Background
Type 2 diabetes is one of the most prevalent and most costly conditions for Medicare; in 2017, more than 200 billion US dollars was spent on direct costs related to diabetes, of which 61% was by older adults ≥65 years of age [
1,
2]. Fortunately, the burden of type 2 diabetes can be substantially reduced through medication and lifestyle interventions [
3‐
5]. Despite this, patients living with type 2 diabetes frequently develop kidney disease, another significant driver of healthcare costs [
3]. These types of diabetes-related complications are some of the key contributors to rising costs in these patients [
3].
There are several possible explanations for the limited success in mitigating these rising costs. First, interventions for type 2 diabetes often focus on patients who have already become costly or poorly controlled, even though these patients may only represent a fraction of those who could benefit from an intervention [
1,
2,
6]. Second, while accurately predicting which individuals are likely to become high cost is essential for targeting interventions [
7‐
12], current approaches to predicting spending generally focus on composite measures (like mean costs) and short time horizons, even though patients with the same condition may have costs and seek care in ways that fluctuate over time [
13,
14]. For example, patients with type 2 diabetes who are hospitalized early in a calendar year may differ meaningfully from those who are hospitalized later in the year in terms of how that patient should be managed, although composite metrics would classify them similarly [
15,
16].
Research in other settings has observed similar healthcare cost dynamics. For instance, Tamang et al. identified a definable group of low-spending patients whose costs “bloomed”, or became costly, in the following year within the general Danish population [
17]. Similarly, Lauffenburger et al. observed seven dynamic patterns of spending among a large sample of US commercially-insured beneficiaries, including individuals whose costs increased rapidly towards the end of the year and another group of relatively high cost individuals for whom spending fell [
18].
These approaches have not yet been applied to specific spending among patients with a chronic disease, such as type 2 diabetes. In specific, little is known about the patterns of spending among patients with type 2 diabetes who are currently low utilizers and how many and when these patients may become costly to Medicare, the US national health insurance program for many US older adults. The ability to better proactively discriminate between patients with diabetes who have increasing spending over time could better target interventions to those at greatest need. Therefore, we used a dynamic, data-driven approach to classify individuals with diabetes by their long-term diabetes-specific spending patterns and assessed the ability to predict membership in these groups.
Discussion
Using data-driven approaches, we identified distinct diabetes spending patterns among a nationally-representative cohort of Medicare patients with type 2 diabetes who were initially low utilizers, including a definable group of patients whose costs began to rise progressively late in the first year of follow-up. These patterns could be predicted using baseline characteristics, including diabetes-specific factors and factors that may be potentially modifiable.
Current efforts to predict healthcare spending largely focus on predicting a composite value, such as total yearly diabetes spending, or a threshold-based measure, like being in the top 5% of spending, both of which collapse spending into a single static value [
6,
9,
12,
40,
41]. Two recently-published approaches offer other cluster-based solutions to elucidate patterns of spending [
40,
42]. For instance, researchers recently identified patients with initially low spending levels whose costs bloomed in the subsequent year using a threshold-based approach [
17]. To our knowledge, neither of these approaches have been applied to patients with specific chronic conditions, such as type 2 diabetes.
Our findings support the conclusion that patients may have dynamic patterns of spending over longer periods of time that can be potentially meaningful, with implications on whom to outreach for intervention as well as when to do so [
13,
14]. The ability to potentially discriminate between patients with differing diabetes spending patterns using variables measured at baseline could better target interventions to those who are at greatest need [
41]. Many healthcare organizations, insurers, researchers, and policymakers make predictions and identify patients for cost-containment interventions using these types of administrative data [
40,
41]. If successful, using these longer time horizons could allow for more time to implement potential cost-containment interventions for type 2 diabetes [
41]. The ability to better leverage these routinely-collected data for predictions with more dynamic cost-modeling methods by chronic condition holds wide potential for possible interventions.
The findings of the most influential predictors also offer several noteworthy suggestions for potential diabetes-specific interventions in patients who previously had low spending levels. First, as observed in prior work, adherence to medication, both all medications and diabetes-specific medications, appears to be an important differentiator of patients in different groups, especially those with progressively rising costs; mean adherence is presented in this manuscript but adherence has been known to have meaningful underlying variations even if means are similar [
43‐
46]. Adherence to medication has been shown in a number of contexts to contribute to the avoidance of poor health outcomes [
44,
45,
47,
48]. Of note, while mean adherence appeared similar at baseline, it is known that there are important variations in adherence that composite metrics such as average adherence do not always represent, which could have explained why it was an important predictor, especially in interactions with other variables [
28,
49]. Number of physician office visits and unique physicians may also be indicators of whether patients are getting sufficient care to prevent future escalation of diabetes problems [
26]. Notably, non-modifiable diabetes factors, such as indicators of clinical progression like presence of neuropathy, nephropathy, or retinopathy, were not particularly influential in the boosted prediction models. Of course, one of the most influential predictors was baseline diabetes spending for several trajectory groups; thus, even though the groups had fairly similar spending in the baseline year (Appendix Figure
4), baseline spending is an important consideration when building prediction models and potentially targeting interventions. Insulin costs in particular could also be a key contributor to healthcare costs in these patients [
1]. Together, these findings suggest that there are possible opportunities for the provision of interventions, such as adherence interventions or interventions that increase access to care, to prevent escalating complications and costs in diabetes. Future work should also explore how to apply these results in interventions and how these results replicate in other population, including electronic tools to build these prediction models.
There are several limitations. First, we examined trajectories from January to December; patients with incomplete enrollment or other policy start and end dates may have different spending patterns. The variables included in the prediction models may also not be exhaustive, and although we used validated algorithms for these variables where possible, they may not be sufficiently sensitive. While we used the 90% threshold from prior work to identify low spenders at baseline, the beneficiaries who were classified in the “high-cost” trajectory may also have had elevated spending to start. Trajectory modeling also provides predicted group membership within a cluster; while beneficiaries were assigned to their closest cluster, there could be some within-group heterogeneity. Given the nature of the data, we also do not have information about patients’ glycemic control at baseline, although this would apply to others using these data as well. These results may also not generalize to non-Medicare Fee-For-Service beneficiaries or younger adults, and the data are from several years ago (owing, in part, to an administrative lag in Medicare data). However, given that costs for diabetes are only continuing to increase and rates of type 2 diabetes are growing progressively in younger populations, we expect that these findings will continue to remain relevant [
1]. Finally, some misclassification of the cohort is possible due to the nature of claims data; however, we used validated algorithms to define diabetes and other comorbidities to the extent possible.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.