Key Points
Question
Are there clinical benefits associated with real-time continuous glucose monitoring (CGM) among patients with insulin-treated diabetes?
Findings
In this retrospective cohort study in a usual care setting that included 5673 patients with type 1 diabetes and 36 080 patients with type 2 diabetes, use of real-time CGM compared with nonuse was associated with significantly lower hemoglobin A1c (difference, −0.40%) and lower rates of emergency department visits or hospitalizations for hypoglycemia (difference, −2.73%) but no significant difference for rates of emergency department visits or hospitalizations for hyperglycemia or for other reasons.
Meaning
Among patients selected by physicians for real-time continuous glucose monitoring use was associated with better glycemic control and lower rates of hypoglycemia.
Abstract
Importance
Continuous glucose monitoring (CGM) is recommended for patients with type 1 diabetes; observational evidence for CGM in patients with insulin-treated type 2 diabetes is lacking.
Objective
To estimate clinical outcomes of real-time CGM initiation.
Design, Setting, and Participants
Exploratory retrospective cohort study of changes in outcomes associated with real-time CGM initiation, estimated using a difference-in-differences analysis. A total of 41 753 participants with insulin-treated diabetes (5673 type 1; 36 080 type 2) receiving care from a Northern California integrated health care delivery system (2014-2019), being treated with insulin, self-monitoring their blood glucose levels, and having no prior CGM use were included.
Exposures
Initiation vs noninitiation of real-time CGM (reference group).
Main Outcomes and Measures
Ten end points measured during the 12 months before and 12 months after baseline: hemoglobin A1c (HbA1c); hypoglycemia (emergency department or hospital utilization); hyperglycemia (emergency department or hospital utilization); HbA1c levels lower than 7%, lower than 8%, and higher than 9%; 1 emergency department encounter or more for any reason; 1 hospitalization or more for any reason; and number of outpatient visits and telephone visits.
Results
The real-time CGM initiators included 3806 patients (mean age, 42.4 years [SD, 19.9 years]; 51% female; 91% type 1, 9% type 2); the noninitiators included 37 947 patients (mean age, 63.4 years [SD, 13.4 years]; 49% female; 6% type 1, 94% type 2). The prebaseline mean HbA1c was lower among real-time CGM initiators than among noninitiators, but real-time CGM initiators had higher prebaseline rates of hypoglycemia and hyperglycemia. Mean HbA1c declined among real-time CGM initiators from 8.17% to 7.76% and from 8.28% to 8.19% among noninitiators (adjusted difference-in-differences estimate, −0.40%; 95% CI, −0.48% to −0.32%; P < .001). Hypoglycemia rates declined among real-time CGM initiators from 5.1% to 3.0% and increased among noninitiators from 1.9% to 2.3% (difference-in-differences estimate, −2.7%; 95% CI, −4.4% to −1.1%; P = .001). There were also statistically significant differences in the adjusted net changes in the proportion of patients with HbA1c lower than 7% (adjusted difference-in-differences estimate, 9.6%; 95% CI, 7.1% to 12.2%; P < .001), lower than 8% (adjusted difference-in-differences estimate, 13.1%; 95% CI, 10.2% to 16.1%; P < .001), and higher than 9% (adjusted difference-in-differences estimate, −7.1%; 95% CI, −9.5% to −4.6%; P < .001) and in the number of outpatient visits (adjusted difference-in-differences estimate, −0.4; 95% CI, −0.6 to −0.2; P < .001) and telephone visits (adjusted difference-in-differences estimate, 1.1; 95% CI, 0.8 to 1.4; P < .001). Initiation of real-time CGM was not associated with statistically significant changes in rates of hyperglycemia, emergency department visits for any reason, or hospitalizations for any reason.
Conclusions and Relevance
In this retrospective cohort study, insulin-treated patients with diabetes selected by physicians for real-time continuous glucose monitoring compared with noninitiators had significant improvements in hemoglobin A1c and reductions in emergency department visits and hospitalizations for hypoglycemia, but no significant change in emergency department visits or hospitalizations for hyperglycemia or for any reason. Because of the observational study design, findings may have been susceptible to selection bias.
This cohort study investigates the effect of real-time continuous glucose monitoring on glycemic control among patients with insulin-treated diabetes.
Introduction
Self-monitoring of blood glucose, which requires fingerstick pricks to obtain a small blood sample for testing, is used by patients with diabetes to intermittently measure blood glucose levels. The development of continuous glucose monitoring (CGM) devices represents a significant advance in convenient, data-driven, diabetes management. A real-time CGM device uses subcutaneous sensors to measure interstitial fluid glucose and provides feedback to the patient about glucose levels and trends. The real-time CGM data can inform self-management (eg, insulin dosing), provide alerts (eg, warnings of hypoglycemia), and can be shared with care teams to improve safety and optimize management. Clinical trials involving patients with type 1 diabetes have demonstrated that real-time CGM improves glycemic control, reduces hypoglycemia risk, and is cost-effective compared with self-monitoring of blood glucose.1,2,3,4 However, less evidence exists regarding the effectiveness of real-time CGM in patients with insulin-treated type 2 diabetes.
Current guidelines recommend CGM for patients with type 1 diabetes or type 2 diabetes who administer multiple daily insulin injections or use an insulin pump (grade A evidence); grade C evidence supports broader CGM use (eg, patients using basal insulin only).5 Therapeutic CGM devices (ie, sufficiently reliable for making insulin adjustments) are covered by Medicare (Centers for Medicare & Medicaid [CMS] Ruling 1682R) for patients with type 2 diabetes who meet strict criteria (eg, self-monitoring of blood glucose ≥4 daily; ≥3 daily insulin injections or insulin pump). Although some trials involving patients with type 2 diabetes have shown a glycemic benefit of CGM use, the effect on hypoglycemia is less clear.2,6 Although clinical trials provide evidence of real-time CGM efficacy, there is a need for estimates of effectiveness from observational studies. This study used a difference-in-differences framework to estimate clinical outcomes associated with real-time CGM initiation among patients with insulin-treated diabetes.
Methods
Study Population
Kaiser Permanente Northern California is an integrated health care delivery system with more than 4.4 million members and a diabetes registry established in 19947 that has included 668 854 patients cumulatively through 2018 with mean follow-up of more than 20 years. Because this study involved secondary research of identifiable private information for which consent is not required, the Kaiser Permanente Northern California Institutional Review Board determined this study to be exempt.
Participants
Patients were identified as real-time CGM initiators if they had a durable medical equipment vendor claim for real-time CGM supplies during January 1, 2015, through December 31, 2019 (Table 1). Baseline for follow-up was defined by the date of the first real-time CGM claim or, for patients without a real-time CGM claim (noninitiators; reference group), a randomly assigned date during the same time frame. Patients whose first claim occurred in 2019 were excluded because they would not have been able to accumulate 12 months of follow-up prior to the end of the study. Patients were included if they maintained continuous health plan membership for 12 months before and after baseline; had pharmacy benefits during the 12 months before baseline; had at least 1 pharmacy dispensing of glucose monitoring test strips during the 12 months before baseline; had at least 1 pharmacy dispensing of insulin during the 6 months before baseline; had not used CGM between January 1, 2014, and baseline; and were diagnosed with type 1 or type 2 diabetes (Figure 1).
Table 1. Baseline Characteristics Before and After Propensity Score Overlap Weighting in 5673 Patients With Type 1 Diabetes, January 1, 2015, Through December 31, 2018.
| Variables | Crude (unweighted)a | Standardized differencec | After overlap weightinga,b | Standardized differencec | ||
|---|---|---|---|---|---|---|
| Real-time CGM initiators (n = 3462) | Noninitiators (n = 2211) | Real-time CGM initiators (n = 3462) | Noninitiators (n = 2211) | |||
| Male | 48.2 | 57.8 | −0.19 | 52.7 | 54.2 | −0.03 |
| Female | 51.8 | 42.2 | 0.19 | 47.3 | 45.8 | 0.03 |
| Age, mean (SD), y | 40.7 (19.6) | 45.2 (18.7) | −0.23 | 42.8 (19.3) | 42.8 (19.0) | 0.00 |
| Age at diabetes onset, mean (SD), y | 20.8 (15.0) | 24.7 (15.8) | −0.26 | 22.1 (15.3) | 22.2 (15.3) | −0.01 |
| No. | 2442 | 1574 | 2442 | 1574 | ||
| Diabetes duration, mean (SD), y | 18.4 (15.5) | 19.9 (15.3) | −0.10 | 19.5 (15.7) | 19.4 (15.7) | 0.01 |
| No. | 2442 | 1574 | 2442 | 1574 | ||
| Race/ethnicityd | ||||||
| White | 74.0 | 66.4 | 0.16 | 70.3 | 71.4 | −0.02 |
| Hispanic | 8.8 | 12.8 | −0.13 | .9 | 9.9 | 0.03 |
| Black | 5.2 | 8.2 | −0.12 | 6.6 | 6.9 | −0.01 |
| Asian | 5.7 | 6.9 | −0.05 | 6.0 | 5.9 | 0.01 |
| Multiracial/multiethnic | 5.7 | 6.9 | −0.05 | 6.0 | 5.9 | 0.01 |
| Other | 0.6 | 1.1 | −0.06 | 1.1 | 0.7 | 0.03 |
| Unknown | 1.6 | 0.9 | 0.06 | 1.3 | 1.4 | −0.01 |
| Preferred spoken language not English8 | 1.1 | 2.2 | −0.09 | 1.7 | 1.2 | 0.05 |
| Neighborhood deprivation indexe | ||||||
| No. | 3430 | 2180 | 3430 | 2180 | ||
| Q1 (least deprived) | 29.3 | 22.7 | 0.15 | 25.6 | 27.0 | −0.03 |
| Q2 | 32.5 | 30.2 | 0.05 | 31.1 | 30.5 | 0.01 |
| Q3 | 25.1 | 28.6 | −0.08 | 26.8 | 27.4 | −0.01 |
| Q4 (most deprived) | 12.2 | 17.1 | −0.14 | 15.4 | 14.0 | 0.04 |
| Insurance typef | ||||||
| Commercial | 92.4 | 90.8 | 0.06 | 90.9 | 90.8 | 0.01 |
| Medicare Advantage | 17.2 | 22.0 | −0.12 | 20.1 | 20.0 | −0.01 |
| Medicaid | 8.6 | .4 | −0.06 | 10.0 | 10.5 | 0 |
| Charlson Comorbidity scoreg | ||||||
| 1 | 46.6 | 45.4 | 0.02 | 45.1 | 45.4 | −0.01 |
| 2 | 33.7 | 28.7 | 0.11 | 32.2 | 32.2 | 0 |
| 3 | 6.0 | 5.4 | 0.03 | 5.5 | 5.8 | −0.01 |
| ≥4 | 13.7 | 20.5 | −0.18 | 17.2 | 16.7 | 0.01 |
| Hypoglycemia risk scoreh | ||||||
| Low | 70.8 | 66.8 | 0.09 | 68.5 | 68.5 | 0 |
| Intermediate | 7.9 | 10.2 | −0.08 | 8.8.7 | 8.2 | 0.02 |
| High | 21.4 | 23.0 | −0.04 | 22.7 | 23.4 | −0.02 |
| Insulin type | ||||||
| Long acting | 62.0 | 72.6 | −0.23 | 74.0 | 75.6 | −0.04 |
| NPHi | 4.5 | 16.7 | −0.41 | 6.6 | 6.6 | 0 |
| Rapid acting | 98.6 | 84.9 | 0.51 | 95.9 | 97.8 | −0.11 |
| Short acting | 3.0 | 14.0 | −0.40 | 5.0 | 4.8 | 0.01 |
| Mixed | 0.3 | 2.6 | −0.20 | 0.5 | 0.6 | −0.02 |
| Glucagonj | 35.5 | 13.2 | 0.54 | 21.6 | 21.6 | 0 |
| Insulin delivery | ||||||
| Pen | 27.4 | 39.3 | −0.25 | 40.3 | 42.6 | −0.05 |
| Pumpk | 57.3 | 15.0 | 0.98 | 30.6 | 30.6 | 0 |
| ≥3 Insulin injections/d | 7.7 | 17.0 | −0.28 | 13.1 | 12.2 | 0.03 |
| Self-monitoring of blood glucose | ||||||
| <1/d | 0.9 | 4.9 | −0.24 | 1.6 | 1.4 | 0.01 |
| ≥1/d | 99.1 | 95.1 | 0.24 | 98.4 | 98.6 | −0.01 |
| HbA1c, mean (SD), %l | 8.16 (1.5) | 8.40 (1.8) | −0.15 | 8.22 (1.6) | 8.33 (1.8) | −0.06 |
| No. | 3114 | 1910 | 3114 | 1910 | ||
| Hypoglycemic eventm | 4.9 | 3.4 | 0.07 | 4.4 | 4.2 | 0.01 |
| Hyperglycemic eventm | 4.8 | 4.8 | −0.003 | 4.7 | 5.2 | −0.02 |
| Any ED visitn | 29.5 | 29.1 | 0.01 | 29.3 | 29.0 | 0.01 |
| Any hospitalizationn | 4.6 | 5.9 | −0.06 | 5.1 | 5.9 | −0.04 |
| No. of visits, mean (SD)n | ||||||
| Outpatient | 5.0 (3.7) | 2.8 (3.0) | 0.66 | 3.8 (3.6) | 3.8 (3.6) | 0 |
| Telephone | 4.2 (5.5) | 3.3 (4.6) | 0.18 | 4.0 (5.6) | 3.5 (4.8) | 0.09 |
Abbreviations: CGM, continuous glucose monitoring; ED, emergency department; HbA1c, hemoglobin A1c; NPH, neutral protamine Hagedorn; Q, quartile.
Column percent unless otherwise specified.
Propensity score models were based only on the variables included in eTable 1 in the Supplement.
The standardized difference (d) compares characteristics for real-time CGM initiators with noninitiators (reference group). An absolute value of d of 0.1 or lower indicates a negligible difference in the mean or in the prevalence of a covariate between groups.9
Self-reported race/ethnicity from the electronic health record. Multiracial/multiethnic means more than 1 of the racial/ethnic categories was selected.
The neighborhood deprivation index10 is a validated, contextual measure of socioeconomic status derived by linking each participant’s geocoded residential address to census tract level socioeconomic indicators as reported in the 2010 American Community Survey.
Kaiser Foundation Health Plan is the provider of all insurance and does not offer Medicare fee-for-service; insurance types are not mutually exclusive.
Based on the modified version of the Deyo Charlson Index score11 using inpatient and outpatient diagnosis and procedure codes. Possible scores ranged from 0 to 20 and represent the number of comorbid conditions identified (excluding diabetes) during the 12 months before baseline.
This validated risk stratification score12 predicts the 12-month risk of hypoglycemia-related utilization using electronic medical records data; the score was categorized as low (<1%), intermediate (1%-5%), and high (>5%).
Insulin types are not mutually exclusive.
Glucagon dispensed from a pharmacy may be an indicator of higher risk of a hypoglycemic event.
Users of both pen and pump were categorized as pump users.
Last HbA1c recorded during the 12 months before baseline. HbA1c was included in propensity score models used for outcomes other than HbA1c, eg, hypoglycemia and hyperglycemia. It appears both in this Table as a baseline measure and as an outcome measure. The same is true of hypoglycemic and hyperglycemic events, ED encounters, hospitalizations, and outpatient visits.
Identified by a principal diagnosis in the inpatient setting or a primary diagnosis in the emergency department during the 12 months before baseline.
Utilization captured during the 12 months before baseline.
Figure 1. Selection and Exclusion of Patients for This Studya .
aPatients were identified as initiators of real-time continuous glucose monitoring (CGM) if they had a durable medical equipment vendor claim for CGM supplies during January 1, 2015, through December 31, 2019. Baseline for follow-up was defined by the date of the first real-time CGM claim or, for noninitiators, a randomly assigned date during the same time frame. Patients with baselines in 2019 were excluded because they lacked 12 months of follow-up prior to the end of the study.
Outcomes
The outcomes included hemoglobin A1c (HbA1c) and emergency department visits or hospital utilizations for hypoglycemia or hyperglycemia. For each participant, the last HbA1c test result during the 12 months before and after baseline were identified. Hypoglycemia or hyperglycemia events were defined by a primary diagnosis from an emergency department encounter or a principal diagnosis from an inpatient admission using International Classification of Diseases, Ninth Revision, Clinical Modification codes (through September 2015) or ICD-10-CM codes (after October 2015). Rates were calculated as the proportion of patients having 1 or more event during the 12 months before and after baseline. The outcomes also included the proportion of participants having HbA1c levels lower than 7%, lower than 8%, or higher than 9% (Healthcare Effectiveness Data and Information Set performance measures), the proportion having 1 or more emergency department visit or hospitalization for any reason, and the number of outpatient or telephone visits.
Exposures and Covariates
Covariates included baseline year; sex; age; self-reported race/ethnicity (based on fixed categories); self-reported preferred spoken language; neighborhood deprivation index10; insurance type; predicted risk of hypoglycemia based on a validated hypoglycemia risk score12; glucagon dispensing from the pharmacy; Deyo Charlson Comorbidity Index score11; indicators for insulin pen, pump, and multiple daily injections; insulin type (neutral protamine Hagedorn or long-, short-, rapid-acting, or mixed insulin), use of noninsulin glucose-lowering medications, age at diabetes diagnosis, and duration since first diabetes diagnosis. Race/ethnicity and preferred spoken language were included as potential confounders given the historically lower uptake of new technology, poorer adherence, and poorer diabetes-related outcomes among some minority groups or among individuals whose preferred spoken language is not English.
Statistical Methods
This was a difference-in-differences study based on a retrospective cohort design to estimate the clinical outcomes of real-time CGM initiation compared with self-monitoring of blood glucose use alone (active comparator reference group).
Difference-in-differences estimates were calculated as the change in values before and after baseline (eg, HbA1c during the 12 months after baseline minus the HbA1c during the 12 months before baseline) among real-time CGM initiators minus the change in values before and after baseline among noninitiators.13,14 The size and sign of a difference-in-differences estimate indicates the magnitude and direction of change in outcome in real-time CGM initiators accounting for changes observed among the noninitiators. The difference-in-differences method assumes outcomes in the exposed group—had they not been exposed to the intervention (ie, counterfactual case)—would be qualitatively similar to the observed outcomes in the unexposed (reference) group. Doubly robust15 model-based difference-in-differences estimates were derived using adjusted and overlap-weighted generalized linear models, assuming a compound symmetry covariance structure to account for nonindependence of the residual error.16 Conservative 95% CIs were estimated using the Huber-White sandwich estimator method.17 Models were not adjusted for multiple comparisons.
In contrast to randomized trials, the validity of comparative effectiveness studies based on observational data is threatened by confounding (eg, by indication).18 Confounding was addressed in the design phase through the use of overlap weighting based on propensity scores.16,19,20 Overlap weighting emulates critical aspects of randomized clinical trials by estimating results that relate to a specified target population, achieving covariate balance between exposed and reference groups, and maximizing precision; this approach avoids limitations associated with the classic propensity score approach of inverse probability of treatment weighting.16,21
To calculate the propensity score, we specified treatment models to predict the probability of real-time CGM initiation using more than 40 candidate predictors and all possible interactions and quadratic terms as inputs (eTable 1 in the Supplement) using a machine-learning approach to stepwise logistic regression with model decisions based on Akaike information criterion. Instrumental variables (eg, cost-sharing for real-time CGM) were excluded from the treatment models.22 Missing data were rare; only 0.4% of the approximately 4.3 million data points were missing. Median values were imputed for missing continuous variables, and we used an indicator for missing categorical variables for the propensity score models. However, imputation was not used for the difference-in-differences models nor in the statistics presented in Table 1 and Table 2. Every propensity score model demonstrated excellent discrimination (C statistic ≥0.98 for each).
Table 2. Baseline Characteristics Before and After Propensity Score Overlap Weighting in 36 080 Patients With Type 2 Diabetes, January 1, 2015, Through December 31, 2018.
| Variables | Crude (unweighted)a | Standardized differencec | After overlap weightinga,b | Standardized differencec | ||
|---|---|---|---|---|---|---|
| Real-time CGM initiators (n = 344) | Noninitiators (n = 35 736) | Real-time CGM initiators (n = 344) | Noninitiators (n - 35 736) | |||
| Male | 52.9 | 50.4 | 0.05 | 53.3 | 48.4 | 0.10 |
| Female | 47.1 | 49.6 | −0.05 | 46.7 | 51.7 | −0.10 |
| Age, mean (SD), y | 59.1 (14.5) | 64.6 (12.1) | −0.41 | 59.3 (14.7) | 61.0 (13.2) | −0.12 |
| Age at diabetes onset, mean (SD), y | 41.6 (12.8) | 48.3 (11.2) | −0.56 | 42.0 (12.8) | 42.9 (12.3) | −0.07 |
| No. | 271 | 29 265 | 271 | 29 265 | ||
| Diabetes duration, mean (SD), y | 17.1 (11.1) | 15.8 (8.8) | 0.13 | 17.0 (10.9) | 17.2 (10.4) | −0.02 |
| No. | 271 | 29 265 | 271 | 29 265 | ||
| Race/ethnicityd | ||||||
| White | 60.8 | 43.1 | 0.36 | 57.6 | 53.8 | 0.08 |
| Hispanic | 8.1 | 21.7 | −0.39 | 9.1 | 12.5 | −0.11 |
| Black | 9.3 | 10.5 | −0.04 | 9.5 | 8.7 | 0.03 |
| Asian | 16.3 | 18.1 | −0.05 | 16.7 | 17.7 | −0.03 |
| Multiracial/multiethnic | 3.8 | 3.9 | −0.00 | 4.8.3 | 4.3.9 | 0.03 |
| Other | 0 | 1.9 | −0.20 | 0 | 1.1 | −0.15 |
| Unknown | 1.7 | 0.9 | 0.08 | 2.3 | 1.9 | 0.03 |
| Preferred language not English8 | 1.2 | 9.5 | −0.38 | 1.5 | 3.4 | −0.12 |
| Neighborhood deprivation indexe | ||||||
| No. | 342 | 35 465 | 342 | 35 465 | ||
| Q1 (least deprived) | 31.4 | 18.4 | 0.30 | 32.5 | 27.9 | 0.10 |
| Q2 | 28.5 | 27.4 | 0.02 | 28.3 | 30.3 | −0.04 |
| Q3 | 28.5 | 30.1 | −0.03 | 26.5 | 24.4 | 0.05 |
| Q4 (most deprived) | 11.1 | 23.3 | −0.33 | 12.0 | 17.0 | −0.14 |
| Insurance typef | ||||||
| Commercial | 95.4 | 96.5 | −0.06 | 94.9 | 95.5 | −0.03 |
| Medicare Advantage | 51.5 | 56.5 | −0.10 | 51.2 | 51.4 | −0.01 |
| Medicaid | 7.0 | 5.4 | 0.06 | 7.4 | 6.6 | 0.03 |
| Charlson Comorbidity Index scoreg | ||||||
| 1 | 22.4 | 21.6 | 0.02 | 19.7 | 18.6 | 0.03 |
| 2 | 27.9 | 22.5 | 0.13 | 26.9 | 27.0 | −0 |
| 3 | 12.5 | 10.0 | 0.08 | 13.4 | 12.5 | 0.03 |
| ≥4 | 37.2 | 45.9 | −0.18 | 40.1 | 41.9 | −0.04 |
| Hypoglycemia risk scoreh | ||||||
| Low | 64.5 | 68.4 | −0.08 | 63.3 | 66.0 | −0.06 |
| Intermediate | 12.8 | 23.6 | −0.28 | 14.1 | 16.2 | −0.06 |
| High | 22.7 | 8.1 | 0.41 | 22.6 | 17.8 | 0.12 |
| Insulin type | ||||||
| Long acting | 61.3 | 13.7 | 1.13 | 62.3 | 56.7 | 0.11 |
| NPHi | 36.1 | 75.5 | −0.86 | 42.7 | 42.9 | −0 |
| Rapid acting | 83.1 | 13.4 | 1.95 | 77.8 | 71.4 | 0.15 |
| Short acting | 20.9 | 30.6 | −0.22 | 24.5 | 23.5 | 0.02 |
| Mixed | 3.8 | 13.3 | −0.35 | 5.0 | 4.6 | 0.02 |
| Glucagonj | 25.0 | 1.2 | 0.76 | 17.4 | 17.4 | 0 |
| Insulin delivery | ||||||
| Pen | 48.3 | 15.0 | 0.77 | 56.6 | 49.1 | 0.15 |
| Pumpk | 22.4 | 0.1 | 0.76 | 6.8 | 6.8 | 0.00 |
| ≥3 Injections/d | 15.1 | 22.1 | −0.18 | 18.4 | 21.6 | −0.08 |
| Use of noninsulin diabetes medicationsl | ||||||
| Biguanide | 45.4 | 61.9 | −0.34 | 47.4 | 50.4 | −0.06 |
| Sulfonylurea | 14.2 | 48.1 | −0.78 | 16.8 | 16.8 | −0.00 |
| Thiazolidinedione | 2.6 | 2.5 | 0.01 | 3.0 | 3.0 | 0 |
| DPP-4 | 1.5 | 1.0 | 0.04 | 1.8 | 2.0 | −0.02 |
| GLP-1 | 4.4 | 0.6 | 0.25 | 3.6 | 4.1 | −0.03 |
| α-Glucosidase inhibitor | 0.6 | 0.2 | 0.07 | 0.8 | 0.3 | 0.07 |
| SGLT-2 | 0.3 | 0.2 | 0.03 | 0.4 | 0.9 | −0.07 |
| Meglitinide | 0.3 | 0.1 | 0.05 | 0.3 | 0.3 | 0.00 |
| Amylin | 0 | 0.03 | −0.03 | 0 | 0.2 | −0.07 |
| No. of diabetes therapeutic classes used | ||||||
| 1 | 49.1 | 26.0 | 0.49 | 46.3 | 43.1 | 0.06 |
| 2 | 35.2 | 35.8 | −0.01 | 36.7 | 39.3 | −0.05 |
| 3 | 13.1 | 36.1 | −0.55 | 13.7 | 14.3 | −0.02 |
| ≥4 | 2.6 | 2.1 | 0.03 | 3.3 | 3.3 | 0.00 |
| Glucose self-monitoring | ||||||
| <1/d | 2.9 | 17.3 | −0.49 | 3.5 | 4.0 | −0.03 |
| ≥1/d | 97.1 | 82.7 | 0.49 | 96.5 | 96.0 | 0.03 |
| HbA1c, mean (SD), %m | 8.20 (1.5) | 8.27 (1.6) | −0.05 | 8.24 (1.6) | 8.08 (1.5) | 0.10 |
| No. | 340 | 34 576 | 340 | 34 576 | ||
| Hypoglycemic eventn | 7.9 | 1.8 | 0.29 | 7.2 | 7.8 | −0.02 |
| Hyperglycemic eventn | 4.0 | 1.1 | 0.19 | 4.1 | 2.0 | 0.12 |
| Any ED visito | 38.4 | 33.7 | 0.10 | 40.1 | 40.9 | −0.02 |
| Any hospitalizationo | 11.9 | 8.1 | 0.13 | 12.7 | 10.0 | 0.09 |
| No. of visits, mean (SD)o | ||||||
| Outpatient | 6.0 (4.4) | 3.7 (3.8) | 0.56 | 5.8 (4.5) | 5.8 (4.3) | 0.00 |
| Telephone | 5.7 (6.5) | 4.9 (6.1) | 0.13 | 5.6 (6.4) | 6.0 (7.4) | −0.05 |
Abbreviations: DPP-4, didpeptidyl peptidase 4; ED, emergency department; GLP-1, glucagon-like peptide 1; HbA1c, hemoglobin A1c; NPH, neutral protamine Hagedorn; Q, quartile; SGLT-2, sodium-glucose cotransporter 2.
Data are presented as percent unless otherwise specified.
Propensity score models were based only on the variables included in eTable 1 in the Supplement.
See footnote c in Table 1 for standard difference definitions.
See footnote d in Table 1 for race/ethnicity definitions.
See footnote e in Table 1 for the neighborhood deprivation index definition.
See footnote f in Table 1 for an explanation on insurance.
See footnote g in Table 1 for Deyo Charlson Score ranges.
See footnote h in Table 1 for risk stratification score ranges.
Insulin types are not mutually exclusive.
See footnote j in Table 1.
See footnote k in Table 1.
Table reports medications at the therapeutic drug class level. Examples of these medications include: biguanide (metformin), sulfonylurea (glipizide), thiazolidinedione (pioglitazone), dipeptidyl peptidase 4 inhibitor (linagliptin), glucagon-like peptide 1 receptor agonist (exenatide), alpha-glucosidase inhibitor (acarbose), sodium-glucose Cotransporter-2 (empagliflozin), meglitinide (repaglinide), amylin (pramlintide).
See footnote l in Table 1 for HbA1c explanation.
Identified by a principal diagnosis in the inpatient setting or a primary diagnosis in the ED during the 12 months before baseline.
Utilization captured during the 12 months before baseline.
Overlap weights were calculated using the formula [overlap weight = 1 −propensity score] for real-time CGM initiators and [overlap weight = propensity score] for noninitiators.16,19,23 Overlap weights give more influence to (ie, upweight) exposed individuals (ie, CGM initiators) predicted to have a low likelihood of being exposed and to unexposed individuals (ie, reference group) predicted to have a high likelihood of being exposed. The overlap weights were used to balance the covariates specified in the treatment model between the exposed and the reference groups (ie, create exchangeability).16
Covariate differences before and after overlap weighting were assessed using the standardized difference (Cohen d),24 which is the difference in means or proportions divided by pooled standard deviation. The d indicates the size and direction of imbalance in the distribution of characteristics among real-time CGM initiators vs noninitiators. A positive d for a given characteristic in this case means that patients were overrepresented among real-time CGM initiators and underrepresented among noninitiators, and vice versa for a negative standardized difference. An absolute value of d less than 0.1 is considered a negligible group imbalance.9
Confounding was also addressed in the analytic phase using statistical adjustment of the weighted models. Analysis of a directed acyclic graph25 of the hypothesized causal framework (eFigures 1 and 2 in the Supplement) was used to identify critical confounders that were then added to the difference-in-differences models: insulin injection frequency, use of rapid- or short-acting insulin, pump use, a history of prior acute metabolic crises, and glycemic control 12 to 24 months prior to baseline (ie, assessed in the 12 months prior to the beginning of the prebaseline period). No violations of model assumptions (ie, experimental treatment assignment,26,27 parallel trends,13 common shock, and no spillover13) were detected. Models were not adjusted for multiple comparisons and were thus exploratory.28
In addition to evaluating the crude and the weighted and adjusted difference-in-differences estimates of the associations between real-time CGM use and outcomes for the overall sample, we tested whether the associations differed significantly by diabetes type by adding interaction terms to the main effect models. We also ran subgroup analyses for each diabetes type since several baseline characteristics and difference-in-differences estimates differed significantly by diabetes type.
As a sensitivity analysis, we assessed a dose-level association comparing before and after changes in outcomes of participants who were persistent real-time CGM users (≥2 claims for CGM transmitters during the 12 months after initiation), nonpersistent real-time CGM users (1 claim), or noninitiators (0 claims).
Analyses were repeated after excluding participants who had initiated an insulin pump in the 12 months before or after baseline given the intractability of disentangling the effects of insulin pump and real-time CGM when initiated simultaneously.
The number needed to treat was calculated using the following formula (number needed to treat = 1/difference-in-differences).29
All statistical analyses were performed using SAS version 9.4 (SAS Institute Inc). P values were 2-sided, with a significance threshold of .05. Because of the potential for type I error due to multiple comparisons, findings for these analyses should be interpreted as exploratory.
Results
This retrospective study included 83 506 person-years of observation on 41 753 patients with insulin-treated diabetes (5673 type 1 diabetes; 36 080 type 2 diabetes) who had not used real-time CGM. From January 1, 2015, through December 31, 2018, a total of 3462 patients with type 1 diabetes (Table 1) and 344 patients with type 2 diabetes (Table 2) initiated real-time CGM. The reference group consisted of the remaining 37 947 patients (2211 type 1 diabetes; 35 736 type 2 diabetes) who did not initiate real-time CGM.
Almost all those who initiated real-time CGM were using basal-bolus insulin. Compared with noninitiators, patients who had initiated real-time CGM were more likely to have type 1 diabetes, be younger at baseline and at diabetes onset, and have fewer comorbidities. Real-time CGM initiators were more likely White and English-language speakers and less likely living in a deprived neighborhood. Compared with noninitiators, real-time CGM initiators at baseline had similar mean HbA1c, higher rates of hypoglycemia and hyperglycemia, and higher predicted risk of hypoglycemia and were more likely to have been dispensed glucagon, insulin pen or pump, and long- or rapid-acting insulin. All patients with type 1 diabetes and 97% with type 2 diabetes who initiated real-time CGM were using basal-bolus insulin.
After overlap weighting, all baseline characteristics in patients with type 1 diabetes were balanced (|d|<0.1) between real-time CGM initiators and noninitiators; the few baseline characteristics in patients with type 2 diabetes that remained unbalanced (|d|≥0.1) were added as covariates in the difference-in-differences models. As a result of overlap weighting, the distribution of the expected (predicted) likelihood of real-time CGM initiation (indicated by the propensity score) was similar between patients who initiated vs those who did not initiate real-time CGM (Figure 2).16,30
Figure 2. Mirrored Density Plot of Propensity Scores in the Full Cohort, Including 3806 Initiators of Real-time Continuous Glucose Monitoring and 37 947 Noninitiators, Before and After Overlap Weightinga.
aPropensity scores were derived from propensity score models that included variables (eTable 1 in the Supplement) used to calculate overlap weights applied to difference-in-differences models for hemoglobin A1c (HbA1c).
Glycemic Control
The mean HbA1c declined from 8.17% to 7.76% (difference, −0.41%) among real-time CGM initiators compared with a decline from 8.28% to 8.19% (difference, −0.09%) among noninitiators, for a difference-in-differences estimate of change in HbA1c associated with real-time CGM of −0.31% (95% CI, −0.36% to −0.27%; P < .001) (Table 3). After weighting and adjustment, the difference-in-differences estimate was a statistically significant change in HbA1c of −0.40% (95% CI, −0.48% to −0.32%; P < .001) (Table 3). There were similar changes in secondary outcomes based on the Healthcare Effectiveness Data and Information Set performance thresholds.
Table 3. Before and After Differences in Outcomes for Patients With Diabetes Treated With Insulina.
| Real-time CGM initiators | Noninitiators | Difference in differences | P value | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Before baseline | After baseline | Differenceb | Before baseline | After baseline | Differenceb | Unweighted and unadjusted (95% CI)c | Weighted and adjusted (95% CI)d | ||
| All participants (real-time CGM group, n = 3806; reference group, n = 37 947) | |||||||||
| Event rate (ED or hospitalization) | |||||||||
| Hypoglycemia | 5.1 | 3.0 | −2.1 | 1.9 | 2.3 | 0.4 | −2.6 (−3.4 to −1.7) | −2.7 (−4.4 to −1.1) | .001 |
| Hyperglycemia | 4.7 | 5.0 | 0.3 | 1.3 | 1.4 | 0.1 | 0.2 (−0.6 to 1.0) | −0.4 (−1.9 to 1.0) | .57 |
| HbA1c, mean, % | 8.17 | 7.76 | −0.41 | 8.28 | 8.19 | −0.09 | −0.31 (−0.36 to −0.27) | −0.40 (−0.48 to −0.32) | <.001 |
| HbA1c <7% | 19.5 | 29.1 | 9.7 | 18.0 | 20.3 | 2.3 | 7.3 (5.8 to 8.8) | 9.6 (7.1 to 12.2) | <.001 |
| HbA1c <8% | 50.2 | 64.1 | 13.9 | 50.8 | 53.2 | 2.4 | 11.5 (9.8 to 13.2) | 13.1 (10.2 to 16.1) | <.001 |
| HbA1c >9% | 21.9 | 14.1 | −7.8 | 25.0 | 23.4 | −1.6 | −6.0 (−7.3 to −4.6) | −7.1 (−9.5 to −4.6) | <.001 |
| No. of visits, mean | |||||||||
| Outpatient | 5.1 | 4.3 | −0.8 | 3.6 | 3.5 | −0.1 | −0.6 (−0.7 to −0.5) | −0.4 (−0.6 to −0.2) | <.001 |
| Telephone | 4.3 | 4.8 | 0.5 | 4.8 | 4.5 | −0.3 | 0.8 (0.6 to 1.0) | 1.1 (0.8 to 1.4) | <.001 |
| Any ED encounter | 30.3 | 28.5 | −1.8 | 33.4 | 35.1 | 1.7 | −3.6 (−5.4 to −1.7) | −2.3 (−5.5 to 0.9) | .17 |
| Any hospitalization | 5.3 | 5.4 | 0.1 | 8.0 | 8.5 | 0.6 | −0.5 (−1.5 to 0.6) | 0.1 (−1.8 to 2.0) | .93 |
| Type 1 diabetes (real-time CGM group, n = 3462; reference group, n = 2211) | |||||||||
| Event rate (ED or hospitalization) | |||||||||
| Hypoglycemia | 4.8 | 2.9 | −1.9 | 3.4 | 4.0 | 0.6 | −2.5 (−3.8 to −1.2) | −2.3 (−4.1 to −0.6) | .01 |
| Hyperglycemia | 4.8 | 5.3 | 0.6 | 4.8 | 5.6 | 0.8 | −0.2 (−1.6 to 1.1) | 0.2 (−1.5 to 2.0) | .78 |
| HbA1c, mean, % | 8.16 | 7.77 | −0.39 | 8.40 | 8.29 | −0.11 | −0.28 (−0.35 to −0.22) | −0.34 (−0.43 to −0.25) | <.001 |
| HbA1c <7% | 19.4 | 28.8 | 9.4 | 18.4 | 19.5 | 1.1 | 8.2 (6.0 to 10.4) | 9.4 (6.6 to 12.3) | <.001 |
| HbA1c <8% | 50.5 | 64.0 | 13.5 | 46.7 | 49.4 | 2.6 | 10.9 (8.3 to 13.6) | 12.1 (8.8 to 15.4) | <.001 |
| HbA1c >9% | 21.7 | 14.3 | −7.4 | 28.2 | 26.1 | −2.1 | −5.0 (−7.2 to −2.9) | −5.6 (−8.3 to −2.8) | <.001 |
| No. of visits, mean | |||||||||
| Outpatient | 5.0 | 4.2 | −0.8 | 2.8 | 2.6 | −0.2 | −0.6 (−0.8 to −0.4) | −0.4 (−0.7 to −0.2) | <.001 |
| Telephone | 4.2 | 4.7 | 0.5 | 3.3 | 3.1 | −0.2 | 0.8 (0.5 to 1.0) | 1.0 (0.7 to 1.4) | <.001 |
| Any ED encounter | 29.5 | 27.8 | −1.7 | 29.1 | 29.9 | 0.9 | −2.6 (−5.4 to 0.3) | −1.2 (−4.5 to 2.4) | .52 |
| Any hospitalization | 4.6 | 5.1 | 0.5 | 5.9 | 5.9 | 0.0 | 0.5 (−1.0 to 2.1) | 1.1 (−0.9 to 3.1) | .28 |
| Type 2 diabetes (real-time CGM group, n = 344; reference group, n = 35 736) | |||||||||
| Event rate (ED or hospitalization) | |||||||||
| Hypoglycemia | 7.8 | 3.2 | −4.6 | 1.8 | 2.2 | 0.4 | −5.1 (−8.3 to −1.8) | −4.0 (−7.8 to −0.2) | .04 |
| Hyperglycemia | 4.1 | 1.4 | −2.6 | 1.1 | 1.2 | 0.0 | −2.7 (−5.0 to −0.3) | −2.5 (−5.1 to 0.2) | .07 |
| HbA1c, mean, % | 8.20 | 7.64 | −0.56 | 8.27 | 8.18 | −0.09 | −0.47 (−0.60 to −0.34) | −0.56 (−0.72 to −0.41) | <.001 |
| HbA1c <7% | 19.7 | 32.2 | 12.5 | 18.0 | 20.3 | 2.3 | 10.2 (5.2 to 15.1) | 10.1 (4.6 to 15.6) | .001 |
| HbA1c <8% | 47.4 | 65.1 | 17.7 | 51.0 | 53.4 | 2.4 | 15.1 (9.7 to 20.5) | 16.2 (9.8 to 22.6) | <.001 |
| HbA1c >9% | 23.8 | 13.0 | −10.8 | 24.8 | 23.3 | −1.5 | −9.3 (−13.7 to −4.8) | −11.5 (−16.7 to −6.2) | <.001 |
| No. of visits, mean | |||||||||
| Outpatient | 6.0 | 5.3 | −0.7 | 3.7 | 3.6 | −0.1 | −0.5 (−1.0 to −0.1) | −0.3 (−0.9 to 0.2) | .24 |
| Telephone | 5.7 | 6.2 | 0.5 | 4.9 | 4.6 | −0.3 | 0.8 (0.0 to 1.6) | 1.3 (0.3 to 2.2) | .008 |
| Any ED encounter | 38.4 | 35.5 | −2.9 | 33.7 | 35.4 | 1.8 | −4.7 (−11.0 to 1.6) | −5.8 (−13.0 to 1.4) | .12 |
| Any hospitalization | 11.9 | 8.1 | −3.8 | 8.1 | 8.7 | 0.6 | −4.4 (−8.6 to −0.2) | −3.4 (−8.5 to 1.6) | .19 |
Abbreviations: CGM, continuous glucose monitoring; ED, emergency department; HbA1c, hemoglobin A1c.
Data are presented as percent unless otherwise indicated.
Difference is calculated as postbaseline value minus prebaseline value.
See the Methods section for difference-in-differences calculation. There are slight discrepancies between the unadjusted and unweighted model-based difference in differences and the crude difference in differences for the HbA1c outcomes due to missing data. Patients with 1 or more HbA1c measure in either the before or after baseline time frame were included in the model.
Overlap weighting was based on propensity score model (eTable 1 in the Supplement) and adjusted for variables identified in the directed acyclic graph (eFigures 1 and 2 in the Supplement) as critical adjusters (prebaseline insulin treatment, glycemic control, and acute metabolic crisis) and baseline covariates with an absolute d greater than .10 after weighting (see Table 1 and Table 2). Conservative 95% CIs for both difference-in-difference estimates were estimated using robust variance estimators.17
The mean HbA1c declined significantly more among patients with type 2 diabetes than in patients with type 1 diabetes (P value for interaction = .003). For patients with type 2 diabetes, the mean HbA1c declined from 8.20% to 7.64% (difference, −0.56%) among real-time CGM initiators compared with a decline from 8.27% to 8.18% (difference, −0.09%) among noninitiators, for a weighted and adjusted net change in HbA1c associated with real-time CGM initiation of −0.56% (95% CI, −0.72% to −0.41%; P < .001). For patients with type 1 diabetes, the mean HbA1c declined from 8.16% to 7.77% (difference, −0.39%) among real-time CGM initiators compared with a decline from 8.40% to 8.29% (difference, −0.11%) among noninitiators, for a weighted and adjusted net change in HbA1c associated with real-time CGM of −0.34% (95% CI, −0.43% to −0.25%; P < .001).
The decline in the prevalence of HbA1c greater than 9% among real-time GCM initiators was significantly greater in patients with type 2 diabetes than in patients with type 1 diabetes (P value for interaction = .04). The proportion of patients with type 2 diabetes whose HbA1c exceeded 9% declined from 23.8% to 13.0% (difference, −10.8%) among real-time CGM noninitiators compared with a decline from 24.8% to 23.3% (difference, −1.5%) among noninitiators, for a weighted and adjusted net change in prevalence associated with real-time CGM initiation of −11.5% (95% CI, −16.7% to −6.2%; P < .001). Similarly, the proportion of patients with type 1 diabetes among real-time CGM initiators whose HbA1c exceeded 9% declined from 21.7% to 14.3% (difference, −7.4%) compared with a decline from 28.2% to 26.1% (difference, −2.1%) among noninitiators, for a weighted and adjusted net change in prevalence associated with real-time CGM initiation of −5.6% (95% CI, −8.3% to −2.8%; P < .001). There was no significant difference in the weighted and adjusted net change in the prevalence of HbA1c lower than 7% and 8% by diabetes type (P = .83 and P = .26 for interaction, respectively).
Hypoglycemia
Crude rates of emergency department visits or hospitalizations for hypoglycemia declined from 5.1% to 3.0% (difference, −2.1%) among real-time CGM initiators but increased from 1.9% to 2.3% (difference, 0.4%) among noninitiators. After weighting and adjustment, real-time CGM initiation was significantly associated with a −2.7% (95% CI, −4.4% to −1.1%; P = .001) change in hypoglycemia rate. Real-time CGM initiation was associated with a weighted and adjusted decrease in the rate of hypoglycemia for patients with type 2 diabetes of −4.0% (95% CI, −7.8% to −0.2%; P = .04) and an estimated decrease for patients with type 1 diabetes of −2.3% (95% CI, −4.1% to −0.6%; P = .01); this difference between diabetes type was not statistically significant (P = .42 for interaction).
Hyperglycemia
Real-time CGM initiation was not associated with statistically significant changes in the rate of emergency department or hospital utilization because of hyperglycemia, overall, or by diabetes type (Table 3).
Other Health Care Use
Real-time CGM initiation was not associated with a statistically significant change in the proportion of patients having any emergency department visit or hospital utilization (overall). However, real-time CGM initiation was significantly associated with a decrease in the mean annual number of outpatient visits (−0.4; 95% CI, −0.6 to −0.2; P < .001) and an increase in telephone visits (1.1; 95% CI, 0.8 to 1.4; P < .001); however, these weighted and adjusted difference-in-differences estimates did not differ by diabetes type in a substantive or statistically significant way (P = .72 and P = .65 for interactions, respectively).
Sensitivity Analyses
In an analysis assessing a dose-level association, findings were consistent with a dose response based on real-time CGM persistence for mean HbA1c (−0.48% among persistent users, −0.38% among nonpersistent users, and −0.09% among noninitiators) and for hypoglycemia rates (−3.4% change among persistent users, −1.6% among nonpersistent users, and 0.4% among noninitiators). A similar dose-response pattern was observed by diabetes type (see eTable 2 in the Supplement).
After excluding 519 patients who initiated the insulin pump during the 12 months before or after baseline, minor differences were found in the estimates but there were no substantive changes that would alter the interpretation of these findings.
Number Needed to Treat
Assuming that the measured observational association was causal, the number needed to treat to avoid 1 hypoglycemic event was 43 (95% CI, 24-175) for patients with type 1 diabetes and 25 (95% CI, 13-476) for patients with type 2 diabetes; the number needed to treat to achieve 1 more person with HbA1c lower than 8% was 8 (95% CI, 6-11) for patients with type 1 diabetes and 6 (95% CI, 4-10) for patients with type 2 diabetes.
Discussion
In this cohort of patients with type 1 diabetes and patients with insulin-treated type 2 diabetes, real-time CGM initiation was associated with significantly better glycemic control and lower rates of hypoglycemia-related emergency department or hospital utilizations. Real-time CGM initiation was not significantly associated with emergency department visits or hospitalizations for hyperglycemia or for any reason. There were minor, but statistically significant, increases in the number of telephone visits and a decrease in outpatient visits. Patients with type 2 diabetes who initiated real-time CGM had higher prebaseline hypoglycemia rates than those who did not, whereas hypoglycemia rates in patients with type 1 diabetes differed little between initiators and noninitiators. In contrast, prebaseline glycemic control was similar between groups of either diabetes type.
These observational findings complement the substantial trial evidence (eg, Gold [Glycaemic Control and Optimisation of Life Quality in Type 1 Diabetes],31 Diamond [Daily Injections and Continuous Glucose Monitoring in Diabetes],32 Silver [Sustained Intensive Treatment and Long-term Effects on HbA1c Reduction],4 HypoDE [Hypoglycemia in Germany],33 Rescue [Reimbursement Study of Continuous Glucose Monitoring in Belgium],34 and WISDM [Wireless Innovation for Seniors with Diabetes Mellitus]1 trials) and meta-analyses35 that demonstrated that real-time CGM improves glycemic control and hypoglycemia in patients with type 1 diabetes.
There have been few efficacy trials of real-time CGM involving patients with insulin-treated type 2 diabetes. Two meta-analyses of trials comparing real-time CGM to self-monitoring of blood glucose in patients with type 2 diabetes reported significant −0.35% and −0.28% changes in HbA1c and shorter time spent in hypoglycemia.6,36 The DIAMOND study, a randomized trial of real-time CGM vs usual care involving 158 patients with insulin-treated type 2 diabetes, reported a decrease in mean HbA1c to 7.7% in the CGM group and 8.0% in the control group, a net difference of −0.3% (95% CI, −0.5% to 0; P = .02) at 24 weeks but no change in hypoglycemia.2
In this observational study, patients selected by clinicians for real-time CGM initiation achieved improved glycemic control and lower hypoglycemia rates. The dose response in the association between real-time CGM adherence (0, 1, or ≥2 claims for CGM transmitters) and the size of the changes in these clinical end points supports the validity of study findings. The improvements in glycemic control were greater for patients with type 2 diabetes than for patients with type 1 diabetes. These results from a usual care setting complement efficacy estimates from trials that may have restrictive eligibility criteria, intensive clinical attention, and levels of treatment adherence that might not be replicable in usual care settings.37
Many patients with type 2 diabetes in the reference group might have been eligible for real-time CGM prescribing under the CMS criteria, but in this study setting there were additional eligibility requirements for non-Medicare members (a history of hypoglycemia, suboptimal glycemic control, or severe dexterity impairment). It is unknown to what extent real-time CGM might benefit patients with type 2 diabetes who do not meet eligibility criteria, eg, those receiving basal-only insulin. This study and others38 have found that underresourced populations were less likely to be prescribed CGM; research is needed to understand what clinician- and patient-level barriers need to be addressed to ensure equitable access. Research is needed to understand how clinicians monitor and use real-time CGM data in clinical decision-making.
Limitations
This study has several limitations. First, although the analysis used robust methods (overlap weighting based on propensity scores and directed acyclic graph-informed adjustment that balanced >40 characteristics in the exposed and reference groups), notable between-group differences (eg, race/ethnicity, neighborhood deprivation index) make it likely that there were unmeasured residual confounders that could have biased the results. Second, because of the potential for type I error due to multiple comparisons, findings for these analyses should be interpreted as exploratory.
Third, this study may not have had adequate power to detect associations between real-time CGM initiation and hyperglycemia. Among patients with type 1 diabetes, the observed association was smaller than expected; whereas among patients with type 2 diabetes the association was clinically substantive but the confidence limits were imprecise due to the small sample size. Fourth, this study was unable to account for advancements in real-time CGM technology (eg, predictive low glucose alerts, no requirement for calibration) that occurred during the study period. Fifth, while this study included a diverse sample from a single integrated health care delivery system, the case mix of real-time CGM initiators may differ from other settings with different prescribing patterns and guidelines.
Sixth, the measures of hypoglycemia and hyperglycemia were based on emergency department visits and hospitalizations and thus missed events that were handled outside of the health care setting.39 Seventh, estimates of number needed to treat are sensitive to case mix and the heterogeneity of treatment effects and assume a causal association that this study design cannot establish; caution is needed when interpreting these estimates and applying them to other populations.
Conclusions
In this retrospective cohort study, insulin-treated patients with diabetes selected by physicians for real-time continuous glucose monitoring compared with noninitiators had significant improvements in hemoglobin A1c and reductions in emergency department visits and hospitalizations for hypoglycemia, but no significant change in emergency department visits or hospitalizations for hyperglycemia or for any reason. Because of the observational study design, findings may have been susceptible to selection bias.
eFigure 1. Directed acyclic graph depicting hypotheses regarding the total effect of real-time continuous glucose monitoring initiation on glycemic control
eFigure 2. Directed acyclic graph depicting hypotheses regarding the total effect of real-time continuous glucose monitoring initiation on acute metabolic crises
eTable 1. Propensity score model specification
eTable 2. Simulated per protocol analysis comparing pre-post change for participants who were persistent users of real time continuous glucose monitoring (≥ claims for continuous glucose monitoring transmitters during the 12 months after initiation, nonpersistent vs noninitiators
References
- 1.Pratley RE, Kanapka LG, Rickels MR, et al. ; Wireless Innovation for Seniors With Diabetes Mellitus (WISDM) Study Group . Effect of continuous glucose monitoring on hypoglycemia in older adults with type 1 diabetes: a randomized clinical trial. JAMA. 2020;323(23):2397-2406. doi: 10.1001/jama.2020.6928 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Beck RW, Riddlesworth TD, Ruedy K, et al. ; DIAMOND Study Group . Continuous glucose monitoring versus usual care in patients with type 2 diabetes receiving multiple daily insulin injections: a randomized trial. Ann Intern Med. 2017;167(6):365-374. doi: 10.7326/M16-2855 [DOI] [PubMed] [Google Scholar]
- 3.Huang ES, O’Grady M, Basu A, et al. ; Juvenile Diabetes Research Foundation Continuous Glucose Monitoring Study Group . The cost-effectiveness of continuous glucose monitoring in type 1 diabetes. Diabetes Care. 2010;33(6):1269-1274. doi: 10.2337/dc09-2042 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Lind M, Ólafsdóttir AF, Hirsch IB, et al. Sustained Intensive Treatment and Long-term Effects on HbA1c Reduction (SILVER Study) by CGM in people with type 1 diabetes treated with MDI. Diabetes Care. 2021;44(1):141-149. doi: 10.2337/dc20-1468 [DOI] [PubMed] [Google Scholar]
- 5.American Diabetes Association . 7. Diabetes Technology: Standards of Medical Care in Diabetes-2021. Diabetes Care. 2021;44(suppl 1):S85-S99. doi: 10.2337/dc21-S007 [DOI] [PubMed] [Google Scholar]
- 6.Dicembrini I, Mannucci E, Monami M, Pala L. Impact of technology on glycaemic control in type 2 diabetes: a meta-analysis of randomized trials on continuous glucose monitoring and continuous subcutaneous insulin infusion. Diabetes Obes Metab. 2019;21(12):2619-2625. doi: 10.1111/dom.13845 [DOI] [PubMed] [Google Scholar]
- 7.Karter AJ, Ferrara A, Liu JY, Moffet HH, Ackerson LM, Selby JV. Ethnic disparities in diabetic complications in an insured population. JAMA. 2002;287(19):2519-2527. doi: 10.1001/jama.287.19.2519 [DOI] [PubMed] [Google Scholar]
- 8.Parker MM, Fernández A, Moffet HH, Grant RW, Torreblanca A, Karter AJ. Association of patient-physician language concordance and glycemic control for limited-English proficiency Latinos with type 2 diabetes. JAMA Intern Med. 2017;177(3):380-387. doi: 10.1001/jamainternmed.2016.8648 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Austin PC. An introduction to propensity score methods for reducing the effects of confounding in observational studies. Multivariate Behav Res. 2011;46(3):399-424. doi: 10.1080/00273171.2011.568786 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Laraia BA, Karter AJ, Warton EM, Schillinger D, Moffet HH, Adler N. Place matters: neighborhood deprivation and cardiometabolic risk factors in the Diabetes Study of Northern California (DISTANCE). Soc Sci Med. 2012;74(7):1082-1090. doi: 10.1016/j.socscimed.2011.11.036 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol. 1992;45(6):613-619. doi: 10.1016/0895-4356(92)90133-8 [DOI] [PubMed] [Google Scholar]
- 12.Karter AJ, Warton EM, Lipska KJ, et al. Development and validation of a tool to identify patients with type 2 diabetes at high risk of hypoglycemia-related emergency department or hospital use. JAMA Intern Med. 2017;177(10):1461-1470. doi: 10.1001/jamainternmed.2017.3844 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Dimick JB, Ryan AM. Methods for evaluating changes in health care policy: the difference-in-differences approach. JAMA. 2014;312(22):2401-2402. doi: 10.1001/jama.2014.16153 [DOI] [PubMed] [Google Scholar]
- 14.Zhou H, Taber C, Arcona S, Li Y. Difference-in-differences method in comparative effectiveness research: utility with unbalanced groups. Appl Health Econ Health Policy. 2016;14(4):419-429. doi: 10.1007/s40258-016-0249-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Funk MJ, Westreich D, Wiesen C, Stürmer T, Brookhart MA, Davidian M. Doubly robust estimation of causal effects. Am J Epidemiol. 2011;173(7):761-767. doi: 10.1093/aje/kwq439 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Thomas LE, Li F, Pencina MJ. Overlap weighting: a propensity score method that mimics attributes of a randomized clinical trial. JAMA. 2020;323(23):2417-2418. doi: 10.1001/jama.2020.7819 [DOI] [PubMed] [Google Scholar]
- 17.Robins JM, Hernán MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology. 2000;11(5):550-560. doi: 10.1097/00001648-200009000-00011 [DOI] [PubMed] [Google Scholar]
- 18.Psaty BM, Koepsell TD, Lin D, et al. Assessment and control for confounding by indication in observational studies. J Am Geriatr Soc. 1999;47(6):749-754. doi: 10.1111/j.1532-5415.1999.tb01603.x [DOI] [PubMed] [Google Scholar]
- 19.Li F, Morgan KL, Zaslavsky AM. Balancing covariates via propensity score weighting. J Am Stat Assoc. 2018;113(521):390-400. doi: 10.1080/01621459.2016.1260466 29930437 [DOI] [Google Scholar]
- 20.Kainz K, Greifer N, Givens A, et al. Improving causal inference: recommendations for covariate selection and balance in propensity score methods. J Soc Social Work Res. 2017;8(2):279-303. doi: 10.1086/691464 [DOI] [Google Scholar]
- 21.Danaei G, García Rodríguez LA, Cantero OF, Logan RW, Hernán MA. Electronic medical records can be used to emulate target trials of sustained treatment strategies. J Clin Epidemiol. 2018;96:12-22. doi: 10.1016/j.jclinepi.2017.11.021 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Schneeweiss S, Rassen JA, Glynn RJ, Avorn J, Mogun H, Brookhart MA. High-dimensional propensity score adjustment in studies of treatment effects using health care claims data. Epidemiology. 2009;20(4):512-522. doi: 10.1097/EDE.0b013e3181a663cc [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Li F, Thomas LE, Li F. Addressing extreme propensity scores via the overlap weights. Am J Epidemiol. 2019;188(1):250-257. doi: 10.1093/aje/kwy201 [DOI] [PubMed] [Google Scholar]
- 24.Cohen J. The statistical power of abnormal-social psychological research: a review. J Abnorm Soc Psychol. 1962;65:145-153. doi: 10.1037/h0045186 [DOI] [PubMed] [Google Scholar]
- 25.Greenland S, Pearl J, Robins JM. Causal diagrams for epidemiologic research. Epidemiology. 1999;10(1):37-48. doi: 10.1097/00001648-199901000-00008 [DOI] [PubMed] [Google Scholar]
- 26.Neugebauer R, van der Laan M. Why prefer double robust estimators in causal inference? J Stat Plan Infer. 2005;129:405-426. doi: 10.1016/j.jspi.2004.06.060 [DOI] [Google Scholar]
- 27.Wang Y, Bangsberg D, Petersen ML, van der Lann MJ. Diagnosing bias in the inverse probability of treatment weighted estimator resulting from violation of experimental treatment assignment. UC Berkley Division of Biostatistics working paper series 211:1-25; 2006.
- 28.Althouse AD. Adjust for multiple comparisons? it’s not that simple. Ann Thorac Surg. 2016;101(5):1644-1645. doi: 10.1016/j.athoracsur.2015.11.024 [DOI] [PubMed] [Google Scholar]
- 29.Jansen JP, Khalid JM, Smyth MD, Patel H. The number needed to treat and relevant between-trial comparisons of competing interventions. Clinicoecon Outcomes Res. 2018;10:865-871. doi: 10.2147/CEOR.S180491 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Rubin DB. Estimating causal effects from large data sets using propensity scores. Ann Intern Med. 1997;127(8 pt 2):757-763. doi: 10.7326/0003-4819-127-8_Part_2-199710151-00064 [DOI] [PubMed] [Google Scholar]
- 31.Seyed Ahmadi S, Westman K, Pivodic A, et al. The association between HbA1c and time in hypoglycemia during CGM and self-monitoring of blood glucose in people with type 1 diabetes and multiple daily insulin injections: a randomized clinical trial (GOLD-4). Diabetes Care. 2020;43(9):2017-2024. doi: 10.2337/dc19-2606 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Roze S, Isitt J, Smith-Palmer J, Javanbakht M, Lynch P. Long-term cost-effectiveness of Dexcom G6 real-time continuous glucose monitoring versus self-monitoring of blood glucose in patients with type 1 diabetes in the UK. Diabetes Care. 2020;43(10):2411-2417. doi: 10.2337/dc19-2213 [DOI] [PubMed] [Google Scholar]
- 33.Heinemann L, Freckmann G, Ehrmann D, et al. Real-time continuous glucose monitoring in adults with type 1 diabetes and impaired hypoglycaemia awareness or severe hypoglycaemia treated with multiple daily insulin injections (HypoDE): a multicentre, randomised controlled trial. Lancet. 2018;391(10128):1367-1377. doi: 10.1016/S0140-6736(18)30297-6 [DOI] [PubMed] [Google Scholar]
- 34.Charleer S, De Block C, Nobels F, et al. ; RESCUE Trial Investigators . Sustained impact of real-time continuous glucose monitoring in adults with type 1 diabetes on insulin pump therapy: results after the 24-month RESCUE study. Diabetes Care. 2020;43(12):3016-3023. doi: 10.2337/dc20-1531 [DOI] [PubMed] [Google Scholar]
- 35.Maiorino MI, Signoriello S, Maio A, et al. Effects of continuous glucose monitoring on metrics of glycemic control in diabetes: a systematic review with meta-analysis of randomized controlled trials. Diabetes Care. 2020;43(5):1146-1156. doi: 10.2337/dc19-1459 [DOI] [PubMed] [Google Scholar]
- 36.Ida S, Kaneko R, Murata K. Utility of real-time and retrospective continuous glucose monitoring in patients with type 2 diabetes mellitus: a meta-analysis of randomized controlled trials. J Diabetes Res. 2019;2019(4684815):4684815. doi: 10.1155/2019/4684815 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Edelman SV, Polonsky WH. Type 2 diabetes in the real world: the elusive nature of glycemic control. Diabetes Care. 2017;40(11):1425-1432. doi: 10.2337/dc16-1974 [DOI] [PubMed] [Google Scholar]
- 38.Lai CW, Lipman TH, Willi SM, Hawkes CP. Racial and ethnic disparities in rates of continuous glucose monitor initiation and continued use in children with type 1 diabetes. Diabetes Care. 2021;44(1):255-257. doi: 10.2337/dc20-1663 [DOI] [PubMed] [Google Scholar]
- 39.Karter AJ, Moffet HH, Liu JY, Lipska KJ. Surveillance of hypoglycemia-limitations of emergency department and hospital utilization data. JAMA Intern Med. 2018;178(7):987-988. doi: 10.1001/jamainternmed.2018.1014 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
eFigure 1. Directed acyclic graph depicting hypotheses regarding the total effect of real-time continuous glucose monitoring initiation on glycemic control
eFigure 2. Directed acyclic graph depicting hypotheses regarding the total effect of real-time continuous glucose monitoring initiation on acute metabolic crises
eTable 1. Propensity score model specification
eTable 2. Simulated per protocol analysis comparing pre-post change for participants who were persistent users of real time continuous glucose monitoring (≥ claims for continuous glucose monitoring transmitters during the 12 months after initiation, nonpersistent vs noninitiators