This retrospective cohort study evaluated diabetes device utilization and the effectiveness of these devices for newly diagnosed type 1 diabetes. Investigators examined the use of continuous glucose monitoring (CGM) systems, self-monitoring of blood glucose (SMBG), continuous subcutaneous insulin infusion (CSII), and multiple daily injection (MDI) insulin regimens and their effects on A1C. The researchers identified 6,250 patients with type 1 diabetes, of whom 32% used CGM and 37.1% used CSII. A higher adoption rate of either CGM or CSII in newly diagnosed type 1 diabetes was noted among White patients and those with private health insurance. CGM users had lower A1C levels than nonusers (P = 0.039), whereas no difference was noted between CSII users and nonusers (P = 0.057). Furthermore, CGM use combined with CSII yielded lower A1C than MDI regimens plus SMBG (P <0.001).
Key Points
≫ About one-third of patients with type 1 diabetes were found to use continuous glucose monitoring (CGM) and/or continuous subcutaneous insulin infusion (CSII) in routine clinical care.
≫ Disparities exist in CGM and CSII adoption, with device use more common in patients of higher socioeconomic status.
≫ Mining clinical narratives with natural language processing techniques can be applied successfully for medical device surveillance and cohort identification for observational studies.
≫ CGM use in conjunction with CSII after type 1 diabetes diagnosis is more effective than other therapy regimens and may translate to improved long-term glycemic control.
Type 1 diabetes is a chronic condition that requires self-monitoring of blood glucose (SMBG) and an intensive insulin therapy regimen. Diabetes technologies, including continuous glucose monitoring (CGM) systems and continuous subcutaneous insulin infusion (CSII, or insulin pump therapy), have been well accepted by patients with type 1 diabetes (1) and used for assessment of glucose levels and insulin delivery. The prevalence of such devices is steadily increasing. In fact, the T1D Exchange clinic network indicates that, between 2010–2012, 50% of its participants used an insulin pump and 6% used CGM (2), whereas between 2016 and 2018, use of insulin pumps increased to 63%, and CGM use increased substantially to 30% (3).
Randomized controlled trials (RCTs) have evaluated the efficacy of CGM and CSII use in improving glycemic control in patients with type 1 diabetes. The results of these trials vary by age-group of participants and intervention studied. Overall, real-time CGM use has been associated with improved A1C levels in patients with diabetes (4,5). The effects of CGM on glycemic control from these RCTs are more favorable and consistent in adults than in the children or the young adult population with type 1 diabetes (6–10). However, these RCTs are limited by their small sample sizes, short follow-up durations, and strict participant recruitment criteria (11). Recent systematic reviews and meta-analyses of RCTs comparing CSII with multiple daily injection (MDI) insulin regimens have concluded that CSII is associated with moderate reduction of A1C in children and adults with type 1 diabetes compared with MDI regimens (11,12). Although these findings provide the basis for clinical assertions, real-world data (RWD) are needed to show the effectiveness of CSII and CGM in routine clinical care, as supported by the 21st Century Cures Act (13).
Recently, device surveillance using RWD from electronic health records (EHRs) has demonstrated the feasibility of investigating these devices in routine clinical care (14). These data include longitudinal patient records and present valuable sources of detailed clinical information from both structured (e.g., laboratory tests and medications) and unstructured (e.g., clinical narrative text) sources. Often, health care providers objectively summarize patient information such as patient history, physical findings, and patient care, including medical device use, in the narrative text. Detection of such information embedded in clinical notes can be used for device surveillance and provides opportunities for real-world effectiveness research to complement RCT data (14–16).
This study aimed to assess the uptake of CGM and CSII using RWD and examine the effectiveness of such technologies on patient outcomes, specifically A1C. Gathering all patient encounters from an academic health care setting, we used natural language processing (NLP) and machine learning (ML) classifiers to identify the prevalence of CGM and CSII use in patients with type 1 diabetes between 2008 and 2019. Specifically, we aimed to achieve the following two objectives: 1) estimate the prevalence of CGM and CSII use in patients with type 1 diabetes and 2) examine the effectiveness of these technologies on glycemic control over a 2-year period. The results from this work can be used to understand gaps in device uptake, and the framework used to generate these results can be validated at other health care systems. Importantly, this work can be used to identify patients eligible for CSII and CGM technologies and potentially aid in both outreach and trial recruitment.
Research Design and Methods
Study Design
This retrospective observational cohort study extracted patient information from EHRs between January 2008 and July 2019 at an academic medical center. The study was approved by the institute’s institutional review board.
Cohort Selection
Patients with diabetes were identified using International Classification of Diseases (ICD) codes for either type 1 or type 2 diabetes: ICD-9: 250.X and ICD-10: E10.X and E11.X. Demographic variables were captured, including age at disease diagnosis, sex, race, ethnicity (Hispanic or non-Hispanic), and health insurance status. Race was collapsed into White and non-White (i.e., Asian, Black, and other), and health insurance was categorized into private and nonprivate (i.e., Medicare, Medicaid, other, and unknown). Patients were considered to have diabetes if they had an ICD code and combined evidence of diabetes-related medication and/or at least one elevated laboratory result (fasting glucose >125 mg/dL, A1C ≥6.5%, or random glucose >200 mg/dL).
Further, we adapted the Klompas algorithm (17) to determine patients’ type of diabetes; patients were categorized as having type 1 diabetes if 1) the ratio of type 1 diabetes codes to type 2 diabetes codes was >0.5, and no oral hypoglycemic medications other than metformin were prescribed or 2) the ratio of type 1 diabetes codes to type 2 diabetes codes was >0.5, and there was a prescription for glucagon. We conducted a manual review of 80 randomly selected patients flagged by the Klompas algorithm as having type 1 diabetes and 140 patients identified as having type 2 diabetes. This phenotype algorithm is publicly available at https://phekb.org/phenotype/type-1-and-type-2-diabetes-mellitus. See the Supplementary Materials for detailed description of the adapted Klompas algorithm.
The diabetes diagnosis date was defined as the first appearance of a type 1 diabetes diagnosis in the EHR. Figure 1 illustrates the cohort selection process of the study. Patients were included in the study if they 1) had A1C results at diagnosis or within 6 months from diagnosis (baseline period), 2) had two or more A1C values in the 2-year period after the baseline period (evaluation period), and 3) initiated CGM and/or CSII within 6 months of diagnosis. We then divided patients into four groups: 1) MDI + SMBG, 2) MDI + CGM, 3) CSII + SMBG, and 4) CSII + CGM.
Device Identification and Classification
We identified patients using CGM and CSII through NLP and ML classifiers. The NLP pipeline was developed on a subset of patients (n = 12,960) with diabetes ICD codes to extract the information from several types of clinical notes, including progress notes, telephone encounters, emergency department provider notes, letters, and consultation notes. A list of terms related to diabetes devices was provided by clinical domain experts (Supplementary Appendix S1), along with a review of 200 notes searching for variations of these terms. These terms were used to extract relevant sentences from clinical notes.
Text cleaning methods were performed, including sentence splitting, duplicate sentence removal, tokenization, conversion of text to lowercase, and number removal. The CLEVER (CLinical EVEnt Recognizer) dictionary (18) was used to map the terms with similar meaning to a standard list of clinical terms. The term frequency-inverse document frequency scores were computed on the extracted sentences for vectorization of the training dataset.
A total of 300 sentences were annotated by a clinical expert. A random selection of 240 notes (80%) served as the gold standard for model training and 60 (20%) for validation. Two separate random forest classifiers (50 weak learners) were trained on the vectorized text to detect glucose monitoring approach (CGM vs. SMBG) and insulin delivery approach (CSII vs. MDI) from narrative text.
Study Measurements
The primary outcome of the study was glycemic control measured by A1C over the 2 years after the baseline period. The dates and values of A1C tests were extracted from both laboratory and point-of-care results. Patients’ baseline A1C was determined by averaging A1C values over 6 months immediately after type 1 diabetes diagnosis. Similarly, A1C values were averaged every 6 months for each patient during the 2-year evaluation period; therefore, a total of four A1C time points were derived for each patient.
Statistical Analysis
Patient characteristics across the four device groups were compared using one-way ANOVA for normally distributed continuous variables. The Kruskal-Wallis test was applied for nonnormally distributed variables. The Pearson χ2 test was used to test differences in proportions among groups. Cumulative proportions of patients using CGM and CSII were plotted over the study period.
To examine the effectiveness of device use on glycemic control, we first examined A1C values over time in CGM users compared with those using SMBG regardless of the insulin delivery approach and also CSII users compared with those using MDI regimens regardless of the glucose monitoring approach. We then compared A1C over time across the four groups. Linear mixed models were used to examine A1C levels over time and by group while adjusting for confounders (i.e., age, sex, race, ethnicity, and health insurance status). Nonsignificant variables were removed to achieve parsimonious models. A group-by-time interaction was included in the model. Least square (LS) mean and standard error (SE) were reported for treatment differences. All analyses were performed using R statistical software, v. 3.5.2, using R studio. P <0.05 was considered statistically significant.
Results
A total of 6,250 patients were identified with type 1 diabetes between January 2008 and July 2019. The Klompas algorithm achieved excellent performance on distinguishing type 1 and type 2 diabetes, with 92.5% precision, 98.7% recall, and 96.8% accuracy. Figure 2A presents the accumulated prevalence of CGM and CSII use in patients with type 1 diabetes (n = 6,250) of all age-groups over time using the NLP pipeline, as well as use of either device between 2008 and 2019. A steady continuous rise of CGM and CSII use was observed during the past decade (Figure 2A), with a more significant rise observed in privately insured patients(Figure 2B), White patients (Figure 2C), and non-Hispanic patients (Figure 2D). As of July 2019, 2,369 patients (38%) with type 1 diabetes used CSII. A total of 2,002 patients with type 1 diabetes (32%) were found to be CGM users by July 2019. Further analysis showed that 764 (38.2%) started CGM, and 1,044 (44.1%) started CSII within 6 months after their diagnosis. Our NLP pipeline to identify the insulin delivery approach achieved an average precision score of 0.93, recall score of 0.92, and F1 score of 0.92, and the glucose monitoring method achieved a precision score of 0.86, recall score of 0.85, and F1 score of 0.85 when compared with 26 manually reviewed cases.
A) Accumulated percentage of CGM and CSII adoption in patients with type 1 diabetes (n = 6,250) from January 2008 to July 2019. B) Device adoption (either CGM or CSII) over time by health insurance group. C) Device adoption over time by racial group. D) Device adoption over time by ethnicity group.
A) Accumulated percentage of CGM and CSII adoption in patients with type 1 diabetes (n = 6,250) from January 2008 to July 2019. B) Device adoption (either CGM or CSII) over time by health insurance group. C) Device adoption over time by racial group. D) Device adoption over time by ethnicity group.
A total of 1,100 patients with type 1 diabetes further met the study inclusion criteria for the longitudinal data analysis. Table 1 presents patient characteristics of four device groups at baseline (comparison between device users and nonusers can be found in Supplementary Table S1). The median age of our cohort at diagnosis was 14.1 years with an interquartile range (IQR) of 17.7 years. There were no between-group differences in sex (P = 0.847) or baseline A1C (P = 0.424). CSII or CGM users had higher proportions of White patients (Ps <0.001), non-Hispanic patients (Ps <0.001) and patients with private insurance (Ps <0.001). Age at diabetes diagnosis was significantly different across the four device groups (P <0.001).
Descriptive Characteristics of the Type 1 Diabetes Cohort at Baseline by Device Type
. | MDI + SMBG (n = 649) . | CSII + SMBG (n = 149) . | MDI + CGM (n = 105) . | CSII + CGM (n = 197) . | P . |
---|---|---|---|---|---|
Age at diagnosis, years, median (IQR) | 13.5 (8.4) | 25.1 (24.9) | 16.8 (25) | 12.2 (19.0) | <0.001 |
Age-group, years | <0.001 | ||||
<26 (n = 818) | 524 (80.7) | 78 (52.3) | 69 (65.7) | 147 (74.6) | |
≥26 (n = 282) | 125 (19.3) | 71 (47.7) | 36 (34.3) | 50 (25.4) | |
Sex | 0.847 | ||||
Male | 331 (51.0) | 75 (50.3) | 58 (55.2) | 99 (50.3) | |
Female | 318 (49.0) | 74 (49.7) | 47 (44.8) | 98 (49.7) | |
Race | 0.037 | ||||
White | 330 (50.8) | 89 (59.7) | 59 (56.2) | 120 (60.9) | |
Non-White | 319 (49.2) | 60 (40.3) | 46 (43.8) | 77 (39.1) | |
Hispanic ethnicity | 162 (25.0) | 24 (16.1) | 16 (15.2) | 23 (11.7) | <0.001 |
Insurance status | <0.001 | ||||
Private | 442 (68.1) | 111 (74.5) | 88 (83.8) | 161 (81.7) | |
Nonprivate | 207 (31.9) | 38 (25.5) | 17 (16.2) | 33 (16.8) | |
Baseline A1C, %, mean ± SD | 8.1 ± 1.6 | 8.2 ± 1.5 | 7.9 ± 1.3 | 8.0 ± 1.3 | 0.424 |
. | MDI + SMBG (n = 649) . | CSII + SMBG (n = 149) . | MDI + CGM (n = 105) . | CSII + CGM (n = 197) . | P . |
---|---|---|---|---|---|
Age at diagnosis, years, median (IQR) | 13.5 (8.4) | 25.1 (24.9) | 16.8 (25) | 12.2 (19.0) | <0.001 |
Age-group, years | <0.001 | ||||
<26 (n = 818) | 524 (80.7) | 78 (52.3) | 69 (65.7) | 147 (74.6) | |
≥26 (n = 282) | 125 (19.3) | 71 (47.7) | 36 (34.3) | 50 (25.4) | |
Sex | 0.847 | ||||
Male | 331 (51.0) | 75 (50.3) | 58 (55.2) | 99 (50.3) | |
Female | 318 (49.0) | 74 (49.7) | 47 (44.8) | 98 (49.7) | |
Race | 0.037 | ||||
White | 330 (50.8) | 89 (59.7) | 59 (56.2) | 120 (60.9) | |
Non-White | 319 (49.2) | 60 (40.3) | 46 (43.8) | 77 (39.1) | |
Hispanic ethnicity | 162 (25.0) | 24 (16.1) | 16 (15.2) | 23 (11.7) | <0.001 |
Insurance status | <0.001 | ||||
Private | 442 (68.1) | 111 (74.5) | 88 (83.8) | 161 (81.7) | |
Nonprivate | 207 (31.9) | 38 (25.5) | 17 (16.2) | 33 (16.8) | |
Baseline A1C, %, mean ± SD | 8.1 ± 1.6 | 8.2 ± 1.5 | 7.9 ± 1.3 | 8.0 ± 1.3 | 0.424 |
Data are n (%) except where otherwise indicated.
The prevalence of diabetes device adoption in the initial 6 months after diagnosis also varied by age-group. Among patients <26 years of age (n = 818), 78 (9.5%) used CSII only, 69 (8.4%) used CGM only, and 147 (18.0%) used both CGM and CSII. In patients ≥26 years of age (n = 282), 71 (25%) used CSII only, 36 (12.8%) used CGM only and 50 (17.7%) used both CGM and CSII.
Mixed-Effect Modeling Results
Figures 3A and 3B show the longitudinal data of A1C across patient groups using a mixed-effects model over the 2-year evaluation period. No group-by-time effects were identified in any model; thus, the interaction terms were removed from the final models. There were significant time effects (Ps <0.001), and A1C worsened during the 2-year period, and these differences varied by CGM users and non-CGM users. After adjusting for ethnicity, insurance status, and baseline A1C, CGM use was associated with lower A1C compared with SMBG use (LS mean ± SE, A1C 7.89 ± 0.07 vs. 8.06 ± 0.04, P = 0.039). A nonsignificant marginal effect was observed between CSII users and MDI users (7.96 ± 0.06 vs. 8.10 ± 0.05, P = 0.057) after adjusting for age, ethnicity, insurance status, and baseline A1C. The details of the two mixed-effects models are available in Supplementary Tables S2 and S3.
A) A1C over time for CGM users and nonusers. B) A1C over time for CSII users and nonusers. C) A1C over time across four groups using CGM or SMBG in combination with CSII or MDI. Models accounted for confounders, including age at disease diagnosis, sex, race, ethnicity, and health insurance.
A) A1C over time for CGM users and nonusers. B) A1C over time for CSII users and nonusers. C) A1C over time across four groups using CGM or SMBG in combination with CSII or MDI. Models accounted for confounders, including age at disease diagnosis, sex, race, ethnicity, and health insurance.
Figure 3C displays the A1C among the four groups during the 2-year evaluation period. Similarly, there was a significant time effect, with A1C increasing over time across the levels of diabetes device groups (LS mean ± SE A1C for time 1,7.82 ± 0.05; time 2,7.97 ± 0.05; time 3,8.04 ± 0.05; and time 4,8.09 ± 0.05; P <0.001), and no group-by-time interactions were found among groups. Among CGM users, CSII users had a significantly lower A1C than MDI users over time (7.78 ± 0.09 vs. 8.14 ± 0.11, P = 0.010). Also, patients using CSII in addition to CGM had a significantly lower A1C compared with patients using MDI and SMBG (7.78 ± 0.09 vs. 8.09 ± 0.05, P <0.001), after adjusting for insurance status and baseline A1C. No impact was observed on A1C levels between CGM and SMBG groups in CSII users (7.78 ± 0.09 vs. 7.91 ± 0.10, P = 0.299) or MDI users (8.14 ± 0.11 vs. 8.09 ± 0.05, P = 0.705). The model results for A1C over the 2-year period by group can be found in Supplementary Table S4.
Discussion
To our knowledge, this study represents the first use of RWD to assess the prevalence of medical devices used to monitor and treat diabetes in routine clinical care. In our population, we found that CGM had superior A1C-lowering effects compared with SMBG, and CGM plus CSII had the best A1C-lowering effect overall. Although CGM and CSII use were increasing over the study period, adoption rates were found to be higher among White and privately insured patients. Furthermore, the NLP pipeline and ML classifiers achieved good performance and can be applied in future research for capturing medical device information.
This study provides evidence on the feasibility of using RWD to assess and monitor the prevalence and effectiveness of medical devices in the real-world setting and offers opportunities for expanding clinical assertions beyond RCTs. Understanding the adoption discrepancy and effects of medical devices in patients with diabetes is essential to addressing the growing disparities in device adoption and clinical outcomes.
In this study, CGM users had a lower A1C than non-CGM users, and the method of insulin delivery (CSII vs. MDI) had no effect on A1C over the 2-year evaluation period. These findings suggest that CGM initiation soon after disease diagnosis may play a more important role than CSII in long-term glycemic control. Recent studies also recommend CGM use over CSII use as the first-line technology for type 1 diabetes treatment, especially soon after diagnosis, considering its effectiveness, safety, lower cost, and improved quality of life for patients with type 1 diabetes (1,19,20).
Our results demonstrate that overall initiation of CGM within the first 6 months of diagnosis improved A1C compared with CGM nonusers; however, further analysis of the four groups shows that the difference may be the result of significantly lower A1C levels observed in the CGM + CSII group than in the SMBG + MDI group. These findings are consistent with results of a previous systematic review of RCTs (11) that patients with type 1 diabetes using CSII in addition to CGM lowered A1C more than patients using MDI and SMBG. These results suggest that initiation of CGM plus CSII, or at a minimum CGM, for glucose monitoring immediately after diagnosis may translate into better blood glucose control over the long term.
The NLP classification algorithm shows high accuracy in capturing information on patients using glucose monitoring and insulin delivery approaches. Our results provide important information on CGM and CSII use in patients with type 1 diabetes in the real-world setting—information that can be essential for clinical trial identification and recruitment. Furthermore, it is exactly this type of information that is demanded by the 21st Century Cures Act (13) for the augmentation of RCTs to address costs and generalizability. The goal of this study was to perform a rigorous assessment of the use of diabetes technologies in routine care. By applying artificial intelligence technologies to the EHR, we provide insights into the adoption and disparities in the use of diabetes technologies in the routine care setting.
In our population, we found the prevalence of CGM and CSII was increasing over time, similar to other observational studies (20). In addition, data from the T1D Exchange clinic network shows that ∼60% of their participants used CSII and 30% used CGM between 2016 and 2018. These estimates are likely higher than in our population because patients in the T1D Exchange clinic registry were seen by an endocrinology specialist, whereas not all patients with type 1 diabetes are treated by a specialist, especially underserved patients.
Importantly, we identified factors associated with the use of CGM and CSII devices after type 1 diabetes diagnosis. We observed that older patients had a higher percentage of CGM or CSII use compared with patients <26 years of age, a finding consistent with a previous study (21). Other studies have reported variations of CSII and CGM adoption related to socioeconomic status, with CGM more common in patients of higher socioeconomic status (20,22), and our findings were consistent with these studies.
Given the additional barriers to optimal diabetes care observed in minority groups (23), addressing the differences in access to CGM and CSII may help reduce disparities in patient outcomes. These data highlight opportunities for targeted quality improvement efforts in settings where certain populations have lower adoption rates of these diabetes medical devices. Understanding the reasons behind the lack of uptake of or barriers to CGM and CSII use across populations may improve patient care and outcomes.
An increase in A1C values was observed for the type 1 diabetes cohort during the 2-year period. This increasing trend in A1C has been observed in other studies showing glucose control becoming more and more challenging over time. A recent study examined the A1C trajectory over 18 months in pediatric patients after type 1 diabetes diagnosis and showed a steadily rising trend in A1C starting around the fifth month (24). Another study of newly diagnosed patients with type 1 diabetes of all age-groups showed a similar increasing trend in A1C from 6 months after diagnosis until 2.5 years after diagnosis (25). These findings suggest that intensifying management targeting the new-onset period may provide an opportunity to flatten the A1C curve and improve long-term glucose control.
Limitations
Several limitations of our study need to be acknowledged. We defined patients’ diagnosis date as the first appearance of a type 1 diabetes diagnosis in the EHR. This information may not be accurate and may include some patients who were diagnosed previously but had past disease management experiences missing from the EHR. This study was performed in a single health care system, and the results may not be generalizable to other institutions. However, our health system serves a wide range of medical facilities, including academic medical centers and community hospitals, with >60 clinics across the greater San Francisco Bay area. Furthermore, we were unable to detect the frequency of CSII and CGM use or whether patients switched or stopped using the insulin delivery or glucose monitoring devices. Our findings favor the use of both CGM and CSII to lower A1C in early type 1 diabetes; further evidence is needed on the effectiveness and safety of combined systems such as sensor-augmented pumps and automated insulin delivery systems as they become increasingly available.
Despite these limitations, this work provided a baseline prevalence for the use of diabetes technologies and established RWD with the support of NLP in identifying such information from clinical notes. Future efforts are needed to validate these results and support continuous evaluation as technologies evolve.
Conclusion
This work uses RWD and innovative methodologies to assess the adoption of diabetes technologies at the population level. The results from this study suggest that CGM initiation early in type 1 diabetes plays an important role in long-term glycemic control. We showed that the initiation of CGM in conjunction with CSII early in the disease course is also effective in lowering blood glucose and may translate to improved long-term glycemic control. Because glucose control becomes more challenging over the disease course, early initiation of CGM at diagnosis should be considered for patients with type 1 diabetes regardless of the insulin delivery approach they use. As increasing evidence supports the use of technologies for glucose monitoring and insulin delivery, more efforts are needed to understand disparities in utilization of these devices, and multi-stakeholder partnerships involving insurance payers and policy makers should be considered to facilitate the adoption of such devices.
Article Information
Acknowledgments
The authors thank Tina Seto, MS, for her support of the research data and clinical note extraction.
Duality of Interest
J.S. is the chief executive officer and an employee of PhysioLogic Devices. No other potential conflicts of interest relevant to this work were reported.
Author Contributions
R.S., J.S., and T.H.-B. designed the study. I.B. and S.S. developed the NLP pipeline and ML classifier. R.S. and J.S. provided the list of terms for diabetes devices. R.S. and J.J. performed chart reviews to check patients’ type of diabetes. R.S. performed the data analysis and wrote the first manuscript draft. All authors critically edited the manuscript and approved the final version. R.S. and T.H.-B. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
This article contains supplementary material online at https://doi.org/10.2337/figshare.13641302.