Predictors of cardio-kidney complications and treatment failure in patients with chronic kidney disease and type 2 diabetes treated with SGLT2 inhibitors

This retrospective observational cohort study was conducted using Optum Clinformatics™ Data Mart (CDM) between April 1, 2012, to June 30, 2019, to identify predictors of cardio-kidney events and predictors of treatment failure among initiators of SGLT2i with DKD. The data source is an administrative health claims database with longitudinal data of patients enrolled in a commercial and Medicare Advantage health plan data in the US. The data contain demographic, medical encounters from inpatient and outpatient settings, pharmacy dispensing, and laboratory results for a subset of patients. The data contain approximately 63 million unique members and are considered representative of the commercially insured US population.

The study population consisted of patients aged 18 years or older with DKD who initiated SGLT2i therapy following a 365-day SGLT2i-naïve period and 365-day continuous enrollment period. The first pharmacy claim for SGLT2i between 1 April 2013 (following approval of SGLT2is in the US) and 30 June 2019 was considered the index date. Follow-up period began 1 day after the index date and ended on the earliest occurrence of the outcome, disenrollment, death, and the end of data on 30 June 2019 (or SGLT2i discontinuation for cardio-kidney outcomes only) (Fig.1). T2D diagnosis was defined according to at least one inpatient ICD-9 or ICD-10 diagnosis code in any position, at least two outpatient ICD-9 and ICD-10 diagnosis codes for T2D in any position 30 to 365 days apart, or at least one prescription claim for a second-line therapy agent (i.e., sulfonylureas, thiazolidinedione (TZD), dipeptidyl peptidase-4 inhibitor (DPP4i), glucagon-like peptide-1 receptor agonist (GLP1ra), insulin, meglitinide, or alpha glucosidase inhibitors) for T2D in the baseline period [16]. CKD diagnosis was defined according to two laboratory values for estimated glomerular rate (eGFR) and/or two laboratory values for urine albumin-to-creatinine ratio (UACR) 90 to 365 days apart within ranges indicating CKD in the baseline period. Patients with baseline claims for conditions other than T2D that may cause kidney disease, such as glomerulonephritis, membranous nephropathy, or malignant kidney tumor were excluded [17]. Full definitions of the patient cohorts, variables, outcomes, and subgroup selection criteria are listed in Additional File 1: Table S1.

Patient demographic characteristics were assessed on index date. Comorbidities, laboratory values (e.g., eGFR, UACR, and hemoglobin A1c [HbA1c] values closest to the index date), and medication use (cardiovascular and antiglycemic agents) were assessed during the 365-day baseline period.

Cardio-kidney events and their predictors were evaluated separately in the study population during the follow-up period. These included CV hospitalization (defined by unstable or stable angina, arrhythmia, coronary intervention or revascularization, transient ischemic attack, heart failure, and peripheral arterial disease), kidney hospitalization defined by CKD diagnosis, and AKI hospitalization. Treatment failure was also evaluated during follow-up, defined as the occurrence of discontinuation of SGLT2i or switch from SGLT2i to another T2D medication class; or addition of another T2D medication class not used in baseline; or initiation of insulin (see Additional File 1: Table S1 for a complete listing of all treatments considered). We applied a conservative grace period of 90 days between two consecutive refills before counting SGLT2i treatment as discontinued [18].

The number and percentage of patients meeting each characteristic were reported. Means (with standard deviations) and medians (with interquartile ranges reported as 25th and 75th percentiles; IQR) were reported for continuous characteristics. Baseline medication use of patients was described according to proportion of patients with at least one dispensation of a drug of interest. Event rates of each outcome per 1000 person-years (PY) were reported, and 95% confidence intervals (CIs) were estimated using Poisson regression with a Normal approximation. For the rate calculations, the numerator consisted of the total number of first occurrences in the population and the denominator included the total person-time (in years) from 1 day after the index date to the date of the outcome or the end of follow-up.

Cox proportional hazard regression was used to identify baseline patient characteristics associated with an increased risk for the outcomes of interest. All results were reported as hazard ratios (HRs) with 95% CIs. Two-sided p values were reported with a pre-specified alpha (ɑ) level of 0.05. Backward selection methods were used to identify a subset of covariates associated with each outcome [19]. Multivariate models including all input variables were initially fitted and then reduced in a stepwise manner to identify the best fit according to the Akaike information criterion (AIC). Prior to the initial model fitting, variables with high collinearity and correlation were removed from the total list of input variables; tests of relative importance were applied when necessary.

All baseline analyses were stratified by insurance type due to the potential differences in T2D management based on commercial versus Medicare insurance plans. Several supplemental analyses were also performed, including the analysis of cardio-kidney outcomes by insurance type subgroups (i.e., commercial or Medicare Advantage; Additional File 2: Figures S1 – S4). Additionally, all baseline characteristics and the CV outcome were analyzed in mutually exclusive cardiovascular risk subgroups based on the level of CVD risk in the baseline period as follows: (1) patients with no inpatient or outpatient CVD diagnosis in baseline (low risk), (2) patients with an outpatient but no inpatient CVD diagnosis in baseline (moderate risk), and (3) patients with inpatient CVD diagnosis in baseline (high risk).

An additional sensitivity analysis explored the effects of several time-dependent SGLT2i-related adverse events (AE) after treatment initiation on treatment failure. These included acidosis, hospitalization for acute kidney injury (AKI), dehydration, diabetic ketoacidosis, genital tract infection, hypotension, lower extremity amputation, urinary tract infections (UTI), and volume depletion. Each of these events was evaluated in 30-day risk-interval windows after the index date until the time of censoring. Baseline factors from the final multivariate treatment failure model along with the time-dependent events of interest were explored univariately and final multivariate models were explored.

All analyses were conducted using the Aetion Evidence Platform® (2020), software for real-world data analysis, which is validated for a range of studies [20].

Table 1 summarizes the patient baseline characteristics. SGLT2i initiators were on average 65.5 (SD 10.6) years of age with an even distribution of use in males and females. CV risk subgroups had similar average age while Medicare enrollees were on average 70.6 (SD7.8) and commercially insured patients were 56.4 (8.7) years of age. Approximately half of all the patients were White; 59.1% of patients resided in the South.

The most prevalent comorbidities (Table 2) in the study population were hypertension (92.3%), hyperlipidemia (87.9%), pain disorders (70.2%), microvascular complications (54.3%), and obesity (43.2%). The average HbA1c level in the study population was 9.2% (SD 3.9) at baseline and ranged from 8.9% (SD 3.6) to 9.4% (SD 4.3) among subgroups. Approximately a third of patients (31.7%) had mild-to-moderately decreased eGFR (stage G3a), the highest percentage being in the high CV risk subgroup (36.1%) and in Medicare patients (35.0%). 43.4% had moderately increased albuminuria (UACR stage A2). Less than 2% of patients had decreased eGFR or kidney failure at baseline (eGFR stage of G4 or G5, respectively). The mean annual rate of eGFR decline during follow-up was 4.4 (SD 6.8) ml/min/1.73 m² per year for patients with treatment failure and 8.7 (SD 13.6) ml/min/1.73 m² per year for patients with cardio-kidney events increasing to up to 12.4 (SD 13.1) among patients with high CV risk.

Approximately 71.1% of patients in the total study population filled a metformin prescription in the year before SGLT2i initiation. The majority of the population filled at least one prescription for a second-line antidiabetic therapy in baseline (86.5%), while nearly half of the patients used at least two second-line antidiabetic therapies at baseline (46.5%). Approximately 82.0% of the total study population filled a statin prescription in the year before SGLT2i initiation. The most commonly filled antihypertensive medications were ACEi/ARBs (82.4%), followed by diuretics (54.2%), beta blockers (48%), and CCBs (39.9%). Percentages were approximately the same among subgroups (Table 3).

The rate of CV hospitalization was 26.0 (95% CI 21.6, 30.4) events per 1000 PY in the overall population assessed over a mean follow-up time of 292.3 (SD 328.5) days; the highest being in the high CV risk subgroup 79.5 (95% CI 40.5, 118.4) and in Medicare patients [32.9 (95% CI 26.4, 39.4), Table 4]. Baseline characteristics associated with higher risk of CV hospitalization included age [multivariate hazard ratio (HR) of 1.03 (95% CI 1.01, 1.05) per year of increased age], atrial fibrillation [HR 2.97 (CI 1.51, 5.84)], peripheral vascular disease (PVD) [HR 1.67 (CI 1.15, 2.42)], and cancer excluding nonmelanoma skin cancer (NMSC) [HR 1.83 (CI 1.14, 2.94)]. The risk of CV hospitalization also increased significantly with the baseline use of alpha blockers [HR 2.07 (CI 1.10, 3.90)], antiplatelets [HR 1.57 (CI 1.04, 2.37)], and nitrates [HR 1.74 (CI 1.09, 2.76), Fig. 3].

The rate of kidney hospitalization was 12.0 (95% CI 9.0, 15.0) events per 1000 PY, evaluated over a mean follow-up time of 294.6 (SD 329.8) days. Baseline characteristics with higher risk of kidney hospitalization included Black race [HR 2.99 (CI 1.61, 5.54)], the provider being a general practitioner or internist [HR 2.79 (CI 1.66, 4.67)], and the number of outpatient encounters [HR 1.02 (CI 1.01, 1.03) per additional encounter]. The risk also increased significantly with baseline evidence of heart failure, hyperkalemia, respiratory failure, and depression, as well as the baseline use of loop diuretics (Fig. 4).

The rate of AKI hospitalization was 22.8 (95% CI 18.7, 26.9) events per 1000 PY, captured over a mean follow-up time of 293.5 (SD 330.0) days. AKI hospitalization risk significantly increased with being a Medicare beneficiary [HR 1.66 (CI 1.04, 2.63)], having baseline evidence of atrial fibrillation [HR 2.74 (CI 1.20, 6.29)], PVD [HR 1.63 (CI 1.08, 2.45)], or cancer excluding NMSC [HR 1.77 (CI 1.06, 2.97)], and the use of loop acting diuretics [HR 1.75 (CI 1.12, 2.74), Fig. 5].

Among the total population, 55.0% of patients overall, and 44.8% and 60% in the commercially insured and Medicare subgroups, respectively, discontinued treatment (90-day grace period) during the follow-up period. The median time to discontinuation of SGLT2is was 204 [IQR 144, 364] days.

The rate of treatment failure was 510.5 (95% CI 492.9, 528.1) events per 1000 PY over an average of 361.5 (SD 366.0) days of follow-up. The reason for treatment failure in more than half (63.6%) of all treatment failure events in follow-up (N = 4802) was attributed to SGLT2i discontinuation or a switch from SGLT2is to another antidiabetic drug class. One-third (31.1%) of treatment failure events were due to the addition of another antidiabetic drug class to therapy, which was not previously used in baseline (Table 4).

The risk of treatment failure increased significantly with the baseline use of DPP4is [HR 1.33 (CI 1.24, 1.43)] and non-significantly with outpatient facility providers [HR 1.31 (CI 0.95, 1.81)]. In contrast, the risk of treatment failure was significantly decreased with baseline use of statins [HR 0.9 (CI 0.82, 0.98)], GLP1ra [HR 0.8 (CI 0.72, 0.89)], basal insulin [HR 0.79 (CI 0.73, 0.86)], or concurrent use of both [HR 0.77 (CI 0.64, 0.94)] (Fig. 6).

Results from the sensitivity analysis of key time-dependent SGLT2i adverse events (Additional file 1: Table S2) showed that, overall, a patient experiencing diabetic ketoacidosis during follow-up was significantly associated with risk of treatment failure [(HR 2.43 (95% CI 1.17, 5.07)] and commercially insured patients had similar results to the overall population. In older Medicare patients, volume depletion was associated with higher risk of treatment failure (Additional file 1: Table S2).

Our study had several limitations. First, administrative claims data carries a potential for misclassification of patients’ diagnoses, since the presence of a diagnosis code may not indicate the presence of a disease, but a rule out code in some instances. To address this limitation in identifying DKD, validated algorithms for identification of T2D patients from inpatient, outpatient, and prescription claims, as well as two laboratory values indicating kidney disease to identify CKD patients, were used. Nonetheless, DKD can only be accurately defined by histopathologic screening of the kidney [27]. Second, not all important comorbidities or risk factors can be accurately ascertained within claims data, such as a patient’s smoking status, body mass index, and other CKD symptoms (e.g., poor appetite, weakness, or itching). As a result, we may not be capturing all relevant factors associated with the outcomes of interest. Third, laboratory results in the database were incomplete since laboratory orders were not always represented in claims. In addition, laboratory results that were done in a hospital setting or directly in the physician office were not represented in the database. While there is no reason to suspect a selection bias as to which patients receive laboratory results, some laboratory results may be missing and so some patients may have been misclassified as incident CKD cases due to unobservable prior lab results that indicate kidney disease. Fourth, dispensing data only covers claims filed for insurance coverage. Generic varieties for drugs (e.g., metformin) are often available and may be affordable enough so that patients never file claims for them. As a result, these dispensings were not captured in this study. Pauly et al. estimate that about 30% of metformin and ACEi fills are done through low-cost generic programs [28]. Our results, therefore, might have slightly underestimated the real use of such drugs during the baseline period, the switching or addition of generic drugs during follow-up and potentially limited the possibility to completely capture reasons for discontinuation. Additionally, therapeutic inertia, although out of the scope of our study population, is an additional reason that could deprive patients of beneficial medications [29]. Fifth, the cardio-kidney outcomes rates estimated in our study might be underestimated, since the follow-up was allowed to be censored on SGLT2i discontinuation; therefore, the risk of occurrence of these outcomes is not included in the estimation (as it would be in a intention-to-treat approach). Lastly, in addition to the studied factors, there might be other characteristics that could predict the studied outcomes. However, based on our study design, we can only draw conclusions of association, and not causation, between the studied factors and the outcomes.