Initiation of Continuous Glucose Monitoring Is Linked to Improved Glycemic Control and Fewer Clinical Events in Type 1 and Type 2 Diabetes in the Veterans Health Administration

Clinical trials and observational studies of continuous glucose monitoring (CGM) have most often been conducted in traditional type 1 diabetes (T1D) patient populations (i.e., those characterized by early onset of disease, autoimmune destruction of islets, and more rapid and complete dependence on insulin). Because most studies have been shorter in duration, they have focused on changes in HbA_1c or metrics such as time in range or time below range (1,2). Although some, but not all, studies have suggested potential reductions in the number of serious hypoglycemia and possibly hyperglycemia events, more data on these clinically relevant outcomes are clearly needed (1–4).

Despite a rapid expansion of CGM use in T2D, even less is known about the nature and duration of benefits of CGM use in this population. Although randomized clinical trials have shown trends toward improvement in HbA_1c, they provide less consistent evidence for reduction in time in hypoglycemia or hyperglycemia and minimal evidence of effects on clinically important hypoglycemia or hyperglycemia events (5–7). In a relatively large retrospective observational study of patients with T1D (with a smaller number of patients with T2D) within the Kaiser health system, Karter et al. (8) confirmed CGM-related improvement in HbA_1c, overall and in both groups separately, but reported more mixed results regarding hospitalization for hypoglycemia or hyperglycemia.

The cost of CGM devices along with the provision of technical and educational support to patients and providers to initiate and maintain broad CGM use in health care institutions is substantial. It is therefore of great importance to understand existing CGM use patterns among both patients with T1D and those with T2D and potential benefits regarding overall glycemic control and adverse events such as emergency room (ER) visits and hospitalizations.

The objective of this study was to use the large population of patients with T1D or T2D and their comprehensive medical records within the Veterans Health Administration (VHA) health care system to determine differences in glucose control and adverse diabetes outcomes over 12 months in those who initiated CGM compared with those using glucose strip–based monitoring. We also examined these outcomes by key baseline characteristics and CGM use patterns.

A retrospective observational cohort design was used to compare glucose control and several clinically relevant events over 12 months of follow-up. The protocol was approved by the Phoenix Veterans Affairs Health Care System Institutional Review Board, which provided a waiver of consent for this analysis of secondary data. We extracted all relevant electronic health record data from the VHA Corporate Data Warehouse, a national repository of clinical and administrative information from VHA encounters that includes both inpatient and outpatient visits, diagnoses, pharmacy and medication use, vital signs records, laboratory measurements, and general patient demographic information.

To identify patients with diabetes initiating CGM from 1 January 2015 to 31 December 2020, all available ICD-9 and ICD-10 diagnostic codes were extracted from medical encounters among U.S. veterans age ≥21 years with at least one diagnosis of diabetes (ICD-9 250 or ICD-10 E10, E11, O24.0×, or O24.1×) between 2002 and 2020. Initiators of CGM (also subsequently referred to as CGM users) were identified as those with their first prescription (defined as index date) between 1 January 2015 and 31 December 2020. CGM prescriptions were identified if they matched a glucose sensor for the devices available in the VHA during this time period (Dexcom, Freestyle Libre, and Medtronic). Initiation of specific CGM sensors for T1D and T2D were, respectively, Dexcom (50% and 34%), Freestyle Libre (10% and 63%), and Medtronic (40% and 3%).

All remaining patients with no record of a filled prescription for a CGM glucose sensor were selected as nonusers and randomly assigned an index date between 1 January 2015 and 31 December 2020 if they were prescribed blood glucose test strips for self-monitoring in the year before their assigned index date. Additional criteria for inclusion in the cohort (detailed in Supplementary Fig. 1) included classification as a patient with T1D or T2D and having sufficient baseline and follow-up participation in VHA care (defined below).

Identification of T1D and T2D followed a modified Klompas algorithm (9) (Supplementary Methods). Individuals included in the cohort had to have at least 2 years of preindex activity within the VHA, at least one clinical encounter in primary care or endocrinology in the year before their index date, and a postindex date interaction within the VHA. To achieve balanced propensity scores (PSs) in T2D, it was necessary to limit both CGM users and nonusers to those using insulin in the year before the index date but not receiving insulin pumps. Because diabetes is an exclusion criterion for enrollment in the military, T1D in the current cohort reflects adult-onset disease.

We used ICD-9 Clinical Modification and ICD-10 Clinical Modification codes to identify prevalent medical conditions or outcomes. Visit types were identified by associated clinic stop codes and were linked with dates, services provided, and diagnoses. Laboratory measurements were captured using LOINC (Logical Observation Identifiers Names and Codes) codes. Calculation of proportion of days covered (PDC) by medication refills was used as an estimate of duration and consistency of medication use.

We extracted outpatient medical records for visits, demographics, medications, diagnoses, and general laboratory measures in the 12 months before the index date. However, baseline HbA_1c was selected as the HbA_1c value closest to the index date (within 6 months), and information for comorbidity indices was collected up to 24 months preindex date. Inpatient admissions or ER visits were identified from inpatient admission records.

A large number of demographic and clinical variables were considered to build PSs for CGM users and nonusers (highlighted in Tables 1 and 2 and Supplementary Methods).

Primary outcomes included 1) glycemic control (change in HbA_1c at 6 and 12 months after the index date), 2) hypoglycemia events (the first postindex date occurrence of an ER or hospital admission if hypoglycemia was listed as one of the diagnostic codes), 3) hyperglycemia events (defined similarly as hypoglycemia), and 4) all-cause hospitalization (the first postindex date occurrence of any inpatient admission). We used a 12-month postindex date observation window to assess the change in HbA_1c and the onset of these outcomes. A secondary outcome included the change in the percentage of patients achieving HbA_1c values below 7%, 8%, and 9% during this time window.

Negative control outcomes assessed potential bias resulting from possible residual confounding factors between CGM users and nonusers (10). They share the same potential sources of bias as primary outcomes but are less plausibly related to CGM use. Two negative control outcomes were considered in outpatient and inpatient settings: 1) specific musculoskeletal diagnoses associated with back, joint, and soft tissue disorders or injuries and 2) upper respiratory infections.

Baseline characteristics are presented as mean (SD), median (interquartile range), or percentage. To handle missing data, multiple imputation by chained equations (11) was performed to create five imputed data sets, which were subsequently combined.

Overlap PS weighting was used to balance baseline covariates between CGM users and nonusers. We calculated PSs by logistic regression models that predicted CGM initiation using >40 candidate predictors (Tables 1 and 2 and Supplementary Methods). All PS models demonstrated excellent discrimination (C statistic ≥0.93 for each). To improve balance between CGM users and nonusers in T2D, we applied downsampling and randomly selected 10% of the nonusers for comparison.

Overlap weights equaled 1 − PS for CGM users and PS for nonusers. Covariate differences before and after overlap weighting were assessed using standardized mean difference (SMD). An absolute SMD value of <0.1 was considered a negligible group imbalance.

We compared the changes in HbA_1c between CGM users and nonusers using linear mixed models (LMMs). We compared the percentage of individuals achieving HbA_1c targets between CGM users and nonusers using generalized estimating equations (GEEs). Difference in differences was then calculated, and the Δ method was used to derive CIs (12).

Our primary clinical outcomes and negative control events were analyzed using Cox proportional hazards models with overlap weighting. Individuals were censored if they were lost to follow-up, died, or had no event in the 12-month follow-up period. The proportionality of all model predictors was confirmed.

Sensitivity analyses were conducted using variations in outcomes and definitions of hypoglycemia and hyperglycemia events and within a subpopulation characterized by narrower inclusion criteria to increase the likelihood of reasonable health and VHA participation at baseline (Supplementary Results). We also tested excluding any follow-up after March 2020 to remove COVID-19 pandemic influences on outcomes. To demonstrate the validity of the PS approach, we also compared CGM effects between groups using Cox proportional hazards models while simply adjusting for key differences. To determine if insulin pump use modified the effects of CGM within T1D, we tested for an interaction between CGM use and insulin pump use and performed analyses excluding insulin pump users to determine the effects of CGM in their absence.

Stratified subgroup analyses within CGM users were conducted based on baseline age, HbA_1c, hypoglycemic risk score, and PDC calculated using CGM sensor fills. LMMs and GEE models were used for these subgroup analyses to compare outcomes pre- and postindex dates. To assess the dose-response effect of CGM use, we assessed the effects of CGM across increments of 25% in PDC.

All statistical analyses were performed using R version 4.1.2 (https://www.r-project.org). The R package MatchThem (13) was used for calculating PSs; survival was used for Cox proportional hazards analyses; geepack was used for GEEs (14,15); lmerTest (16) was used for LMM analyses. A two-sided test with P < 0.05 was considered statistically significant, and we did not adjust for multiple testing.

Individual patient data are protected by the federal government and Veterans Affairs, and the data are not available for sharing. Detailed information on the study protocol and code is available upon request from the corresponding author.

Supplementary Figure 1 illustrates the selection of patients into the T1D (Supplementary Fig. 1A) and T2D (Supplementary Fig. 1B) cohorts. Applying exclusion criteria (as outlined in Research Design and Methods) resulted in 5,015 CGM initiators and 3,815 nonusers with T1D. Applying similar criteria among nonusers with T2D, and after downsampling to 10%, 15,706 CGM initiators and 29,912 nonusers with T2D were included in the same period. At least 1 year of follow-up within the VHA was available for >90% of patients with T1D and T2D.

Tables 1 and 2 show the baseline characteristics of the T1D and T2D cohorts before and after PS overlap weighting. Initiation of CGM for T1D increased over the 5-year interval from 2015 to 2020. Compared with nonusers with T1D, CGM users tended to be younger and less frequently male, have more frequent visits with primary care and endocrinology providers, and have greater use of insulin pumps and glucagon (Table 1). African American and Hispanic American groups had lower percentages of CGM users than nonusers, whereas the percentage of CGM users was higher for White patients with T1D. General indicators of health, such as indices of hypoglycemia risk, comorbidities, and diabetes complications, were relatively similar between groups. After overlap weighting, baseline characteristics demonstrated small SMDs (<0.1), indicating well-matched groups (Table 1 and Supplementary Fig. 2A).

In comparison with those with T1D, patients with T2D (Table 2) were on average older and more obese, used far more noninsulin diabetes medications and less glucagon, and had lower rates of hypoglycemia. Initiation of CGM in T2D increased dramatically from 2015 through 2020. Similar disparities, as noted in T1D, were seen in percentages of CGM users among racial/ethnic groups. CGM users (vs. nonusers) tended to have more clinical visits, higher HbA_1c values, lower estimated glomerular filtration rate, greater use of insulin and glucagon, greater use of other diabetes medications, more visits to ERs, and higher risk and comorbidity scores. After PS overlap weighting, baseline features demonstrated small SMDs (<0.1) between CGM initiators and nonusers (Table 2 and Supplementary Fig. 2B).

Declines in HbA_1c (differences in difference) were significantly greater in CGM users with T1D compared with nonusers at 6 (−0.26%; 95% CI −0.31, −0.21%) and 12 (−0.26%; 95% CI −0.33, −0.19%) months (Table 3). The percentages of patients with T1D achieving HbA_1c <8 and <9% after 12 months were greater in CGM users, with a nearly 10–percentage point increase among CGM users achieving a <9% target.

The declines in HbA1c were also significantly greater in CGM users with T2D compared with nonusers at 6 (−0.39%; 95% CI −0.42, −0.36%) and 12 (−0.35%; 95% CI −0.40, −0.31%) months (Table 3). The percentages of patients with T2D achieving <8 and <9% after 12 months were greater in CGM users than in nonusers. Of note, declines in HbA_1c in the smaller subsets of women with T1D or T2D were in line with overall cohort results (Supplementary Table 1).

In patients with T1D, CGM initiation was associated with a significantly reduced risk of hypoglycemia (Table 4), with a hazard ratio (HR) of 0.69 (95% CI 0.48, 0.98). If a hypoglycemia event included an outpatient blood glucose level <54 mg/dL, this difference remained significant (HR 0.72; 95% CI 0.57, 0.91). Risk reduction for admission for hyperglycemia was not statistically significant (HR 0.83; 95% CI 0.65, 1.06). All-cause hospitalization was significantly reduced in CGM users (HR 0.75; 95% CI 0.63, 0.90).

In patients with T2D (Table 4), no difference in risk of admission for hypoglycemia was seen between CGM users and nonusers, but there was a reduction in risk of hyperglycemia in CGM users (HR 0.87; 95% CI 0.77, 0.99). The risk of all-cause hospitalization was reduced in CGM users (HR 0.89; 95% CI 0.83, 0.97).

For T1D and T2D, sensitivity analyses (Supplementary Table 2) excluding less healthy individuals, censoring those whose follow-up included years in which COVID was active, or using alternative definitions of hypoglycemia and hyperglycemia (Supplementary Table 3) or excluding patients with T1D using insulin pumps at baseline resulted in reduced cohort sizes but did not change the pattern of results meaningfully. In T1D, we found no significant interaction (at P value <0.05) between CGM use and insulin pump use for any outcome.

For T1D and T2D, neither outpatient nor inpatient diagnosis of a musculoskeletal disorder or upper respiratory infection (as negative control) indicated there was a bias favoring CGM users resulting from unmeasured confounding factors (Supplementary Table 4).

As shown in Supplementary Figure 3A–D, certain subsets seemed to gain more benefit over 6 to 12 months of CGM use. In patients with T1D, both declines in HbA_1c and risks of admission for hyperglycemia and all-cause hospitalization were greater in those age <65 than in those age ≥65 years. Declines in HbA_1c, but not admissions for hypoglycemia, hyperglycemia, or any causes, were greater in those with higher baseline HbA_1c values (vs. those with lower HbA_1c values). In contrast, in those with intermediate or high risk of hypoglycemia (vs. low risk) at baseline, declines in HbA_1c were similar, but declines in risk of admissions for hyperglycemia, hypoglycemia, and all causes were significantly greater. Those with excellent compliance with CGM throughout the 12 months, defined as a PDC score of ≥0.80, demonstrated greater declines in HbA_1c as well as in risk of admissions for hypoglycemia and hyperglycemia. In T2D, generally very similar patterns of response to CGM initiation were seen in these same subsets (Supplementary Figure 3A–D). Patients with T2D with excellent compliance with CGM throughout the 12 months showed greater improvements in all outcomes. As shown in Supplementary Table 5, declines in HbA_1c increased across increments of PDC.

This study indicates, in a large national cohort, that initiation of CGM was associated with meaningful improvements in HbA_1c in both patients with T1D and those with T2D using insulin. Importantly, these improvements in glycemic control were accompanied by a general trend toward reduced risk of admission to an ER or hospitalization for hypoglycemia or hyperglycemia. There was an even more consistent reduction in risk of all-cause hospitalization in both T1D and T2D.

The declines in HbA_1c (−0.26 to −0.39%) are in line with those seen in most randomized clinical trials of real-time CGM in T1D (1,2). Some variation in published results in T2D likely results from heterogeneous medication regimens and initial HbA_1c levels in the cohorts studied. In one relatively large and well-conducted trial of CGM in patients with T2D using basal insulin and with HbA_1c levels >8% (17), CGM use was associated with a 0.4% greater decline in HbA_1c.

However, most randomized clinical trials of CGM effects were shorter in duration (3–6 months), and, therefore, the current results provide a more intermediate-/longer-term assessment of the effects of CGM on glycemic control. Finding similar declines in HbA_1c values at 6 and 12 months indicates that the benefit of initiating CGM seems relatively stable for at least 12 months. This is consistent with the persistently lower HbA_1c in T1D after 16 months of follow-up reported in the SILVER (Sustained Intensive Treatment and Long-Term Effects on HbA1c Reduction) trial after randomization to CGM for 26 weeks (18). CGM initiation in the current study also increased the percentage of individuals achieving HbA_1c values of <8 and <9% in both T1D and T2D at 12 months, targets at or below the inflection point for greater risk of diabetes complications (19–22). Because baseline HbA_1c levels were relatively similar between patients with T1D and those with T2D, the greater decline in T2D certainly highlights that CGM use in this group can also be quite effective.

The large size of these cohorts allowed us to compare effects of CGM initiation on changes in HbA_1c among unique subsets of individuals. Individuals with T1D or T2D who were younger (age <65 years), had higher baseline HbA_1c levels, or had excellent CGM compliance (PDC >0.8) all demonstrated greater reductions in HbA_1c. Because the rationale for starting CGM in those with high hypoglycemic risk is primarily to prevent additional low-glucose episodes, it is not surprising that this baseline determinant did not influence change in HbA_1c value in either T1D or T2D.

The decline in HbA_1c associated with CGM initiation seemed clinically meaningful, particularly in patients with T2D, in whom changes approached 0.4%. Moreover, some subsets of individuals, such as those with higher baseline HbA_1c values, showed robust declines in HbA_1c ranging from −0.71% in those with T1D to −1.14% in those with T2D.

Reductions in risk of ER or hospital admission for hyperglycemia, hypoglycemia, or any cause ranged from 17 to 30% in T1D. Overall, these results suggest a fairly robust effect of CGM initiation in T1D to reduce relatively common but potentially serious diabetes complications that can increase the need for costly ER and hospital admissions.

Generally similar, but less overall favorable, patterns for clinical events were seen in CGM initiators with T2D. Although risk of admission related to hypoglycemia was not reduced, the risk of hyperglycemia declined. And again, CGM initiation in T2D was linked to a significant 11% decline in risk of all-cause hospitalization. These results raise the possibility that decreases in health care use could offset the excess costs of supporting CGM initiation and use within appropriate patient groups.

As seen with declines in HbA_1c, certain subgroups of CGM users seemed to achieve greater risk reduction for ER or hospitalization admissions from initiation of CGM. However, this did not follow the same patterns as those seen for glycemic control. Indeed, for several subgroups, admission for hypoglycemia or hyperglycemia or all-cause hospitalization did not appear related to changes in glycemic control. For example, those with higher baseline hypoglycemia risk, who had shown little difference in change in HbA_1c values versus those at lower hypoglycemia risk, had lower risk of all three clinical outcomes. This further highlights the potential value of initiating CGM in those with recurrent hypoglycemia. More consistent use of CGM during the 12 months of follow-up (higher PDC) was also an important determinant of admissions related to hypoglycemia and hyperglycemia, with less consistent trends toward decreases in risk of all-cause hospitalization for T1D and T2D. This relationship between PDC and hospitalization suggests efforts to educate patients about consistent use of CGM may have a positive impact on reducing diabetes-related hospitalizations.

We speculated that CGM benefits may be an indirect function of increased frequency of health care interactions, but several factors suggest that this is not the case. First, effects of CGM on glycemic control were in line with results from randomized clinical trials. Second, compliance with CGM, as measured by PDC, was directly related to improvements in both glycemic control and outcomes. Third, although we recognize that simple visit frequency may not capture the potential benefit of health care visits for all patients, accounting for any excess visits postinitiation of CGM in analysis models did not seem to be an important mediator of CGM effects on change in HbA_1c (Supplementary Materials). Overall, these findings suggest that it is the direct effects of CGM (in comparison with self-monitoring of blood glucose) to provide better information about glucose levels and facilitate shared decision making between patients and providers about diabetes management that improve outcomes.

Although the VHA has a relatively diverse racial and ethnic patient population and growing female enrollment, the number of women included in our cohorts was low. However, CGM-associated changes in HbA_1c in women were consistent with overall results. We also excluded patients with T2D not receiving insulin, limiting generalizability to the broader T2D population. Although overlap weighting provided excellent balancing of characteristics between groups, there remains the possibility that CGM users and nonusers may not have been well matched on variables not considered. We also acknowledge that ICD codes associated with hospital visits may not always reliably identify hypoglycemia- and hyperglycemia-related admissions. However, several sensitivity analyses using alternative definitions, including more objective laboratory data, did not meaningfully alter the results. The current study included CGM users who met inclusion criteria across a 6-year period (2015–2020), which likely resulted in patients using multiple types of CGM devices during a period in which newer CGM models have improved accuracy and usability compared with older models. CGM devices also vary in their frequency of glucose measurements, need for calibration, ability to provide alerts at high or low glucose levels, and other attributes. Our analyses do not provide insight into the potential effects these different types of CGM devices or features may have on outcomes.

The VHA cohorts of CGM users and nonusers are the largest reported to date worldwide and permitted examination of CGM effects on not just changes in HbA_1c but also clinically relevant outcomes such as ER visits and hospitalizations. This also allowed study of two groups, those with T1D developing in adulthood and those with T2D, who have been less commonly evaluated, thus providing novel and complementary information to previously published studies. Importantly, the current findings demonstrate that benefits in these groups were generally similar to those reported in the more frequently studied childhood-onset T1D. We also examined key subgroups of CGM users, demonstrating the types of benefits that might be anticipated based on baseline characteristics and with consistent use of CGM. An additional strength was the number of additional sensitivity analyses conducted to exclude potential confounding explanations for the associations found between CGM and outcomes. Importantly, several of the study results were consistent with those from a retrospective observational analysis conducted within Kaiser Permanente suggesting consistency of CGM effects across large health care systems (8). Because the VHA health care system provides relatively comprehensive and inexpensive medical care for all enrolled veterans, access to providers and devices such as CGM and the resulting outcomes should be less influenced by disparities in finances or health care availability.

In conclusion, this large real-world study demonstrated CGM initiation in T1D and T2D was linked to clinically meaningful and sustained improvements in glucose control and reductions in risk of hypoglycemic and hyperglycemic events contributing to ER or hospital admissions as well as to all-cause hospitalizations. Study results also highlight that patient baseline characteristics and consistency of CGM use seem important determinants of the types of outcomes most likely to improve with initiation of CGM.

Funding. This work was supported using resources and facilities of the Department of Veterans Affairs (VA), Phoenix VA Health Care System, and VA Informatics and Computing Infrastructure VA Health Services Research and Development 13-457. J.J.Z. was partly supported by National Institutes of Health grants R01HG006139 and R21HL150374 and National Science Foundation grants DMS-2054253 and IIS-2205441.

The contents of this article do not represent the views of the Department of VA or the U.S. Government.

Duality of Interest. This work was supported by pilot funding from Dexcom, Inc. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. P.D.R., M.N., X.Z., and J.J.Z. had full access to all data. P.D.R., M.N., and J.J.Z. drafted the first manuscript. P.D.R. and J.J.Z. were responsible for the study concept and design. M.N., S.R., X.Z., and J.J.Z. prepared and contributed to the data extraction and analysis. All authors provided input on the statistical analysis and interpretation of results, and all authors revised the manuscript and approved the final version. M.N. 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.

Prior Presentation. Preliminary parts of this study were presented in abstract form as a poster presentation at the 82nd Scientific Sessions of the American Diabetes Association, New Orleans, 3–7 June 2022.