Risk of Severe Hypoglycemia With Newer Second-line Glucose-lowering Medications in Older Adults With Type 2 Diabetes Stratified by Known Indicators of Hypoglycemia Risk

Abstract

Background

Severe hypoglycemia is associated with adverse clinical outcomes. We evaluated the risk of severe hypoglycemia in older adults initiating newer glucose-lowering medications overall and across strata of known indicators of high hypoglycemia risk.

Methods

We conducted a comparative-effectiveness cohort study of older adults aged >65 years with type 2 diabetes initiating sodium-glucose cotransporter 2 inhibitors (SGLT2i) versus dipeptidyl peptidase-4 inhibitors (DPP-4i) or SGLT2i versus glucagon-like peptide-1 receptor agonists (GLP-1RA) using Medicare claims (3/2013–12/2018) and Medicare-linked-electronic health records. We identified severe hypoglycemia requiring emergency or inpatient visits using validated algorithms. After 1:1 propensity score matching, we estimated hazard ratios (HR) and rate differences (RD) per 1,000 person-years. Analyses were stratified by baseline insulin, sulfonylurea, cardiovascular disease (CVD), chronic kidney disease (CKD), and frailty.

Results

Over a median follow-up of 7 (interquartile range: 4–16) months, SGLT2i was associated with a reduced risk of hypoglycemia versus DPP-4i (HR 0.75 [0.68, 0.83]; RD −3.21 [−4.29, −2.12]), and versus GLP-1RA (HR 0.90 [0.82, 0.98]; RD −1.33 [−2.44, −0.23]). RD for SGLT2i versus DPP-4i was larger in patients using baseline insulin than in those not, although HRs were similar. In patients using baseline sulfonylurea, the risk of hypoglycemia was lower in SGLT2i versus DPP-4i (HR 0.57 [0.49, 0.65], RD −6.80 [−8.43, −5.16]), while the association was near-null in those without baseline sulfonylurea. Results stratified by baseline CVD, CKD and frailty were similar to the overall cohort findings. Findings for the GLP-1RA comparison were similar.

Conclusions

SGLT2i was associated with a lower hypoglycemia risk versus incretin-based medications, with larger associations in patients using baseline insulin or sulfonylurea.

Graphical Abstract

Graphical Abstract.

Hypoglycemia is one of the most common adverse events related to glucose-lowering medications in patients with diabetes (1,2), and is associated with an increased risk of adverse health outcomes and mortality (3–5). It is a leading cause of hospital admissions and emergency department visits in older adults 65 years and older (1,6,7). The American Diabetes Association’s (ADA) Standards of Care emphasize the risk of hypoglycemia as a critical consideration in the management of type 2 diabetes (T2D) and recommend incorporating newer glucose-lowering medication classes with lower hypoglycemia risk—sodium-glucose cotransporter 2 inhibitors (SGLT2i), dipeptidyl peptidase-4 inhibitors (DPP-4i), and glucagon-like peptide-1 receptor agonists (GLP-1RA)—and adjusting therapies when using medication classes with known hypoglycemia risk—insulin and sulfonylureas (SU), particularly in older patients (8).

There are, however, several barriers to implementing such guidelines in routine care. First, evidence on the hypoglycemia risk of glucose-lowering medications is mainly derived from placebo-controlled randomized clinical trials (RCT). To date, there are no trials directly comparing the risk among newer glucose-lowering medication classes. Since both SGLT2i and GLP-1RA have demonstrated cardioprotective effects (8), this comparative evidence on safety could help balance their cardiovascular benefits versus the risk of hypoglycemia. Second, treatment regimens used in trials may not be generalizable to patients receiving routine care since trials closely monitor the participants’ glucose-lowering regimens, while real-world drug utilization patterns are less closely monitored. Third, there is evidence that the risk of hypoglycemia varies not only by the individual treatment regimen but also by the combination of specific regimens (9,10). For example, evidence from RCTs showed that adding DPP-4i in patients with baseline SU increases the risk of hypoglycemia compared to placebo (11).

To address these gaps in evidence, we compared the risk of hypoglycemia among patients initiating SGLT2i relative to those initiating DPP-4i or GLP-1RA in routine care, overall and stratified by baseline use of insulin or SU, or other known indicators of hypoglycemia risk.

Method

We conducted a comparative-effectiveness cohort study using the U.S. Medicare fee-for-service claims data. The Medicare claims data contain the longitudinal history of patient demographics, enrollment start and end dates, dispensed medications, diagnosis/procedure codes and their associated dates and settings for older adults >65 years. The study protocol was approved by the Institutional Review Board at Mass General Brigham (MGB).

Study Population and Exposure Assessment

The study populations included 2 pairwise comparisons of older adults aged >65 years newly prescribed with (i) SGLT2i versus DPP-4i or (ii) SGLT2i versus GLP-1RA during 2013–2018. These medication initiators were identified after requiring a 12-month washout period, during which they could be treated with glucose-lowering medications other than the ones being compared, that is, SGLT2i or DPP-4i in the first comparison, and SGLT2i or GLP-1RA in the second comparison.

Cohort entry was the date of the first prescription of any of the above drug classes, and the 12-month washout period on or before cohort entry is defined as the baseline period during which all cohort eligibility criteria were assessed (Supplementary Figure 1). During the baseline period, we required patients to have continuous enrollment in Medicare plan parts A, B, and D, and at least one diagnosis code for T2D.

We excluded patients with any diagnosis or procedure codes for type 1 or secondary diabetes, end-stage renal disease, renal replacement therapy, human immunodeficiency virus, solid organ transplant, malignancy, nursing home admissions, missing age or sex, or Saxenda, a high-dose liraglutide approved for weight loss (Supplementary Figures 2 and 3).

Follow-up started from one day after cohort entry in an as-treated manner until the earliest occurrence of one of the following events: discontinuation of the index drug, switching to the comparator drug, a gap in Medicare enrollment (>30 days), death, end of the study period (December 31, 2018), or the occurrence of the study outcomes. When patients were censored for drug discontinuation or switching to the comparator drug, the exposure assessment period ended on the last day of the days-supply of the last prescription of the index exposure drug plus a grace period of 60 days. We added this grace period to account for the potential delayed effect of the treatment or to allow time for the prescription refill.

Outcomes

We identified severe hypoglycemic events resulting in inpatient or emergency care using validated claims-based algorithms (positive predictive values [PPV] [78%–89%])—requiring at least one diagnosis code of hypoglycemia (International Classification of Diseases, 9th and 10th Revisions, Clinical Modification) in the primary discharge position in an inpatient, or in any position in an emergency care settings (12,13). Recurrent outcomes were not considered in this analysis, and we examined the outcomes in inpatient and emergency care settings separately in secondary analyses.

Covariates

We measured ~150 potential confounders during the 12-month baseline, including patient demographics, calendar time of cohort entry, comorbidities, a modified Charlson/Elixhauser combined comorbidity score (14), proxies of diabetes severity (diabetes complications, number of glucose-lowering medications on cohort entry, diabetes medication use during the baseline period or on the cohort entry date, number of hemoglobin A1c (HbA1c) or glucose tests ordered) (15), validated claims-based algorithms for obesity, and tobacco use disorders (16,17), chronic disease medications, a claims-based frailty index (18,19), indicators of healthcare utilization as a proxy for the overall disease state and intensity of care, and the history and timing of hypoglycemic events during baseline. We assessed the clinical/biochemical parameters in a subset of Medicare data linked to electronic health records (EHR) in an external validation study (see below for more details).

The analyses were stratified by indicators of high hypoglycemia risk: (i) baseline insulin use (yes/no), (ii) baseline SU use (yes/no), (ii) combinations of baseline insulin and SU use (both insulin and SU, insulin but no SU, no insulin but SU, and no insulin nor SU), (iv) baseline history of hypoglycemia (yes/no), and (v) levels of the claims-based frailty index (robust, pre-frail, and frail) (3,5,20,21). We further stratified analyses by the history of baseline cardiovascular diseases (CVD) (atherosclerotic CVD or heart failure [yes/no]) and chronic kidney disease (CKD) (yes/no) to reflect the individualized guideline recommendations within these subgroups (8).

Statistical Analyses

Using the measured covariates, we predicted the probability of initiating SGLT2i versus each comparator (propensity score [PS]) with logistic regression (22). To adjust for confounding, we then matched SGLT2i versus each comparator in a 1:1 ratio by the PS using a greedy algorithm without replacement within a 1% caliper of the PS (23). After matching, we assessed balance in the measured covariates between SGLT2i and the comparators using the standardized mean differences (SMD) (ranging from 0 to 1, with lower values indicating better balance) (24), and the post-matched c-statistics of the model predicting SGLT2i conditional on covariates (0.5 indicating excellent balance) (25). For subgroup analyses, PS was re-estimated, and the matching was reperformed within each subgroup.

We estimated hazard ratios (HR) with 95% CI based on the Cox proportional hazards model, and rate differences (RD) based on the weighted least squares regression model (26). The heterogeneity across subgroups was evaluated using the Wald test for homogeneity. We present the risk of hypoglycemia over the follow-up using nonparametric cumulative incidence function (CIF) plots, while accounting for the competing risk of mortality (27). Numbers of patients needed to treat (NNT) to prevent one case of hypoglycemia were also calculated based on cumulative risk differences at 2 years of follow-up (28).

Sensitivity Analyses

We conducted pre-specified sensitivity analyses to evaluate the robustness of our findings. First, we conducted an intent-to-treat analysis, which did not censor patients for treatment changes, up to one year of maximum follow-up. This method accounts for informative censoring by disregarding treatment discontinuation or switching during follow-up, though it is prone to treatment misclassification. Second, we adjusted for both time-fixed and time-varying predictors of treatment changes (eg, medications and hospitalizations due to CVD) via the inverse probability of treatment and censoring weighted analyses (29). This method accounts for informative censoring by creating a pseudopopulation, in which censoring is independent of both measured time-fixed and time-varying predictors. Third, to examine the potential misclassification of the exposure definition, we reduced the grace period in the exposure-assessment period from 60 to 30 days before censoring for treatment discontinuation/switching.

Fourth, to reduce residual confounding, we restricted our cohort to patients with a baseline dispensing record for metformin (recommended first-line medication among patients without severely reduced kidney function) (8). Fifth, we conducted bias analyses evaluating the estimates obtained after hypothetically adjusting for strong unmeasured confounders (estimated glomerular filtration rate [eGFR] and HbA1c) (30). We used evidence from prior studies as parameters for this bias assessment—relative risks (for hypoglycemia) of 1.3 per 10 mL/min/1.73 m² decrease in eGFR (31), and 1.2 per percentage increase in HbA1c above 7.0 (9). Sixth, we attempted to replicate the known positive associations of SGLT2i with diabetes ketoacidosis (DKA) and genital infections, and the expected null associations of SGLT2i with herpes zoster (32–35).

We conducted post-hoc sensitivity analyses to assess the potential confounding due to baseline or time-varying use of SU or insulin. First, to examine the potential time-varying confounding due to the differential use of insulin or SU between treatment groups over the follow-up, we compared the prevalance of insulin or SU use during 3-monthly incremental intervals over the follow-up between SGLT2i versus comparator initiators. Second, to assess the bias due to initiation/discontinuation of SU over the follow-up within each baseline SU subgroup, we repeated the subgroup analyses by baseline SU, censoring for SU initiation/discontinuation over the follow-up. Third, we examined the distributions of individual SU agents and their baseline dosage patterns in patients with baseline SU dispensing.

External Validation

The Medicare-Research Patient Data Registry (RPDR) contains EHRs from various hospital systems within the MGB healthcare system linked to the Medicare claims data (36–38). The EHR database contains information on patient demographics, diagnosis/procedure codes, medications, vital signs, laboratory data, and various clinical notes, and reports. To assess the balance in unmeasured risk factors (HbA1c, eGFR, blood pressure, lipid profile, and body mass index) between patients initiating SGLT2i versus DPP-4i/GLP-1RA in the primary Medicare cohorts, we identified separate cohorts in the Medicare RPDR data (2014–2017) using the similar eligibility criteria as the primary Medicare cohorts, and repeated PS matching adjusting only for claims-based variables. We then assessed the balance in EHR variables between SGLT2i and each matched comparator using SMDs.

Analyses were performed using Aetion Evidence Platform (2021), a software for real-world data analysis, which has been validated for a range of studies (Aetion, Inc., Gloucester, MA) (39,40), and SAS 9.4 Statistical Software (SAS Institute, Inc., Cary, NC).

Results

We identified (i) 93,195 SGLT2i versus 371,102 DPP-4i initiators in the first comparison, and (ii) 137,486 SGLT2i versus 128,700 GLP-1RA initiators in the second comparison who fulfilled cohort eligibility criteria. After 1:1 PS matching, our final study cohorts included 82,994 and 88,726 matched pairs in the first and second comparisons, respectively.

We present baseline characteristics before and after matching in Supplementary Table 1 and Table 1, respectively. Prior to matching, compared to DPP-4i initiators, SGLT2i initiators were younger (mean age 71.6 vs 74.4 years), more likely to be white (82.7% vs 75.0%), more likely to be on metformin, and on insulin, and to have a lower combined comorbidity score and a lower burden of baseline CVD and CKD. History of baseline SU use and hypoglycemia were similar. Compared to GLP-1RA initiators, SGLT2i initiators were similar in age (72.2 vs 71.7 years) before matching, although they were slightly more likely to have used metformin, second-generation SU, and insulin, and had a slightly lower burden of baseline CVD and diabetes complications. All baseline characteristics were similar between SGLT2i and comparators after PS matching. Distributions of baseline laboratory parameters were balanced between treatment comparisons in the Medicare-linked-EHR data (Supplementary Table 2).

Table 1.

Selected Baseline Characteristics of Patients Initiating SGLT2i Versus DPP-4i or GLP-1RA After 1:1 Propensity Score Matching

Characteristic	SGLT2i N = 82,994	DPP-4i N = 82,994	SMD	SGLT2i N = 88,726	GLP-1RA N = 88,726	SMD
Demographics
Age, mean (SD)	71.8 (5.1)	71.8 (5.1)	0.009	71.8 (5.2)	71.9 (5.2)	0.005
Sex—Female	41,348 (49.8%)	41,458 (50.0%)	0.003	47,243 (53.2%)	47,187 (53.2%)	0.001
Race categories
White; n (%)	67,988 (81.9%)	67,959 (81.9%)	0.001	73,132 (82.4%)	72,956 (82.2%)	0.005
Black; n (%)	6 064 (7.3%)	6 126 (7.4%)	0.003	6 824 (7.7%)	6 722 (7.6%)	0.004
Asian; n (%)	2 675 (3.2%)	2 737 (3.3%)	0.004	2 327 (2.6%)	2 468 (2.8%)	0.010
Hispanic; n (%)	2 484 (3.0%)	2 438 (2.9%)	0.003	2 617 (2.9%)	2 704 (3.0%)	0.006
Other or unknown; n (%)	3 783 (4.6%)	3 734 (4.5%)	0.003	3 826 (4.3%)	3 876 (4.4%)	0.003
Burden of comorbidities or frailty
Combined comorbidity score, mean (SD)	1.1 (1.9)	1.1 (1.9)	0.002	1.3 (2.0)	1.3 (1.9)	0.004
Frailty Index, mean (SD)	0.2 (0.1)	0.2 (0.1)	0.007	0.2 (0.1)	0.2 (0.1)	0.003
Diabetes-related conditions
Diabetic nephropathy	9 759 (11.8%)	9 684 (11.7%)	0.003	13,090 (14.8%)	13,267 (15.0%)	0.006
Diabetic retinopathy	10,817 (13.0%)	10,801 (13.0%)	0.001	12,789 (14.4%)	12,778 (14.4%)	0.000
Diabetic neuropathy	21,231 (25.6%)	21,157 (25.5%)	0.002	25,210 (28.4%)	25,247 (28.5%)	0.001
Diabetic foot	2 398 (2.9%)	2 437 (2.9%)	0.003	2 854 (3.2%)	2 867 (3.2%)	0.001
Hypoglycemia (−60 days to cohort entry date)	3 679 (4.4%)	3 671 (4.4%)	0.000	4 203 (4.7%)	4 176 (4.7%)	0.001
Hypoglycemia (−180 to −61 days)	3 261 (3.9%)	3 320 (4.0%)	0.004	3 768 (4.2%)	3 824 (4.3%)	0.003
Hypoglycemia (−365 to −181 days	3 963 (4.8%)	3 974 (4.8%)	0.001	4 528 (5.1%)	4 544 (5.1%)	0.001
Hyperglycemia	29,278 (35.3%)	29,584 (35.6%)	0.008	34,516 (38.9%)	34,358 (38.7%)	0.004
Diabetic ketoacidosis	278 (0.3%)	307 (0.4%)	0.006	326 (0.4%)	316 (0.4%)	0.002
Hyperosmolar hyperglycemic nonketotic syndrome	694 (0.8%)	754 (0.9%)	0.008	826 (0.9%)	815 (0.9%)	0.001
Diabetes treatment
Number of glucose-lowering medications on cohort entry, mean (SD)	1.3 (0.9)	1.3 (0.8)	0.001	1.5 (1.0)	1.5 (1.0)	0.003
Glucose-lowering medications use on the day of cohort entry:
Metformin	52,620 (63.4%)	52,820 (63.6%)	0.005	52,483 (59.2%)	52,384 (59.0%)	0.002
Sulfonylureas—second generation	29,425 (35.5%)	29,682 (35.8%)	0.006	32,867 (37.0%)	32,607 (36.8%)	0.006
Thiazolidinediones (TZD)	6 022 (7.3%)	6 066 (7.3%)	0.002	6 833 (7.7%)	6 773 (7.6%)	0.003
GLP-1RA	4 283 (5.2%)	3 755 (4.5%)	0.030	20,813 (23.5%)	21,336 (24.0%)	0.014
Insulins	16,327 (19.7%)	16,415 (19.8%)	0.003	21,238 (23.9%)	21,328 (24.0%)	0.002
Lifestyle factors
Obesity	28,018 (33.8%)	28,409 (34.2%)	0.010	33,824 (38.1%)	33,622 (37.9%)	0.005
Smoking	17,237 (20.8%)	17,378 (20.9%)	0.004	18,892 (21.3%)	18,803 (21.2%)	0.002
Alcohol abuse or dependence	1 013 (1.2%)	1 020 (1.2%)	0.001	967 (1.1%)	943 (1.1%)	0.003
Cardiorenal and other systemic comorbidities
Acute MI	1 827 (2.2%)	1 801 (2.2%)	0.002	1 848 (2.1%)	1 865 (2.1%)	0.001
Unstable angina	2 531 (3.0%)	2 566 (3.1%)	0.002	2 653 (3.0%)	2 632 (3.0%)	0.001
Coronary procedure	1,956 (2.4%)	1,933 (2.3%)	0.002	1,887 (2.1%)	1 868 (2.1%)	0.001
Heart failure	9 456 (11.4%)	9 399 (11.3%)	0.002	10,920 (12.3%)	11,065 (12.5%)	0.005
Ischemic stroke	9 472 (11.4%)	9 533 (11.5%)	0.002	10,365 (11.7%)	10,261 (11.6%)	0.004
PAD and generalized/unspecified atherosclerosis	10,741 (12.9%)	10,726 (12.9%)	0.001	12,296 (13.9%)	12,436 (14.0%)	0.005
Acute kidney injury	2 621 (3.2%)	2 614 (3.1%)	0.000	3 549 (4.0%)	3 589 (4.0%)	0.002
CKD Stage 1–2	3 336 (4.0%)	3 306 (4.0%)	0.002	4 054 (4.6%)	4 083 (4.6%)	0.002
CKD Stage 3–4	7 109 (8.6%)	7 077 (8.5%)	0.001	10,591 (11.9%)	10,710 (12.1%)	0.004
Electrolyte disorders	5 822 (7.0%)	5 837 (7.0%)	0.001	6 860 (7.7%)	6 766 (7.6%)	0.004
Disorders of fluid balance	2,716 (3.3%)	2 704 (3.3%)	0.001	3 238 (3.6%)	3 254 (3.7%)	0.001
COPD	10,396 (12.5%)	10,607 (12.8%)	0.008	11,735 (13.2%)	11,777 (13.3%)	0.001
Obstructive sleep apnea	13,932 (16.8%)	14,150 (17.0%)	0.007	16,537 (18.6%)	16,559 (18.7%)	0.001
NASH/NAFLD	4 200 (5.1%)	4 298 (5.2%)	0.005	4 777 (5.4%)	4 792 (5.4%)	0.001
Dementia	3 449 (4.2%)	3 528 (4.3%)	0.005	4 024 (4.5%)	4 037 (4.5%)	0.001
Other medications use
ACEI and ARBs	65,165 (78.5%)	65,198 (78.6%)	0.001	70,897 (79.9%)	70,873 (79.9%)	0.001
Beta blockers	38,932 (46.9%)	38,872 (46.8%)	0.001	42,773 (48.2%)	42,718 (48.1%)	0.001
Calcium channel blockers	27,620 (33.3%)	27,691 (33.4%)	0.002	30,303 (34.2%)	30,416 (34.3%)	0.003
Loop diuretics	14,720 (17.7%)	14,818 (17.9%)	0.003	17,651 (19.9%)	17,706 (20.0%)	0.002
Mineralocorticoid receptor antagonist	3 771 (4.5%)	3 800 (4.6%)	0.002	4 339 (4.9%)	4 425 (5.0%)	0.004
Anticoagulants (oral)	8 065 (9.7%)	8 071 (9.7%)	0.000	8 735 (9.8%)	8 701 (9.8%)	0.001
Antiplatelet agents	11,168 (13.5%)	11,219 (13.5%)	0.002	12,038 (13.6%)	12,126 (13.7%)	0.003
Statins	64,371 (77.6%)	64,415 (77.6%)	0.001	69,953 (78.8%)	70,001 (78.9%)	0.001
Corticosteroids (oral)	14,283 (17.2%)	14,337 (17.3%)	0.002	15,660 (17.6%)	15,514 (17.5%)	0.004
Opioids	28,738 (34.6%)	28,954 (34.9%)	0.005	32,263 (36.4%)	32,168 (36.3%)	0.002
Measures of healthcare utilization
Internist (–30 days to CED)	52,793 (63.6%)	53,194 (64.1%)	0.010	55,850 (62.9%)	55,775 (62.9%)	0.002
Endocrinologist (–30 days to CED)	10,099 (12.2%)	9,815 (11.8%)	0.011	13,376 (15.1%)	13,425 (15.1%)	0.002
Cardiologist (–30 days to CED)	10,831 (13.1%)	10,864 (13.1%)	0.001	11,104 (12.5%)	11,031 (12.4%)	0.002
Nephrologist (–30 days to CED)	907 (1.1%)	1 069 (1.3%)	0.018	1 273 (1.4%)	1 447 (1.6%)	0.016
HbA1c test order (number of tests), mean (SD)	2.8 (1.3)	2.8 (1.4)	0.002	2.9 (1.4)	2.9 (1.4)	0.002
Glucose test and monitoring	27,339 (32.9%)	27,936 (33.7%)	0.015	31,161 (35.1%)	31,816 (35.9%)	0.015
Hospitalization event (number of), mean (SD)	0.2 (0.5)	0.2 (0.5)	0.004	0.2 (0.5)	0.2 (0.5)	0.000
LOS (–30 days to CED), mean (SD)	0.1 (0.9)	0.1 (0.8)	0.000	0.1 (0.9)	0.1 (0.9)	0.002
ED event (number of), mean (SD)	0.7 (1.7)	0.7 (1.7)	0.004	0.7 (1.8)	0.7 (1.8)	0.002
Distinct medications ALL, mean (SD)	13.0 (5.8)	13.0 (5.8)	0.007	13.9 (6.1)	13.9 (5.9)	0.002

Note: ACEI = angiotensin converting enzyme inhibitors; ARB = angiotensin receptor blockers; BB = beta blockers; CCB = calcium channel blockers; CED = cohort entry date; CKD = chronic kidney disease; COPD = chronic obstructive pulmonary diseases; CVD = cardiovascular disease; DM = diabetes mellitus; ED = emergency department; DPP4i = dipeptidyl peptidase-4 inhibitors; GLP-1RA = glucagon-like peptide-1 receptor agonist; LOS = length of stay; MI = myocardial infarction; NASH = non-alcoholic steatohepatitis; NAFLD = non-alcoholic fatty liver disease; PAD = peripheral arterial diseases; PSA = prostate surface antigen; SD = standard deviation; SGLT2i = sodium-glucose cotransporter 2 inhibitors; SMD = standardized mean difference; SU = sulfonylurea; TZD = thiazolidinediones.

Over a median follow-up of 7 months (interquartile range: 4–16), after PS matching, SGLT2i initiators were associated with a lower hypoglycemia risk versus DPP-4i initiators, with incidence rates of 9.8 and 13.0 per 1,000 person-years, respectively, which corresponded to a HR (95% CI) of 0.75 (0.68, 0.83), and RD (95% CI) of −3.21 (−4.29, −2.12) per 1,000 person-years (Table 2). Relative to GLP-1RA, over a similar duration of follow-up, SGLT2i had a lower incidence of hypoglycemia per 1,000 person-years (11.1 vs 12.5), with a corresponding HR (95% CI) of 0.90 (0.82, 0.98) and RD (95% CI) of −1.33 (−2.44, −0.23) per 1,000 person-years (Table 2). The most common reason for censoring in both comparisons was the discontinuation of the index exposure (Supplementary Table 3). Results for secondary outcomes were similar to the main findings (Table 2).

Table 2.

Number of Events and Incidence of Hypoglycemia in 1:1 PS Matched Cohorts of SGLT2i and DPP-4i (GLP-1RA) Initiators

Comparison	Outcomes	Number of Matched Pairs	SGLT2i Initiators	Comparator Initiators	Rate Differences Per 1 000 PY (95% CI)	Hazard Ratios (95% CI)
			No. of events (rate/1 000 PY)	No. of events (rate/1 000 PY)
SGLT2i vs DPP-4i	Hypoglycemia (inpatient + ED)	82,994	692 (9.8)	987 (13.0)	−3.21 (−4.29, −2.12)	0.75 (0.68, 0.83)
	Hypoglycemia (inpatient)	82,994	77 (1.1)	96 (1.0)	−0.17 (−0.52, 0.18)	0.87 (0.65, 1.18)
	Hypoglycemia (ED)	82,994	689 (9.7)	986 (13.0)	−3.24 (−4.33, −2.15)	0.75 (0.68, 0.83)
SGLT2i vs GLP-1RA	Hypoglycemia (inpatient + ED)	88,726	848 (11.1)	898 (12.5)	−1.33 (−2.44, −0.23)	0.90 (0.82, 0.98)
	Hypoglycemia (inpatient)	88,726	92 (1.2)	75 (1.0)	0.17 (−0.17, 0.51)	1.16 (0.85, 1.57)
	Hypoglycemia (ED)	88,726	842 (11.1)	896 (12.4)	−1.38 (−2.49, −0.28)	0.89 (0.81, 0.98)

Note: DPP4i = dipeptidyl peptidase-4 inhibitors; ER = emergency room; GLP-1RA = glucagon-like peptide-1 receptor agonist; PS = propensity-score; PY = person-years; SGLT2i = sodium-glucose cotransporter 2 inhibitors.

The association of SGLT2i versus DPP-4i with hypoglycemia was similar between patients with baseline insulin use (HR: 0.70 [0.62, 0.80]), and those without it (HR: 0.65 [0.55, 0.75]). The absolute risk reduction with SGLT2i was greater in patients on baseline insulin (RD −8.81 [−11.90, −5.73]) than those not (RD −2.83 [−3.81, −1.86]; Figure 1). In patients on baseline SU, SGLT2i was associated with a lower hypoglycemia risk relative to DPP-4i (HR [95% CI]: 0.57 [0.49, 0.65]), while the risks were similar in those not on baseline SU (HR [95% CI]: 0.90 [0.78, 1.02]). When cross-stratified by baseline use of insulin and SU, we observed a lower risk of hypoglycemia in SGLT2i versus DPP-4i initiators in patients on both insulin and SU; and similar risk between SGLT2i and DPP-4i in patients on neither insulin nor SU (Supplementary Figures 4 and 5). Subgroup analyses by history of hypoglycemia, CVD, CKD and frailty (Supplementary Figure 4) demonstrated findings consistent with the primary analyses, though the absolute RDs were larger in patients with history of hypoglycemia, CKD and frailty. Subgroup analyses for the SGLT2i versus GLP-1RA comparison produced generally similar findings but estimates were more attenuated toward the null (Figure 1 and Supplementary Figure 5).

Figure 1.

Association of SGLT2i versus DPP-4i or GLP-1RA with risk of hypoglycemia by baseline use of insulin or SU. DPP4i = dipeptidyl peptidase-4 inhibitors; GLP-1RA = glucagon-like peptide-1 receptor agonist; HR = hazard ratios; PS = propensity-score; RD = rate difference; SGLT2i = sodium-glucose cotransporter 2 inhibitors; SU = sulfonylurea.

NNT at 2 years to prevent one case of severe hypoglycemia were much lower in patients using baseline SU compared to those not: 75 versus 390 for the SGLT2i versus DPP-4i, and 177 versus 2585 for SGLT2i versus GLP-1RA. Similarly, NNTs were much lower for patients using baseline insulin versus those not: 56 versus 182 for SGLT2i versus DPP-4i initiators, and 246 versus 538 for SGLT2i versus GLP-1RA initiators.

In the CIF plots, the risk of hypoglycemia was lower in SGLT2i versus comparators over the follow-up, particularly in those with baseline SU use (Figure 2 and Supplementary Figure 6). Sensitivity analyses varying the exposure, follow-up and cohort definitions were similar to the main findings (Supplementary Table 4). Results for the control outcomes were consistent with the literature (Supplementary Table 5). Analyses censoring for SU changes did not change the primary findings (Supplementary Figures 4 and 5), and proportions of time-varying insulin or SU dispensing were similar between SGLT2i versus either comparator (Supplementary Table 6). Most baseline SU agents were glipizide (25.7%) and glimepiride (25.0%) (Supplementary Tables 7 and 8). Censoring weighted estimates were also similar to the primary findings (Supplementary Table 9) although the estimate for GLP-1RA was attenuated toward the null. Bias analyses showed that our findings were fairly robust to unmeasured confounders (Supplementary Figure 7) (30).

Figure 2.

Cumulative incidence of hypoglycemia between PS-matched initiators of SGLT2i versus DPP-4i/GLP-1RA over as-treated follow-up—overall and by baseline SU subgroup. DPP4i = dipeptidyl peptidase-4 inhibitors; GLP-1RA = glucagon-like peptide-1 receptor agonist; PS = propensity-score; SGLT2i = sodium-glucose cotransporter 2 inhibitors; SU = sulfonylurea

Discussion

In this comparative effectiveness study of older adults, we observed a risk reduction of severe hypoglycemic events requiring inpatient or emergency care among SGLT2i versus DPP-4i (and to a lesser extent GLP-1RA) initiators. Subgroup analysis by history of insulin use showed estimates similar to the overall cohort, though the absolute risk reduction was greater (with smaller NNT to avoid one hypoglycemia event) in subgroups of patients with baseline SU or insulin use. On both relative and absolute scales, the risk of severe hypoglycemic events was similar between SGLT2i versus comparators in patients without baseline use of SU or insulin, suggesting the observed risk reduction with SGLT2i may be more pronounced among patients receiving SU (or insulin) at baseline. Results from the subgroup analyses by baseline CVD, CKD, history of hypoglycemia, and levels of frailty were similar to the overall cohort findings, highlighting again the potential importance of co-therapies with SU and insulin in modifying the hypoglycemia risk.

Although newer glucose-lowering medication classes are increasingly recommended due to low hypoglycemia risk and cardiovascular benefits (for SGLT2i and GLP-1RA) (8), trials directly comparing SGLT2i versus incretin-based medications with respect to important adverse effects are lacking. Overall, our finding on an incremental benefit of SGLT2i regarding the risk of hypoglycemia compared to DPP-4i (and GLP-1RA to a lesser extent) was in line with prior evidence suggesting that SGLT2i could induce glucagon release from pancreatic alpha islet cells, which could ameliorate the risk of severe hypoglycemia (41).

However, such risk reduction with SGLT2i versus incretin-based medications appeared particularly sensitive to baseline SU use. This finding is consistent with evidence from meta-analyses of placebo-controlled RCTs (11,42), warnings on the product label (43), and the 2022 ADA guidelines (8), which emphasize stopping or reducing current treatment regimens when initiating new therapies (9). Given the popular use of SU as a first or second-line therapy (44), our findings suggest the need for adjustment or close monitoring of treatment regimens when adding DPP-4i (and to a lesser extent GLP-1RA) in patients using SU. Future studies should explore the potential for and extent of this drug interaction between DPP-4i and SU in routine care.

Regarding baseline insulin use, evidence from RCTs suggested that initiation of DPP-4i might increase hypoglycemia risk relative to placebo in patients using baseline insulin (43,45–48). Our findings showed that the absolute benefit of SGLT2i versus DPP-4i was larger in patients on baseline insulin, compared to those not on it, but the HRs were similar in patients with and without baseline insulin.

In a subgroup of patients without baseline insulin or SU, HR for hypoglycemia benefit with SGLT2i was attenuated toward the null, which suggests that hypoglycemia risk emerges most strongly in those with baseline SU and insulin use. This finding is consistent with expectations from known mechanisms of action of these insulinotropic medications and placebo-controlled trials.

Our study has limitations. First, our outcome definition is not confirmed by laboratory results—we instead relied on severe hypoglycemic events reported in hospitalized and emergency care settings, using a validated algorithm with high PPV (12,13)^. Rates of hypoglycemia-related emergency and hospital visits observed in our study were comparable with prior studies based on electronic health care databases (1,44). Since the risk of severe hypoglycemic may be under-represented in trials due to strict protocols and tight monitoring of blood glucose, our findings complement trials evidence by providing data on severe hypoglycemic events in clinical practice, although we may not capture milder hypoglycemic events not requiring emergency or hospitalized care. Second, the potential for unmeasured confounding cannot be excluded. For example, Medicare does not report eGFR or HbA1c test results, although our bias analyses and external validation suggested robust findings to such unmeasured confounders (30). A prior study with a new-user, active-comparator design similar to ours showed that adjustment for a rich set of confounders measured in claims data could balance the biomarkers and clinical parameters that are only available in structured EHR (15). We were also able to replicate the known associations of SGLT2i with the risks of DKA and genital infections, and the expected null association of SGLT2i with the risk of herpes zoster infection in prespecified sensitivity analyses. We further attempted to reduce confounding by frailty by using an active comparator as well as adjusting for markers of frailty and the validated claims-based frailty index (18,19). Regarding the bias due to time varying treatment regimens, our extensive sensitivity analyses suggest that their potential to materially change the primary findings is low, although we do not capture the dose of insulin/SU treatments over the follow-up which could have biased our findings. Third, follow-up time was brief due to the real-world drug utilization patterns, however, the characterization of a short-term risk of severe hypoglycemia in routine care may nevertheless have important clinical implications for prevention.

Conclusion

In this comparative effectiveness study of older adults, the initiation of SGLT2i was associated with a lower risk of severe hypoglycemia relative to incretin-based glucose-lowering agents. This risk reduction appeared more pronounced among patients treated with SU or insulin at baseline, while other subgroup analyses by baseline CVD, CKD, history of hypoglycemia, and levels of frailty provided similar results to the overall cohort findings. Unmeasured confounding could not be excluded entirely.

Funding

This work was supported by a research grant (DB-2020C2-20326) from the Patient Centered Outcomes Research Institute. The authors had full control of the design and conduction of the study and interpretation of the study’s 
findings. The authors retained the right of publication and determined the final wording of the manuscript. P.T.H. was previously supported by a training grant (5T32DK007527) from the National Institute of Diabetes and Digestive and Kidney Diseases and is currently supported by the post-doctoral grant (4-22-PDFPM-15) from the American Diabetes Association. E.P. was supported by a career development grant (K08AG055670) from the National Institute on Aging, and by a research grant (DB-2020C2-20326) from the Patient-Centered Outcomes Research Institute, and the Food and Drug Administration (5U01FD007213).

Conflict of Interest

P.T.H. previously worked at Johnson & Johnson on an unrelated work. D.J.W. reports serving on data monitoring commitees for Novo Nordisk. S.C.K. has received research grants to the Brigham and Women’s Hospital from AbbVie, Roche, Pfizer, and Bristol-Myers Squibb for unrelated studies. The other authors declare no conflict of interest.

Author Contributions

P.T.H. contributed to the study design, data analysis, and the development of the manuscript. P.T.H. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. E.P. contributed to the study conception, design, and critical review of the manuscript. D.J.W., S.C.K., J.M.P., D.H.K., and E.A. contributed to the study conception and critical review of the manuscript. All authors approved the final version of the manuscript.