Development of predictive risk models for major adverse cardiovascular events among patients with type 2 diabetes mellitus using health insurance claims data

Type 2 diabetes may cause complications of microvascular origin, including nephropathy, neuropathy, and retinopathy, or macrovascular origin, including peripheral artery disease and cardiovascular disease (CVD) [1, 2]. Although diabetes clinical practice guidelines are intended to reflect consensus and evidence-based best medical practices, different entities have some conflicting recommendations, and providing high-quality and detailed guidelines for specific patient subgroups remains challenging [3]. For example, relative to non-diabetic patients, patients with type 2 diabetes have a two- to threefold higher risk of suffering from a CVD event, including a higher risk of myocardial infarction (MI), stroke, unstable angina, and congestive heart failure [4‐7], and a higher rate of CVD-related death [8]. Therefore, certain patients with type 2 diabetes could benefit from specialized care that both improve glycemic control and mitigate the risk of CVD.

Thus, having reliable tools making use of readily available data to predict the risk of cardiovascular events among patients with type 2 diabetes may allow healthcare resources to be directed towards patients at high risk, and help healthcare providers meet new quality standard of care. In fact, in 2016, the National Committee for Quality Assurance (NCQA) implemented a new Healthcare Effectiveness Data and Information Set (HEDIS) performance measure based on the rates of hospitalization for potentially preventable complications [9]. More specifically, this measure, which is used by over 90% of health plans in the US [9], targets, among other complications, diabetes short- and long-term complications, including CVD events leading to hospitalization [10]. This means that higher rates of adverse cardiovascular events among patients with type 2 diabetes may negatively affect the NCQA ratings of healthcare providers. Moreover, given the high costs incurred by patients with both CVD and diabetes [11], using such tool efficiently may translate into significant cost savings.

Several of the predictive CVD risk models that have been developed for the general population include diabetes as a risk factor, with models derived from the Framingham Heart Study being among the most well-known [12‐14]. Scores based on the Framingham risk models assign weights to risk factors in order to predict cardiovascular events separately for men and women. Risk factors identified for CVD include older age, smoking status, treated and untreated systolic blood pressure, total cholesterol and high-density-lipoprotein cholesterol levels, and diabetes [12‐14]. However, the Framingham risk models were not developed for patients with diabetes, and were shown to systematically underestimate CVD risk in this population [15]. In fact, the characteristics of patients enrolled in the Framingham study may differ from real-world populations with diabetes in several ways, including the proportion of minorities, socioeconomic determinants of health, and comorbidity burden [16]. Thus, other risk models have been developed for this population, but all of them rely on data from medical records [17‐23]. For example, risk models derived from the United Kingdom Prospective Diabetes Study (UKPDS) identified several risk factors that cannot be used as quantitative predictors using health insurance claims, such as duration of type 2 diabetes, glycated hemoglobin (HbA1c) levels, systolic blood pressure, and cholesterol/high-density lipoprotein ratio [21, 23]. Similarly, the ADVANCE study identified age at diabetes diagnosis, known duration of diabetes, pulse pressure, treated hypertension, HbA1c, urinary albumin/creatinine ratio, and non-HDL cholesterol among risk factors for CVD events; these risk factors cannot be assessed using health insurance claims [22]. Consequently, these models cannot be used to predict CVD risk by entities, like payers, that do not have access to information derived from medical records or laboratory results.

As the face of healthcare provision changes and population management evolves, entities such as public and private payers are moving toward a capitated system of reimbursement, with payments made based on value rather than volume of care. It is thus important for both payers and providers to be able to assess the risks in a given population. Therefore, a CVD risk assessment tool based solely on accessible medical data such as health insurance claims would be relevant for payers to help identify patients with type 2 diabetes at high risk of CVD events. In fact, rationally allocating resources towards these patients by, for example, including CVD risk models in a tool made available to healthcare providers may result in reduced morbidity, mortality, and cost savings. Thus, this study aimed to develop new predictive models and assess their performance in predicting the risk of cardiovascular events in patients with type 2 diabetes based solely on information available in medical health insurance claims. More specifically, models were developed for patients without prior CVD events (hereinafter referred to as the primary prevention population) and for patients with prior CVD events (hereinafter referred to as the secondary prevention population).

A total of 181,619 patients with type 2 diabetes were included in the study; 136,544 (75.2%) in the primary prevention population and 45,075 (24.8%) in the secondary prevention population (Fig. 1). Among patients in the training set and in the primary prevention population, the proportions of patients with MACE, MACE-plus, and CVD-related death during the at-risk period were 4.7%, 6.5%, and 1.8%, respectively (Additional file 3). In the secondary prevention population, the same proportions were 16.5%, 24.9%, and 8.2%, respectively (Additional file 3). The median duration of the at-risk period following the index date in the training set of the primary prevention population was 12 months (range 1–109 months), with 5.4% of patients having a follow-up longer than 60 months. The median duration of the at-risk period in the training set of the secondary prevention population was 11 months (range 1–108 months), with 3.9% of patients having a follow-up longer than 60 months.

Patients with a CVD event during the at-risk period were older and had higher aDCSI scores compared to patients without CVD events for both the primary and the secondary prevention populations (primary prevention population: mean age = 72.7 vs. 66.4 years, mean aDCSI = 1.9 vs. 1.1, respectively; secondary prevention population: mean age = 75.0 vs. 71.4 years, mean aDCSI = 4.1 vs. 3.2, respectively; Additional file 3). Most patients (> 75%) had a recorded diagnosis for hypertension and/or hyperlipidemia in both the primary and secondary prevention populations. Moreover, compared to patients without CVD events, patients with a CVD event during the at-risk period were more likely to have a recorded diagnosis for select baseline comorbidities—such as infections (primary prevention population: 53.8% vs. 48.8%; secondary prevention population: 69.1% vs. 61.5%, respectively) chronic pulmonary disease (primary prevention population: 22.7% vs. 15.6%; secondary prevention population: 44.5% vs. 31.4%, respectively), and peripheral vascular disorders (primary prevention population: 19.0% vs. 9.3%; secondary prevention population: 34.3% vs. 26.1%, respectively) (Additional file 3).

This study developed and validated models that predict the risk of adverse cardiovascular events in patients with type 2 diabetes using exclusively information derived from health insurance claims. The main risk factors identified in the primary prevention population included age, diabetes-related hospitalizations, and recorded diagnosis for coagulopathy and chronic pulmonary disease. In the secondary prevention population, age, prior CVD diagnoses, diabetes-related hospitalizations, and recorded diagnosis for chronic pulmonary disease had the most important impact on the risk of having a CVD event. Overall, the models reliably predicted the cardiovascular events for the primary and secondary prevention populations, as illustrated by the C-statistics ranging between 0.70 and 0.81.

The finding that age was one of the most important risk factor in predicting cardiovascular events is consistent with findings in previous studies that primarily focused on a diabetes population, such as the UKPDS risk engine [21, 23] and studies that focused on a general population, such as the Framingham Heart Study [12]. However, a major difference between the models developed in the current study and previous ones is that the latter included risk factors derived from laboratory results and medical records [12, 17, 18, 20‐22], which are often not available to national quality of care organizations and payers. In contrast, the current study used only information that is readily available from medical claims data.

Nonetheless, claims-based information can be used as a proxy for risk factors derived from laboratory results and medical records. For example, blood pressure measurements were not available in claims data, but hypertension-identified based on a recorded diagnosis in a medical claim—was included in the models. Similarly, recorded diagnosis for hyperlipidemia was used as a proxy for high-density lipoprotein cholesterol and low-density lipoprotein cholesterol levels, although it was not included in any models. Yet, certain risk factors identified in the Framingham and UKPDS models tend to be underreported in medical claims, and thus, may have limited predictive accuracy in claims-based models. For example, although diagnosis codes for smoking do exist, this condition is typically underreported in medical claims. Therefore, smoking was not included in any of the claims-based models. However because our study was limited to risk factors available in insurance claims data, certain risk factors identified in other studies were not available for selection in the models. In particular, several studies pointed to a link between glycemic markers and CVD [33‐37], but given that HbA1c measures are not available in insurance claims data, this potential risk factor could not be included in the models.

This study also found that obesity was associated with a lower risk of MACE and CVD-related death in the secondary prevention population. Several previous studies found obesity to be associated with better survival in patients with chronic or cardiac diseases, hence the term “obesity paradox” to describe this counterintuitive phenomenon [38]. Several explanations have been proposed, including the advantages of fat reserves during illness, biases or confounding in observational studies (e.g., more intensive management), or weight loss due to illness in the reference group [39]. However, due to the observational nature of the current study, no causal relationship can be inferred.

Regardless of the aforementioned differences in the risk factors identified in the current study versus previously published models, the models developed here performed well in predicting the risk of cardiovascular events in a population with two well-defined risk factors, namely patients with type 2 diabetes and above 50 years of age. Overall, the predictive accuracies of the models presented in the current study are comparable to those of previously published models. For example, the Framingham risk score, which included diabetes as a predictor, yielded C-statistics of 0.76 and 0.79 for men and women in the general population, respectively [12]. However, when evaluated in an older diabetic cohort and in patients without prior CVD, the Framingham risk score had a C-statistic of 0.65 [19]. The performances of the claims-based models presented here were also comparable to those of previously developed risk models specific to the diabetic population, such as the UKPDS risk engine [21, 40]. Although C-statistics were not reported in the UKPDS original publications, subsequent validations in other diabetic cohorts yielded C-statistics ranging from 0.61 to 0.73 [19, 41]. The ADVANCE model, developed in a population of diabetic patients at risk of cardiovascular events similar to the secondary prevention population in this study, also presented comparable C-statistics of 0.69-0.70 [22]. Moreover, several other multivariate risk models were published and reported C-statistics ranging between 0.64 and 0.70 [17, 18, 20]. A comprehensive external validation study would be needed to evaluate the performance of the different models on the same cohort of patients [42].

The Framingham and UKPDS models were not developed and tested for patients with a prior history of CVD (i.e., the secondary prevention population), meaning that their predictive accuracy may be lower in this subpopulation [12, 21]. Therefore, another advantage of the models developed in the current study over several previous ones is their ability to predict CVD risk in patients with prior history of CVD, who represented almost a quarter of the sample population. More generally, the reliability of this claims-based approach is perhaps best illustrated by the limited incremental predictive accuracy conferred by the additional inclusion of variables derived from medical records or laboratory results.

In light of the HEDIS performance measure that targets hospitalization for potentially preventable complications, rationally allocating healthcare resources to patients with type 2 diabetes at higher risk of cardiovascular complications may help healthcare providers meet quality of care standards, and lead to reductions in morbidity, mortality, and cost savings. With growing evidence suggesting that certain types of diabetes treatments—such as SGLT2 inhibitors or GLP-1 receptor agonists—may mitigate cardiovascular risk in addition to improving glycemic control, the potential dual purpose of these diabetes medications could be considered-despite their higher cost—to optimize treatment decisions in patients with type 2 diabetes at high risk of CVD [24‐27]. Patients receiving these game-changing treatments were excluded from the present study due to the potential for indication bias: the use of SGLT2 inhibitors or GLP1 receptor agonists could effectively reduce the risk of CVD, but may appear as risk factors associated with a higher risk of CVD if these agents are preferentially prescribed to higher-risk patients. Such counterintuitive phenomena are common in observational studies. Another potential clinical application of the models developed here would be to identify patients at high risk of CVD events within a certain time window in order to provide preventive care. The threshold used for this high-risk group could be rationally determined using the risk that maximizes the sum of the model sensitivity and specificity. For example, using this method, the high-risk threshold in the primary prevention population would be 2.5%, 3.5%, and 1.0% for MACE, MACE-plus, and CVD-related death, respectively (sensitivity ranging from 67 to 73%, and specificity ranging from 67 to 76%). In the sample population used in the current study, applying these thresholds would result in approximately one out of three patients classified at high-risk of having MACE or MACE-plus within a 1-year window, and one out of four patients at high-risk of CVD-related death. In the secondary prevention population, the same thresholds would be 12.5%, 18.0%, and 5.0% for MACE, MACE-plus, and CVD-related death, respectively (sensitivity ranging from 62 to 78%, and specificity ranging from 66 to 72%), resulting in approximately one out of three patients with a high-risk of having any cardiovascular event within a 1-year window.

In summary, this study developed risk models that could reliably identify patients with type 2 diabetes at risk of MACE, MACE-plus, and CVD-related death based on information available in health insurance claims. Ultimately, stakeholders—such as quality of care organizations and payers—may use these models to identify diabetic patients at high risk of cardiovascular events and potentially improve their clinical management, thereby preventing a significant part of the disease burden and associated costs.