern Medicine system ¹⁰. Established in 2007, the NMEDW contains records for over 3.8 million patients, with most EHR data going back to at least 2002, and with some billing claims data going back to 1998 or even earlier ¹⁰. We included patients who met the following criteria: 1) at least one prescription to an ADM 2) at least one diagnosis code for type 2 diabetes (see supplemental Table 1 for type 2 diabetes diagnosis codes)¹¹ 3) no excluded diagnoses: pregnancy, type 1 diabetes mellitus 4) at least one year of records in the database before their first ADM exposure and 5) at least three years of continuous ADM prescription records after first ADM exposure. Given that the first SGLT-2 inhibitor was approved by FDA for use in the United States in March 2013, we only included patients who had their first ADM exposure after 2013-03-01. Figure 1 shows the flow chart of the patient selection process.

• Download figure
• Open in new tab

Figure 1. Cohort selection flowchart.

Abbreviations: T2DM, type 2 diabetes mellitus; ADM, antidiabetic medication; DX, diagnosis; T1D, type 1 diabetes; SU, sulfonylureas; DPP-4 inhibitor, dipeptidyl peptidase 4 inhibitors; GLP-1RA, glucagon-like peptide receptor agonists; SGLT-2 inhibitor, sodium-glucose cotransporter 2 inhibitor; TZD, Thiazolidinediones.

We defined an ADM sequence as the chronological series of ADM prescription orders a patient received from the date of their first ADM prescription to one year later (e.g., “Metformin, Glipizide, Glipizide, Glucophage, Sitagliptin, Glipizide”). We then mapped each generic name and brand name to its respective drug class in Anatomical Therapeutic Chemical Classification (e.g., “Biguanides, SU, SU, Biguanides, DPP-4 inhibitor, SU”) ¹². Repeated listings of the same medication in each sequence were understood to be refilled prescription. Therefore, in this example, only the first occurrence of a medication is kept in the final sequence (e.g., “Biguanides, SU, DPP-4 inhibitor”). During the 1-year period, only patients who were treated with biguanides as their first line of medication and had a second medication of either DPP-4 inhibitor, SGLT-2 inhibitor, GLP-1RA, SU, TZD, or Insulin were included in the study.

Exposure

The exposure in this study was the medication sequence during a patient one-year exposure period as described above. The sequences we were interested in include ‘Biguanides, SU’, ‘Biguanides, DPP-4 inhibitor’, ‘Biguanides, Insulin’, ‘Biguanides, GLP-1RA’, ‘Biguanides, SGLT-2 inhibitor’ and ‘Biguanides, TZD’. A patient might switch to a third medication during the one-year exposure time, we considered them in the same group as the patients who did not. For example, ‘Biguanides, SU, DPP-4’ was considered in the same second-line ADM group as ‘Biguanides, SU’. Patients who were only on Biguanides during the 1-year drug exposure period were excluded from the study. Index date was defined as one year after first the exposure to ADM.

Renal outcomes

Our primary outcome is CKD incidence identified by the first appearance of an associated International Classification of Diseases, 9^th/10^th revision (ICD9/10) diagnosis code (see supplemental Table 1)⁷. Patients who have CKD, ESRD or macroalbuminuria before index date are excluded in the analysis for this outcome. In addition to the incident CKD event, we evaluated ADMs’ associations with CKD hospitalization. CKD hospitalization were identified based on ICD9/10 codes. For any of the outcomes, patients were followed up from their index date until meeting criteria of one of the above outcomes, last hospital visit date, end of 5-year follow up period or end of study date (2019-10-29) whichever comes first.

Covariates

We included 5 major categories of covariates: demographics, laboratory tests related to CKD, diagnoses from medical history before index date, medications that are known to affect CKD related outcome, insurance status and smoking status. Demographics includes age, gender, race were collected at baseline (index date). Race categories included White, Black, Asian, and other races. We used White race as the reference group. Insurance status categories included Medicaid, self-pay, Medicare, commercial insurance. Commercial insurance was used as the reference group. Laboratory test included hemoglobin A1C (HbA1c), high density lipoprotein cholesterol (HDL), total cholesterol (TOTCHL), serum creatinine, body mass index (BMI). Medications include angiotensin-converting enzyme (ACE) inhibitors, aldosterone receptor antagonists (ARA), angiotensin II receptor blocker (ARB), antiplatelet drugs, beta-blockers, calcium channel blockers, diuretics, statins, and other lipid modifying drugs (loop and thiazide diuretics). Medical history includes cardiovascular disease (CVD), congestive heart failure (CHF), hypertension, vascular disease, vascular complication of diabetes (including skin ulcer), diabetic neuropathy, diabetic oculopathy, dyslipidemia, and other diabetic complications (diabetic nephropathy, nephrotic syndrome, nephritis/nephropathy, lower extremity amputations). For CKD hospitalization outcome, prior CKD hospitalization is also included as a covariate.

Data cleaning

We included the medical history and medication use covariates for any time before the index date. For lab tests, we used the measurement closest to the index date in the prior 2 years. We treated All the lab tests (HbA1c, HDL, TOTCHL, BMI, serum creatinine) as continuous variables. We excluded extreme values that were likely to be erroneous values (e.g. HDL > 100 mg/dL, BMI< 8 kg/m, BMI> 100 kg/m, HbA1c>20% (195 mmol/mol), TOTCHL > 1000 mg/dL) from the study. A patient may not have any laboratory test during our 2-year searching time window. In addition, some patients also have missing values in race and insurance status in EHR. To deal with the missing data, we used multiple imputation by chained equation ¹³. In our analysis, we used predictive mean matching as the imputation method for continuous variables and polytomous regression imputation for unordered categorical data. We created 10 multiple imputed datasets. For each imputation, we set the number of iterations as 20. We fit cox regression model on each imputed dataset. We then used Rubin’s rule to pool coefficient estimates from 10 cox regression models ¹⁴.

Statistical analysis

We generated a simple statistical summary and stratified patients by second-line ADM medication. For categorical variables, we conducted Chi-square test for differences among second-line ADM classes. For continuous variables, we performed t-test to test the differences among second-line ADM classes at baseline. To evaluate the association of ADM drug classes and three different CKD-related outcomes, we developed a series of cox proportional hazard regression models. In our baseline model, we included only ADM class variables. In our basic demographic model, we included both ADM class and basic demographic information. In our demographic-medical history model, we included ADM class variables, basic demographics, and medical history. In the full model, we included all the mentioned variables in the Covariates section. Cox regressions were conducted using ‘survival’ package in R, version 3.6.0. Multiple data imputation was performed using ‘MICE’ package in R, version 3.6.0. Estimate pooling from 10 imputed datasets were performed using ‘Hmisc’ in R, version 3.6.0. Descriptive data statistics were generated using ‘TableOne’ module in python, version 3.7.3.

Sensitivity analysis

We conducted a sensitivity analysis to evaluate the robustness of our association findings. In the sensitivity analysis, we excluded patients with missing race, insurance status, HbA1c, total cholesterol, BMI, SBP, eGFR, and serum creatinine instead of imputing the missing data. The same cox regression models were performed to assess the association of ADMs with the three above mentioned outcomes.

Risk for incident CKD and CKD hospitalization outcome in primary analysis

Table 2 showed the number of events and hazard ratio of incident CKD outcome for each second-line ADM class using SU as reference group in a fully adjusted model in the primary analysis (see Supplemental Table 3, 4 for hazard ratio in baseline model, basic demographics model, basic demographics/medical history model in primary analysis)⁷. For CKD incidence, both SGLT-2 inhibitor (HR, 0.43; 95% CI, [0.22;0.87]; P=0.02) and DPP-4 inhibitor (HR, 0.71; 95% CI, [0.53;0.96]; P=0.03) are significantly associated with lower CKD incidence event. GLP-1RA was associated with reduced risk for CKD incidence (HR, 0.52; 95% CI, [0.21;1.29]; P=0.16), but the P-value is not significant. For CKD hospitalization outcome, DPP-4 is significantly associated with lower risk for CKD hospitalization (HR, 0.60; 95% CI, [0.37;0.96]; P=0.03). Use of the other ADMs also did not show significant difference in CKD hospitalization outcome compared to use of SU.

Sensitivity analysis

Table 3 showed hazard ratios for ADMs relative to use of SU in the sensitivity analysis in fully adjusted model (see Supplemental Table 5, 6 for hazard ratio in baseline model, basic demographics model, basic demographics/medical history model in sensitivity analysis)⁷. Our sensitivity analyses included only those patients with complete data. For the incident CKD event, after excluding patients with missing race, insurance status, HbA1c, insurance status, HDL, total cholesterol, BMI, baseline serum creatinine, and baseline eGFR, 2364 patients remained. Our analysis showed that in the fully adjusted model, SGLT-2 inhibitor use (HR, 0.42; 95% CI, [0.19, 0.92], P=0.03) was significantly associated with lower incidence of CKD. DPP-4 inhibitor use (HR, 0.75; 95% CI, [0.52;1.08]; P=0.12) was associated with lower CKD incidence but the P value was not statistically significant. For the CKD hospitalization outcome, after excluding patients with missing values, there were 2499 patients left. we did not observe any significant association between a particular second line ADM class relative to SU.

Approaches in Diabetes

A Comparative Effectiveness Study (GRADE) (54) is the only clinical trial (finished or ongoing) that we are aware of that evaluates the comparative effectiveness of newer ADMs vs older ADMs on glucose control ¹⁶. Further, most clinical trials treated second line ADMs as a monotherapy ^3,17. Therefore, it is unclear how the effect changes when the ADMs are used as add-on therapy on top of metformin, which is the most common case in the primary care setting. Secondly, patients in clinical trials are better positioned to adhere to medication and therapy than in real clinical settings. Therefore, the results drawn from clinical trials may not apply to real clinical settings. Observational studies using EHR data like ours could potentially close these gaps by providing evidence from the real-world data of general populations.

In this study, we used EHR data to investigate the association of second-line ADMs with renal outcomes among 3403 type 2 diabetes patients who initiated treatment with metformin. We found that both the use of SGLT-2 inhibitor and DPP-4 inhibitor showed significant association with less incident CKD as compared to use of SU as a second-line medication. However, in the sensitivity analysis, only SGLT-2 inhibitor remained significantly associated. For the CKD hospitalization event, DPP-4 inhibitor appeared protective for CKD hospitalization in the primary analysis, but the association was not statistically significant in the sensitivity analysis. The above observations are largely in line with results from placebo-controlled cardiovascular outcomes trials, which showed empagliflozin, canagliflozin, dapagliflozin, and Linagliptin all had beneficial effects on indices of CKD ^18,19. For example, the CREDENCE (Canagliflozin and Renal Events in Diabetes with Established Nephropathy Clinical Evaluation) trial showed that type 2 diabetes patients with prior albuminuria assigned canagliflozin (a SGLT-2 inhibitor) had lower risk of composite renal events compared to placebo group ¹⁸. Cooper et al. found Linagliptin (a DPP-4 inhibitor) significantly reduced the risk of kidney disease events compared to placebo among type 2 diabetes patients ¹⁹. Groop et al ²⁰ found linagliptin on top of renin-angiotensin-aldosterone system significantly reduced albuminuria among patients with type 2 diabetes and renal disfunction. We did not find statistically significant difference between GLP-1RA and SU, Insulin and SU, TZD and SU in either primary analysis or sensitivity analysis regarding CKD incidence and CKD hospitalization event. For GLP-1RA, previous studies showed inconclusive evidence of its effect in reduction of renal disease ²¹. The LEADER ²¹ study showed GLP-1RA had benefit in reduction of new onset albuminuria while the LIRA-RENAL study showed no improvement in urine ACR, and eGFR change among renal impaired T2DM patients compared to the placebo group ²². For TZD vs SU, Insulin vs SU, we did not find evidence of one is superior to the other in renal disease outcome in the literature. Our study has several strengths. First, our clearly defined exposure time starting from first prescription of metformin, with additional prescription of a second line ADM within a year: this ensured comparison of patients who were in a clinically similar diabetes stage. Second, we only included patients who started ADMs after 2013-03-01 which is the time when the first SGLT-2 inhibitor was approved by FDA for use. This ensures that patients using different second-line ADMs have comparable follow-up times and avoids the immortal time bias ²³. Third, we used a rigorously defined patient cohort: we excluded any patients who had type 1 diabetes diagnosis or gestational diabetes. In addition, we only included patients who had 3 continuous years of ADM prescriptions since their first ADM exposure. This makes sure that each patient has sufficient and comparable data depth. Last while there could be confounders that were not included in the study, our list of covariates is relatively comprehensive based on a review of the literature. This list contains basic demographic information, laboratory tests, and a wide range of diagnoses that might not only be associated with renal disease, but also suggest disease severity.

Our study has several limitations. Firstly, for simplicity, we could not distinguish between a switch of medication from an add-on medication. For example, patients who have a DPP-4 inhibitor as the second medication in their sequence could either: 1) have switched entirely from metformin to a DPP-4 inhibitor, or 2) began using metformin and a DPP-4 inhibitor together. Secondly, while we controlled for as many variables/confounders as we could, it is possible that there remain other covariates we did not capture, which could affect prescription bias or are still associated with the renal outcomes. Therefore, conclusions drawn from our study should be interpreted with caution. Thirdly, as EHR data is primarily designed for administrative purpose, it is well known that EHR’s have missing data issues. Because this is a single site study, patients may have sought care elsewhere. Therefore, it is likely that there were data points that we were unable to capture in this single site study. Additionally, this was a single center study of an academic hospital in a major US city, limiting its generalizability to other populations. Finally, diabetes duration is an important factor that affects renal outcomes. In our study, we did not control diabetes duration as a covariate because it was unavailable in our dataset. That said, we collected medication sequences at first exposure to ADMs, suggesting a clinically recent disease onset time. This makes the impact of not being able to include diabetic duration minimum.