Pharmacologic Treatment of Type 2 Diabetes in the U.S., Sweden, and Israel

Pharmacotherapy and lifestyle modification are the cornerstone interventions for diabetes management. In the past two decades, several new glucose-lowering medications have been developed with demonstrated cardiorenal benefits (1–3). Glucagon-like peptide 1 receptor agonists (GLP1RA) and sodium–glucose cotransporter 2 inhibitors (SGLT2i) are now recommended in clinical guidelines for the management of diabetes in patients at high risk, defined as those with atherosclerotic cardiovascular disease, heart failure, or chronic kidney disease (4–6).

Documenting the differences and trends in diabetes treatment across countries is of interest for health care providers and patients, as well as policy makers, and can be helpful for identification of areas for improvement. In a previous study with use of medication sales data, investigators reported different uptake of new diabetes medications across European countries (7). However, no such comparison has been performed between the U.S. and other developed countries. The comparison is of interest given that the U.S. has a unique health care structure (e.g., no universal health care coverage) compared with other developed countries. In addition, there is a lack of detailed data for identification of factors that are associated with specific medication choice. In this study, we characterized the pharmacologic treatment of type 2 diabetes and identified predictors of the use of specific medications in a nationally representative sample of the U.S. population and two large health systems in Sweden and Israel.

We used data from the National Health and Nutrition Examination Survey (NHANES), the Stockholm CREAtinine Measurements (SCREAM) project, and Maccabi Health Services (Maccabi). NHANES uses a complex, stratified, and multistage probability-cluster sampling design. The study population is a nationally representative sample of the noninstitutionalized civilian U.S. population (8). The SCREAM study population is comprised of a health care use cohort of all residents in the region of Stockholm, Sweden, which has a population of ∼3 million, for whom universal health care is provided within a single unified health system (9). For SCREAM, there is complete information on demographics and medication dispensations at pharmacies, with no loss to follow-up. Maccabi is a health maintenance organization covering 25% of the Israeli population. The electronic health record data from Maccabi include information on demographics, health care encounters, prescriptions, and laboratory results.

The study population in NHANES included participants with type 2 diabetes from 2007 through 2018 who were 20 years of age or older. Diabetes was defined according to self-reported diagnosis of diabetes from a physician, fasting glucose >126 mg/dL, hemoglobin A_1c (HbA_1c) ≥6.5%, or self-reported use of glucose-lowering medications. We chose this definition to mirror the validated algorithm for diabetes applied in electronic health records in SCREAM and Maccabi (10,11). Patients who were diagnosed with diabetes at ages <30 years and received only insulin were excluded, as these patients may have type 1 diabetes (12).

In SCREAM and Maccabi, type 2 diabetes was ascertained with use of a validated algorithm based on diagnosis codes, prescription of glucose-lowering medication (excluding medication use for conditions such as polycystic ovarian syndrome and gestational diabetes mellitus), elevated HbA_1c, and elevated fasting glucose (10,11). Patients with type 1 diabetes were excluded. We constructed a series of period prevalent cohorts with 2-year intervals from 2007–2008 to 2017–2018 to be comparable with NHANES 2-year survey cycles. Each 2-year period prevalent cohort included patients with prevalent diabetes and incident diabetes. For example, the 2017–2018 period prevalent cohort included patients with prevalent diabetes (diagnosed before or on 1 January 2017) and those with incident diabetes (diagnosed after 1 January 2017 and before or on 31 December 2018). To make sure that patients were engaged with health systems, we required patients to have at least one serum creatinine measurement during the 2-year cohort period. This study was approved by the Stockholm Ethics Review Board and Maccabi and Johns Hopkins University institutional review boards.

In NHANES, medication information was extracted from prescription data according to the 3-level nested category system of Multum Lexicon (13). Participants were asked during the home interview if they had taken any prescription medications in the past month. Those who answered “yes” were asked to show the containers of all medications to interviewers. When a container was unavailable, participants reported the name of the medication. All medications were converted to a standard generic drug name.

Medication information was ascertained with outpatient dispensation records in SCREAM and outpatient prescription records in Maccabi. Ever use of any glucose-lowering medication as well as medication classes during 2-year cohort period was examined. Glucose-lowering medications were classified as follows: metformin, insulin, sulfonylureas, thiazolidinedione (TZD), dipeptidyl peptidase 4 inhibitor (DPP4i), SGLT2i, and GLP1RA. Use of combination therapy was included. For example, if a patient was on metformin and SGLT2i, the patient would be classified as both metformin user and SGLT2i user.

A priori, we selected potential predictors that might influence prescribing patterns of glucose-lowering medications. For this analysis, we limited the analysis to the population who received at least one glucose-lowering medication. We used the most recent data: 2015–2018 cycles in NHANES (patient flowchart in Supplementary Fig. 1A) and 2017–2018 period prevalent cohorts from SCREAM (Supplementary Fig. 1B) and Maccabi (Supplementary Fig. 1C). We included two survey cycles for NHANES to obtain a larger sample size.

Information on age and sex was self-reported in NHANES and abstracted from the electronic health record in SCREAM and Maccabi. Hypertension, coronary heart disease (CHD), heart failure, and stroke were defined according to self-reported physician diagnosis in NHANES and according to presence of ICD-9 and ICD-10 diagnosis codes in SCREAM and Maccabi. The diagnosis required one code if from an inpatient encounter or problem list or at least two relevant codes within 2 years if from a clinical encounter other than the inpatient setting any time before the end of cohort period (31 December 2018) (Supplementary Table 1). Information for concurrent statin use was extracted from prescription data in NHANES and was ascertained from outpatient dispensation/prescription records in SCREAM and Maccabi.

Serum creatinine, HbA_1c, BMI (only in NHANES and Maccabi), and urine albumin-to-creatinine ratio (ACR) were included for analysis. In NHANES, HbA_1c was measured in whole blood and calibrated to account for changes in laboratory methods over time (14). In SCREAM and Maccabi, the mean values of measurements during the 2-year cohort period were used for analysis. When measured ACR was not available, we used converted ACR from urine protein-creatinine ratio or dipstick proteinuria (15). The 2021 Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation without race coefficient was used to calculate estimated glomerular filtration rate (eGFR) based on serum creatinine level (16). eGFR and ACR were categorized as ≥60, 45–59, 30–44, and <30 mL/min/1.73 m² and <30, 30–300, and ≥300 mg/g, respectively, according to the Kidney Disease: Improving Global Outcomes (KDIGO) guidelines (17). HbA_1c was categorized as <7.5%, 7.5% to <9%, and ≥9%. BMI was categorized as <30, 30 to <40, and ≥40 kg/m².

We analyzed each data source separately. Participant characteristics in 2015–2018 NHANES data and 2017–2018 period prevalent cohorts of SCREAM and Maccabi were presented as mean (95% CI) for continuous variables and percentage (95% CI) for categorical variables. We first examined trends over time in any glucose-lowering medication use among all individuals with type 2 diabetes and then trends in each class of medication use among individuals who used at least one glucose-lowering medication.

We fit multivariable logistic regression models to identify independent predictors for metformin, insulin, SGLT2i, and GLP1RA use, respectively. In NHANES, ∼14% of patients had at least one missing variable and we performed complete case analysis. In SCREAM and Maccabi, missing data were primarily for ACR (30.9% missing for SCREAM and 34.4% for Maccabi). We used multiple imputation by chained equations to impute missing ACR, BMI, and HbA_1c levels (18). Patient characteristics; previous values of ACR, BMI, and HbA_1c; and use of medication (outcomes) were included in the imputation model. Rubin rules were used to combine estimates from a logistic regression model in each imputed data set for analysis in SCREAM and Maccabi (19). We further examined the proportion of patients who received SGLT2i and/or GLP1RA who had high risk of adverse cardiorenal events and those who did not. Patients at higher risk was defined as those with CHD, stroke, heart failure, ACR >300 mg/g, or eGFR <60 mL/min/1.73 m² based on American Diabetes Association (ADA) guidelines (6). The analysis in NHANES accounted for the complex survey design and incorporated survey weights (20). All statistical analyses were conducted with Stata, version 15 (StataCorp, College Station, TX), and R (www.R-project.org/) (21).

A total of 1,345 patients from NHANES, 71,185 patients from SCREAM, and 118,374 patients from Maccabi were included (Table 1). Mean age was 61.5 years (95% CI 60.4, 62.8) in NHANES, 59.6 years (58.8, 60.5) in SCREAM, and 64.4 years (64.3, 64.4) in Maccabi. Patients in the Maccabi system had a greater prevalence of CHD but lower prevalence of heart failure and stroke compared with NHANES and SCREAM. The prevalence of statin use was substantially higher in Maccabi than in NHANES and SCREAM (60.2% in NHANES, 65.3% in SCREAM, 83.3% in Maccabi). Patients in Maccabi had better glycemic control (64.4% in NHANES, 64.2% in SCREAM, and 72.0% in Maccabi had HbA_1c <7.5%). Patients in Maccabi were less likely to have obesity and had lower eGFR than those in NHANES and SCREAM.

In all periods, use of glucose-lowering medications among patients with type 2 diabetes was substantially higher in Maccabi than in NHANES and SCREAM (66.0% in NHANES, 68.4% in SCREAM, and 88.1% in Maccabi in 2017–2018) (Fig. 1A).

Among patients on at least one glucose-lowering medication, greater use of metformin was observed in Maccabi (e.g., in 2017–2018, 74.1% of patients in NHANES, 75.9% in SCREAM, and 92.6% in Maccabi) (Fig. 1B), but use increased in NHANES and SCREAM over time (58.6% in 2007–2008 and 75.1% in 2017–2018 in NHANES and 64.0% in 2007–2008 and 75.9% in 2017–2018 in SCREAM). Use of insulin remained stable over time, with greater use in SCREAM than in NHANES and Maccabi. Use of sulfonylureas decreased within each of the three cohorts; however, such use remained relatively more common in NHANES (31.5% in NHANES, 16.0% in SCREAM, and 14.9% in Maccabi in 2017–2018) (Fig. 1C). Even though TZDs were commonly used in earlier years in NHANES (22.0% in 2007–2008), a rapid decrease in use was observed from 2007–2008 to 2013–2014. In 2017–2018, use of TZDs was low in all three cohorts (3.9% in NHANES, 0.5% in SCREAM, and 4.7% in Maccabi).

Newer glucose-lowering medications were increasingly used over the years, with more rapid uptake in Maccabi than in NHANES and SCREAM. For example, use of DPP4i was more common and increased more rapidly in Maccabi than in NHANES and SCREAM (8.7% in NHANES, 14.4% in SCREAM, and 37.0% in Maccabi in 2017–2018) (Fig. 1D). Similarly, uptake of SGLT2i was more rapid in Maccabi than in NHANES and SCREAM (6.7% in NHANES, 8.3% in SCREAM, and 21.5% in Maccabi in 2017–2018) (Fig. 1E). We also observed increased uptake of GLP1RA over time, with more common use in Maccabi, particularly compared with NHANES (2.1% in NHANES, 10.5% in SCREAM, and 12.7% in Maccabi in 2017–2018) (Fig. 1F).

Across the three cohorts, CHD, stroke, heart failure, and lower eGFR were associated with lesser use of metformin (e.g., odds ratio [OR] for heart failure 0.70 [95% CI 0.31, 1.57] in NHANES, 0.61 [0.58, 0.65] in SCREAM, 0.65 [0.59, 0.72] in Maccabi) (Fig. 2A). In contrast, patients with hypertension, statin use, and higher BMI had greater use of metformin across cohorts. Older patients were more likely to use metformin in NHANES and Maccabi but less likely in SCREAM (1.49 [1.03, 2.15] in NHANES, 0.71 [0.68, 0.75] in SCREAM, 1.19 [1.13, 1.26] in Maccabi).

Comorbidities, higher HbA_1c, lower eGFR, and higher ACR were generally associated with greater insulin use across the three cohorts (Fig. 2B). Older age was associated with greater use of insulin in SCREAM but lesser use in Maccabi.

Older age and lower eGFR were associated with lesser use of SGLT2i across the three cohorts (e.g., OR for age ≥65 vs. <65 years 0.19 [95% CI 0.04, 0.89] in NHANES, 0.56 [0.52, 0.59] in SCREAM, 0.62 [0.60, 0.64] in Maccabi) (Fig. 2C). Hypertension and CHD were generally associated with greater use of SGLT2i. Patients with higher HbA_1c were more likely to use SGLT2i in SCREAM and Maccabi but less likely to use SGTL2i in NHANES. Similarly, patients with higher BMI were more likely to use SGLT2i in Maccabi but may have been less likely to use SGLT2i in NHANES. We did not find a consistent association of stroke, heart failure, or ACR with use of SGLT2i across cohorts.

Older age and lower eGFR were associated with lesser use of GLP1RA (Fig. 2D). We did not find a consistent association between cardiovascular disease and use of GLP1RA across cohorts: CHD, stroke, and heart failure were not associated with GLP1RA use in NHANES, while stroke and heart failure were associated with lesser use in SCREAM and Maccabi. Higher ACR was associated with greater use of GLP1RA in Maccabi (OR 1.14 [95% CI 1.09–1.19] for ACR 30–300 mg/g and 1.27 [1.17, 1.36] for ACR >300 mg/g) but not in NHANES or SCREAM. Higher BMI was associated with greater GLP1RA use in Maccabi (3.26 [3.13, 3.40] for BMI 30–39 kg/m² and 4.88 [4.54, 5.23] for BMI ≥40 kg/m² vs. BMI <30 kg/m² in Maccabi) but not associated with GLP1RA use in NHANES.

Across the cohorts, there was little difference in use of SGLT2i by risk of cardiorenal events (Fig. 3). Among patients with type 2 diabetes and heart failure, 7.5%, 6.8%, and 18.9% used SGLT2i in the U.S., Sweden, and Israel, respectively, compared with 4.7%, 9.4%, and 20.5% among patients with no guideline indication for SGLT2i. Use of SGLT2i was also low among patients with stroke and patients with eGFR <60 mL/min/1.73 m² (4.0% in NHANES, 5.7% in SCREAM, and 18.3% in Maccabi with stroke and 3.9% in NHANES, 11.0% in SCREAM, and 19.8% in Maccabi with eGFR <60 mL/min/1.73 m²). Similarly, there was little difference in use of GLP1RA by risk of cardiorenal events.

In this study from the U.S., Sweden, and Israel, we found substantial differences by country in real-world use of glucose-lowering medications and limited personalization of prescribing practices. The U.S. had lesser use of metformin, the first-line therapy, and persistently greater use of sulfonylureas than health systems in Sweden and Israel. Uptake of newer medications such as SGLT2i and GLP1RA was also slower in the U.S. Across all systems, however, there was little apparent personalization of medication prescription, with similar rates of SGLT2i or GLP1RA use by presence of cardiovascular disease or kidney disease.

Among traditional glucose-lowering medications, we observed lesser use of metformin and greater use of sulfonylureas in the U.S. compared with Israel. Metformin is recommended as first-line therapy and should be continued even after initiation of other therapies (including insulin) unless contraindicated or not tolerated (4–6). Although use of metformin increased over time in NHANES and SCREAM, use was lower compared with that in Maccabi. Interestingly, we found that use of metformin was lower among patients with cardiovascular disease in all the three cohorts. This was consistent with results of a previous study with use of data from multiple countries (22). The reason is unclear. Overall weak evidence of cardiovascular benefits of metformin may have contributed to this finding (23).

We also observed differential uptake of DPP4i, SGLT2i, and GLP1RA across the cohorts. Use of these newer glucose-lowering medications increased more rapidly in Maccabi than in NHANES and SCREAM. Although overall use of SGLT2i and GLP1RA increased across the three cohorts, limited personalization of medication use was observed. The ADA released revised 2018 guidelines, recommending empagliflozin and liraglutide for patients with type 2 diabetes and established atherosclerotic cardiovascular disease (level A evidence) (24). The 2018 guidelines also suggested that canagliflozin may be considered for these populations (level C evidence) (24). The results of our study showed that the presence of atherosclerotic cardiovascular disease and kidney disease was not consistently associated with greater use of SGLT2i or GLP1RA in any of the cohorts around this time. At the time, limited data were available: the BI 10773 (Empagliflozin) Cardiovascular Outcome Event Trial in Type 2 Diabetes Mellitus Patients (EMPA-REG OUTCOME) trial showed that empagliflozin reduces the risk of the primary cardiovascular outcome and the secondary renal outcome in 2016 (25,26), Trial to Evaluate Cardiovascular and Other Long-term Outcomes With Semaglutide in Subjects With Type 2 Diabetes (SUSTAIN-6) and Liraglutide Effect and Action in Diabetes: Evaluation of Cardiovascular Outcome Results (LEADER) showed primary cardiovascular outcome benefits with some renal benefit signals in albuminuria (27,28), and SGLT2i was contraindicated in patients with chronic kidney disease for various eGFR cut points: e.g., it was recommended that canagliflozin be avoided in the case of eGFR <45 mL/min/1.73 m² (29). This may explain the lesser use of these medications among patients with chronic kidney disease in our study. With cumulative evidence of the cardiorenal benefits of SGLT2i and GLP1RA, current clinical guidelines recommend these medications for patients with atherosclerotic cardiovascular disease, heart failure, or chronic kidney disease (5,6). Although the adoption of SGLT2i and GLP1RA increased after 2018, more recent data from 2019–2020 show that personalized use of SGLT2i or GLP1RA for patients at high risk remains limited (22,30). Our data add knowledge to the current literature, emphasizing the need for provider education and support in identifying and treating the individuals at highest risk.

Variation in medication use and differential uptake in new medications by country have previously been observed in Europe and other countries (7,31). Differences in medication use are likely multifactorial involving differences in country-specific guidelines, medical service delivery, and institutional and economic constraints on the use of specific medications or medication classes. During the study period of 2007–2018, guidelines from the ADA and European Association for the Study of Diabetes (EASD) for glucose-lowering medication use were broadly similar, with metformin recommended as first-line treatment and SGLT2i and GLP1RA recommended for those expected to benefit from prevention of adverse cardiac outcomes, starting in 2018 guidelines (24,32–35). Access to care and medication cost likely played important roles in the observed differences. Sweden and Israel provide universal health coverage, while the U.S. does not. In addition, U.S. patients have more limited access to specialty care than patients in other developed countries (36), which may limit the uptake of new medications, as patients are more likely to get new medications from specialty care. Prices for medications are substantially higher in the U.S. than other developed countries, with larger gaps for brand-name originator drugs (37). The persistently higher use of sulfonylureas in the U.S. may have been driven by the lower costs for sulfonylureas compared with other therapies, such as lower co-payments (38). In addition, insurance in the U.S. may require prior authorization for prescription of newer medications, and this increased administrative burden may deter adoption of newer medications (39).

Our study is among the first to systematically characterize glucose-lowering medication use by individual-level characteristics. Granular data from three different cohorts allowed us to comprehensively examine predictors of medication use and to report on consistencies across countries. Our study also has several limitations to be acknowledged. First, across the cohorts disparate data sources were used that may not be entirely comparable. Second, the SCREAM and Maccabi data sets, while large, may not be fully representative of the populations from each country. Third, medication use was captured through self-report in NHANES, prescription records in Maccabi, and pharmacy dispensation in SCREAM and may not always accurately reflect actual use. Fourth, data were available only to 2018 and our results are more reflective of early adoption of medications with cardiorenal benefits prior to the widespread guideline recommendations of SGLT2i and GLP1RA in populations at high risk. Indeed, the adoption of SGLT2i and GLP1RA increased as knowledge and evidence of the cardiorenal benefits of SGLT2i and GLP1RA have grown. However, limited personalization of therapy persists (22,30).

In conclusion, our study demonstrates large differences in real-world use of glucose-lowering medications across the U.S., Sweden, and Israel. Uptake of newer medications was faster in Israel than the other two countries, but use of SGLT2i and GLP1RA in individuals at high risk was limited in all three countries. Together with more recent data, these results suggest that additional education and support for guideline-recommended therapies are needed.

Funding. J.-J.C. acknowledges support from the Swedish Research Council (2019-01059). E.S. was supported by National Institutes of Health (NIH)/National Heart, Lung, and Blood Institute (NHLBI) grant K24HL152440. E.L.F. acknowledges support by a Rubicon Grant of the Netherlands Organisation for Scientific Research (NWO). M.E.G. and B.L. were supported by NIH/National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) grant R01DK115534, and M.E.G. was supported by NIH/NHLBI grant K24HL155861. J.-I.S. was supported by NIH/NIDDK grant K01DK121825.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. B.L. researched data and wrote the manuscript. Y.S. researched data and reviewed and edited the manuscript. E.S., A.R.C., G.C.A., C.M.C., J.C., V.S., G.C., A.K., J.-J.C., E.L.F., Y.X., and M.E.G. reviewed and edited the manuscript. J.-I.S. researched data and reviewed and edited the manuscript. B.L. and J.-I.S. were guarantors, taking responsibility for the integrity of the data and the accuracy of the data analysis.