This article requires a subscription to view the full text. If you have a subscription you may use the login form below to view the article. Access to this article can also be purchased.
Abstract
Background and objectives Patients on hemodialysis have high 30-day unplanned readmission rates. Using a national all-payer administrative database, we describe the epidemiology of 30-day unplanned readmissions in patients on hemodialysis, determine concordance of reasons for initial admission and readmission, and identify predictors for readmission.
Design, setting, participants, & measurements This is a retrospective cohort study using the Nationwide Readmission Database from the year 2013 to identify index admissions and readmission in patients with ESRD on hemodialysis. The Clinical Classification Software was used to categorize admission diagnosis into mutually exclusive clinically meaningful categories and determine concordance of reasons for admission on index hospitalizations and readmissions. Survey logistic regression was used to identify predictors of at least one readmission.
Results During 2013, there were 87,302 (22%) index admissions with at least one 30-day unplanned readmission. Although patient and hospital characteristics were statistically different between those with and without readmissions, there were small absolute differences. The highest readmission rate was for acute myocardial infarction (25%), whereas the lowest readmission rate was for hypertension (20%). The primary reasons for initial hospitalization and subsequent 30-day readmission were discordant in 80% of admissions. Comorbidities that were associated with readmissions included depression (odds ratio, 1.10; 95% confidence interval [95% CI], 1.05 to 1.15; P<0.001), drug abuse (odds ratio, 1.41; 95% CI, 1.31 to 1.51; P<0.001), and discharge against medical advice (odds ratio, 1.57; 95% CI, 1.45 to 1.70; P<0.001). A group of high utilizers, which constituted 2% of the population, was responsible for 20% of all readmissions.
Conclusions In patients with ESRD on hemodialysis, nearly one quarter of admissions were followed by a 30-day unplanned readmission. Most readmissions were for primary diagnoses that were different from initial hospitalization. A small proportion of patients accounted for a disproportionate number of readmissions.
- dialysis
- end stage kidney disease
- mortality
- Readmission
- Comorbidity
- depression
- Depressive Disorder
- hospitalization
- Humans
- hypertension
- Kidney Failure, Chronic
- Logistic Models
- Myocardial Infarction
- Odds Ratio
- Patient Discharge
- Patient Readmission
- renal dialysis
- Retrospective Studies
- Software
- Substance-Related Disorders
- Surveys and Questionnaires
Introduction
Patients with ESRD requiring hemodialysis (HD) are twice (35% versus 16%) as likely to be readmitted within 30 days of discharge compared with general patients on Medicare (1). This contributes to the overall economic burden, because inpatient costs account for approximately 40% of total Medicare expenditures for dialysis. Although national epidemiology of unplanned readmissions has been described in other diseases, there are limited data on patients with ESRD on HD, especially those with non-Medicare insurance (2,3).
For the majority of patients on HD, the first medical encounter after an inpatient discharge is at the outpatient HD unit. As of 2017, outpatient dialysis units will be indirectly penalized by the Centers for Medicare and Medicaid (CMS) for excessive readmissions as the standardized readmissions ratio (SRR) becomes part of the ESRD Quality Incentive Program. The SRR compares the observed number with the expected number of readmissions on the basis of multivariable adjustment for patient demographics, patient socioeconomic factors, and discharge hospital characteristics (4). Excluded from the SRR are index admissions with discharge against medical advice (AMA), certain malignancy index admission diagnoses, and readmission within the first 3 days after discharge. Although the SRR is intended to incentivize reducing readmissions, there is a lack of information regarding characteristics and predictors of readmission.
Using a large, nationally representative, all-payer database, we examined the national readmission rate in patients on HD, examined the concordance of primary diagnosis on index admissions and readmissions, and determined independent factors associated with readmissions.
Materials and Methods
Data Sources
This is a retrospective cohort study using the Nationwide Readmission Database (NRD) from the Healthcare Cost Utilization Project (HCUP) (5). This database includes discharge data from 21 geographically dispersed states (Supplemental Table 1) and accounts for 49% of hospitalizations in the noninstitutionalized United States population. Data are drawn from the HCUP State Inpatient Databases (SIDs), with verified patient linkage numbers so that a person can be tracked across hospitals within a state. These states were chosen, because they are geographically dispersed and have verified patient linkage numbers on at least 90% of adult discharges. The NRD only contains data from community hospitals (nonfederal, short-term, general, and other specialty hospitals, excluding hospital units of institutions). It does not include data on rehabilitation or long-term acute care facilities. The database has a weighting variable that allows for national estimation, and it estimates approximately 36 million discharges in the United States. The weighting variable is derived from total hospitalizations from SIDs and numbers of hospitalizations in the sampling frame while taking into account hospital characteristics and patient characteristics. The NRD does not capture index-readmissions pairs that occurred in different states and does not have data on out of hospital mortality. We selected the year 2013, because we had full access to the database, and it was the most recent public release.
Study Population and Design
We queried the NRD using Internal Classification of Disease, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes to identify ESRD admissions. We excluded admissions with the primary diagnosis code of ESRD and those with AKI codes, because these patients were likely initiated on dialysis during index admission. A list of codes used to identify these admissions is in Supplemental Table 2. We excluded patients younger than 18 years old, on peritoneal dialysis, or with a renal transplant. Patients on peritoneal dialysis were excluded, because they represent a substantially different population with potentially different reasons for admissions. We also excluded those index admissions related to pregnancy and for chemotherapy. Readmissions that occurred the same day as a discharge from an index admission were considered a continuation of the index admission, because these readmissions were likely due to unresolved issues from the index admission. Finally, we excluded patients who died during the index admission, index admissions in December (lacking 30 days for accrual for readmissions), and those with missing discharge disposition, discharge date, and hospital length of stay. Two sensitivity analyses were done: the first one included only hospitalizations where the primary payer was Medicare and excluded those discharged AMA, and the second one included admissions that occurred the same day as a discharge. A study flow diagram is included in Supplemental Figure 1.
Definition of Index Admissions and Readmissions
Index hospitalizations were admissions without any hospitalizations in the preceding 30 days. Readmissions were any admissions within 30 days of an index hospitalization (Supplemental Figure 2). Unplanned readmissions were identified as any admissions that were not flagged as an elective admission using a database variable. A hospitalization within 30 days of a readmission was considered a readmission, and therefore, an index admission may be associated with multiple readmissions as long as they were <30 days apart. Readmission rates were calculated between index and first readmission, and they did not include readmissions between the first readmission and the second readmission, etc. This level of comparison was maintained throughout the analyses, except for the comparison of patient characteristics among those with one, two, and three or more readmissions, which was on an individual patient level. Although multiple readmissions per patient may be correlated, a study evaluating 30-day readmissions comparing methods accounting for correlation (generalized estimating equations and deduplication) did not find that they were superior to basic modeling (6).
Statistical Analyses
Descriptive statistics were generated using chi-squared tests for categorical variables, t tests for normally distributed continuous variables, Wilcoxon rank sum tests for non-normally distributed continuous variables, and ANOVA for analysis of two or more groups; observations with missing variables were excluded. Weighted numbers were used for analysis. We calculated the crude number and proportion to assess the most common primary diagnoses for index admissions and readmissions. We grouped primary diagnoses of index admissions and readmissions by Clinical Classification Software (CCS) codes for clinically meaningful comparisons (7). CCS categories were used to calculate the percentage concordance (those with the same clinical category) versus discordance (those with a different clinical category) between primary diagnosis of index admissions and the first associated readmission. The top ten CCS diagnoses were chosen as a representative sample.
We used survey logistic regression to assess the relationship between patient and admission characteristics and the odds of the first 30-day unplanned readmission. This is an appropriate analysis for data with nested, weighted observations, such as the NRD, which is inherently stratified in clusters to produce national estimates (8). We adjusted for patient demographics (age, sex, and median household income category for the patient’s zip code), primary payer (Medicare, Medicaid, commercial insurance, self-pay, or no charge), and index admission characteristics (hospital size, hospital type, private versus government control, and discharge disposition). Discharge disposition was grouped into routine or self-care, short-term hospital nursing facility, home health care, AMA, and alive (unknown). We used validated All Patient Refined Diagnosis-Related Group (APR-DRG) scores to account for severity of illness (9). Because the database did not have any information regarding out of hospital mortality, this could not be included in our analysis.
Comorbidities were defined using the NRD-defined comorbidity measures and included in our model using the CCS classification system, which aggregates ICD-9-CM diagnosis and procedure codes into mutually exclusive categories as covariates. Because the number of chronic conditions is limited to, at most, one per CCS category, a CCS category count is an indicator of medical multimorbidity (10). We also used another risk adjustment model using paired APR-DRG as a categorical variable in conjunction with severity of illness scores (11).
To measure comorbidity effect on readmission risk at the population level, which depends on both the adjusted odds ratio (aOR) and the comorbidity prevalence, we calculated population-attributable fractions (PAFs), taking into account sampling weights using the survey package in R, version 3.2.2 (R Foundation for Statistical Computing, Vienna, Austria). The PAF is calculated using the aOR, controlling for factors as described earlier, and comparing a hypothetical scenario with the outcome being readmissions, where none of the patients had the selected comorbidities, with the actual outcomes from the data; 95% CIs were calculated by bootstrapping (12,13).
All significance levels were two sided, with a P value <0.01 considered to be statistically significant. Analyses were done using SAS 9.2 (SAS Institute Inc.). Because we used publically available deidentified data, the study was considered to be institutional review board exempt.
Results
Readmission Rates and Characteristics
There were 390,627 index hospitalizations in patients on HD identified from the NRD; 7861 readmissions were flagged as elective and therefore, excluded from our analytic set, and 87,302 (22%) index admissions were followed by at least one readmission (Supplemental Table 3). On separate sensitivity analysis with only non-AMA patients on Medicare and including readmissions that occurred the same day as discharge found similar overall readmission rates 21% and 22%, respectively (Supplemental Figure 3).
Table 1 presents the baseline characteristics of admissions without readmissions and those with one or more readmission. Although readmissions were statistically different than those without readmissions on several patient demographics and comorbidities, there were small absolute differences. Index hospitalizations with readmissions were less likely to have commercial insurance (6% versus 8%; P<0.001).
Patient-, admission-, and hospital-level characteristics for patients with ESRD without and with at least one readmission
There were significant differences in the index admission characteristics between groups. Those with readmissions were more likely to be admitted through the emergency department (83% versus 78%; P<0.001), have nonelective admissions (92% versus 88%; P<0.001), have higher APR-DRG severity scores (70% versus 64% with major/extreme loss of function; P<0.001), and have longer length of stay (median, 3.8 days; interquartile range [IQR], 1.9–7.1 versus median, 3.3 days; IQR, 1.7–6.2; P<0.001).
Primary Diagnosis on Index Admissions and Concordance on Readmission
The top ten primary diagnoses on index admission were for complications of vascular access (which included admissions for arteriovenous fistulas/grafts/central venous catheters; 11%), hypertension (8%), septicemia (6%), congestive heart failure (6%), fluid disorder (6%), diabetes mellitus (5%), pneumonia (5%), cardiac dysrhythmias (2%), acute myocardial infarction (AMI; 2%), and chest pain (2%), respectively (Figure 1); this represented 54% of all index admissions. The absolute number of patients within these diagnoses stratified by readmission categories is shown in Supplemental Figure 4. AMI index admissions had the highest percentage of readmissions (25%), whereas hypertension had the lowest percentage of readmissions (20%) (Figure 1).
High readmission rate among the top ten reasons for index admission. Black bars represent the percentage of index admissions for each primary diagnosis compared with all index admissions for patients with ESRD on hemodialysis. Gray bars represent the percentage of index admissions with any readmission for each primary diagnosis. AMI, acute myocardial infarction; CHF, congestive heart failure.
The primary diagnosis on index admission and associated first readmission were discordant for an average of 80% of admissions within the top ten index admission diagnosis. The percentage of discordance varied by admission diagnoses: the highest for chest pain (94%) and the lowest for diabetes mellitus (68%) (Figure 2).
Low concordance between index admission diagnosis and readmission diagnosis. Black bars represent the percentage of readmissions diagnoses that were different from index admission diagnoses. Gray bars represent the percentage of readmission diagnoses that were the same as index admission diagnosis. AMI, acute myocardial infarction; CCS, Clinical Classification Software; CHF, congestive heart failure; Dx, diagnosis.
Characteristics of Multiple Readmissions
Patients with three or more readmissions, referred to as high utilizers, constituted only 2% of all patients but accounted for 20% of all readmissions (Supplemental Figure 5). High utilizers compared with patients with only two readmissions or one readmission were younger (median age, 55 years old; IQR, 44–67 versus median age, 62 years old; IQR, 50–82 versus median age, 63 years old; IQR, 53–73, respectively; P<0.001), were more often men (56% versus 50% versus 50%, respectively; P<0.001), and had multiple comorbidities, notably AIDS (2% versus 0.9% versus 0.7%, respectively; P<0.001), depression (13% versus 11% versus 11%, respectively; P<0.001), and drug abuse (8% versus 5% versus 4%, respectively; P<0.001). They were less likely to have insurance (6% versus 2% versus 2%, respectively; P<0.001) and live in zip codes within the 76%–100% for median income (13% versus 14% versus 15%, respectively; P<0.001). They were more likely to be admitted to metropolitan teaching hospitals (55% versus 41% versus 40%, respectively; P<0.001) (Table 2).
Patient-, admission-, and hospital-level characteristics for patients with ESRD stratified by one, two, or at least three readmissions
Significant Predictors of Unplanned Readmission
On multivariate analysis, the following demographics were significantly associated with increased readmission: younger age of 18–34, 35–49, and 50–64 years old (aOR, 1.35; 95% CI, 1.26 to 1.45; aOR, 1.09, 95% CI, 1.05 to 1.14; aOR, 1.06, 95% CI, 1.02 to 1.10, respectively) compared with ≥65 years old and women (aOR, 1.08; 95% CI, 1.05 to 1.11). Within insurance categories, commercial insurance had the lowest odds of readmission compared with Medicaid (aOR, 0.70; 95% CI, 0.66 to 0.74). The comorbidities with the highest aORs for readmissions were drug abuse (aOR, 1.41; 95% CI, 1.31 to 1.51), liver disease (aOR, 1.21; 95% CI, 1.14 to 1.27), and chronic pulmonary disease (aOR, 1.16; 95% CI, 1.12 to 1.20). Of note, depression was highly associated with readmissions (aOR, 1.10; 95% CI, 1.05 to 1.15) (Table 3). Index admissions characteristics associated with increased readmission odds included extreme loss of function on the APR-DRG severity scale (aOR, 1.46; 95% CI, 1.31 to 1.63), length of stay ≥7 days (aOR, 1.21; 95% CI, 1.16 to 1.26), and discharge AMA (aOR, 1.57; 95% CI, 1.45 to 1.70).
Multivariable analysis of predictors for unplanned 30-day readmissions
We calculated PAFs for comorbidities with increased adjusted odds of readmission (Supplemental Table 4). The highest partial PAFs were for hypertension (3.7%), chronic pulmonary disease (2.5%), and heart failure (1.9%). However, the PAFs for drug abuse (0.9%) and depression (0.8%) were on par with diabetes mellitus (0.9%) and higher than those for AIDS (0.06%) and liver disease (0.7%).
Discussion
Using a nationally representative all-payer database, we elucidated the top ten diagnoses for admissions in patients with ESRD on HD. The overall readmission rate was 22%, and on average, 80% of readmissions diagnoses were discordant to index admission. Interestingly, 2% of patients on HD accounted for 20% of all readmissions, and these high utilizers differed in many aspects from other patients on HD. Finally, we identified patient and hospital characteristics independently associated with increased odds of readmission.
Although the readmission rate that we report is low compared with that of the US Renal Data System (USRDS; 37%), it is similar to that of other studies (20%–24%) (1,14–17). Studies using the Dialysis Outcomes and Practice Patterns (DOPPS) data report a similar readmission rate, whereas Canadian studies show a substantially lower readmission rate: 15%–17% (18–21). Although the USRDS reports readmissions at the admission level and only includes patients on Medicare, we reported readmissions at the patient level and included patients not on Medicare (20%). Analysis of Medicare non-AMA hospitalizations revealed no substantial difference in readmission rates. The rate may also be affected, because the database does not capture readmissions that occurred across states. Lastly, we are unable to determine the number of patients who were discharged and died outside of the hospital, which may have contributed to the lower readmission rate seen in this dataset. Despite the lower readmission rates in this study, nearly one quarter of all admissions were followed by an unplanned readmission. Although the CMS has been tracking readmissions for multiple diagnoses since 2007, only recently has the CMS decided to track ESRD readmissions (22). Although currently, only Medicare payments will be penalized, other insurances are likely to follow Medicare’s example (23,24).
The discordance between index and readmission diagnoses is similar to that in other chronic disease (i.e., congestive heart failure, AMI, and pneumonia) hospitalizations, which report 19%–37% concordance (2,25). In contrast, an analysis of the DOPPS found nearly 50% concordance; this may have been due to differences in methodology, including diagnoses classification and the inclusion of chemotherapy-related hospitalizations, which showed the highest concordance (18). We categorized diagnoses using the CCS, which may lead to loss of granularity; however, it enhances the likelihood that concordance categorization is accurate.
The high readmission rates and the discordance with admission diagnosis may be reflective of the high disease burden in patients with ESRD rather than poor care at index admission. An analysis of patients on Medicare found that additional adjustments not included in the Medicare risk-adjustment model explained a substantial proportion of the difference in readmission risk between hospitals (26). Research using more granular data is needed to determine whether the readmissions due to a different clinical reason are preventable and whether focusing on the overall care of the multimorbid patient with ESRD rather than individual diagnoses might improve readmission rates.
Interventions to reduce ESRD readmissions have not targeted the subset of patients with psychosocial comorbidities (27–29). Although the PAFs were low due to the low prevalence, psychosocial comorbidities are often under-recognized during acute hospitalizations and underdiagnosed in the ESRD population, despite evidence that they are associated with readmissions in both patients on HD and patients not on HD (30–33). Additionally, these characteristics are associated with increased odds of emergency room utilization (34). Therefore, research should focus on identifying interventions that can improve care in patients with these psychosocial comorbidities, because improvements will have a global effect on not only hospital readmissions but also, emergency department utilization.
An unexpected finding is that younger age was associated with readmissions, and the high-utilizer group was nearly 10 years younger than those with one readmission. Potential explanations may be due to the higher proportion of patients without insurance; however, this needs to be further explored. Others have speculated on the effect of higher rates of nonadherence and underlying psychosocial factors in younger patients (3,35) Unfortunately, nonadherence and psychosocial factors are often under-reported and difficult to capture in electronic medical records. Although discharge alive destination unknown was associated with lower odds of readmission, the meaning of this finding is unclear, because only 20 hospitalizations were under this category.
A potential area of interest is collaborative mental health care in outpatient dialysis units or improvement in referrals to psychiatric professionals, which has been shown to reduce health care utilization and improve outcomes in other chronic diseases (36,37). Another potential intervention in the dialysis setting is increasing the number of nephrology visits, because one additional visit in the month after hospital discharge was estimated to reduce 30-day readmissions by 3.5% in patients with ESRD (16). Special focus on risk conditions identified in our analysis may help identify patients to target for more frequent visits. The high readmission rates, regardless of initial admission diagnoses, suggest that effective interventions to reduce admissions would have a widespread effect across admission diagnoses.
Our study should be interpreted in the context of several limitations. First, the database used in this analysis did not capture certain granular patient-level data, such as dialysis vintage, outpatient dialysis unit, laboratory data, shortened and/or skipped HD treatments, vascular access type, and race, which are important risk predictors. The use of administrative codes is likely to misclassify patients; a systematic review shows a sensitivity ranging from 40% to 80% but a specificity of >90% (38). Second, the broad groupings of diagnoses codes used to assess discordance between index admissions and readmissions could misclassify some admission-readmission pairs. However, the CCS schema was designed to group diagnoses into clinically meaningful categories; therefore, misclassification may not be clinically relevant. Third, this study did not address discharge factors, such as discharge date over the weekend, that were recently shown to also be predictive of 30-day readmission in patients on HD (39). The relevance of discharge factors to 30-day readmission supports shared accountability between the dialysis clinic and the acute care hospital, which would, in turn, foster care coordination between these settings. Fourth, we recognize that the overall prevalence of multiple significant predictors is low; however, these predictors may help identify a small group of patients who would benefit most from effective interventions to reduce readmissions. Despite these limitations, the benefit of using a large, representative, all-payer sample affords a truly national perspective, and these findings are highly generalizable to the United States ESRD population, including the 20% of the patients who do not have Medicare and thus, are not captured by the USRDS.
In conclusion, this study highlights the high proportion of readmissions in the ESRD population and how a small minority of patients with ESRD are responsible for a disproportionate number of readmissions. We show that most readmission diagnoses are different from the index admission and that several predictors of readmissions are psychosocial in nature. If we were to attempt to improve readmissions in the vulnerable ESRD population, perhaps a good starting place would be to institute interventions targeted at high utilizers and create a validated risk score incorporating psychosocial factors (40). However, further studies are necessary to determine if interventions decrease 30-day readmission rates and improve patient outcomes.
Disclosures
None.
Acknowledgments
L.C. is supported, in part, by National Institutes of Health (NIH) grant 5T32DK007757–18. B.F. is supported, in part, by American Heart Association grant 16MCPRP31030016. G.N.N. is supported, in part, by the NIH grant 1K23DK107908-01A1.
Footnotes
Published online ahead of print. Publication date available at www.cjasn.org.
See related editorial, “Thirty-Day Hospital Readmissions in the Hemodialysis Population: A Problem Well Put, But Half-Solved,” on pages 1566–1568.
This article contains supplemental material online at http://cjasn.asnjournals.org/lookup/suppl/doi:10.2215/CJN.02600317/-/DCSupplemental.
- Received March 8, 2017.
- Accepted June 26, 2017.
- Copyright © 2017 by the American Society of Nephrology
References
In this issue
Jump to section
More in this TOC Section
Original Articles
Maintenance Dialysis
Cited By...
Similar Articles
Related Articles
Keywords
- dialysis
- end stage kidney disease
- mortality
- Readmission
- Comorbidity
- Depression
- Depressive Disorder
- hospitalization
- Humans
- hypertension
- Kidney Failure, Chronic
- Logistic Models
- Myocardial Infarction
- Odds Ratio
- Patient Discharge
- Patient Readmission
- renal dialysis
- Retrospective Studies
- Software
- Substance-Related Disorders
- Surveys and Questionnaires