Predicting Severe Diabetes Complications Using Administrative Claims Data in Maryland

ABSTRACT

Objective: To describe operations and performance of a large-scale predictive model of severe type 2 diabetes complications (DC) for Medicare beneficiaries in Maryland.

Study Design: Retrospective longitudinal multivariable regression.

Methods: Using Medicare fee-for-service (FFS) claims from March 2021 to July 2024, we created an analytic data set of 219 candidate risk factors spanning 12,611,063 person-months. Multivariable discrete-time survival modeling was used to assess the relation between risk factors and the risk of incurring a future hospitalization due to severe DC. Using stepwise variable selection to retain only statistically significant risk factors, DC risk scores were created by applying training coefficients to risk factors from the most recently available month of data. Risk scores for 346,614 individuals were released on June 7, 2024. Predictive performance of scores was assessed using actual events in the month following the release of the scores, compared with the performance of Hierarchical Condition Category (HCC) scores.

Results: As of April 2024, the model retained 95 significant risk factors. The mean risk score was 0.0124, and utilization- and condition-based risk factors primarily accounted for the top 15 risk factors. The June 2024 risk scores were strongly predictive of true events: Individuals with the top 10% of DC risk scores accounted for 56.9% of severe DC events in the following month. This significantly outperformed HCC scores, which accounted for 37.5% of true events in the same period.

Conclusions: A risk prediction model based on administrative claims can predict severe DC events for the Medicare FFS population in Maryland.

Am J Manag Care. 2026;32(3):e66-e70. https://doi.org/10.37765/ajmc.2026.89898

Takeaway Points

Diabetes is costly and prevalent, affecting 1 in 9 individuals in the US. Maryland has prioritized diabetes care in its Statewide Integrated Health Improvement Strategy, and in this study, we document the development and performance of a predictive model of severe diabetes complications currently in production for Medicare fee-for-service beneficiaries in Maryland.

• This predictive model is based on administrative claims and publicly available social determinants of health data.
• The model has strong predictive performance, underscoring the potential value of incorporating such models in public health initiatives.
• This model outperforms hierarchical condition category scores, a more general, widely available risk metric.

Diabetes is both prevalent and costly. Diabetes affects 1 in 9 individuals in the US, and it is estimated that, as of 2021 in the US, 38.1 million adults had diabetes, 23% of whom were undiagnosed.¹ Moreover, in 2022, the estimated total cost of diabetes was $413 billion. This consisted of $307 billion in direct medical costs and $106 billion in indirect costs.¹ Prior research data show that, when not effectively managed, diabetes can result in serious complications such as peripheral neuropathy, nephropathy, retinopathy, and macrovascular complications.^2,3

An important aspect of many diabetes complications (DC) is that they are, in theory, preventable. Preventable conditions are those “for which timely and effective outpatient care can help to reduce the risks of hospitalization by either preventing the onset of an illness or condition, controlling an acute episodic illness or condition, or managing a chronic disease or condition.”⁴ If a patient with diabetes receives appropriate and timely care in an outpatient setting and adheres to recommended treatment regimens, the patient should have minimized their chances of developing complications. In reality, however, diabetes is underdiagnosed, and those who do receive a diagnosis often develop complications requiring inpatient care; potential drivers for this include limited financial resources, health literacy, and access to care.⁵ Primary care plays a central role in diabetes care, with primary care providers often serving as a first-line resource for managing a patient’s treatment regimen and monitoring their risk for developing severe complications.^6,7 This makes primary care settings ideal touchpoints for interventions aimed at improving health outcomes for patients with diabetes.

As part of Maryland’s Statewide Integrated Health Improvement Strategy, which was active from 2021 to 2024, the state aimed to address diabetes-related health needs.⁸ The Maryland Primary Care Program (MDPCP), a key element of the Maryland Total Cost of Care model, supports the delivery of advanced primary care through the use of precision event risk models for the community-dwelling Medicare fee-for-service (FFS) population.⁹ One such model estimates the risk of incurring a severe complication from type 2 diabetes within the next month. Participating practices have access to this risk score for all Medicare FFS patients in their panels, as well as the main risk factors driving each patient’s risk level, through the state’s health information exchange, Chesapeake Regional Information System for our Patients (CRISP). With this information, care managers and providers can identify patients at risk of severe DC events and theoretically intervene proactively before those complications occur. Importantly, these risk scores are designed to support, not replace, provider experience and judgment when making clinical decisions. This predictive model has been in production for approximately 350,000 Medicare FFS beneficiaries each month since October 2022.

Notably, although other diabetes-related predictive models exist, this is the first such model in production for a large-scale, publicly insured population of which we are aware.^10-14 This study describes the development and performance of this model. We believe that other states can benefit from this use case of precision medicine based on commonly available administrative data.

METHODS

The development and validation of DC event risk scores occurs in 3 distinct steps: training, scoring, and assessment. The model training occurs quarterly and is based on 35 months of Medicare FFS parts A, B, and D claims data for individuals attributed to MDPCP-participating primary care practices. These claims data are transformed into 219 person-month risk factors, which are then used to predict emergency department (ED) visits and inpatient stays related to severe complications of type 2 diabetes the following month. The risk factors are based on an extensive literature review and include demographic factors, environmental social determinants of health, utilization- and condition-based factors, and pharmacy utilization–based factors.^15,16 For a full list of risk factors, see eAppendix Table 1 (eAppendix available at ajmc.com).

The outcome for the model is a binary (0/1) variable indicating whether a beneficiary incurred a severe DC event in the following month, which was based on the severe complications defined by the Diabetes Complication Severity Index (DCSI), an index of retinopathy, nephropathy, cerebrovascular complications, cardiovascular complications, peripheral vascular disease, and metabolic complications.^17,18 The traditional DCSI includes both claims and laboratory data to measure DC severity. However, our predictive model uses administrative data that do not include laboratory results. Prior research has validated using only International Classification of Diseases, Ninth Revision or International Statistical Classification of Diseases, Tenth Revision claims to measure DC severity.^19,20

We used a discrete-time survival model trained on 80% of the sample in our analytic person-month data set to estimate coefficients for each risk factor. During model training, we used a stepwise variable selection approach so that only statistically significant risk factors (defined as P values < .000579) are retained. The production process uses 4 separate models to account for incomplete information for certain beneficiaries, but we focused on the primary specification for this study, which is defined as the model for individuals with at least 12 months of claims history and valid geographic identifiers (99.0% of beneficiaries in the scoring cohort).

Risk scores are generated monthly. Using the same definitions as in the model training phase, we created risk factors for the most recently available month of data and applied the most recent available set of coefficients from the training phase to create beneficiary-level risk scores. The risk scores are then sent to primary care practices via Maryland’s health information exchange, CRISP. MDPCP-participating practices receive regular training regarding how to access the DC scores and incorporate them into their clinical workflow. Through CRISP, primary care providers and care managers can access risk scores for Medicare FFS beneficiaries seen in their practice. The methods and processes have been described in greater detail elsewhere.^21,22

Finally, to evaluate the performance of our predictive model, we assessed the predictive value using the concentration curve. This graphically presents the cumulative share of outcome events—in this case, ED visits or inpatient admissions due to severe DC—by descending levels of risk. A concentration curve is an appropriate model performance metric for models that perform risk ranking as opposed to risk classification, where an area under the curve performance metric is commonly used.^23,24 Although this can be calculated in the holdout sample (which, in this model, is 20% of beneficiaries), for this study, we calculated this in the production environment by linking risk scores to actual severe DC events in the following month.

As an additional benchmark for the DC event model performance, we compared the performance of our DC event model against a widely available risk metric, the CMS Hierarchical Condition Category (HCC) score. HCC scores were developed as a capitated payment risk adjustment methodology for Medicare Advantage participants and are designed to index general health needs and future health care costs.^25,26 The HCC score for each beneficiary is provided by CMS and presented in the same CRISP reporting suite as the DC event model risk scores, so providers have access to both risk scores. HCC scores are calculated and reported on a quarterly basis using a rolling 12-month lookback period of claims data. To allow for sufficient claims runout, there is a 1-year gap from the end of the lookback period before scores are calculated and reported for use. For our comparison, we used the HCC scores provided in April 2024.

For the model training conducted in April 2024, the analytic data set consisted of 12,611,063 person-month observations. The coefficients generated during the training were used to create the scores released on June 7, 2024, for 346,614 individuals. We assessed model performance against actual events from June 8, 2024, to July 7, 2024. This work was exempt from institutional review board review because it met the criteria for exemption 4 (secondary research that does not require consent) as it involved secondary use of identifiable health information when use is regulated by the Health Insurance Portability and Accountability Act as “health care operations,” “research,” or “public health activities and purposes” (45 CFR 46.104(d)(4)(iii)) and presented no risk or minimal risk to participants. The risk scores were calculated using SAS 9.4 (SAS Institute Inc), and the analyses for this study were conducted using Stata 18 (StataCorp LLC).

RESULTS

The DC event model converged normally, and 95 of the potential 219 risk factors were considered statistically significant (defined as P < .000579) and thus retained. Table 1 shows the risk factors with the highest ORs from model training in April 2024. The top predictor was based on the outcome of previous hospitalizations, followed by other utilization-based and diagnostic risk factors. See eAppendix Table 2 for complete results.

The mean risk score across all beneficiaries was 0.0124, representing an approximate likelihood of incurring a severe DC event in the following month of 1.24%. This is consistent with the actual percentage of beneficiaries in the sample who experienced DC events (1.25%) in the month used to assess model performance (6/8/2024-7/7/2024). Summary statistics of the risk scores stratified by demographic groups can be found in Table 2. On average, male, Black, and older beneficiaries had higher risk scores. As would be expected, considering the low frequency of DC events in a given month, there were relatively few beneficiaries with high absolute levels of risk. Approximately 2% of beneficiaries had risk scores greater than 0.1, indicating a greater than 10% likelihood of incurring a severe DC event in the next month, and fewer than 0.5% of beneficiaries had risk scores greater than 0.5.

The assessment phase demonstrated strong predictive performance of the model. Following the release of the risk scores on June 7, 2024, we used administrative claims data to identify actual severe DC events (ie, the outcome predicted by the model) that occurred in the following month to analyze the predictive performance of the model in temporally distinct data. Crucially, these events were neither used in model training nor in the calculation of the scores; as such, this quantifies model performance in the production environment. We evaluated whether individuals with higher DC event model risk scores in June 2024 accounted for a greater fraction of true DC events in the following month than would be expected due to chance.

We found that the predictive model for DC events performed well, both in absolute terms and relative to HCC scores, a widely used risk measure designed to index general health needs and future health care costs. In the Figure, we plot the concentration curve for both the DC event model risk scores from June 2024 and the HCC scores from April 2024. The solid curve represents the performance of the DC event model risk scores, and the dashed blue line represents the performance of the CMS HCC scores during the same period. During the evaluation period of June 8, 2024, to July 7, 2024, there were 4010 DC events incurred by beneficiaries in the sample. The top 10% highest-risk individuals identified by the DC event model risk score accounted for 2283 (56.9%) of the observed DC events, and the top 20% accounted for 2987 (74.5%) of DC events. The HCC scores underperformed the DC event model scores when predicting DC events: The top 10% highest-risk individuals defined by HCC scores accounted for 1503 (37.5%) of true DC events, and the top 20% accounted for 2272 (56.7%). These statistics are important not only because they are indicators of predictive performance, but also because, in practice, individuals in the highest-risk group are prioritized for intervention and care coordination purposes when care management resources are limited.

DISCUSSION

This study describes the development and performance of a predictive model of severe DC events in production for approximately 350,000 Medicare FFS beneficiaries in Maryland. The mean risk score was 0.0124, and the most salient risk factor was “prior hospitalization discharge status: other,” followed by other utilization-based and diagnostic risk factors. This model performed well in a production environment: The top 10% highest-risk individuals according to scores released on June 7, 2024, accounted for 56.9% of true severe DC events in the following month. Further, the DC event model scores outperformed HCC scores, a more general index of anticipated health care need, when predicting severe DC events by a considerable margin. Although this is expected, because the DC event model is optimized to predict this specific outcome, we believe that these results underscore the value of the DC event model, as well as its risk scores, as a tool for supporting diabetes management in a primary care setting.

Individual-level risk scores may help primary care providers and staff statistically triage their patient panels and efficiently allocate limited time and care management resources to those with the greatest need. Results from this analysis suggest that if care managers focus on beneficiaries in the highest 10% of DC event risk scores, they will be able to reach almost 60% of patients who will have a severe DC event within the next month. Additionally, based on feedback from primary care providers receiving risk scores, providers receive each individual’s risk score and the top “reasons for risk” that drive the risk score. Notably, individuals with similar risk scores can have different risk drivers. For example, one individual’s risk may be driven by a history of heart failure and hypertension, whereas another individual’s risk might be driven by a recent hospitalization with discharge status of “other.” Although the risk factor coefficients—and reasons for risk—are not causal drivers of the DC outcome, they may be of use to providers and care managers in designing personalized interventions to mitigate the risk of incurring the DC outcome.

More broadly, the specificity of the DC event model scores relative to other, more general risk scores may facilitate efforts to meet policy-driven quality goals related to diabetes care at the practice level and improve population health outcomes. For example, a theoretical intervention program designed to provide care management resources to patients with a high risk of DC, focusing on the top 10% of highest-risk individuals as defined by the event DC scores, would reach more patients with actual severe DC events than if they focused on the top 20% as defined by the HCC scores.

Limitations

This model has 3 primary limitations. First, we used only administrative claims data and publicly available data on social determinants of health; we did not use clinical data. Although certain fields—for example, hemoglobin A_1c levels—would almost certainly improve model performance, they were not available for this population. However, the relatively standardized structure of administrative claims data used in the development of this model implies that this model could, in principle, be adapted for other populations (eg, state Medicaid programs or private payers). Relatedly, as of the time of writing, the state of Maryland is planning to deploy this model in 3 additional publicly insured populations to support the implementation of the recently approved AHEAD (Achieving Healthcare Efficiency through Accountable Design) Model: Medicare FFS non-MDPCP beneficiaries, Medicaid managed care beneficiaries, and Medicaid FFS beneficiaries.²⁷

Second, although we can monitor the model’s statistical performance, we have limited visibility into how providers use the risk scores. A recent evaluation of MDPCP found that the introduction of the program led to reductions in inpatient and ED utilization, which may in part be due to the predictive risk scores available to providers, but did not attempt to isolate the effects of the predictive model risk scores alone.²⁸ We encourage future researchers to further explore the ways in which practices incorporate predictive modeling risk scores into decision-making processes regarding patient care.

Finally, we calculated risk scores for the entire MDPCP-attributed population, even though only a portion of it has been diagnosed with diabetes. Although this is operationally expedient, this is also intentional: It is, in principle, possible for an individual to have undiagnosed diabetes or not have a recent claim in which their diabetes diagnosis is included, which may make it appear the individual did not have diabetes. However, this individual may still be at high risk of incurring a severe DC event. We sought to avoid circumstances in which such individuals “fall through the cracks” by estimating risk scores for all.

CONCLUSIONS

In this study, we illustrated the development and operation of a predictive model for severe DC events currently deployed in the Maryland Medicare FFS population attributed to providers participating in the MDPCP. The prevention of DC is a public health priority in Maryland, and this study provides evidence that predictive models of severe DC events based on administrative data may be useful in helping providers to proactively optimize care, inform care management practices, and eventually reduce the incidence of adverse health events and thereby contain medical costs. 

Acknowledgments

The authors would like to thank Chad Perman for his valuable feedback.

Author Affiliations: The Hilltop Institute, University of Maryland, Baltimore County (LG, DB, FH, MAH), Baltimore, MD; Department of Psychology, University of Maryland, Baltimore County (RS), Baltimore, MD.

Source of Funding: Funding for the project came from the Maryland Department of Health and the Maryland Primary Care Program.

Author Disclosures: The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Authorship Information: Concept and design (LG, FH, RS, MAH); analysis and interpretation of data (LG, DB, FH, RS, MAH); drafting of the manuscript (LG, DB, RS, MAH); critical revision of the manuscript for important intellectual content (LG, MAH); statistical analysis (LG, FH, MAH); and administrative, technical, or logistic support (LG, DB, MAH).

Address Correspondence to: Leigh Goetschius, PhD, University of Maryland, Baltimore County, 1000 Hilltop Circle, Sondheim Hall, Third Floor, Baltimore, MD 21250. Email: lgoetschius@hilltop.umbc.edu.

REFERENCES

1. Diabetes statistics. National Institute of Diabetes and Digestive and Kidney Diseases. January 2024. Accessed January 6, 2025. https://www.niddk.nih.gov/health-information/health-statistics/diabetes-statistics

2. Cloete L. Diabetes mellitus: an overview of the types, symptoms, complications and management. Nurs Stand. 2022;37(1):61-66. doi:10.7748/ns.2021.e11709

3. Papatheodorou K, Banach M, Edmonds M, Papanas N, Papazoglou D. Complications of diabetes. J Diabetes Res. 2015;2015:189525. doi:10.1155/2015/189525

4. Billings J, Zeitel L, Lukomnik J, Carey TS, Blank AE, Newman L. Impact of socioeconomic status on hospital use in New York City. Health Aff (Millwood). 1993;12(1):162-173. doi:10.1377/hlthaff.12.1.162

5. Nam S, Chesla C, Stotts NA, Kroon L, Janson SL. Barriers to diabetes management: patient and provider factors. Diabetes Res Clin Pract. 2011;93(1):1-9. doi:10.1016/j.diabres.2011.02.002

6. Kushner PR, Cavender MA, Mende CW. Role of primary care clinicians in the management of patients with type 2 diabetes and cardiorenal diseases. Clin Diabetes. 2022;40(4):401-412. doi:10.2337/cd21-0119

7. Shrivastav M, Gibson W Jr, Shrivastav R, et al. Type 2 diabetes management in primary care: the role of retrospective, professional continuous glucose monitoring. Diabetes Spectr. 2018;31(3):279-287. doi:10.2337/ds17-0024

8. Statewide Integrated Health Improvement Strategy Proposal. Maryland Health Services Cost Review Commission. December 2020. Accessed January 6, 2025. https://hscrc.maryland.gov/Documents/Modernization/SIHIS%20Proposal%20-%20CMMI%20Submission%2012142020.pdf

9. Maryland’s Total Cost of Care model. Maryland Health Services Cost Review Commission. Accessed January 6, 2025. https://hscrc.maryland.gov/Pages/tcocmodel.aspx

10. Dagliati A, Marini S, Sacchi L, et al. Machine learning methods to predict diabetes complications. J Diabetes Sci Technol. 2018;12(2):295-302. doi:10.1177/1932296817706375

11. Huang J, Yeung AM, Armstrong DG, et al. Artificial intelligence for predicting and diagnosing complications of diabetes. J Diabetes Sci Technol. 2023;17(1):224-238. doi:10.1177/19322968221124583

12. Ljubic B, Hai AA, Stanojevic M, et al. Predicting complications of diabetes mellitus using advanced machine learning algorithms. J Am Med Inform Assoc. 2020;27(9):1343-1351. doi:10.1093/jamia/ocaa120

13. Rubin DJ, Shah AA. Predicting and preventing acute care re-utilization by patients with diabetes. Curr Diab Rep. 2021;21(9):34. doi:10.1007/s11892-021-01402-7

14. Schiborn C, Schulze MB. Precision prognostics for the development of complications in diabetes. Diabetologia. 2022;65(11):1867-1882. doi:10.1007/s00125-022-05731-4

15. Pelser C, Henderson M, Stockwell I. Risk Factors for Potentially Avoidable Hospital Admissions and Emergency Department Visits: A Literature Review. The Hilltop Institute, UMBC; 2019. Accessed January 6, 2025. https://hilltopinstitute.org/wp-content/uploads/publications/MDPCPAvoidableHospitalizationLiteratureReview-April2019.pdf

16. Goetschius LG, Henderson M, Han F, et al. Assessing performance of ZCTA-level and Census tract-level social and environmental risk factors in a model predicting hospital events. Soc Sci Med. 2023;326:115943. doi:10.1016/j.socscimed.2023.115943

17. Glasheen WP, Renda A, Dong Y. Diabetes Complications Severity Index (DCSI)-update and ICD-10 translation. J Diabetes Complications. 2017;31(6):1007-1013. doi:10.1016/j.jdiacomp.2017.02.018

18. Young BA, Lin E, Von Korff M, et al. Diabetes complications severity index and risk of mortality, hospitalization, and healthcare utilization. Am J Manag Care. 2008;14(1):15-23.

19. Chang HY, Weiner JP, Richards TM, Bleich SN, Segal JB. Validating the adapted Diabetes Complications Severity Index in claims data. Am J Manag Care. 2012;18(11):721-726.

20. Wicke FS, Glushan A, Schubert I, et al. Performance of the adapted Diabetes Complications Severity Index translated to ICD-10. Am J Manag Care. 2019;25(2):e45-e49.

21. Henderson M, Han F, Perman C, Haft H, Stockwell I. Predicting avoidable hospital events in Maryland. Health Serv Res. 2022;57(1):192-199. doi:10.1111/1475-6773.13891

22. Goetschius LG, Barefoot D, Han F, Henderson M. Risk Score Specifications and Codebook for The Hilltop Institute’s Pre- Models (Version 3). The Hilltop Institute, UMBC; 2025. Accessed January 6, 2025. https://www.hilltopinstitute.org/wp-content/uploads/publications/RiskScoreSpecificationsCodebookForHilltopPre-Models-Version3-Jan2025.pdf

23. Llorca J, Delgado-Rodríguez M. Visualising exposure-disease association: the Lorenz curve and the Gini index. Med Sci Monit. 2002;8(10):MT193-MT197.

24. Mauguen A, Begg CB. Using the Lorenz curve to characterize risk predictiveness and etiologic heterogeneity. Epidemiology. 2016;27(4):531-537. doi:10.1097/EDE.0000000000000499

25. Report to Congress: Risk Adjustment in Medicare Advantage. CMS. December 2018. Accessed January 6, 2025. https://www.cms.gov/medicare/health-plans/medicareadvtgspecratestats/downloads/rtc-dec2018.pdf

26. Pope GC, Kautter J, Ellis RP, et al. Risk adjustment of Medicare capitation payments using the CMS-HCC model. Health Care Financ Rev. 2004;25(4):119-141.

27. AHEAD Model. Maryland Health Services Cost Review Commission. Accessed January 6, 2025. https://hscrc.maryland.gov/Pages/ahead-model.aspx

28. Evaluation of the Maryland Primary Care Program. Maryland Health Services Cost Review Commission; 2024:31-114. January 6, 2025. https://www.medchi.org/Portals/18/2024%20JCR-%20Evaluation%20of%20MDPCP.pdf