Patient context is represented as a balance between workload and capacity. This balance must be optimized to ensure care effectiveness and improve outcomes. In turn, the outcomes achieved feed back to affect the workload-capacity balance.
Size of the data marker corresponds to the relative weight assigned in the pooled analysis using random-effects models. RR indicates relative risk.
eAppendix 1. Hospital Readmissions Search Strategy
eAppendix 2. Excluded Full Text Articles and Rationale
eTable 1. Activity-Based Coding of Interventions
eTable 2. Risk of Bias of Individual Studies
eFigure 1. Summary of Evidence Search and Selection
eFigure 2. Summary of Risk of Bias Across Included Studies
eFigure 3. Funnel Plot: Publication Bias Plot Suggestive of Underpublication of Small Negative Trials
Leppin AL, Gionfriddo MR, Kessler M, Brito JP, Mair FS, Gallacher K, Wang Z, Erwin PJ, Sylvester T, Boehmer K, Ting HH, Murad MH, Shippee ND, Montori VM. Preventing 30-Day Hospital ReadmissionsA Systematic Review and Meta-analysis of Randomized Trials. JAMA Intern Med. 2014;174(7):1095-1107. doi:10.1001/jamainternmed.2014.1608
Copyright 2014 American Medical Association. All Rights Reserved. Applicable FARS/DFARS Restrictions Apply to Government Use.
Reducing early (<30 days) hospital readmissions is a policy priority aimed at improving health care quality. The cumulative complexity model conceptualizes patient context. It predicts that highly supportive discharge interventions will enhance patient capacity to enact burdensome self-care and avoid readmissions.
To synthesize the evidence of the efficacy of interventions to reduce early hospital readmissions and identify intervention features—including their impact on treatment burden and on patients’ capacity to enact postdischarge self-care—that might explain their varying effects.
We searched PubMed, Ovid MEDLINE, Ovid EMBASE, EBSCO CINAHL, and Scopus (1990 until April 1, 2013), contacted experts, and reviewed bibliographies.
Randomized trials that assessed the effect of interventions on all-cause or unplanned readmissions within 30 days of discharge in adult patients hospitalized for a medical or surgical cause for more than 24 hours and discharged to home.
Data Extraction and Synthesis
Reviewer pairs extracted trial characteristics and used an activity-based coding strategy to characterize the interventions; fidelity was confirmed with authors. Blinded to trial outcomes, reviewers noted the extent to which interventions placed additional work on patients after discharge or supported their capacity for self-care in accordance with the cumulative complexity model.
Main Outcomes and Measures
Relative risk of all-cause or unplanned readmission with or without out-of-hospital deaths at 30 days postdischarge.
In 42 trials, the tested interventions prevented early readmissions (pooled random-effects relative risk, 0.82 [95% CI, 0.73-0.91]; P < .001; I2 = 31%), a finding that was consistent across patient subgroups. Trials published before 2002 reported interventions that were 1.6 times more effective than those tested later (interaction P = .01). In exploratory subgroup analyses, interventions with many components (interaction P = .001), involving more individuals in care delivery (interaction P = .05), and supporting patient capacity for self-care (interaction P = .04) were 1.4, 1.3, and 1.3 times more effective than other interventions, respectively. A post hoc regression model showed incremental value in providing comprehensive, postdischarge support to patients and caregivers.
Conclusions and Relevance
Tested interventions are effective at reducing readmissions, but more effective interventions are complex and support patient capacity for self-care. Interventions tested more recently are less effective.
Early hospital readmissions have been recognized as a common and costly occurrence, particularly among elderly and high-risk patients. One in 5 Medicare beneficiaries is readmitted within 30 days, for example, at a cost of more than $26 billion per year.1 To encourage improvement in the quality of care and a reduction in unnecessary health expense, policymakers, reimbursement strategists, and the US government have made reducing 30-day hospital readmissions a national priority.2- 4 Achieving this goal, however, requires a more complete understanding of the underlying causes of readmission.
The cumulative complexity model (CuCoM)5 is a framework developed by our research group that conceptualizes patient context as a balance between workload and capacity (Figure 1). Workload consists of all the work of being a patient and includes efforts to understand and plan for care, to enroll the support of others, and to access and use health care services.6,7 Capacity is determined by the quality and availability of resources that patients can mobilize to carry out this work (physical and mental health, social capital, financial resources, and environmental assets). The CuCoM is novel in its consideration of the effects of treatment burden on patient context, and it illustrates how infeasible, unsupported, and context-irreverent care can lead to poor health outcomes and reduced health care effectiveness. Because patients recently discharged from the hospital are in a state of extreme physiologic and psychological vulnerability,8 their capacity for enacting self-care is low. The CuCoM predicts that, unless sufficient support is given to enhance patient and caregiver capacity to carry out the work of patienthood, placing highly burdensome discharge demands on these patients will lead to poor outcomes and hospital readmission.
To evaluate the validity of the CuCoM and provide hypothesis-generating work in the understanding of patient context, we chose to synthesize the evidence on the efficacy of interventions to reduce early hospital readmissions. In particular, we sought to determine the degree to which a number of intervention characteristics—including their impact on patient capacity and workload—might account for differences in their effectiveness.
A registered protocol (PROSPERO CRD42013004773) guided the conduct of this review,9 which we report in adherence to the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) Statement.10
Eligible studies were randomized trials reported in English or Spanish, since 1990, that assessed the effectiveness of peridischarge interventions vs any comparator on the risk of early (ie, within 30 days of discharge) all-cause or unplanned readmissions with or without out-of-hospital deaths. The intervention had to focus its efforts on the hospital-to-home transition, permit patients across arms to have otherwise similar inpatient experiences, and be generalizable to contexts beyond a single patient diagnosis. Adult patients had to be admitted from the community to an inpatient ward for at least 24 hours with a medical or surgical cause. Studies including obstetric or psychiatric admissions or only including discharges to skilled nursing or rehabilitation facilities were excluded.
In collaboration with an experienced research librarian (P.J.E.), we searched in April of 2013 the following databases: PubMed, Ovid MEDLINE, Ovid EMBASE, EBSCO CINAHL, and Scopus. The complete search strategy is reported in eAppendix 1 (in Supplement). Two reviewers (T.S. and A.L.L.) hand-searched the bibliographies of included studies and recent reviews. Experts in the field were asked to identify additional references.
Four reviewers (A.L.L., M.R.G., J.P.B., and T.S.) worked independently and considered the eligibility of candidate articles by examining their titles and abstracts, and then the full version of articles identified as potentially eligible by at least 1 reviewer. Conflicts about the eligibility of full articles were resolved by discussion and consensus. Eligibility was delayed for studies reporting outcomes incompletely, pending author contact.
After creating and piloting a standardized form, the reviewers (A.L.L., M.R.G., and J.P.B.), working independently and in duplicate and using a web-based program (DistillerSR), abstracted details about the patient population, the interventions compared, and the outcomes reported.
We abstracted details of the interventions tested verbatim from either the trial report or a cited protocol, limiting our focus to the period of hospitalization until 30 days after discharge, and identifying the “net intervention” by selecting out activities that occurred in the intervention arm but not in the control arm. These activities were coded using a taxonomy adapted from Hansen et al11 (Table 1). We also noted the number of meaningfully involved individuals participating in the intervention’s delivery and the number of meaningful interactions these individuals had with patients. Meaningfully involved individuals played a structured and requisite function in the delivery of central aspects of the intervention (eg, a physician who might be contacted only as needed would not be considered meaningfully involved). Similarly, meaningful patient interactions were defined as those that were the proposed sources of the intervention’s effectiveness (eg, a nurse visiting a patient only to deliver educational materials but not to actually engage in educational activity would not be considered a meaningful interaction). Two team members (A.L.L. and M.R.G.) created summary descriptions of the interventions in a standardized format; these were shared with each author to confirm their fidelity to what happened in the trial.
After calibrating judgments on a pilot sample, 2 raters familiar with the CuCoM (F.S.M. and K.G.), not involved in data collection and blinded to trial results, evaluated each standardized intervention description on a scale of 1 (substantially decrease) to 4 (no effect) to 7 (substantially increase) to reflect the degree to which the intervention was likely to affect patient workload and patient capacity for self-care. The impact on patient capacity was rated with perfect agreement 50% of the time and within 1 point of difference in 42% of cases (8% differed by 2 points). Because no interventions were rated to decrease patient capacity and all mean ratings fell within the range of 4.0 to 5.5, we elected to dichotomize the variable (threshold of ≥5 for increasing capacity) for analysis. Workload was more difficult to assess reliably: perfect agreement and minor disagreement (±1 point) were seen in 29% and 44% of cases, respectively, with 27% of cases differing by 2 or more points. This variable was divided into 3 categories (increase, decrease, no change).
For each included trial, we extracted or computed the risk of early readmission for each arm, analyzing patients as randomized (intention to treat analysis). We used the number randomized as the denominator except when the number of patients discharged was reported and differed from the number randomized. We selected the outcome to extract on the basis of an ad hoc hierarchy of outcomes of interest, with priority given to unplanned readmissions, then to all-cause readmissions, and finally to the composite end points of unplanned and all-cause readmissions plus out-of-hospital deaths, respectively. Outcomes were extracted and analyzed at the longest period of follow-up, up to 30 days from discharge. Examination of trials reporting the effect of interventions on more than 1 of these outcomes revealed that treatment effects were consistent across them (data not shown).
Two raters (A.L.L. and M.K.) worked independently and in duplicate to determine the extent to which each trial was at risk of bias using a standardized form based on the Cochrane Collaboration’s tool.12 The assessment considered the quality of the randomization sequence generation, allocation concealment, blinding of outcome assessors, the potential for missing outcomes (ie, likelihood of missing readmissions to other hospitals), and the proportion of patients lost to follow-up. For missing outcomes, “high risk of bias” was assigned when the readmissions data came from internal health system records only. To assess for publication bias, we examined a funnel plot for asymmetry and conducted asymmetry regression according to Sterne and Egger13 and determined the associated P value.
We used random-effects meta-analyses to estimate pooled risk ratios and 95% confidence intervals for early readmission.14,15 We tested for heterogeneity of effect on this outcome using the Cochran Q χ2 test16 and estimated between-trial inconsistency not due to chance using the I2 statistic.17
To explore the effects of patient, intervention, and outcome characteristics on the impact of measured intervention effectiveness, we conducted planned subgroup analyses, testing variables 1 at a time.
Patient characteristics tested were age (mean ≥65 years or not), diagnosis (heart failure or other), and hospital ward (general medical or other). Intervention characteristics tested included the number of unique activities involved in the intervention, the number of unique individuals or roles meaningfully involved in its delivery, the minimum number of meaningful patient interactions occurring within 30 days, the location of the intervention activity (ie, whether it occurred entirely during the inpatient stay, after discharge, or as a combination that “bridged” the transition), whether the intervention was rated to increase or decrease patient workload, and whether the intervention was rated to increase patient capacity (no intervention was found that decreased patient capacity for self-care). Ad hoc variables tested were year of publication and type of outcome reported (ie, unplanned readmissions vs other).
Informed by the findings of the exploratory subgroup analyses and our initial hypotheses, we constructed a post hoc metaregression model to test a variable that reflected the degree to which discharge interventions provided comprehensive patient and caregiver support. This “comprehensive support” variable could return values within a range of 0 to 4 “points” on the basis of whether the intervention (1) was rated to increase patient capacity, (2) had at least 5 (75th percentile of distribution) unique intervention activities, (3) had at least 5 (75th percentile of distribution) meaningful patient contacts, and (4) had at least 2 (75th percentile of distribution) individuals involved in its delivery. We created 3 categories for this variable: interventions with zero points (category 1), interventions with 1 or 2 points (category 2), and interventions with 3 or 4 points (category 3). To control for changes in standard care delivery over time, we adjusted on the basis of the year of publication variable.
Our initial database search generated 1128 reports (eFigure 1 in Supplement). Through abstract and title screening, 256 reports were identified for full-text review. During full-text screening (agreement, 89%), 24 were selected for inclusion and 39 were set aside for author contact prior to making a decision. Of 7 potentially eligible studies identified from bibliographies and expert consultation, 2 were included and 1 was set aside for author contact. Of the 40 trials requiring author contact for a final eligibility decision, 21 were deemed eligible. Of the 48 apparently eligible trials, 1 was found ineligible after the author confirmed that readmission data were collected only for readmissions related to the index diagnosis.18 The final sample therefore comprised 47 trials from 46 reports.19- 64
Of the 47 eligible trials, 42 contributed data for the primary meta-analysis, and 5 (those that reported numbers of readmissions rather than the number of patients readmitted) were analyzed separately.31,45,50,55,61 A complete list of excluded full-text studies with rationale for exclusion is available in eAppendix 2 (in Supplement).
Table 2 describes the included trials. Many were single-center trials taking place in academic medical centers, enrolling few patients (eg, 22 trials enrolled <200 patients), and reporting 30-day readmissions. Most interventions tested took place in both the inpatient and outpatient settings. The coded activity analysis is reported in eTable 1 (in Supplement). In general, interventions included anywhere from 1 to 7 unique activities. Case management, patient education, home visits, and self-management support were commonly present in net activity descriptions (eTable 1 in Supplement). Trial authors responded to confirmation requests for 34 of the 47 net intervention descriptions. Three authors requested minor modifications and 1 author made major modifications to these descriptions.
Most studies were at low risk of bias (eTable 2 and eFigure 2 in Supplement). The most common methodological limitation of these trials was the lack of a reliable method for dealing with missing data.
In the 42 trials reporting readmission rates, the overall pooled relative risk (RR) of readmission within 30 days was 0.82 (95% CI, 0.73-0.91; P < .001) (Figure 2). Inconsistency across trials was low (I2 = 31%). Funnel plot examination showed asymmetry suggestive of publication bias in the context of smaller studies (eFigure 3 in Supplement), and the Egger test was significant (P = .02). The 5 trials reporting number of readmissions (rather than number of patients with readmissions) had a pooled relative risk of readmission of 0.93 (95% CI, 0.72-1.20; I2 = 23%; P = .59). Although this result was consistent with the risk found in trials reporting readmission rates (interaction P = .38), we opted not to include these trials in subgroup analyses.
Subgroup analyses failed to find an interaction between trial results and patient characteristics or outcome measured (Table 3). A number of intervention characteristics, however, did interact with measured effectiveness. These include whether the intervention was rated to augment patient capacity for self-care (RR, 0.68 [95% CI, 0.53-0.86] when it was and RR, 0.88 [95% CI, 0.80-0.97] when it was not; interaction P = .04), whether the intervention had at least 5 unique, component activities (RR, 0.63 [95% CI, 0.53-0.76] when it did and RR, 0.91 [95% CI, 0.81-1.01] when it did not; interaction P = .001), and whether the intervention had at least 2 individuals involved in delivery (RR, 0.69 [95% CI, 0.57-0.84] when it did and RR, 0.87 [95% CI, 0.77-0.98] when it did not; interaction P = .05). Studies testing interventions more recently were associated with reduced effectiveness (RR, 0.89 [95% CI, 0.81-0.97] when published in 2002 or later and RR, 0.56 [95% CI, 0.40-0.79] when published prior to 2002; interaction P = .01). Other characteristics of the interventions, such as their rated effect on patient workload and the site of delivery, had no significant interaction with the intervention effect.
Despite potential colinearity of the contributing variables, metaregression showed a significant and incremental effect of “comprehensive support” on reducing readmissions (Table 4). Category 3 comprised 7 interventions.28,30,37,47,56,58,64 Compared with category 1 interventions, these were associated with a relative risk of readmission of 0.63 (95% CI, 0.43-0.91; P = .02). Category 3 interventions used a consistent and complex strategy that emphasized the assessment and addressing of factors related to patient context and capacity for self-care (including the impact of comorbidities, functional status, caregiver capabilities, socioeconomic factors, potential for self-management, and patient and caregiver goals for care). These interventions coordinated care across the inpatient-to-outpatient transition and involved multiple patient interactions; all but 128 involved patient home visits.
The body of randomized trial evidence shows a consistent and beneficial effect of tested interventions on the risk of 30-day readmissions. Exploratory subgroup analyses suggest that effective interventions are more complex and seek to enhance patient capacity to reliably access and enact postdischarge care. In addition, interventions tested more recently are, in general, less efficacious when compared with controls.
Our findings are consistent with the CuCoM in their suggestion that providing comprehensive and context-sensitive support to patients reduces the risk of early hospital readmission; however, we could not identify an effect of rated intervention workload on this risk.
Many studies in this review were conducted in single, academic centers; this raises questions about applicability. Also, the scales that we used to evaluate intervention effects on patient workload and capacity relied on global judgments (rather than criterion-based judgments) and are original to this work. To our knowledge, no validated scale exists to assess the potential of an intervention to impose patient workload or treatment burden and/or affect a patient’s capacity for self-care. Although our raters were consistent in their assessments of interventions’ effect on patient capacity, their judgment of impact on patient workload was less reliable. Particularly, raters believed that some burdensome interventions could be beneficial if the patient had the capacity and resources to access and enact the care. Because the experience of treatment burden is not constant between patients, an ideal analysis of its effects would be based on patient-reported assessments of intervention workload. Indeed, many eligible patients declined enrollment into some studies,23,28,44,50,53 often because they did not wish to take on the perceived burden of the intervention; evaluating the effect of intervention-imposed workload in such samples is of limited applicability. In general, these assessments should be regarded as hypothesis-generating and the inferences made on the basis of subgroup analyses must be viewed as tentative (given the potential for chance findings from testing multiple hypotheses and the possibility that some variables are correlated). Finally, despite robust efforts to obtain unpublished data, there was evidence of publication bias. The overall effect of this on our findings is not known.
This review also has many strengths. First, it provides, to our knowledge, the largest and most comprehensive assessment of discharge interventions and their effect on 30-day readmissions, including 47 randomized trials at low risk of bias. This is a stronger and less heterogeneous body of evidence than previously assembled,11,65 and it includes unpublished data from 18 trials. Our study used an activity-based coding method designed to ensure appropriate characterization of each intervention and the net difference in activity between intervention and control arms. This method contributes to the field and can be applied to future assessments of complex interventions. To our knowledge, this is also the first use of the CuCoM5 to analyze the impact of health care delivery interventions on patients as an explanation for their relative efficacy.
We identified 31 more randomized clinical trials than were accumulated in the most recent review of discharge intervention effects on 30-day readmission rates,11 and we provide the first meta-analysis on this topic. Although previous studies and reviews have suggested that “bundled” interventions are of greater value,11,65 this meta-analysis provides objective support for this claim. In addition, our study adds to and enhances the body of evidence related to the importance of patient contextual factors in affecting health outcomes.66
In this analysis, interventions that used a complex and supportive strategy to assess and address contextual issues and limitations in patient capacity were most effective at reducing early hospital readmissions. Many of these contacted the patient frequently, used home visits, and reported cost savings. This information can be used to guide the design and testing of future interventions. The CuCoM may also have value in helping to conceptualize the effects of health care interventions across diverse patient contexts, but we were unable to characterize a consistent effect of rated intervention workload on outcomes. Finally, we found that more recently tested interventions were less effective. We hypothesize that this may represent (1) a general improvement over time in the standard of care that was not fully appreciated in control descriptions, (2) an increased effort over time to test simpler and less comprehensive interventions, (3) a higher likelihood over time of more diverse interventions to measure and report 30-day readmission rates (eg, including those less focused on reducing early readmissions), and/or (4) a general shift away from interventions stressing human interaction toward those more high tech in nature. Additional study is needed to determine the implications of this finding.
Our results suggest that most interventions tested are effective in reducing the risk of early readmissions. Some features, however, may enhance the effect of these programs. In particular, we found value in interventions that supported patients’ capacity for self-care in their transition from hospital to home. Future work intended to improve the effectiveness of health care delivery may benefit from consideration of the demands that health care interventions place on recently discharged patients and their caregivers and the extent to which these demands are offset by comprehensive support for implementation.
Accepted for Publication: March 8, 2014.
Corresponding Author: Victor M. Montori, MD, Knowledge and Evaluation Research Unit, Department of Medicine, Mayo Clinic, 200 First St SW, Plummer Bldg, Rochester, MN 55905 (email@example.com).
Published Online: May 12, 2014. doi:10.1001/jamainternmed.2014.1608.
Author Contributions: Drs Leppin and Montori had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
Study concept and design: Leppin, Gionfriddo, Mair, Gallacher, Erwin, Murad, Shippee, Montori.
Acquisition, analysis, or interpretation of data: Leppin, Gionfriddo, Kessler, Brito, Mair, Gallacher, Wang, Sylvester, Boehmer, Ting, Murad, Montori.
Drafting of the manuscript: Leppin, Mair, Gallacher, Boehmer, Murad, Montori.
Critical revision of the manuscript for important intellectual content: All authors.
Statistical analysis: Wang, Murad, Montori.
Administrative, technical, or material support: Leppin, Kessler, Brito, Mair, Gallacher, Erwin, Sylvester, Boehmer, Montori.
Study supervision: Leppin, Montori.
Conflict of Interest Disclosures: None reported.
Funding/Support: This publication was made possible by Clinical and Translational Science Award grant UL1 TR000135 from the National Center for Advancing Translational Sciences, a component of the National Institutes of Health.
Role of the Sponsors: The funding source had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.
Disclaimer: The contents are solely the responsibility of the authors and do not necessarily represent the official view of the National Institutes of Health.
Additional Contributions: The following individuals provided unpublished data, conducted secondary analyses, assisted with study identification, and/or provided guidance and support: Agneta Björck Linné, MS, PhD, and Hans Liedholm, MD, PhD (Malmö University Hospital, Sweden); Marcia E. Leventhal, RN, MSN, Sabina De Geest, PhD, RN, and Kris Denhaerynck, PhD, RN (Institute of Nursing Science, University of Basel, Switzerland); Lars Rytter, MD (Glostrup University Hospital, Denmark); Gillian A. Whalley, PhD (University of Auckland, New Zealand); David Maslove, MD, FRCPC (University of Toronto, Canada); Judith Garcia-Aymerich, MD, PhD (Universitat Pomeu Fabra-Barcelona, Spain); Bonnie J. Wakefield, PhD, RN (Iowa City Veterans Affairs Healthcare System); Kathleen Finn, MD (Department of Medicine, Massachusetts General Hospital, Boston); Jon C. Tilburt, MD, MPH (Mayo Clinic); Christiane E. Angermann, MD (Universitätsklinikum Würzburg, Denmark); Felipe Atienza, MD, PhD (Hospital General Universitario Gregorio Maranon-Madrid, Spain); Dan Gronseth, BS (Mayo Clinic); Michael W. Rich, MD (Washington University, St Louis); Andrew Masica, MD, MSCI (Baylor Health Care System); Karen B. Hirschman, PhD, and Mary D. Naylor, PhD (University of Pennsylvania School of Nursing); James F. Graumlich, MD (University of Illinois College of Medicine at Peoria); Anna Strömberg, RN, PhD (Linköping University Hospital, Sweden). These contributors were not compensated for their contributions.