Kidney function decline and increased urine protein with GLP-1 drugs compared to insulin in type 2 diabetes patients

The MHS database includes around 180,000 patients with type 2 diabetes among its over 2 million registrees. We included adults with type 2 diabetes who initiated a GLP-1 RA (exenatide [introduced in Israel in 12.2007], liraglutide [2.2010], lixisenatide [1.2015], exenatide extended-release [11.2015], dulaglutide [4.2016], and semaglutide [8.2019]) or basal insulin between February 2010 to December 2019. The day of drug initiation was defined as the index date, and the year preceding the index date was defined as the baseline period. We selected basal insulin as a comparator to ensure comparability between the groups, because during these years in Israel both drugs were mainly used as injectable drugs for glycemic control in advanced stages of diabetes. Accordingly, we included patients treated with at least two other glucose-lowering agents (GLAs) at the baseline period, reflecting the common use at Israel at the time. Only those with at least one eGFR measurement at the baseline period were included. We excluded patients with type 1 diabetes, eGFR < 15 mL/min/1.73 m², an indication of kidney transplantation or dialysis treatment, or those treated with the comparator drug within the year prior index date. Patients with an indication of pregnancy within 9 months before the index date were also excluded. To reduce bias associated with physicians' preference to treat severely ill patients with familiar and less costly drugs, we excluded patients with a diagnosis of dementia; history of organ transplantation; in MHS’ cancer (within the past 5 years) or heart failure registers; or those hospitalized for ≥ 5 consecutive days within the past 180 days (Additional file 1: Figure S1).

In the protocol, we defined two follow-up periods. In the intention to treat (ITT) analysis, follow-up continued until the end of data availability, death, or October 2021. In the as-treated (AT) analysis, follow-up was censored also at exposure discontinuation (added by 180 days of grace period) or the initiation of the comparator. In addition, we performed a sensitivity analysis censoring the ITT follow-up for all patients after 4 years of follow-up (ITT-48mo). The rationale behind this analysis was to terminate follow-up when a large portion of the participants was still exposed to the study drugs. Four years cut-off was also selected to emulate the follow-up duration of the LEADER trial [19].

The study was approved by the institutional review board (IRB) at MHS. Due to the de-identified nature of the data, the IRB did not require obtaining informed consent from the participants.

Validated MHS registries were used to identify patients with type 2 diabetes, cardiovascular disease (including heart failure), hypertension, or cancer [20–22]. The relevant International Classification of Diseases-9 diagnosis codes, Anatomical Therapeutic Chemical medications codes, and MHS registries are presented in Additional file 1: Table S1. Blood and urine samples included in this study were collected in community settings and were measured in the MHS-certified central laboratory. eGFR was calculated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation [23]. Residential socioeconomic status (SES) was ranked on a 1 (lowest) to 10 (highest) scale. This score was derived by Points Location Intelligence Ltd, combining geographic and socioeconomic information for each neighborhood (e.g., expenditures related to retail chains, credit cards, and housing). This score is highly correlated with the SES measured by the Israeli Central Bureau of Statistics. This parameter was categorized into 4 groups (low [1–3], low-medium [4, 5], medium [6, 7] and high [8–10]) [24].

The main kidney outcome was a composite of confirmed  ≥ 40% eGFR reduction from baseline or new ESKD. Additional outcomes were confirmed or single-measurement eGFR reductions of  ≥ 30, ≥ 40, ≥ 50, or  ≥ 57% (corresponding to a doubling of serum creatinine) or new ESKD alone. We also assessed albuminuria outcomes: (1) a categorical increase of urine albumin-to-creatinine ratio (UACR) for the following categories:  < 30, 30- < 300 or  ≥ 300 mg/g; (2) or new-onset macroalbuminuria (UACR ≥ 300) among those with UACR ≤ 230 mg/g at baseline (resembling a ≥ 30% increase in UACR). The albuminuria-based outcomes were assessed either as a single- or as confirmed -measurement. In addition, we compared the eGFR slopes between the groups.

In addition to the entire study population, analyses were performed in subgroups defined by sex, age (< 60 or ≥ 60 years), presence of CVD, years in diabetes registry (≤ 10 or > 10 years), body mass index (< 30 or ≥ 30 kg/m²), HbA1c (< 8, or ≥ 8%), UACR (< 30, 30- < 300, or ≥ 300 mg/g), treatment with ACEi/ARBs, and treatment with SGLT2i. Patients were also divided into subgroups by their baseline eGFR (≥ 90 and < 90 mL/min/1.73 m²); this threshold was selected owing to the relatively preserved kidney function of the study population. As a sensitivity analysis, we also divided patients into three eGFR subgroups (≥ 90, 60- < 90, or < 60 mL/min/1.73 m²).

Participants were propensity-score (PS) matched in 1:1 ratio using greedy matching, as previously described [25]. The model included 88 baseline parameters, including demographic variables (including SES), medical history, concomitant medications, and laboratory values (see the complete list in the Additional file 1). Continuous variables were categorized, and missing values were defined as a distinct ‘missing’ category to allow all patients to be matched. The PS matching was carried out by layers of baseline eGFR (> 90, 60–90, 45–60, 30–45, and 15–30 mL/min/1.73 m²).

Baseline values were described using mean and standard deviation (continuous variables with approximately normal distribution), median and IQR (continuous variables with skewed distribution), or proportions (categorical variables). Standardized difference (STD) was used to assess differences between the GLP-1 RAs and basal insulin group, with values of  < 10% considered negligible.

Cumulative incidence functions were used to describe the incidence of the outcomes in each group. Cox proportional hazard regression models were applied to estimate hazard ratios, confidence intervals, and p-value. The models were adjusted for the competing risk of mortality using cause-specific hazard models for the cumulative incidence functions and by using sub-distribution hazard functions for the Cox model [26].

The eGFR change from baseline at different time points was estimated using mixed models for repeated measures. We defined time windows of 6 months during the first 3 years and each year thereafter. At each time window, we included for each patient the eGFR measurement closest to the end of the period. Differences between groups at different time points were estimated using mixed-effect models with repeated measures. In randomized controlled trials (RCTs), eGFR slopes are also estimated using mixed-effect models with repeated measures; however, this approach may not fit the irregular sampling in real-world settings, where the times of measurements vary between patients. Therefore, to assess eGFR slopes, we fitted a linear regression model per patient (with time from index date as the independent variable and eGFR values as the dependent variable), enabling us to use all available eGFR measurements. The between-group difference in the linear slope estimates were compared using a t-test. We included in this analysis only patients with  ≥ 2 eGFR measurements with  ≥ 180 days between the first and last evaluations.

This study did not include formal hypothesis testing, and the p-values are presented for descriptive purposes only. No correction for multiple testing was performed. We consider a p < 0.05 as statistically significant. Analyses were performed using SAS version 9.4.

The study was funded by Novo Nordisk. The funder was involved in the study design, data analysis, data interpretation, writing of the report, and the decision to submit the paper for publication. This report was written according to a predefined protocol, including main and additional outcomes.

Overall, 11,634 and 22,598 patients initiated GLP-1 RAs or basal insulin, respectively. After applying the inclusion and exclusion criteria and PS-matching, there were 3424 participants in each group (Additional file 1: Figure S1). The mean age at baseline was 59.4 years (SD 9.4), and 44.8% were female. Participants at baseline had a mean diabetes duration of 9.7 years (4.6), body mass index of 33.4 (5.4) kg/m², and HbA1c of 9.0% (1.4). Cardiovascular disease was prevalent in 21.0%, mean eGFR was 90.6 mL/min/1.73 m² (19.3), and median [IQR] UACR was 14.6 mg/g [0.0–54.7]. At baseline, 74.9% were treated with ACEi/ARBs and 13.9% were treated with SGLT2 inhibitors. Baseline characteristics were well balanced between the groups following PS-matching. (Table 1 and Additional file 1: Tables S2 and S3).

In the ITT follow-up, during a median of 81.1 months [IQR 50.9–110.0], the median number of eGFR measurements per patient was 14 and 13 [7–22 and 7–20] for GLP-1 RAs and basal insulin group, respectively (Additional file 1: Tables S4 and S5). The composite kidney outcome (≥ 40% eGFR loss or ESKD) occurred overall in 631, 199, and 194 patients in the ITT, ITT-48mo, and AT follow-ups, respectively (Fig. 1 and Additional file 1: Figure S2). The hazard ratios of this outcome with GLP-1 RAs compared to basal insulin, were 0.96 ([95% CI 0.82–1.11]; p = 0.566), 0.85 ([0.64–1.12]; p = 0.246), and 0.71 ([0.54–0.95]; p = 0.020) in the ITT, ITT-48mo, and AT follow-ups, respectively (Fig. 1). There was no evidence that the effect varies across most tested subgroups, including by baseline treatment with SGLT2i (Fig. 2). The treatment effect seemed to be more pronounced in patients with eGFR < 90 compared to eGFR ≥ 90 mL/min/1.73 m² (p-interaction = 0.025 [ITT] and 0.058 [AT]). This association was not significant when those with eGFR < 90 were further divided into eGFR < 60 or 60- < 90 mL/min/1.73 m², although the direction was similar (p-interaction = 0.083 [ITT] and 0.129 [AT]) (Additional file 1: Table S6).

The risk of confirmed ≥ 40% eGFR reduction was lower with GLP-1 RA compared with basal insulin only in the AT follow-up analysis (HR 0.73 [95% CI 0.54–0.97]), but not in the ITT (HR 0.96 [0.82–1.12]) or ITT-48mo (0.83 [0.63–1.10) analyses. Similar trends were observed for other categorical eGFR thresholds, whether defined as single- or confirmed-measurement (Fig. 3). The risk of ESKD was not different between the treatment groups (HR 0.92 [0.64–1.32] in the ITT analysis). All-cause mortality with GLP-1 RA versus basal insulin occurred at an incidence of 1.1 vs 1.5 and 0.5 vs 1.3 events per 100 patient’s years in the ITT and AT analyses, respectively. In the ITT analysis, GLP-1 RA’ initiation was associated with a lower risk for single- (HR 0.90 [0.83–0.97]) and confirmed—(0.89 [0.80–0.99]) categorical progression of UACR, compared with basal insulin (Fig. 3). The risk for single-measurement new onset macroalbuminuria was also lower with GLP-1 RA (0.87 [0.75–0.997]) but not the risk of confirmed new macroalbuminuria (0.95 [0.78–1.16]). Similar associations between initiation of GLP-1 RA versus basal insulin with albuminuria progression outcomes were observed for the ITT-48mo and AT analyses (Fig. 3 and Additional file 1: Figure S2).

Compared with basal insulin, initiators of GLP-1 RAs had mitigated eGFR loss in the AT analysis starting at 12 months and afterward at almost all time points (p ≤ 0.027). eGFR loss mitigation was observed in the ITT analysis only at time points 12, 24, and 30 months (p ≤ 0.036), but not in others (Fig. 4). Use of GLP-1 RAs was associated with a less steep eGFR slope compared with basal insulin in the AT analysis (mean annual between-group difference of 0.42 mL/min/1.73 m²/year [95% CI 0.11–0.73]; p = 0.008). No significant differences were observed in the whole study population in the ITT (0.08 mL/min/1.73 m²/year [−0.06 to 0.23]; p = 0.258) or ITT-48mo analyses (0.14 mL/min/1.73 m²/year [−0.03 to 0.30]; p = 0.103) (Figs. 4 and 5).

The between-group difference in the AT analysis was observed in patients with eGFR < 90 mL/min/1.73 m² but not those with eGFR ≥ 90 mL/min/1.73 m² (0.95 mL/min/1.73 m²/year [95% CI 0.38–1.52] and 0.09 mL/min/1.73 m²/year [-0.26–0.43], respectively). (Fig. 5 and Additional file 1: Table S6). In the AT analysis, the between group-differences in eGFR decline with GLP-1 RAs versus basal insulin was 0.53 [−0.23 to 1.28] and 0.41 mL/min/1.73 m²/year [0.07–0.74] in those with or without SGLT2i therapy at baseline (Fig. 5).

This is an observational study, and while we used different approaches to emulate an RCT setting, no causation can be concluded. Real-world analyses pose several challenges requiring careful expertise. In this study, we followed recent recommendations [37]. Namely, we used a predefined protocol with new-initiators design, defined specific outcomes, used an insulin comparator that is also indicated in advanced stages of type 2 diabetes, and applied PS matching to balance the cohorts. However, the baseline characteristics of the study cohorts before matching were unbalanced, and no statistical method is capable to completely emulate randomization, thus residual confounding cannot be excluded. Also, PS matching limits the generalizability of the results to populations that were dropped out in the process. We also predefined both ITT and AT analyses, each considers different aspects of real-world effectiveness of drugs. The ITT follow-up had a long median duration of 81.1 months. However, during most of the follow-up period, patients were not treated with the index drug alone—either stopping it or initiating the comparator drug, along with potential changes in other drugs during follow-up. Therefore, previous studies favored an AT analysis to assess treatment effects of eGFR slopes [38]. On the other hand, the AT analysis is prone to biases due to imbalanced censoring. For example, the event rate (per 100 patients-years) for mortality in the AT follow-up with basal insulin and GLP-1 RAs was 1.3 and 0.5, respectively. This difference is much more than shown in RCTs [12]. It may not be attributed directly to the treatment allocation, and is possibly explained by discontinuation of GLP-1 RA or initiation of insulin among patients who experience a pronounced deterioration in their clinical status. To address these challenges, we performed an additional analysis censoring the ITT follow-up at 4 years, emulating the LEADER study [13, 19]. This analysis still follows the ITT principle, but censoring is stopped when most participants are still on index treatment. The number of events in this analysis was relatively low owing to the low baseline risk of the population, reducing the power to detect treatment effects. However, the treatment effects were closer to those seen in the AT analysis. Finally, the dissimilarity between the baseline characteristics of the pre-matched cohorts, required extensive matching. Finaly, this study includes one healthcare organization, limiting the study's external validity to other populations or healthcare systems.

The study was approved by the institutional review board (IRB) at MHS. Due to the de-identified nature of the data, the IRB did not require obtaining informed consent from the participants.

Not applicable.

MS reports travel support from Novo Nordisk and AstraZeneca through Haddasah Medical Center. CMC, AF, DRS, and GC have no conflict of interest to declare. IY and AR receive hourly payment from AstraZeneca through Hadassah Medical Center and from Novo Nordisk. AC and JL are employees of NovoNordisk. TJA is an employee and a shareholder of NovoNordisk. AK has received research grants and speaking honoraria from AstraZeneca, Novo Nordisk, and Boehringer Ingelheim. Starting May 1st 2023, I am an employee of Regeneron Pharmacutical. OM reports Advisory Board: Novo Nordisk, Eli Lilly, Sanofi, Merck Sharp & Dohme, Boehringer Ingelheim, AstraZeneca, BOL Pharma. Research grant support through Hadassah Hebrew University Hospital: Novo Nordisk, AstraZeneca. Speaker's Bureau: AstraZeneca, Novo Nordisk, Eli Lilly, Sanofi, Merck Sharp & Dohme, Boehringer Ingelheim.

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request and pending study protocol and regulations.