Glucose-Lowering Agents and the Risk of Hypoglycemia: a Real-world Study

Abstract

Background

Sodium-glucose cotransporter-2 inhibitors (SGLT2i) and glucagon-like peptide-1 receptor agonists (GLP1RA) are increasingly recommended in type 2 diabetes. Hypoglycemia is a serious adverse effect of glucose-lowering agents. Real-world comparison of hypoglycemic risks among SGLT2i, GLP1RA, dipeptidyl peptidase-4 inhibitor (DPP4i), and sulfonylureas is limited.

Objective

Quantify the risk of hypoglycemia associated with SGLT2i, GLP1RA, DPP4i, and sulfonylureas (the primary reference group).

Design

Retrospective cohort study conducted using electronic health records from Geisinger Health, Pennsylvania (2015–2019).

Participants

A total of 10,713 patients with type 2 diabetes who newly received SGLT2i (n=1487), GLP1RA (n=1241), DPP4i (n=2938), or sulfonylureas (n=5047). Propensity score–based inverse probability of treatment weighting was used to balance patient characteristics across four treatment groups simultaneously.

Main Measures

Hypoglycemia was defined as capillary blood glucose <70 mg/dL; severe hypoglycemia was defined as capillary blood glucose <54 mg/dL. A weighted Cox proportional hazards regression model was used to estimate the risk of outcomes for pairwise comparisons of SGTL2i, GLP1RA, DPP4i, and sulfonylureas.

Key Results

Median follow-up was 21.3 months. Compared with sulfonylureas, the risk of hypoglycemia was lower with SGLT2i (hazard ratio 0.60 [95% confidence interval 0.48–0.75]), GLP1RA (0.49 [0.34–0.69]), and DPP4i (0.60 [0.48–0.78]). The risk of severe hypoglycemia was also lower with SGLT2i (0.43 [0.35–0.74]), GLP1RA (0.50 [0.28–0.87]), and DPP4i (0.64 [0.46–0.90]) compared to sulfonylureas. The risks of hypoglycemia and severe hypoglycemia were similar across the SGLT2i, GLP1RA, and DPP4i groups (SGLT2i vs. DPP4i: 0.95 [0.67–1.34]; GLP1RA vs. DPP4i: 0.81 [0.55–1.19]; SGLT2i vs. GLP1RA: 1.17 [0.76–1.82] for hypoglycemia).

Conclusion

SGLT2i and GLP1RA confer a lower risk of hypoglycemia compared with sulfonylureas and similar risk compared with DPP4i. Given the known cardiovascular benefits associated with SGLT2i and GL1PRA, our results suggesting the safety of SGLT2i and GL1PRA further support their use.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11606-022-07726-8.

INTRODUCTION

Diabetes is a complex chronic disease that often necessitates pharmacotherapy for optimizing glycemic control and risk reduction of macro- and microvascular complications. Sulfonylureas remain the most commonly used noninsulin glucose-lowering agents after metformin in adults with type 2 diabetes.¹ However, newer classes of agents such as sodium-glucose cotransporter-2 inhibitors (SGLT2i) and glucagon-like peptide-1 receptor agonists (GLP1RA) are now recommended, with evidence from randomized clinical trials demonstrating reduction in adverse cardiovascular events among adults with diabetes.^2–6 Dipeptidyl peptidase-4 inhibitors (DPP4i) are commonly used, with real-world evidence for cardiovascular safety despite initial concern for increased risk of hospitalization for heart failure.^7–9

With all diabetes pharmacotherapy, hypoglycemia is a serious treatment-limiting adverse reaction. Hypoglycemia is associated with increased risks of adverse cardiovascular outcomes, mortality, and healthcare costs.^10–13 Although modest size clinical trials showed similar rates of hypoglycemia with newer classes of glucose-lowering agents — SGLT2i, GLP1RA, and DPP4i^14–18 — there has been limited evaluation in real-world practice. Estimation of the hypoglycemic risks in routine care settings may inform clinical practice and support decision making, especially when multiple classes of medications are available for similar indications. Therefore, we performed a population-based cohort study to compare the risk of hypoglycemia associated with SGLT2i, GLP1RA, DPP4i, and sulfonylureas.

METHODS

Design and Setting

This new-user, active-comparator cohort study was conducted using electronic health record data from Geisinger Health between January 1, 2015, and January 30, 2019.¹⁹ Geisinger Health is a tertiary health system that serves a stable patient population in 45 counties of central and northeastern Pennsylvania, USA, with only 1% outmigration rate.²⁰ The electronic health records contain information on patient demographic characteristics, prescriptions, diagnoses, laboratory findings, and healthcare encounters. The study was approved by the institutional review boards of Geisinger Medical Center and Johns Hopkins University.

Study Population

Eligible records from adults aged ≥18 years with type 2 diabetes who received a new prescription for SGLT2i, GLP1RA, DPP4i, or sulfonylureas were included. The date of the first study glucose-lowering agent prescription served as the index date. We required 1-year wash-out period (i.e., no records of study medication) prior to the index date to ensure a new user design. Patients with prescriptions for >1 study medication class on the index date were excluded. Diabetes was defined using a validated algorithm with a positive predictive value of 87.0% (Appendix Table 1).²¹ Those with gestational diabetes, polycystic ovary syndrome, or type 1 diabetes were excluded.²² Patients were required to be free of end-stage kidney disease and have a prior healthcare encounter (outpatient visit, hospitalization, prescription, or laboratory test) with Geisinger Health ≥1 year preceding the cohort entry (Appendix Figure 1).

Exposures

As metformin is the preferred first-line pharmacotherapy in type 2 diabetes, the exposures of interest were other glucose-lowering agents (SGLT2i, GLP1RA, DPP4i, and sulfonylureas) which may be combined with metformin.² Using outpatient prescription information, patients were categorized by their first exposure to a study medication class per an intention-to-treat approach.^{23, 24} We chose sulfonylureas as the primary reference group because it is the most commonly used oral glucose-lowering agent following metformin for adults with diabetes.¹ We chose DPP4i as the secondary reference group because SGLT2i and GLP1RA have proven cardiovascular benefits while DPP4i do not.^2–6 We also compared between SGLT2i and GLP1RA.

Outcomes

Hypoglycemia and severe hypoglycemia were defined as capillary blood glucose <70 mg/dL (3.9 mmol/L) and <54 mg/dL (3.0 mmol/L), respectively.^{17, 25} In Geisinger Health, a point-of-care capillary blood glucose test is performed when patients exhibit signs of hypoglycemia during healthcare encounters. Patients were followed from the index date until the date of the defined hypoglycemic event, last encounter with Geisinger Health, or end of study period (January 30, 2019), whichever came first.

Covariates

Demographic characteristics (age, sex, and race) and health insurance status were extracted from the electronic health record. Race was self-reported or identified by healthcare providers. We queried information on medical comorbidities that signify diabetes complications (cerebrovascular disease, coronary artery disease, heart failure), as well as those that predispose to hypoglycemia (chronic liver disease, alcohol use disorder, and acute kidney injury). Comorbidities were captured by International Classification of Diseases, Ninth and Tenth Revisions (ICD-9 and ICD-10) codes (Appendix Table 2). We calculated the Charlson comorbidity index to summarize overall comorbidity burden.²⁶

Table 2.

Glucose-lowering agents and the risk of hypoglycemia using capillary blood glucose measurements*

Glucose-lowering agent class (unweighted N)	Unweighted no. of events	Incidence rate, per 100 person-years (95% CI) †	Hazard ratio (95% CI) †
SGLT2i (N = 1487)	49	2.1 (1.6–2.7)	0.60 (0.48–0.75)
GLP1RA (N = 1241)	52	1.7 (1.2–2.1)	0.49 (0.34–0.69)
DPP4i (N = 2938)	117	2.0 (1.7–2.5)	0.60 (0.48–0.78)
Sulfonylureas (N = 5047)	331	3.3 (2.8–3.5)	1 (reference)

Abbreviations: CI, confidence interval; SGLT2i, sodium-glucose cotransporter-2 inhibitor; GLP1RA, glucagon-like peptide-1 receptor agonist; DPP4i, dipeptidyl peptidase-4 inhibitor

*Hypoglycemia was defined as capillary blood glucose <70 mg/dL.

†Estimated using inverse probability of treatment weighted Cox proportional hazards regression.

We included the most recent laboratory and vital measurements within the year prior to the index date. Given that chronic kidney disease predisposes patients to hypoglycemia,²⁷ we included estimated glomerular filtration rate (eGFR; derived from serum creatinine using the Chronic Kidney Disease Epidemiology Collaboration equation²⁸) and urinary albumin-to-creatinine ratio (ACR). When ACR measurement was not available, we derived ACR measurement from urine protein-creatinine ratio or urine protein dipstick.²⁹ Hemoglobin A1c (HbA1c), body mass index (BMI), and systolic blood pressure measurements were also included.

We identified coprescription information on other glucose-lowering agents (metformin, insulin, meglitinide, and thiazolidinedione), medications for diabetes complications (statin and angiotensin-converting-enzyme inhibitor/angiotensin II receptor blocker), and medications that may predispose to hypoglycemia (beta-blockers, nonsteroidal anti-inflammatory drugs, and fluoroquinolones).^{30, 31} We further accounted for other antihypertensive medications. We obtained information on the number of outpatient visits and hospitalizations, and capillary blood glucose measurements in the year preceding the index date.

Multiple Imputation and Inverse Probability of Treatment Weighting

At least one covariate was missing in 41.8% of the patients, with urinary ACR (36.4%) and BMI (28.8%) being the most commonly missed. We used multiple imputation by chained equation to impute 40 datasets.^{31, 32} The aforementioned patient characteristics, exposure, and outcomes were included in the imputation model.

We used propensity score–based inverse probability of treatment weighting (IPTW) to control for confounding. Within each imputed dataset, we used multinomial logistic regression to derive propensity scores for different study glucose-lowering agents simultaneously.³³ The aforementioned patient characteristics were included in the multinomial logistic model. Subsequently, stabilized IPTW based on the propensity scores were used to balance the patient characteristics among the four exposure groups.³⁴ Estimated weights were trimmed at 1st and 99th percentiles to mitigate influence of outliers. We used pairwise standardized mean differences (SMD) before and after IPTW to assess balance in covariates.^{35, 36} SMD <0.10 (10%) indicated good balance of covariates in the pairwise comparisons.

Statistical Analyses

We described baseline characteristics using means ± standard deviations (SD) or medians (interquartile intervals, IQI) for continuous variables and percentages for categorical variables from one randomly selected imputed dataset. We used the weighted Kaplan-Meier curves and log rank test to depict outcome-free survival by the treatment groups. We estimated IPTW hazard ratios (HRs) with 95% confidence intervals (CIs) for outcomes using Cox proportional hazards regression. HRs were estimated in each imputed dataset and combined using Rubin’s rules. We tested the proportional hazard assumption by checking Schoenfeld’s partial residuals. Covariates that remained unbalanced after applying IPTW were included as additional adjustments in the Cox model. We performed prespecified subgroup analyses to examine whether the risk of hypoglycemia differed by age (<65 vs. ≥65 years), eGFR (<60 vs. ≥60 mL/min/1.73 m²), HbA1c (<7 vs. ≥7%), or concomitant insulin use (yes vs. no). All statistical analyses were done using SAS 9.4 (Cary, NC) and R (www.R-project.org).³⁷ A two-sided p-value <0.05 was considered statistically significant.

Sensitivity Analyses

We performed a series of sensitivity analyses. First, we did an “as-treated” analysis and censored at study drug discontinuation (defined as a gap in prescription of initial drug longer than 60 days^{20, 38}). Second, to explore whether the results of the intention-to-treat analysis were confounded by differential insulin use during the follow-up period, we compared insulin initiation across the four exposure groups among patients who did not use insulin at baseline. Third, we investigated the risk of hypoglycemia-related hospitalizations captured by validated ICD-9 and ICD-10 codes.³⁹ Fourth, we repeated the analyses after excluding users of first-generation sulfonylureas, given the higher risk of hypoglycemia with the first-generation vs. second-generation sulfonylureas.⁴⁰ Fifth, we performed a negative control outcome analysis to address potential unmeasured/residual confounding.⁴¹ Bleeding-related hospitalization was selected as the negative control outcome, an outcome thought to be unaffected by different glucose-lowering agents.

Because greater glycemic variability is associated with increased risk of hypoglycemia,⁴² we assessed the association between the exposure groups and variability in HbA1c within a subcohort of patients with ≥3 HbA1c measurements during the follow-up period. We used the standard deviation of HbA1c levels as the outcome because this metric has been associated with increased risks of hypoglycemia-related hospitalizations, mortality, and cardiovascular diseases.^42–44

RESULTS

Baseline Characteristics

We identified 10,713 adults with type 2 diabetes who newly received SGLT2i (n=1487), GLP1RA (n=1241), DPP4i (n=2938), or sulfonylureas (n=5047) (Appendix Table 3). The study population had a mean age (SD) of 58.4 (14.2) years, 49.4% were female, and 1.5% had a history of hypoglycemia in the year prior to cohort entry (Appendix Table 4). Concomitant metformin and insulin use were 58% and 12%. After applying IPTW, baseline characteristics were well balanced across the exposure groups (Table 1), except the mean age was lower in the GLP1RA group (56.3±12.8 years) compared to the sulfonylurea group (58.4±14.6 years; SMD=0.14; Appendix Table 5).

Table 1.

Weighted baseline characteristics of the study population by glucose-lowering agent class

Variables*	SGLT2i	GLP1RA	DPP4i	Sulfonylurea
Unweighted N	1487	1241	2938	5047
Age, years	57.5±11.8	56.3±12.8	58.3±14.3	58.4±14.6
Female, %	49.1	52.7	49.3	49.5
White, %	92.3	91.1	91.3	91.7
Year of prescription, %
2015	22.5	24.8	24.3	24.9
2016	25.4	26.4	25.3	25.8
2017	27.0	24.5	25.8	25.9
2018	25.1	24.3	24.5	23.3
Uninsured, %	1.7	1.5	1.7	1.7
Comorbidities, %
Cerebrovascular disease	6.4	6.3	6.9	6.7
Coronary artery disease	18.0	19.4	20.7	20.3
Heart failure	6.2	6.9	8.0	7.8
Chronic liver disease	14.5	14.0	13.7	14.1
Alcohol use disorder	3.4	2.9	3.9	3.9
Hypoglycemia in the past year	1.1	1.6	1.6	1.4
Acute kidney injury in the past year	4.6	5.3	5.6	5.2
Charlson comorbidity index	4.0±2.4	4.1±2.8	4.3±2.8	4.3±2.8
eGFR, mL/min/1.73 m2	86.6±22.2	87.0±24.2	84.9±24.4	85.2±24.8
eGFR categories, %
≥90 mL/min/1.73 m2	52.0	53.4	48.3	48.8
60–89 mL/min/1.73 m2	34.0	31.4	34.8	34.4
30–59 mL/min/1.73 m2	12.8	13.5	15.2	15.3
<30 mL/min/1.73 m2	1.2	1.7	1.7	1.5
Hemoglobin A1c, %	8.4±1.7	8.4±1.9	8.4±1.8	8.3±1.7
Body mass index, kg/m2	36.2±8.0	37.0±8.3	36.2±8.9	36.0±8.7
Urinary albumin-to-creatinine ratio, mg/g	11.0 (4.5–34.0)	10.5 (4.0–36.0)	11.0 (4.0–40.0)	11.0 (4.0–40.0)
Systolic blood pressure, mmHg	128.6±14.8	129.0±14.9	129.1±15.9	129.1±16.3
Coprescriptions, %
Metformin	55.6	57.2	58.0	57.3
Insulin	13.4	12.7	12.3	11.7
Meglitinide	1.5	1.7	1.4	1.4
Thiazolidinedione	1.9	1.9	1.7	1.8
Statin	47.2	46.5	48.0	48.3
ACEi/ARB	46.8	44.3	47.8	47.3
Beta-blocker	26.7	26.0	28.6	28.5
Other antihypertensive agents	37.2	37.3	38.4	37.2
Nonsteroidal anti-inflammatory drug	21.2	20.9	19.8	20.0
Fluoroquinolone	2.0	2.5	2.3	2.3
No. of outpatient visits in the past year	3 (1–7)	4 (2–7)	4 (2–7)	4 (2–8)
No. of hospitalizations in the past year	0 (0–1)	0 (0–1)	0 (0–1)	0 (0–1)
No. of capillary blood glucose measurements in the past year	0 (0–1)	0 (0–1)	0 (0–1)	0 (0–1)

Abbreviations: SGLT2i, sodium-glucose cotransporter-2 inhibitor; GLP1RA, glucagon-like peptide-1 receptor agonist; DPP4i, dipeptidyl peptidase-4 inhibitor; eGFR, estimated glomerular filtrate rate; ACEi, angiotensin-converting-enzyme inhibitor; ARB, angiotensin II receptor blocker

*All baseline covariates listed in this table were included in the propensity score model. Continuous variables were presented as weighted means ± standard deviations or medians (interquartile intervals). Categorical variables were presented as weighted proportions.

Risk of Hypoglycemia

The median (IQI) follow-up was 21.3 (10.1–33.6) months. The weighted incidence rate of hypoglycemia was 2.1 events per 100 person-years with SGLT2i, 1.7 events per 100 person-years with GLP1RA, 2.1 events per 100 person-years with DPP4i, and 3.2 events per 100 person-years with sulfonylureas. The weighted Kaplan-Meier curves for the risk of hypoglycemia by study medication class are illustrated in Figure 1. Compared to sulfonylureas, the risk of hypoglycemia was lower with SGLT2i (HR 0.60 [95% CI 0.48–0.75]), GLP1RA (HR 0.49 [95% CI 0.34–0.69]), and DPP4i (HR 0.60 [95% CI 0.48–0.78]; Table 2). The risk of hypoglycemia was similar across the SGLT2i, GLP1RA, and DPP4i groups (HR 0.95 [95% CI 0.67–1.34] for SGLT2i vs. DPP4i; HR 0.81 [95% CI 0.55–1.19] for GLP1RA vs. DPP4i; HR 1.17 [95% CI 0.76–1.82] for SGLT2i vs. GLP1RA). The findings suggest that 20 patients would need to switch from sulfonylureas to SGLT2i or GLP1RA to prevent one hypoglycemic event in the next 5 years.

Figure 1.

Weighted Kaplan-Meier curves by glucose-lowering agent class

Risk of Severe Hypoglycemia

The weighted incidence rate of severe hypoglycemia was 0.6 events per 100 person-years with SGLT2i, 0.7 events per 100 person-years with GLP1RA, 0.9 events per 100 person-years with DPP4i, and 1.3 events per 100 person-years with sulfonylureas. Compared to sulfonylureas, the risk of severe hypoglycemia was lower with SGLT2i (HR 0.43 [95% CI 0.35–0.74]), GLP1RA (HR 0.50 [95% CI 0.28–0.87]), and DPP4i (HR 0.64 [95% CI 0.46–0.90]; Table 3). The risk of severe hypoglycemia was similar across the SGLT2i, GLP1RA, and DPP4i groups (HR 0.66 [95% CI 0.36–1.20] for SGLT2i vs. DPP4i; HR 0.77 [95% CI 0.42–1.41] for GLP1RA vs. DPP4i; HR 0.86 [95% CI 0.41–1.81] for SGLT2i vs. GLP1RA).

Table 3.

Glucose-lowering agents and the risk of severe hypoglycemia using capillary blood glucose measurements*

Glucose-lowering agent class (unweighted N)	Unweighted no. of events	Incidence rate, per 100 person-years (95% CI) †	Hazard ratio (95% CI) †
SGLT2i (N = 1487)	14	0.6 (0.5–0.8)	0.43 (0.35–0.74)
GLP1RA (N = 1241)	20	0.7 (0.5–0.8)	0.50 (0.28–0.87)
DPP4i (N = 2938)	50	0.9 (0.8–1.1)	0.64 (0.46–0.90)
Sulfonylureas (N = 5047)	137	1.3 (1.2–1.5)	1 (reference)

Abbreviations: CI, confidence interval; SGLT2i, sodium-glucose cotransporter-2 inhibitor; GLP1RA, glucagon-like peptide-1 receptor agonist; DPP4i, dipeptidyl peptidase-4 inhibitor

*Severe hypoglycemia was defined as capillary blood glucose <54 mg/dL.

†Estimated using inverse probability of treatment weighted Cox proportional hazards regression.

Subgroup Analyses

A greater reduction in the risk of hypoglycemia associated with SGLT2i vs. sulfonylureas was observed (i) among patients aged ≥65 years than those aged <65 years (interaction p =0.02) and (ii) among patients with eGFR <60 mL/min/1.73 m² than those with eGFR ≥60 mL/min/1.73 m² (interaction p =0.04; Figure 2). A larger reduction in the risk of hypoglycemia associated with GLP1RA vs. sulfonylureas was observed (i) among patients with eGFR <60 mL/min/1.73 m² than those with eGFR ≥60 mL/min/1.73 m² (interaction p <0.01) and (ii) among patients who did not use insulin than those who used insulin at baseline (interaction p =0.01).

Figure 2.

Weighted hazard ratios (95% confident intervals) for hypoglycemia comparing (a) SGLT2i vs. sulfonylureas, (b) GLP1RA vs. sulfonylureas, and (c) DPP4i vs. sulfonylureas by subgroups of age, eGFR, HbA1c, and concomitant insulin use

Sensitivity Analyses

The results were consistent in the as-treated analysis (Appendix Table 6). Within the subcohort of patients who did not use insulin at baseline, SGLT2i, GLP1RA, and DPP4i users were more likely to initiate insulin during follow-up compared to sulfonylurea users (Appendix Table 7). The assessment of hypoglycemia-related hospitalizations using diagnosis codes yielded similar results, although the small number of events limited the precision in some of the estimates (Appendix Table 8). The results were congruent in the analysis exclusive of first-generation sulfonylurea users (5 chlorpropamide users; Appendix Table 9). In the negative control outcome analysis, the risk of hospitalization with bleeding was similar across the SGLT2i, GLP1RA, DPP4i, and sulfonylurea groups (Appendix Table 10).

For 4,901 (45.7%) patients included in the assessment of glycemic variability, the median (IQI) number of HbA1c measurements was 5 (4–7). Compared to sulfonylureas, the variability in HbA1c was lower with SGLT2i (coefficient −0.09 [95% CI −0.17 to −0.01]) and GLP1RA (coefficient −0.13 [95% CI −0.21 to −0.05]) (Appendix Table 11). The results were consistent before and after controlling for the number of HbA1c measurements during follow-up.

DISCUSSION

In this real-world study of adults with type 2 diabetes, SGLT2i, GLP1RA, and DPP4i were associated with lower risks of hypoglycemia compared to sulfonylureas. The risks of hypoglycemia associated with SGLT2i and GLP1RA were comparable to DPP4i. Similar patterns were observed with the risk of severe hypoglycemia. Moreover, these associations were supported by various sensitivity analyses.

Small clinical trials (approximately 1000 or fewer participants) showed that newer glucose-lowering agents were well tolerated with respect to hypoglycemia, and the rates of hypoglycemia were comparable between SGLT2i and DPP4i,^{16, 18} and between GLP1RA and DPP4i.^{14, 15, 17} Our study of more than 10,000 adults with type 2 diabetes provided additional real-word evidence for safety of the newer glucose-lowering agents in terms of hypoglycemic risk. Our findings suggest that only 20 patients would need to switch from sulfonylureas to prevent one hypoglycemic event in the next 5 years. The population-level impact of reducing the use of sulfonylureas would be significant since sulfonylureas are prescribed to approximately 6.5 million US adults.¹ Hypoglycemia poses a substantial burden to healthcare systems, with an approximate mean cost of $20,000 per hospitalization.¹¹ Together with the cardiovascular benefits of SGLT2i and GLP1RA,^2–6 our results suggest that diabetes clinical practice should incorporate greater use of these agents.

Older age, lower eGFR, concomitant insulin use, and HbA1c <7% are associated with increased risk of hypoglycemia and may modify the association between the glucose-lowering agents and hypoglycemia. The point estimates from our subgroup analyses suggest the risks of hypoglycemia are likely to remain lower for SGLT2i, GLP1RA, and DPP4i compared to sulfonylureas among patients with older age, lower eGFR, tight glycemic control, or concomitant insulin use. However, the small sample size in each subgroup may have limited our evaluation.

With varying clinical presentations that may or may not prompt a capillary blood glucose measurement during healthcare encounters, the exact incidence of hypoglycemia can be challenging to measure. Reassuringly, we found similar relative risks in the analysis with hypoglycemia-related hospitalization captured by validated ICD codes. In addition, the assessment of insulin initiation by exposure medication class suggested the lower risks of hypoglycemia observed with SGLT2i and GLP1RA were not attributable to lower insulin use during follow-up. Moreover, the smaller variabilities in HbA1c associated with SGLT2i and GLP1RA vs. sulfonylureas support the observed associations with hypoglycemia.

Our study has several strengths. Real-world data on the risks of hypoglycemia associated with SGLT2i, GLP1RA, and DPP4i is extremely limited, and our study may be the first study to quantify such risks in routine patient care settings. We used capillary blood glucose, the most frequently used measurement in diabetes care, to ascertain hypoglycemia. Assessment of hypoglycemia using capillary blood glucose <70 mg/dL has been used in randomized trials to assess safety of oral glucose-lowering agents.^{16, 17} We confirmed the robustness of our findings using an extensive set of sensitivity analyses, including the as-treated analysis, assessment of hypoglycemia using diagnosis codes, and evaluation of negative control outcome. We used multiple imputation to address missing covariates. Propensity score–based IPTW was employed to balance the exposure groups in carefully selected patient characteristics for a valid risk assessment.

However, our study also has limitations. As with all observational studies, our results are subject to residual confounding. While confounding by indication is less likely with use of active comparators with similar indications, indication for using one glucose-lowering agent over another may have predisposed patients to our study outcomes.¹⁹ For example, physicians might have prescribed the newer glucose-lowering agents vs. sulfonylurea to patients who were at a higher risk of hypoglycemia. In this scenario, our estimates of hypoglycemia risk reduction with SGLT2i and GLP1RA would be an underestimate, compared to the true risk reduction. Second, the incidence of hypoglycemia may be underestimated in our study because patients with asymptomatic or uncommon symptoms during clinical encounters or those who experience hypoglycemia outside the clinical settings (e.g., home) are less likely to be captured. Our study population was derived from a single US tertiary health system and composed mostly of white adults with health insurance. Although we designed our study to include new glucose-lowering agent users (i.e., no study medication prescription records for at least 1 year while other health encounters existed in Geisinger), it may be plausible that patients received the study medications from another health system prior to the first prescription in Geisinger. We used prescription records to determine study drug exposure, and actual usage could not be guaranteed. Finally, HbA1c variability could be affected by non-glycemic factors, such as timing and frequency of measurements, anemia, and kidney dysfunction.⁴⁵

In conclusion, SGLT2i and GLP1RA, the agents that reduce adverse cardiovascular events, also confer lower risks of hypoglycemia compared with sulfonylureas in real-world practice. Efforts for greater use of SGLT2i and GLP1RA rather than sulfonylureas are needed.

Funding

Dr. Morgan E. Grams was supported by grant number R01DK115534 and K24HL155861 (Principal Investigator: M.G.) from the National Institute of Diabetes and Digestive and Kidney Diseases.

Dr. Jung-Im Shin was supported by grant number K01DK121825 (Principal Investigator: J.S.) from the National Institute of Diabetes and Digestive and Kidney Diseases.

The funding sources had no role in study design; data collection, analysis, and interpretation; writing of the manuscript; and decision to submit the manuscript for publication.

Declarations

Conflict of interest

Dr. Lyu has nothing to disclose. Dr. Hwang has nothing to disclose. Dr. Selvin reports grants from National Institutes of Health (NIH), during the conduct of the study; personal fees from Novo Nordisk, other from Wolters Kluwer, outside the submitted work. Dr. Jameson reports speakers bureau Sanofi Aventis — glargine, Soliqua, Boehringer Ingelheim — Jardiance. Dr. Chang reports grants from Novo Nordisk, personal fees from Reata, personal fees from Novartis, outside the submitted work. Dr. Grams reports grants from National Institute of Health, during the conduct of the study; grants from NIDDK, grants and other from NKF, other from ADA, non-financial support and other from KDIGO, other from USRDS, outside of submitted work. Dr. Shin reports grants from NIH, during the conduct of the study; grants from Merck, outside the submitted work.