A Risk Stratification Approach to Allocating Diabetes Education and Support Services

Abstract

The objective of this study was to describe a predictive modeling approach to risk stratify people with type 2 diabetes for diabetes self-management education and support (DSMES) services. With data from a large health system, a predictive model including age, glycated hemoglobin (HbA1c), and insulin use among other factors, was developed to assess risk of future high HbA1c. The model was retrospectively applied to a cohort of people who received DSMES over a 2-year period to assess the impact of DSMES on glycemia by risk strata. Of 6934 eligible people, 4014 (58%) were in the composite low-risk group and 2604 (38%) were in the composite high-risk group. Mean HbA1c change after DSMES was −0.38% in the low-risk group and −0.84% in the high-risk group. This analysis demonstrates the potential application of predictive modeling as one approach to target DSMES resources to people who will benefit most.

Introduction

Diabetes self-management education and support (DSMES) is the ongoing process of facilitating the knowledge, skill, and ability necessary for diabetes self-care and is fundamental to achieving treatment goals for people with diabetes mellitus.^1,2 It has been well established that DSMES improves diabetes care quality and glycemia,^3,4 self-management behaviors,^5,6 and psychosocial outcomes such as diabetes distress.⁷ Although the benefits of DSMES are clear, access to this resource can be a challenge due to a number of reported barriers that include low referral rates, poor reimbursement, location of service delivery, and limited appreciation of DSMES services.^8
–10 This has influenced DSMES program closings, which can affect access and job opportunities with fewer health care professionals available to provide this service. These barriers are complicated by the rapid growth in the number of available therapeutics and personal device technology that require effective diabetes self-management skills, as well as a growing number of people with multiple comorbid conditions that increase the complexity of diabetes care.^11,12 Given these dynamics, population health management (PHM) strategies become even more relevant to direct limited DSMES resources to people who are at high risk of diabetes complications.

PHM is one approach to optimize health outcomes and resource utilization that has been applied to diabetes care delivery.¹³ PHM includes identification of a target population with clear criteria, health assessment using a “data warehouse” that integrates clinical data for populations of interest, data-driven risk stratification to prioritize patients, and delivery of interventions tailored to the needs of the target population.¹⁴ PHM is gaining widespread attention in determining processes for systematic improvements in health outcomes and value-based care and can be integrated into many components of diabetes care delivery, including registries of people with diabetes, proactive identification of those at highest risk for services, team-based diabetes care to provide patient-centered interventions, and decision support tools to deliver guideline-based care.¹³

Predictive modeling is one technique that has recently been introduced and integrated into clinical information systems to support PHM. This mathematical process uses data and statistical analysis to identify patterns for risk stratification and the prediction of future outcomes. By relying on data and clinical information systems to help stratify a large population of patients, predictive modeling has the potential to help to prioritize people with diabetes who could benefit most from receipt of DSMES and other services. The purpose of this study was to test the applicability of a PHM data-driven predictive modeling approach for allocating DSMES services to those at highest risk of poor diabetes outcomes.

Methods

Model development

Members of the University of Pittsburgh Medical Center (UPMC) Division of Endocrinology and Metabolism partnered with the Clinical Data Analytics team to first develop a predictive modeling approach. The predictive model was designed to predict the risk of future elevated glycated hemoglobin (HbA1c) for people with diabetes. The model utilized data from the patient's electronic health record stored in the UPMC Clinical Data Warehouse.

Patients with a documented diagnosis of type 2 diabetes mellitus (T2DM), as well as any patient with two HbA1c results >6.5% (48 mmol/mol) within 1 year, were included in the cohort of the 38,173 patients used to develop the model. In the final model, the three variables with the greatest weight included patient HbA1c (at the time of risk stratification), age, and insulin use. Higher HbA1c, younger age, and prior insulin use all predicted higher risk of future elevated HbA1c. The model was then applied retrospectively to risk stratify a distinct cohort of patients with T2DM who had received DSMES within the UPMC system. Although this group of patients may have included some of the same patients as the predictive modeling cohort, this was not a direct subset of that cohort.

Model application to DSMES

Patients with T2DM who received DSMES at a UPMC-affiliated primary care or endocrinology practice from March 2019 to February 2021were included in the cohort for retrospective risk assessment. The predictive model was applied using clinical data available for each patient immediately before their initial DSMES session during the study period to stratify patients by risk of high HbA1c in the future. Information including vital signs and laboratory values such as HbA1c, diabetes-related medication prescriptions, and comorbid diagnoses was collected to describe patient characteristics.

Glycemia was assessed based on the change in HbA1c values before and after DSMES over the study period. DSMES was defined as a clinical encounter with a certified diabetes care and education specialist documented in the electronic medical record. Encounters could be face-to-face or through telemedicine and were primarily transitioned to telemedicine after March of 2020 because of the coronavirus disease 2019 (COVID-19) pandemic. This study received approval from the UPMC Quality Improvement Review Committee (Project ID #3246), and thus IRB oversight was waived.

Results

Of the total population of patients with T2DM who were seen within the UPMC system during the study period, 12.0% had received DSMES. Among those who were stratified as high risk for future high HbA1c, this percentage increased to 18.7%. For retrospective analysis, 6934 patients had both received DSMES during the study period and had complete data available for risk assessment. These patients had a total of 26,209 encounters of any type for diabetes care over the 24-month period; 63.9% of them had a visit with an endocrinologist, whereas the remainder received diabetes treatment from their primary care provider. Mean age of this cohort was 59.7 years, body mass index (BMI) of 34.5, and mean HbA1c of 8.3% (67 mmol/mol). Rates of comorbidities reported in this patient population were hypertension (75.9%), depression (19.0%), coronary artery disease (22.4%), congestive heart failure (10.9%), cancer diagnosis (10.7%), chronic kidney disease (6.7%), and end-stage renal disease (3.3%). Long-acting insulin was prescribed to 56.3% of patients, and 46.9% were prescribed rapid-acting insulin.

Of the patients with T2DM included in the retrospective risk assessment, 4014 were stratified as low risk (composite category of low, lower, or lowest risk strata) based on their clinical data before their initial DSMES visit during the study period. With a mean HbA1c of 7.5% (58 mmol/mol), 68.6% of these people received diabetes care from an endocrinologist (see detailed characteristics by risk strata in Table 1). Rates of long-acting insulin (43.2%) and rapid-acting insulin (48.8%) use within this patient population were similar to those of the overall cohort. There were 2604 patients categorized as high risk (composite category of high, higher, and highest risk strata), with mean HbA1c of 9.6% (81 mmol/mol). Of these patients, 56.8% received diabetes care from an endocrinologist (Table 1). More of the patients in this group were prescribed insulin, with 65.1% receiving long-acting insulin and 50.7% receiving rapid-acting insulin.

Table 1.

Patient Characteristics

Risk strata	Lowest	Lower	Low	Medium	High	Higher	Highest
N	2276	551	1187	316	2262	284	58
Mean HbA1c
%	7.3	8.6	7.1	9.6	9.3	10.5	11.7
mmol/mol	56	70	54	81	78	91	104
Mean age	63.5	57.9	63.8	56.4	56	52.8	50.4
Mean BMI	33.9	35.5	34.1	35.1	34.8	35.1	34.2
Seen by endocrinology	71.9%	69.7%	62.3%	70.3%	54.4%	74%	60.8%

HbA1c, glycated hemoglobin.

For patients stratified as low- risk for future high HbA1c, the mean change in HbA1c after DSMES was −0.38% (−4 mmol/mol, see detailed HbA1c change by risk strata in Fig. 1). Sixty-eight percent of people in this group had stable or improved HbA1c after DSMES. In comparison, among patients classified as high risk, mean change in HbA1c after DSMES was −0.84% (−9 mmol/mol). Eighty-two percent of these people had stable or improved HbA1c after DSMES.

FIG. 1.

HbA1c change after DSMES by risk strata. DSMES, diabetes self-management education and support; HbA1c, glycated hemoglobin.

Discussion

This study illustrates the applicability of predictive modeling to identify and risk stratify patients with T2DM who are most likely to benefit from DSMES services. Despite many advances in diabetes therapies over the past 15 years, there has been little progress in increasing the proportion of people with diabetes who meet treatment goals.¹⁵ In this study, people in the high-risk group had higher mean baseline HbA1c and BMI and had greater improvements in HbA1c after DSMES. This is consistent with a prior review of the impact of DSMES on glycemic outcomes when people were stratified by baseline HbA1c levels, which demonstrated that those with higher HbA1c values had the greatest improvement and sustained reduction in HbA1c results.¹⁶ With the growing population of people with diabetes, data-driven strategies to target resources toward populations with the highest risk for poor diabetes outcomes will be increasingly important.

Using data and clinical information systems as a resource to review patient clinical characteristics and treatment also affords the opportunity to tailor care to the needs of each patient. Predictive models have been used to forecast future health care needs in studies that have predicted the initial onset of T2DM, development of microvascular complications based on patient-specific clinical factors, and cost of diabetes-related complications based on HbA1c change at a population level.^17
–19

This study presents a novel application of predictive modeling to allocating DSMES based on risk of future poor diabetes outcomes. In this study, people in the high-risk group were prescribed insulin more frequently. DSMES can be especially important for these people given the complexity and risk profile of insulin treatment, as well as the significant education and ongoing support required for successful self-management.²⁰ In addition, with growing evidence of the utility of advanced diabetes technologies to help with glycemic management, people with T2DM are now being introduced to tools such as continuous glucose monitors, insulin pumps, and smart pens.^21,22 Although offering many advantages to people including convenience, flexibility, and additional data, these advanced treatment modalities also require an increasing amount of provider expertise and time to train people, provide self-management education, and monitor treatment. This study demonstrates the potential of using a data-driven technique to target DSMES services to those who will reap the most benefit when resources are limited.

This approach can be adapted by diabetes care programs in a variety of ways. In settings wherein access to DSMES is limited, patients with known risk factors for poor diabetes outcomes, such as those presented in this study, can be prioritized for these services. Protocols with automatic DSMES referrals for high-risk patients can be implemented to overcome the barrier of low referral rates. Increasingly, clinical analytics are also becoming available to guide PHM approaches to improve care and outcomes for people with diabetes. Relying on risk stratification to identify those who could most benefit from specialized diabetes services offers a unique and innovative opportunity for the provision of patient-centered high-quality care.

This study has multiple strengths and limitations. Although we recognize that all institutions do not have tools that may be available in a large integrated health system to develop an approach to tackle access to specialty services, these findings are the first to demonstrate future opportunities to apply predictive modeling to DSMES. In addition, the study presented utilizes real-world evidence from a large cohort of people with diabetes with large variations in glycemia, comorbidities, and treatment, and thus may represent a more realistic estimate of the impact of DSMES as the population includes many people who would be excluded from clinical trials.

However, as it is a retrospective cohort study performed in the context of a clinical operations initiative, there is no control group for comparison of changes in HbA1c for patients of different risk strata who did not receive DSMES. Second, due to the COVID-19 pandemic, DSMES delivery transitioned from all face-to-face to all majority telemedicine after March 2020. We were unable to associate the frequency or method of DSMES delivery with change in HbA1c, and thus are not able to assess for a possible relationship between intensity of DSMES or virtual delivery with improvement in glycemia in this study.

Conclusion

With an increasing focus on improving quality and value in chronic disease care, PHM strategies that harness the power of health care data to direct resources will be in high demand. Although the benefits of DSMES are well established, those at highest risk of poor diabetes outcomes are not consistently prioritized for these services. It is time to consider the use of novel tools such as predictive modeling to maximize the benefits of this critical component of diabetes care for people with high-risk diabetes. Innovative approaches like the model presented in this study will have a growing role in the future of individualized high-value diabetes and chronic disease care.

Footnotes

Authors' Contributions

M.F.Z., J.K., J.M.N., and L.S. conceived of and designed the study. K.C. and O.M. contributed to design of predictive model and acquisition of data. M.F.Z. and J.M.N. performed data analysis. M.F.Z., J.K., J.M.N., and L.S. contributed to interpretation of data and drafted the article. All authors critically reviewed the article for intellectual content.

Author Disclosure Statement

M.F.Z., K.C., and O.M. have no relevant disclosures to report. J.K. reports research support from Sanofi Aventis and Becton, Dickinson and Company. J.M.N. reports research support from Sanofi Aventis. L.S. reports research support from Sanofi Aventis and Becton, Dickinson and Company.

Funding Information

Research funding from Becton, Dickinson and Company supported this study. Dr. Zupa was supported by a T32 grant from NIDDK (5T32DK007052-45).

References

Powers

, Bardsley

, Cypress

, et al.: Diabetes self-management education and support in adults with type 2 diabetes: a consensus report of the American Diabetes Association, the Association of Diabetes Care & Education Specialists, the Academy of Nutrition and Dietetics, the American Academy. Diabetes Care, 2020; 43:1636–1649.

Funnell

, Brown

, Childs

, et al.: National standards for diabetes self-management education. Diabetes Care, 2007; 30:1630–1637.

Chrvala

, Sherr

, Lipman

: Diabetes self-management education for adults with type 2 diabetes mellitus: a systematic review of the effect on glycemic control. Patient Educ Counsel, 2016; 99:926–943.

Brunisholz

, Briot

, Hamilton

, et al.: Diabetes self-management education improves quality of care and clinical outcomes determined by a diabetes bundle measure. J Multidiscip Healthc, 2014; 7:533–542.

Tang

, Funnell

, Oh

: Lasting effects of a 2-year diabetes self-management support intervention: outcomes at 1-year follow-up. Prev Chronic Dis, 2012; 9:E109.

Adjei Boakye

, Varble

, Rojek

, et al.: Sociodemographic factors associated with engagement in diabetes self-management education among people with diabetes in the United States. Public Health Rep, 2018; 133:685–691.

Sperl-Hillen

, Beaton

, Fernandes

, et al.: Comparative effectiveness of patient education methods for type 2 diabetes: a randomized controlled trial. Arch Intern Med, 2011; 171:2001–2010.

Alsayed Hassan

, Curtis

, Kerver

, Vangsnes

: Diabetes self-management education and support: referral and attendance at a patient-centered medical home. J Prim Care Commun Health, 2020; 11:2150132720967232.

Carey

, Agarwal

, Horne

, et al.: Exploring organizational support for the provision of structured self-management education for people with Type 2 diabetes: findings from a qualitative study. Diabetic Med, 2019; 36:761–770.

10.

Siminerio

, Hamm

, Kanter

, et al.: A diabetes education model in primary care: provider and staff perspectives. Diabetes Educ, 2019; 45:498–506.

11.

Daly

, Hovorka

: Technology in the management of type 2 diabetes: present status and future prospects. Diabetes Obes Metabl, 2021; 23:1722–1732.

12.

Piette

, Kerr

: The impact of comorbid chronic conditions on diabetes care. Diabetes Care, 2006; 29:725–731.

13.

Schmittdiel

, Gopalan

, Lin

, et al.: Population health management for diabetes: health care system-level approaches for improving quality and addressing disparities. Curr Diabetes Rep, 2017; 17:31.

14.

Struijs

, Drewes

, Heijink

, Baan

: How to evaluate population management? Transforming the Care Continuum Alliance population health guide toward a broadly applicable analytical framework. Health Policy, 2015; 119:522–529.

15.

Kazemian

, Shebl

, McCann

, et al.: Evaluation of the cascade of diabetes care in the United States, 2005-2016. JAMA Internal Medicine. 2019; 179:1376–1385.

16.

Siminerio

, Koshinsky

, Payne

, et al.: Improving Glycemia in Primary Care: A Model Integrating Diabetes Education into Practice. Boston, MA: Poster presented at American Diabetes Association, 2015.

17.

Mani

, Chen

, Elasy

, et al.: Type 2 diabetes risk forecasting from EMR data using machine learning. AMIA Annu Symp Proc, 2012; 2012:606–615.

18.

Dagliati

, Marini

, Sacchi

, et al.: Machine learning methods to predict diabetes complications. J Diabetes Sci Technol, 2018; 12:295–302.

19.

Fortwaengler

, Parkin

, Neeser

, et al.: Description of a new predictive modeling approach that correlates the risk and associated cost of well-defined diabetes-related complications with changes in glycated hemoglobin (HbA1c). J Diabetes Sci Technol, 2017; 11:315–323.

20.

Yehl

: AADE practice paper in brief: diabetes educators play a critical role in successful insulin management. AADE Pract, 2018; 6:36–37.

21.

Kompala

, Neinstein

: A new era: increasing continuous glucose monitoring use in type 2 diabetes. Am J Manag Care, 2019; 25(4 Spec No.):SP123–SP126.

22.

Freckmann

, Buck

, Waldenmaier

, et al.: Insulin pump therapy for patients with type 2 diabetes mellitus: evidence, current barriers, and new technologies. J Diabetes Sci Technol, 2021; 15:901–915.