 To evaluate the impact of enrollment in a consumer-directed health plan (CDHP) on out-of-pocket (OOP) spending and on the financial burden associated with healthcare utilization.

 Using commercial claims data from 2011 through 2013, we estimated difference-in-differences models that compared changes in outcomes for individuals who switched to CDHPs (CDHP group) with outcome changes for individuals who remained in traditional plans (traditional plan group).

 We estimated the impact of CDHP enrollment on OOP spending at the point of care and on having high financial burden, defined as whether an enrollee spent 3% or more of household income on OOP spending. Additionally, we assessed these outcomes for 2 subgroups: those with lower household income and those with chronic conditions.

 Within the first year of CDHP enrollment, CDHP enrollees experienced a mean marginal increase in OOP spending of $285 (41% increase; 95% CI, $271-$299; 

<.001) relative to traditional plan enrollees. The lower-income and chronic conditions subgroups experienced mean marginal increases in OOP costs of $306 (44% increase; 95% CI, $257-$353; 

<.001) and $387 (56% increase; 95% CI, $339-$435; 

<.001), respectively. The probability of an enrollee having excessive financial burden increased by 4.3 percentage points (95% CI, 4.0-4.6; 

<.001) for the full CDHP sample. These effects were about 3 times larger for the lower-income subgroup (12.3 percentage points; 95% CI, 10.7-13.8; 

<.001) and 2 times larger for the chronic conditions subgroup (8.0 percentage points; 95% CI, 6.9-9.1; 

 CDHP enrollment led to a significant increase in financial burden associated with healthcare utilization, especially for those with lower incomes and those with chronic conditions.

Consumer-directed health plan (CDHP) enrollment increases the financial burden associated with healthcare utilization, especially for those with lower incomes, those with chronic conditions, and those at the higher end of the healthcare spending distribution. 

Due to the unpredictable nature of healthcare utilization, it is unclear in advance who will end up with high healthcare utilization and significant financial burden in a given year. 

More effort is needed to make individuals aware of the potentially significant out-of-pocket spending and financial burden that could be incurred after enrollment in a CDHP.

Consumer-directed health plans (CDHPs) are high-deductible health plans (HDHPs) coupled with a health savings account (HSA) or a health reimbursement arrangement (HRA). The proportion of covered workers with employer-sponsored insurance enrolled in a CDHP increased from 4% in 2006 to 29% in 2016.

In 2016, 56% of covered workers were employed by a firm that offered a CDHP and, among the 41% of covered workers who were offered only 1 type of plan, nearly one-third were offered only a CDHP.

CDHPs are designed to make patients more cost-conscious and to encourage value-based decision making. Although previous work has consistently demonstrated that CDHPs reduce healthcare spending and utilization,

 they may also increase financial burden when patients utilize healthcare. Previous research has shown that CDHP enrollees are more likely to have difficulty accessing necessary healthcare and to incur high medical bills and medical debt,

 particularly among lower-income individuals and those with chronic conditions.

However, many existing studies are based on findings from cross-sectional surveys and thus might be subject to potential selection bias. We therefore investigated the impact of CDHP enrollment on the financial burden of healthcare utilization at the point of service, using longitudinal private insurance claims data from January 1, 2011, to December 31, 2013. We estimated the effects of CDHP enrollment on out-of-pocket (OOP) costs both at the mean and across the distribution of healthcare spending. Additionally, we analyzed the effect of CDHP enrollment on the probability of an enrollee having excessive financial burden from OOP spending. Finally, given evidence that lower-income individuals and the chronically ill may be more likely to forgo or delay care due to cost, coupled with the fact that the RAND Health Insurance Experiment suggests that high cost sharing may lead to impaired health outcomes among the poor and sick,

 we examined the effects of CDHP enrollment on financial burden for these vulnerable populations. 

We used claims data from the OptumInsight (a subsidiary of UnitedHealth Group) Clinformatics Data Mart from 2011 to 2013 to analyze a 25% random sample of the insurer’s commercially insured subscribers and dependents who were continuously enrolled over the full 3-year period in both employer-sponsored and nongroup plans. These claims data also include measures of socioeconomic status, including categorical household income, predicted by a demographic-based analytical model using Census block group—level data where the unit population size (~600-3000) is smaller than that of a Census tract (~1200-8000) or a zip code tabulation area.

 The data also include measures of education level, predicted by the median level of education attained by individuals 25 years or older at the Census block level, and race/ethnicity code, derived from a combination of sources including public records, self-reported surveys, and a proprietary ethnic code algorithm.

We included 2 groups of enrollees. The CDHP group included enrollees who were in a traditional plan in 2011 and in a CDHP in 2012 and 2013 (switch date: January 1, 2012) and enrollees who were in a traditional plan in 2011 and 2012 and in a CDHP in 2013 (switch date: January 1, 2013); switching was split relatively evenly across the 2 years in the study sample (

]). The traditional plan control group included enrollees who were in a traditional plan throughout all 3 years. We excluded enrollees who were 65 years or older (5.7% of the sample) and enrollees with negative OOP costs (0.009% of the sample). The lower-income subgroup included enrollees with an estimated annual household income of less than $40,000 (the lowest category defined in the data). The chronic conditions subgroup was defined as enrollees with at least 1 chronic condition in the baseline year, defined by the Charlson Comorbidity Index (CCI).

 The data were deidentified, and this study was approved by the University of Southern California University Park Institutional Review Board.

We analyzed OOP spending and a binary indicator of excessive financial burden, defined by OOP spending being greater than or equal to 3% of household income, based on prior research.

 We also used alternative thresholds (5% and 10%) to examine the sensitivity of our findings to this definition.

 In all cases, OOP spending refers only to spending incurred at the point of care and represents the sum of co-payments, deductibles, and coinsurance paid by an enrollee for all healthcare services utilized in the given year. Because estimated household income was a categorical variable in our data, we calculated the excessive financial burden indicator using the midpoint of each income interval, capped at $100,000 (the highest category defined in the data). 

test to compare baseline characteristics between the CDHP group and the traditional plan group. We then performed descriptive analyses of the trends in mean OOP spending and the percentage of enrollees having excessive financial burden before and after CDHP enrollment. We used difference-in-differences (DID) regression analysis to compare changes in the outcomes for individuals who switched to CDHPs (CDHP group) with those for individuals who remained in traditional plans (traditional plan group). Compared with a standard DID model in which the participants in the treatment group usually experience a 1-time shift, we analyzed enrollees who switched to a CDHP at 2 different time points to enable estimation of both short-term (1-year) and medium-term (2-year) effects of CDHP enrollment. In all models, the analysis was at the enrollee-year level, and the primary independent variables were the indicators for the first and second year (where applicable) of CDHP enrollment, adjusted for group fixed effects, year fixed effects, age, gender, race, education, Census division, and CCI score.

For each population, we first estimated the impact of CDHP enrollment on mean OOP spending. Because OOP spending was highly skewed, we used a generalized linear model (GLM) with log link function and gamma distribution. We calculated the average marginal effects (incremental effects) of first- and second-year CDHP enrollment. Standard errors were clustered at the enrollee level. We also investigated the impact of CDHP enrollment on the distribution of differences in OOP spending (relative to traditional plan enrollees) using the linear quantile DID (QDID) model,

 which outputs the treatment effects of CDHP enrollment at each decile of OOP spending. We used the QDID regression coefficients to predict traditional plan enrollees’ adjusted OOP spending distribution in 2013 and the traditional plan group’s counterfactual OOP spending distribution (ie, what the OOP spending distribution of the traditional plan group would have been had they enrolled in a CDHP). Based on the adjusted and counterfactual OOP cost distribution, we examined the change in the percentage of enrollees who would have very high OOP spending (eg, higher than $2000) due to CDHP enrollment. Finally, we analyzed the impact of CDHP enrollment on the indicator of whether an enrollee had excessive financial burden (OOP spending to household income ratio ≥3%, 5%, or 10%) in a given year using GLM with logit link function and Bernoulli distribution. Standard errors were clustered at the enrollee level. Statistical analyses were performed using SAS version 9.4 (SAS Institute Inc; Cary, North Carolina) and Stata version 14 (StataCorp; College Station, Texas). We used a significance level of 

 ≤.05 and all statistical tests were 2-sided.

There were a total of 689,542 enrollees in our sample, consisting of 36,387 enrollees in the CDHP group and 653,155 enrollees in the traditional plan group. The baseline characteristics of each group are shown in 

 test indicates that CDHP enrollees differed from traditional plan enrollees in terms of age, education, gender, health status, region, and race and ethnicity. There were no statistically significant differences in predicted annual household income between the 2 groups.

Trends in OOP Spending and Financial Burden

Both the levels and trends in our key outcome variables were similar for the CDHP and the traditional plan group during the pre-CDHP enrollment period of up to 2 years (

). Within the first year of CDHP enrollment, the percentage of CDHP enrollees having excessive financial burden rose from 9.7% (95% CI, 9.4%-10.0%) to 16.0% (95% CI, 15.7%-16.4%) for the full sample, compared with a slight increase from 9.0% (95% CI, 8.9%-9.1%) to 9.8% (95% CI, 9.7%-9.9%) among traditional plan enrollees (eAppendix Figure 2).

The impact of CDHP enrollment on financial burden was more pronounced among the lower-income and chronic conditions subgroups: The percentage of CDHP enrollees having excessive financial burden among these subgroups rose from 32.9% (95% CI, 31.4%-34.5%) to 47.7% (95% CI, 46.1%-49.4%) and from 25.3% (95% CI, 24.1%-26.5%) to 33.9% (95% CI, 32.7%-35.2%), respectively (eAppendix Figure 2). The effects of CDHP enrollment persisted for 2 years; OOP spending increased among CDHP enrollees while OOP spending for traditional plan enrollees remained stable (eAppendix Figure 1).

The effects of CDHP enrollment on mean OOP spending are shown in 

. In the first year of CDHP enrollment, the mean marginal increase in OOP spending was $285 (41% increase; 95% CI, $271-$299; 

<.001). CDHP enrollment resulted in larger marginal increases in OOP spending for the lower-income ($306; 44% increase; 95% CI, $257-$354; 

<.001) and chronic conditions ($387; 56% increase; 95% CI, $339-$435; 

<.001) subgroups. By the second year of CDHP enrollment, mean marginal OOP spending increased by $306 among the full sample (44% increase; 95% CI, $286-$325; 

<.001), $364 (53% increase; 95% CI, $300-$429; 

<.001) among the lower-income subgroup, and $428 (62% increase; 95% CI, $364-$492; 

<.001) among the chronic conditions subgroup. The larger increases in OOP spending in the second year of CDHP enrollment are consistent with prior work which has found that utilization increases in the second year of CDHP enrollment.

presents the QDID estimates of the impact of CDHP enrollment at each decile across the distribution of OOP spending. Generally, the marginal effect of CDHP enrollment increased with higher levels of OOP spending. For example, among the full sample, CDHP enrollment led to a $162 (95% CI, $150-173; 

<.001) marginal increase in OOP spending at the median in the first year compared with a $726 (95% CI, $678-$773; 

<.001) marginal increase in OOP spending at the 90th percentile. That is, the marginal impact of CDHP enrollment on OOP spending (compared with traditional plan enrollees) was greater for those at the higher end of the OOP spending distribution. For the lower-income and chronic conditions subgroups, the marginal increases at the median were $241 (95% CI, $200-$282; 

<.001), respectively, and the marginal increases at the 90th percentile were $715 (95% CI, $558-$871; 

The adjusted and counterfactual OOP spending distributions for the traditional plan group are shown in the 

. As expected, the counterfactual OOP spending distributions fall above the predicted OOP spending distribution, suggesting that traditional plan enrollees would have had higher OOP spending if they had switched to a CDHP (Figure). As a sensitivity analysis, we also separately estimated the magnitude of these counterfactual spending estimates if the traditional plan group switched to CDHP/HRAs compared with CDHP/HSAs (based on separate underlying QDID regressions for the CDHP/HRA and CDHP/HSA populations). We found that increased OOP spending after enrollment in a CDHP tended to be larger for enrollees in CDHP/HRAs versus CDHP/HSAs across nearly all of the OOP spending distributions (

). This is consistent with prior work, which has found that enrollees in CDHP/HSAs had significantly greater reductions in utilization than did enrollees in CDHP/HRAs, potentially because employees with HSAs have a stronger incentive to save money because HSA account balances are owned by the employee whereas HRA account balances are owned by the employer.

CDHP enrollment also increased the probability of having very high OOP spending. For example, the percentage of enrollees spending more than $2000 increased by nearly 10 percentage points for the full sample and for the lower-income subgroup and by about 15 percentage points for the chronic conditions subgroup (Figure). These findings regarding the impact of CDHPs on the probability of having very high OOP spending are supported by the results of the impact of CDHP enrollment on the probability of having excessive financial burden (Table 2b). For the entire sample, enrollment in a CDHP increased the probability of an enrollee being exposed to excessive financial burden (defined as OOP spending being ≥3% of household income) by 4.3 percentage points (95% CI, 4.0-4.6; &#8203;

<.001; baseline, 9.7%). The effects were about 3 times larger for the lower-income subgroup (12.3 percentage points; 95% CI, 10.7-13.8; 

<.001; baseline, 32.9%) and 2 times larger for the chronic conditions subgroup (8.0 percentage points; 95% CI, 6.9-9.1; 

<.001; baseline, 25.3%). These effects persisted in the second year of CDHP enrollment and were robust to alternative definitions of excessive financial burden (Table 2b).

We estimated the impact of CDHP enrollment on financial burden due to cost sharing incurred at the point of service when utilizing care. We found that CDHP enrollment led to a significant increase in OOP spending and that this higher financial burden was particularly pronounced for the lower-income and chronic conditions subgroups. For example, more than half of the lower-income subgroup and more than one-third of the chronic conditions subgroup faced excessive financial burden from OOP spending after enrollment in a CDHP. Moreover, we found that CDHP enrollment increased the probability of having very high OOP spending. 

Our sample was subject to potential selection concerns because we were unable to evaluate whether CDHP enrollees actively chose a CDHP among a menu of plan options or whether it was the only option made available to them by their employer, although the fact that the level and trend of the key outcome variables were similar between the CDHP and traditional plan groups in the baseline period reduces these concerns (eAppendix Figures 1 and 2). Additionally, the household income data were based on a prediction model and thus may be subject to measurement error, although the predictions were based on very local geography. We were also unable to link dependents to subscribers, so we could not assess financial burden due to OOP spending at the household level; however, because the income levels represent household (rather than individual) income whereas we assessed individual-level OOP spending, our estimates may thus represent a lower bound of financial burden on OOP spending pooled across all household members. Finally, our data did not include additional detail, such as plan premiums or whether and how much the employer contributed to the individual’s HSA or HRA, so we could not directly evaluate the impact of CDHP enrollment on overall spending.

Our findings, and in particular, the variation in the impact of CDHP enrollment on OOP spending across the distribution, should be considered in the broader context of overall healthcare spending, including premiums and potential HSA/HRA contributions. Nationally representative data on employer-sponsored health benefits indicate that, in 2013, the average annual worker contribution to premiums for single coverage was $1058 for CDHPs with an HRA, $726 for CDHPs with an HSA, and $1027 for non-CDHP plans. This suggests that enrollees in CDHPs with an HSA saved $301 ($1027 minus $726) in employee-paid premium contributions compared with those enrolled in non-CDHPs, on average. Additionally, 34% of covered workers in CDHPs with an HSA worked for an employer that did not make an HSA contribution but, among those that did, the average annual HSA contribution for a single employee was $950.

At the lower end of the distribution, we found very modest increases in OOP spending at the point of care among CDHP enrollees compared with traditional plan enrollees. Thus, it is likely that these individuals may actually be financially better off by enrolling in CDHPs, as the savings in employee-paid premium contributions and/or the benefits of employer contributions to a tax-advantaged savings account likely more than offset the increased OOP spending at the point of care. However, at the high end of the distribution, the increased OOP spending would still heavily outweigh the reduced spending on employee contributions to premiums, particularly if the individual was among the one-third of covered workers who did not receive an HSA contribution from their employer. Moreover, even if those individuals at the lower end of the actual spending distribution may have lower total expenditures on premiums plus OOP spending at the point of care in a given year, due to the unpredictable nature of catastrophic health events, they still face the potential risk of having very large OOP spending under a high deductible, which may be particularly problematic for lower-income individuals with limited assets.

The potential consequences of CDHP enrollees experiencing high financial burden are significant. Previous work suggests that high financial burden may cause CDHP enrollees to delay or skip needed healthcare.

 Additionally, individuals who face financial burden often have to change their employment or lifestyle, or make other sacrifices, to make ends meet.

 We do not know the extent to which individuals considered potential downstream OOP costs when enrolling in CDHPs. However, prior research suggests that many individuals make poor health plan choices and frequently choose worse plans when alternative choices that have both lower financial risk of OOP spending and lower premiums are available.

 A possible explanation for these poor choices is that many individuals do not understand the basics of benefit design, such as the meaning of deductibles and co-payments.

 Moreover, this problem of poor health insurance literacy is more acute among the low-income population,

 suggesting that these individuals may not understand the potential financial burden associated with enrollment in CDHPs. Although prior work suggests that providing information on OOP costs can improve health plan choice,

 more work is needed to design effective interventions for improving plan choice. Additionally, further research on the impact of financial burden associated with CDHP enrollment on the use of appropriate care, unmet medical need, and other health outcomes is warranted.

School of Pharmacy (XZ), Price School of Public Policy (ET, NS), and Schaeffer Center for Health Policy and Economics (XZ, ET, NS), University of Southern California, Los Angeles, CA.

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. 

Concept and design (XZ, ET, NS); acquisition of data (NS); analysis and interpretation of data (XZ, ET, NS); drafting of the manuscript (XZ, NS); critical revision of the manuscript for important intellectual content (XZ, ET); statistical analysis (XZ); administrative, technical, or logistic support (ET); and supervision (NS). 

Neeraj Sood, PhD, Price School of Public Policy, University of Southern California, 635 Downey Way, Los Angeles, CA 90089-3333. Email: nsood@healthpolicy.usc.edu.

1. Kaiser Family Foundation and Health Research & Educational Trust. 

Employer Health Benefits: 2016 Annual Survey

. Menlo Park, CA: Kaiser Family Foundation; 2016. files.kff.org/attachment/Report-Employer-Health-Benefits-2016-Annual-Survey. Accessed March 7, 2018. 

2. Chen S, Levin RA, Gartner JA. Medication adherence and enrollment in a consumer-driven health plan. 

3. Fronstin P, Sepulveda MJ, Roebuck MC. Medication utilization and adherence in a health savings account-eligible plan. 

4. Beeuwkes Buntin M, Haviland AM, McDevitt R, Sood N. Healthcare spending and preventive care in high-deductible and consumer-directed health plans. 

5. Finkelstein A, Taubman S, Wright B, et al; Oregon Health Study Group. The Oregon health insurance experiment: evidence from the first year. 

6. Haviland AM, Marquis MS, McDevitt RD, Sood N. Growth of consumer-directed health plans to one-half of all employer-sponsored insurance could save $57 billion annually. 

. 2012;31(5):1009-1015. doi: 10.1377/hlthaff.2011.0369.

7. Haviland AM, Eisenberg MD, Mehrotra A, Huckfeldt PJ, Sood N. Do “consumer-directed” health plans bend the cost curve over time? 

2016;46:33-51. doi: 10.1016/j.jhealeco.2016.01.001.

Articles

WEB EXCLUSIVE

Enrollment in a consumer-directed health plan increases the financial burden associated with healthcare utilization, especially for those with lower incomes and with chronic conditions.

Financial Burden of Healthcare Utilization in Consumer-Directed Health Plans

 The association between the progression of diabetes severity and risk of severe hypoglycemia is unknown. This study aimed to evaluate the association between the progression of diabetes severity and severe hypoglycemia in patients with diabetes.

 A 13-year population-based retrospective cohort study of patients with diabetes in Taiwan.

 Diabetes progression was evaluated by the adapted Diabetes Complications Severity Index (aDCSI) score from index date to end of follow-up. The progression of diabetes severity was divided into 3 categories: slow, moderate, and rapid increase in aDSCI score. We further compared those 3 categories and evaluated the risk of first hospitalization due to severe hypoglycemia.

 A total of 330,831 patients with diabetes were recruited. The mean age of patients in this study was 56.8 years, and mean follow-up duration was 9.3 years. The mean initial aDCSI score was 0.7, whereas the mean aDCSI score at the event date or end date was 2.9. A rapid increase in aDCSI score was associated with higher risk of severe hypoglycemia compared with a slow increase (hazard ratio, 4.91; 95% CI, 4.65-5.18). The incidence densities of severe hypoglycemia (per 1000 person-years) for slow, moderate, and rapid increase in aDCSI score were 2.3, 2.5, and 11.4, respectively.

 This study demonstrated that rapid progression of diabetes complications was associated with higher risk of severe hypoglycemia. It is imperative that treating physicians identify patients with acute worsening of diabetes severity and provide proper hypoglycemia education and prevention care.

Rapid progression of diabetes complications was associated with higher risk of severe hypoglycemia. 

Being 45 years or older or having a prescription for insulin or sulfonylureas was associated with increased risk of severe hypoglycemia, whereas having a prescription for biguanides, α-glucosidase inhibitors, thiazolidinediones, or dipeptidyl peptidase-4 inhibitors was associated with a lower risk of hypoglycemia. 

Patients with metabolic complications had the highest risk of severe hypoglycemia.

The treatment of diabetes involves striking a delicate balance between attaining good glycemic control and avoiding hypoglycemic events. When ideal glycemic control is achieved, a favorable outcome can be expected due to the decreased frequency of microvascular and cardiovascular complications.

 However, overtreatment with insulin or oral antidiabetic drugs (eg, a sulfonylurea)can lead to severe hypoglycemia, which has negative consequences due to its association with increased cardiovascular events,

 Although mild hypoglycemia can be easily resolved by consuming carbohydrates, when hypoglycemia occurs rapidly or goes untreated, confusion, seizure, or irreversible cognitive impairment can occur.

Severe hypoglycemia, which is a diabetes emergency, is defined as having low blood glucose that requires the assistance of another person to treat. The frequency of severe hypoglycemia is common. It has been estimated that about 7% to 25% of patients with type 2 diabetes who use insulin experience at least 1 severe episode annually.

 Nevertheless, the distribution of severe hypoglycemia in patients with diabetes is skewed, with a small proportion (5%) of patients accounting for the majority (54%) of severe hypoglycemic events.

 Therefore, it is important to improve the methods by which patients with a high risk of hypoglycemia are identified and managed. Previous study findings have demonstrated several risk factors for severe hypoglycemia in patients with diabetes, including antidiabetic medication prescription,

Moreover, the timing and progression of diabetes severity might play an important role in the development of severe hypoglycemia. However, few studies in the literature address the association between progression of diabetes severity and risk of severe hypoglycemia. Therefore, we conducted a 13-year population-based cohort study using Taiwan’s National Health Insurance Research Database (NHIRD) to evaluate the association between the progression of diabetes severity and severe hypoglycemia in patients with newly diagnosed diabetes. We hypothesized that a rapid progression of adapted Diabetes Complications Severity Index (aDCSI) score is associated with higher risk of severe hypoglycemia.

This observational retrospective cohort study was conducted using the Longitudinal Cohort of Diabetes Patients database, which was obtained from Taiwan’s National Health Insurance (NHI) program. This program is a compulsory insurance system that is regulated by Taiwan’s government, providing coverage for almost 99% of the population. For data protection and privacy reasons, the NHIRD releases only about 75% of data from the diabetic population for research purposes every year. This data set contains detailed information pertaining to patients’ disease diagnoses, drug prescriptions, medical expenses, hospital admissions, and discharge diagnoses. A detailed description of patient recruitment and sampling procedures is available on the NHIRD website.

 The information on disease diagnoses, drug prescriptions, and hospitalizations contained in the NHIRD is of high quality, and numerous studies have been published based on data obtained from this database.

 The accuracy of diabetes diagnoses in this NHI claims database has been validated with a good positive predictive value.

 All patients’ data in the NHIRD are encrypted and scrambled before being released to the public for academic research. Data for this study were extracted and analyzed by 1 independent reviewer. This study was reviewed and approved by the Institutional Review Board of Chung Shan Medical University Hospital (Taichung, Taiwan).

Patients were enrolled in the study by random selection of 360,000 newly diagnosed patients with diabetes between 1999 and 2001. Diabetes was defined as either an inpatient hospitalization for diabetes diagnosed in accordance with the 

International Classification of Diseases, Ninth Revision, Clinical Modification 

) code 250.x, at least 2 outpatient physician visits with 

code 250.x, or 1 outpatient physician visit with 

code 250.x plus a prescription for an oral antihyperglycemic agent within 365 consecutive days. All patients with diabetes diagnosed during this study period were included except those with gestational diabetes. Patients were tracked from January 1, 1999, to December 31, 2011, with a total follow-up duration of 13 years. Overall, 330,831 patients were included in the retrospective cohort analyses.

Progress of Diabetes Severity and aDCSI Score

The Diabetes Complications Severity Index (DCSI), which was developed by Young et al in 2008, measures the severity of diabetes-related complications and is used to predict the risk of diabetes-related hospitalizations, healthcare utilization,

and mortality. A DCSI score ranges from 0 to 13 according to the presence of 7 categories of diabetes complications: retinopathy, nephropathy, neuropathy, stroke, cardiovascular disease, peripheral vascular disease, and metabolic complications. Each diabetes complication category is scored on a range of 0 to 2, determined by the presence and severity of complications (0 = not present, 1 = some abnormality, 2 = severe abnormality), except for neuropathy, which is given a score of 0 or 1 (0 = not present, 1 = abnormal). For example, within the cardiovascular category, patients with angina pectoris would be given a score of 1, whereas patients with myocardial infarction would have a score of 2. When multiple complications within the same category exist, the highest score within that category should be used (score of 2). It should be noted that the presence of severe hypoglycemia was identified as a disease event, rather than a DCSI item (

The patients’ original DCSI scores were calculated using a combination of their diagnosis code (

code) and laboratory data (within nephropathy category). However, a recently modified version of the DCSI developed by Chang et al, known as aDCSI, which omits laboratory data from the nephropathy category, was demonstrated to be as effective as the original DCSI for estimating diabetes severity and predicting hospitalizations.

 The aDCSI was validated in a study that analyzed data from Taiwan’s NHIRD, and the results showed that the performance of the aDCSI in predicting the risk of hospitalization was similar to that of the original DCSI found in the study by Young et al.

All patients with diabetes and their complications were identified from the index date, and the patients were followed until the event date (severe hypoglycemia) or the end date of the study (disenrollment, mortality, or December 31, 2011). An observation for each patient was stopped after the first episode of severe hypoglycemia. The highest aDCSI score occurring in the first 6 months after diagnosis of diabetes was used as the initial aDCSI score (

). When diabetes complications within the same category appeared multiple times during this study period, they were calculated only once. The progression of diabetes severity was defined as the increase in aDCSI score per year, which was calculated by subtracting the initial aDCSI score from the total aDCSI score during this study period and dividing by the number of observation years. For instance, patient A had retinopathy at the initial diagnosis of diabetes and thus had a DCSI score of 1. In the fifth year, 1 additional diabetes complication developed (diabetic neuropathy with aDCSI score = 1) without the occurrence of severe hypoglycemia. In the seventh year, he experienced severe hypoglycemia, and during that time, he also had a first stroke (aDSCI score = 2), so his total aDCSI score was 4. Thus, the final score of progression of diabetes severity was 0.43 per year, which was obtained by subtracting the initial aDCSI score of 1 from the total aDCSI score of 4 and then dividing by 7, the number of observation years. We further classified the progression of diabetes severity into 3 categories (approximately 110,000 patients in each group): slow (increased scores of <0.1 per year), moderate (increased scores of 0.1-0.3 per year), or rapid increase in aDCSI score (increased scores of >0.3 per year). Thus, in this example, patient A should be categorized in the third group with an increase in aDCSI score of more than 0.3 per year.

The date of diabetes diagnosis was defined as the index date. Patients were followed until the first event of severe hypoglycemia, which served as the event date, or the end date, which was designated asdisenrollment; mortality; or December 31, 2011. Severe hypoglycemia was defined as a diagnosis of patients who required hospitalization or a visit to an emergency department (ED) for treatment with an 

code of 250.3, 251.0, 251.1, 251.2, or 962.3. Because of the possibility of ambiguity with 

code 250.3 (diabetes with other coma) between diabetic ketoacidosis (DKA) or hyperosmolar hyperglycemic syndrome (HHS) with hypoglycemic coma,

 we manually identified and analyzed our data for patients who were diagnosed with 

code 250.3. However, none of the patients in our study population appeared to have DKA/HHS with hypoglycemic coma. Disease comorbidities that potentially increased the risk of severe hypoglycemia, such as hypertension (

codes 414, 429.2), chronic liver disease (

codes 296.2, 296.3), chronic obstructive pulmonary disease (

codes 430-437), were identified, as were prescriptions for antidiabetic drugs, such as biguanides, sulfonylureas, α-glucosidase inhibitors, thiazolidinediones, glinides, dipeptidyl peptidase-4 (DPP-4) inhibitors, and insulin. Antidiabetic medication prescription was defined as a prescription with a drug exposure window of at least 180 days during this study period. A cutoff of 180 days was applied to avoid an irregular pattern of drug use, which might be related to patients’ nonadherence or initial antidiabetic drug adverse effects. The primary end point of this study was the first hospitalization due to severe hypoglycemia.

Hazard ratios (HRs) and 95% CIs for the Cox proportional hazards model were used to analyze the data and determine the risk of severe hypoglycemia. One-way analysis of variance was used to compare data among the 3 groups. The Kaplan-Meier method was used to estimate severe hypoglycemia cumulative incidences. A 2-sided 

value <.05 was considered to be statistically significant. All statistical analyses were performed using the SPSS Statistical Package, version 18 (SPSS; Chicago, Illinois).

shows the baseline demographic characteristics of the patients in this study. A total of 330,831 patients were included in this cohort from January 1, 1999, through December 31, 2011. The mean age was 56.8 years, and mean follow-up duration was 9.3 years. The mean initial aDCSI score was 0.7, whereas the mean aDCSI score at the event date or end date was 2.9. Patients with rapid increase in aDSCI score per year generally were older, were more likely to be male, and had more disease comorbidities and complications and a shorter follow-up duration.

shows that the risks of severe hypoglycemia were associated with progression in aDCSI score. Compared with the patients with diabetes with slow progression of diabetes severity (increase in aDCSI score of <0.1 per year), the HRs of patients with moderate progression (increase of 0.1-0.3) and rapid progression (increase of >0.3) increased in proportion to diabetic progression, with HRs of 1.08 (95% CI, 1.02-1.14) and 4.91 (95% CI, 4.65-5.18), respectively. The incidence densities of severe hypoglycemia (per 1000 person-years) for slow, moderate, and rapid increase in aDCSI score were 2.3, 2.5, and 11.4, respectively.

shows the risk of severe hypoglycemia with different diabetes complications. Before adjustment for potential confounders, all diabetes complications were associated with higher risk of severe hypoglycemia. However, after adjustment for age, gender, and antidiabetic medication, patients with metabolic complications showed the highest risk of severe hypoglycemia (HR, 1.96; 95% CI, 1.87-2.06). Retinopathy (HR, 1.11; 95% CI, 1.07-1.15), nephropathy (HR, 1.33; 95% CI, 1.28-1.37), and stroke (HR, 1.21; 95% CI, 1.17-1.25) were also associated with severe hypoglycemia. Meanwhile, cardiovascular disease (HR, 0.90; 95% CI, 0.87-0.93) was associated with lower risk of severe hypoglycemia.

The eAppendix Figure shows the Kaplan-Meier curves of the cumulative incidence of severe hypoglycemia in the 3 groups of aDCSI score increase per year. The Kaplan-Meier curves showed that the risk of severe hypoglycemia seemed to emerge in the first year, and it can be seen that there was marked divergence of aDCSI score increase among the 3 groups over the 13-year follow-up period.

shows various model adjustments that were performed to evaluate the risk of severe hypoglycemia according to the progression of diabetes complications. Despite adjustment for age, gender, disease comorbidities, and prescriptions for antidiabetic drugs, the progression of diabetes complications still had a significant effect on the risk of developing severe hypoglycemia.

This study found that progression of diabetes severity was associated with an increased incidence of severe hypoglycemia. To the best of our knowledge, this is the first study to examine the association between progression of diabetes severity and risk of severe hypoglycemia. An increase in aDCSI score per year was positively related to risk of severe hypoglycemia. Furthermore, we found that being older than 45 years or having a prescription for insulin or a sulfonylurea was associated with increased risk of severe hypoglycemia, whereas having a prescription for biguanides, α-glucosidase inhibitors, thiazolidinediones, or DPP-4 inhibitors was associated with a lower risk of hypoglycemia.

The progression of diabetes complications is difficult to quantify. The severity of diabetes complications increases with time after diabetes onset, and patients with diabetes often exhibit different patterns of progression and complications. Previous study findings have demonstrated that patients with older age,

 have a higher risk of hypoglycemia. In this study, we further explored the importance of timing and severity of these complications. We used increase in aDCSI score per year to measure the progression of diabetes complications. Compared with patients with slow progression of diabetes complications, severe hypoglycemia occurred more frequently in patients with complications that progressed rapidly within a short period of time. Furthermore, we demonstrated that severe hypoglycemia risk increased in proportion to the progression of diabetes complications (Table 4).

Surprisingly, patients with metabolic complications, such as DKA or HHS, had the highest risk of developing severe hypoglycemia. In these patients, a high degree of glucose variability might contribute to the mechanism of severe hypoglycemia.

 The potential mechanism of hypoglycemia might be characterized by deficient counterregulatory hormone release, especially in older patients, and a diminished autonomic response.

 Regarding antidiabetic medications, our findings showed that sulfonylurea prescription (HR, 2.50; 95% CI, 2.39-2.61) was associated with the highest risk of hypoglycemia, whereas DPP-4 inhibitor prescription was correlated with the lowest risk of hypoglycemia (HR, 0.22; 95% CI, 0.22-0.25). This finding was consistent with results reported in a previous study.

This study has several strengths, most notably its large sample size and long-term follow-up. Our data are representative of Taiwan’s general population. To minimize the risk of selection bias and overdiagnosis of severe hypoglycemia, we included only patients who required hospitalization or an ED visit with a primary diagnosis of hypoglycemia. 

The potential limitations of this study should also be noted. First, because the NHIRD is primarily maintained for reimbursement purposes, there were no data on patients’ personal information, such as smoking history, body weight, and body mass index; laboratory data, including glucose level, glycated hemoglobin, and lipid profile; and dosage of medications, such as insulin, which may each potentially influence the risk of hypoglycemia. Second, patients with self-reported severe hypoglycemia who did not have an ED visit or hospital admission were not included in this study. Third, the retrospective nature of the study itself may have potentially resulted in selection bias. Finally, this study included only Taiwanese patients, and thus the results might not be generalizable to other populations.

Because the data set used in the present study was derived from the whole population without applying a sampling procedure, potential selection bias related to sampling error was not a concern. Although disease diagnosis might have been misclassified, the likelihood of such an occurrence was low because miscategorized diagnoses influence reimbursement, which would not be tolerated by hospitals and clinics. Taiwan’s NHI program also conducts stringent monitoring of claims and imposes strict penalties to avoid overdiagnosis and prevent financial irregularities.

The current guideline of the American Diabetes Association recommends frequent self-monitoring of blood glucose for patients receiving intensive insulin treatment (multiple-dose insulin or insulin pump) to monitor and prevent hypoglycemia. Patients should receive education to increase their awareness of certain situations that might increase the risk of hypoglycemia, such as vigorous exercise, fasting for medical tests or procedures, and during sleep.

Finally, this study demonstrated that patients with rapid progression of diabetes complications had a higher risk of developing severe hypoglycemia. Therefore, it is important for treating physicians to slow the progression of diabetes complications and identify patients with increases in aDCSI score per year. At-risk patients should be given proper hypoglycemia education, prevention, and treatment.&ensp;

The authors thank Yu-Hsun Wang of Chung Shan Medical University Hospital for his support and contribution. 

The interpretation and conclusions contained herein do not represent those of the Bureau of National Health Insurance, the Ministry of Health and Welfare, or the National Health Research Institutes.

Division of Endocrinology and Metabolism, Department of Internal Medicine (EK, Y-SY, S-CL, C-NH), and Department of Pharmacy (Y-RL), Chung Shan Medical University Hospital, Taichung, Taiwan; School of Health Policy and Management (J-YC), Institute of Medicine of Chung Shan Medical University (EK, Y-SY, S-CL, C-NH), Taichung, Taiwan; Institute of Biotechnology, HungKuang University (C-HP), Taichung, Taiwan.

This work was supported by grants from the Chung Shan Medical University Hospital (CSH-2013-C-002).

Concept and design (EK, Y-SY, S-CL, C-HP, Y-RL, J-YC, C-NH); acquisition of data (EK, Y-SY, S-CL, C-HP, Y-RL, J-YC, C-NH); analysis and interpretation of data (EK, Y-SY, S-CL, C-HP, Y-RL, J-YC, C-NH); drafting of the manuscript (C-NH); critical revision of the manuscript for important intellectual content (EK, Y-SY, S-CL, C-HP, C-NH); statistical analysis (EK, Y-SY, S-CL, Y-RL, J-YC, C-NH); provision of patients or study materials (Y-RL, J-YC); obtaining funding (EK, Y-SY, S-CL, Y-RL, J-YC, C-NH); administrative, technical, or logistic support (EK, Y-SY, S-CL, Y-RL, J-YC, C-NH); and supervision (J-YC, C-NH). 

Chien-Ning Huang, MD, PhD, or Jeng-Yuan Chiou, PhD, Chung Shan Medical University Hospital, 110th, 1st Section, Jianguo North Rd, South District, Taichung, Taiwan. Email: tom@csmu.edu.tw.

1. UK Prospective Diabetes Study (UKPDS) Group. Intensive blood-glucose control with sulphonylureas or insulin compared with conventional treatment and risk of complications in patients with type 2 diabetes (UKPDS 33) [erratum in 

. 1998;352(9131):837-853. doi: 10.1016/S0140-6736(98)07019-6.

Rapid progression of diabetes complications was associated with higher risk of severe hypoglycemia.

Progress of Diabetes Severity Associated With Severe Hypoglycemia in Taiwan

 To determine if electronic health record (EHR) tools and patient engagement can improve the quality of chronic kidney disease (CKD) care.

 We enrolled 153 primary care physicians caring for 3947 high-risk and 3744 low-risk patients with stage III CKD across 13 ambulatory health centers in eastern Massachusetts. Intervention physicians received a set of electronic alerts during office visits recommending risk-appropriate CKD care. Patients of intervention physicians also received tailored educational mailings. For high-risk patients, we assessed for a visit with a nephrologist and prescription of an angiotensin-converting enzyme (ACE) inhibitor or angiotensin receptor blocker (ARB) during the 12-month study period. For low-risk patients, we assessed for a urine microalbumin screening and prescription of an ACE inhibitor or ARB during the 12-month study period.

 Among high-risk patients, those in the intervention arm were significantly more likely to have an office visit with a nephrologist compared with those in the control arm (45% vs 34%; 

<.001). Among low-risk patients, those in the intervention arm were significantly more likely than those in the control arm to have received urine microalbumin testing (45% vs 21%; 

<.001). There was no difference between the intervention and control arms in rates of prescription of an ACE inhibitor or ARB in either the high-risk patient group (76% vs 79%; 

= .17) or the low-risk patient group (64% vs 65%; 

 A combined program of EHR tools and patient engagement improved some areas of CKD care, but substantial gaps remain.

We randomized 153 primary care physicians caring for nearly 8000 patients with chronic kidney disease (CKD) to receive an intervention that combined electronic decision support tools, patient engagement materials, and collaboration between primary care and nephrology. 

Awareness of CKD is low among patients and physicians.

The intervention was favorably received by physicians and patients. 

We developed a practical population-based approach to assist accountable care organizations that are seeking to engage collaboration between primary care and specialty care to improve the quality of CKD care.

Chronic kidney disease (CKD) affects more than 25 million Americans, or more than 10% of the adult population.

 Effective management of moderate-stage CKD is needed to reduce the high mortality rates and extensive costs associated with progression to more advanced kidney failure.

Many challenges exist to improving care for CKD,

 which remains a frequently unrecognized condition by both primary care physicians (PCPs)

 Just 12% of patients with stage III or IV CKD are aware of their diagnosis, and just 63% of PCPs can correctly identify the presence of CKD.

 However, monitoring for disease progression, using appropriate medications, and involving nephrologists early can improve CKD outcomes.

This highlights the need for healthcare systems to develop a systematic approach to treating this condition that supports primary care providers and nephrologists.

 Electronic health records (EHRs) present an opportunity to deliver appropriate care by identifying patients with CKD, stratifying the patient population, and facilitating tailored treatment and care coordination among patients, primary care, and nephrology.

 We conducted a randomized controlled trial to assess the impact of a comprehensive set of EHR tools and patient engagement materials to improve the management of CKD.

This 18-month trial was conducted from 2011 to 2013, with patient enrollment occurring during the initial 6 months and all patients followed for 12 months after enrollment. We randomly assigned PCPs to receive alerts within the EHR during office visits for patients with CKD and mailed educational materials to patients of physicians in the intervention arm. The Human Studies Committee at Brigham and Women’s Hospital approved the study protocol, and a waiver of informed consent was approved for physicians and patients. The trial was registered at ClinicialTrials.gov (ID NCT01203813).

We conducted our study at Harvard Vanguard Medical Associates, an integrated multispecialty group practice in eastern Massachusetts caring for approximately 300,000 adult patients. The system has significant experience in population health management, such as participation as a Pioneer Accountable Care Organization. The practices use a common EHR (Epic Systems; Verona, Wisconsin) that captures clinical notes, electronic diagnosis codes, specialty referrals, medication prescriptions, and laboratory test results. This system has delivered automated reporting of estimated glomerular filtration rate (eGFR), computed using the Modification of Diet in Renal Disease (MDRD) Study equation. The EHR does not provide decision support for patients with CKD. Nephrology services are provided by 8 nephrologists employed by the group practice.

We enrolled 153 physicians across 13 health centers and 7691 patients aged 18 to 80 years with an established diagnosis of stage III CKD (

). The diagnosis was based on meeting each of the following criteria: 1) the presence of an office visit with a PCP within the group practice within the prior 18 months, 2) the presence of at least 2 eGFR results between 30 and 60mL/min/1.73m

 within the prior 5 years, 3) the qualifying abnormal eGFR results were separated by at least 90 days, and 4) the most recent eGFR was less than 60 mL/min/1.73m

The intervention was randomized at the individual physician level. Within each health center, we paired clinicians based on their number of eligible patients with CKD and then randomly assigned 1 physician in each pair to the intervention arm.

Based on local consensus and emerging data on the importance of both eGFR and albuminuria,

 we stratified patients with stage III CKD according to their risk of complications and identified relevant treatment targets. Our local consensus was achieved prior to the publication of recent guidelines by the Kidney Disease: Improving Global Outcomes (KDIGO) initiative

 and involved gathering input from both primary care and nephrology leadership within our multispecialty group practice using current data on predictors of mortality for patients with CKD. “High-risk” patients were defined as those with either 1) at least 1 eGFR of less than 45 mL/min/1.73m

 in the prior 5 years or 2) at least 1 eGFR of at least 45 mL/min/1.73m

, in combination with the presence of diabetes or albuminuria (urine microalbumin >30 mcg/mg or spot protein to creatinine ratio >0.15 mcg/mg). All other patients—specifically, those with an eGFR of at least 45 mL/min/1.73m

 and no history of diabetes or albuminuria—were considered “low-risk”. We identified our treatment targets using the same process, and for high-risk patients, we recommended use of an angiotensin-converting enzyme (ACE) inhibitor or angiotensin receptor blocker (ARB), as well as referral to a nephrologist. For low-risk patients, we recommended use of an ACE inhibitor or ARB, as well as annual monitoring of urine microalbumin to assess for disease progression and risk stratification.

All EHR alerts were displayed when physicians accessed the electronic ordering module of the patient chart. During office visits with high-risk patients, PCPs received up to 2 alerts. The first alert recommended a referral to a nephrologist if no such specialty office visit had occurred in the prior 12 months (

]). The second alert recommended prescription of an ACE inhibitor or ARB if the patient carried a diagnosis of hypertension or albuminuria, had not been prescribed the medication in the last 12 months, and had no documented allergy to such medication. 

During office visits with low-risk patients, PCPs received up to 2 alerts. The first was the same ACE inhibitor alert used for high-risk patients. The second recommended overdue annual laboratory tests, including those for urine microalbumin, serum creatinine, low-density lipoprotein (LDL) cholesterol, 25-hydroxy (OH) vitamin D, parathyroid hormone, calcium, phosphorus, and hemoglobin.

Our intervention included a mailed outreach program to promote patient engagement. We encouraged PCPs in the intervention arm to enroll patients using 2 strategies. First, all electronic reminders also prompted physicians to enroll patients in the mailed outreach program (

). Second, for those physicians who did not respond to the request to enroll a patient via the electronic alert, we sent a follow-up postcard within 1 month of the office visit requesting them to enroll the patient. We required physicians to enroll patients to ensure that patients received a diagnosis of CKD from their care team prior to receiving any mailings.

The outreach program consisted of quarterly mailings to patients that provided tailored treatment recommendations based on detailed extracts from their EHR. The mailings were based on educational materials developed by the National Kidney Disease Education Program (

 These mailings provided recommendations specific to CKD for managing blood pressure, appropriate use of ACE inhibitors or ARBs, education on current medications, and recommendations for overdue laboratory tests or follow-ups on previously abnormal results.

Physicians in the control group received no EHR alerts, and their patients were not eligible to receive the mailed outreach program.

The primary study end points were based in part on the KDIGO guidelines and extracted from the EHR. Among high-risk patients, the primary end points included an office visit to a nephrologist during the 12-month study period and the prescription of an ACE inhibitor or ARB during the 12-month study period for those with hypertension and/or microalbuminuria and no documented allergy. The primary study end points among low-risk patients included the presence of a urine protein test during the 12-month study period and the prescription of an ACE inhibitor or ARB for those with hypertension and/or microalbuminuria and no documented allergy. We also assessed secondary outcomes of rates of annual serum creatinine, LDL cholesterol, hemoglobin, phosphorus, 25-OH vitamin D, calcium, and parathyroid hormone testing. Physicians and patients were not blinded to intervention status, although all outcomes data were collected without respect to intervention status.

We surveyed all patients in the intervention arm who were enrolled in the outreach program (n = 1002) by their PCP. Patients used a 4-point ordinal scale from “definitely yes” to “definitely no” to report on whether the mailings gave them choices to think about for treating CKD, helped them set specific CKD treatment goals, and helped them understand their medications for CKD. Patients also reported on whether their doctor or another health professional had told them that they had weak or failing kidneys, and they used a 5-point ordinal scale from “strongly agree” to “strongly disagree” to report agreement with their diagnosis of CKD. Finally, patients used a 5-point scale from “excellent” to “poor” to rate the CKD care they received. The survey was administered via a single mailing at the end of the intervention and achieved a 24% (n = 242) response rate. 

We surveyed 153 study physicians at the completion of the intervention. Physicians used a 5-point ordinal scale from “always” to “never” to report on the frequency with which they informed patients of a new diagnosis of CKD once they recognized it was present. Physicians also reported on the eGFR threshold at which they felt comfortable informing their patients of a diagnosis of CKD. Intervention physicians also rated the effectiveness of the electronic alerts, patient mailings, and collaboration with nephrology on improving the quality of CKD care among their patients (“very effective,” “somewhat effective,” or “not effective”). The survey was implemented via an initial paper mailing, followed by a reminder email to nonresponders and a final paper mailing at 4 weeks, achieving a 73% (n = 111) response rate.

Balance between patient demographic characteristics in the intervention and control arms was checked using a 

 test for patient age, Fisher exact tests for binary variables, and χ

 tests for categorical variables. We analyzed the impact of the intervention by fitting logistic regression models using the generalized estimating equation approach to account for clustering of patients within clinics, with performance of each of our prespecified outcomes as the dependent variable and intervention status as the primary independent variable. The models were implemented using the GENMOD procedure in SAS version 9.3 (SAS Institute; Cary, North Carolina).

Decision support tools, disease registries, and patient engagement materials can improve population-based chronic kidney disease care.

April 2018

Financial Burden of Healthcare Utilization in Consumer-Directed Health Plans

Progress of Diabetes Severity Associated With Severe Hypoglycemia in Taiwan

Physician and Patient Tools to Improve Chronic Kidney Disease Care

Limited Distribution Networks Stifle Competition in the Generic and Biosimilar Drug Industries

Provider and Patient Burdens of Obtaining Oral Anticancer Medications

Delivering on the Value Proposition of Precision Medicine: The View From Healthcare Payers

Cost Sharing and Branded Antidepressant Initiation Among Patients Treated With Generics

Care Coordination for Children With Special Needs in Medicaid: Lessons From Medicare

The Well-Being of Long-Term Cancer Survivors

A Payer–Provider Partnership for Integrated Care of Patients Receiving Dialysis