The current literature on preventable readmissions in the US contains evidence from a variety of populations, locations, settings, designs, and conditions. If a single common set of consistent patient-level risk factors can be distilled from this review it would include a variety of measures of poor-health or frailty: co-morbidities [5
], increasing severity class [23
], increasing age [5
], general poor health [44
], or high previous utilization of the healthcare system [5
]. In addition, some studies highlighted racial/ethnic disparities in preventable readmission for diabetics [31
], patients with pulmonary embolism [23
], and other preventable conditions [2
]. However, these potential risk factors are common to other investigations of hospitalization. In Jencks et al.
] recent examination of rehospitalizations (where they make no claim to preventability), they identified similar indicators of patient ill-health and disparities by race, socio-economic status, and geography. Other types of healthcare utilization show similar patterns: disparities according to race/ethnicity [56
] and risks based on age [57
] for hospitalizations due to ambulatory care sensitive hospitalizations, and those with poor health are more likely to be frequent users of emergency departments [58
The combined results of encounter level factors run along similar lines. Across multiple conditions, encounters covered by Medicaid [2
] or self-pay [2
] were indicators of increased odds of subsequent preventable readmissions; these are again probably proxies for either socio-economic status or access to primary care issues. In addition, while length of stay is encounter-specific and identified as an associated factor in multiple studies [5
], it may in part reflect underlying patient health [59
]. The same may be true for those studies that indicated discharge to some other care facility or supplemental care were associated with readmission [23
Intuitively and from a few studies in this review, we know that the admitting hospital may make a difference on subsequent readmissions. We cannot definitively say why or how. We do not know if the admitting hospital actually exerts some effect (through structures, policies, and procedures), or if it is merely variation for which examinations must account. Several studies documented that hospitals are different [23
], but very few looked for organizational-level factors. Even when organizational factors are explicitly examined, we are still uncertain about the magnitude or validity of the effect because statistical assumptions were violated [5
In similar fashion, the results of factors at the environment level are, on balance, more suggestive than informative at this point. Living in a private residence [49
], difficulty in getting care givers [52
], or lack of social support [45
] are really features of the patient's environment. However, only the study by Schwarz [45
] used multivariate statics, theoretically linked the index and readmission, and ensured adequate patient follow-up. Even then, the study focused on a small, narrowly defined population. Ferraris et al.
] found a patient's zip code associated with unplanned readmissions, but knowing what these results means is obscured because we know nothing about the resources or socioeconomics of the areas, and the modeling fails to account for multilevel measurement. By specifically modeling the zip code, Ferraris et al
. were asserting that the environment has an effect. Likewise, Weeks et al.
] found effects for rural residence. The result is intriguing, but the questions about the underlying mechanism accounting for the risk it raises are more logically answered by features of the environment: is it access to specialists, primary care, or rehabilitation and preventative services? While residence could be considered a patient-level variable, we would argue that rurality is more about the patients' context, and less about their own characteristics and behaviors.
The current research is missing in-depth examinations of more than one aspect of preventable readmissions. While it is fairly clear that patients with markers of general poor health are more likely to come back to the hospital, our knowledge about encounter-level factors is predominately related to length of stay and payer. Variance in the former depends substantially upon condition, and the latter is confounded by socioeconomic status, access, and a host of other factors. Few studies ventured to examine organizational and environmental factors. Fortunately, these gaps can be readily addressed. All multi-facility investigations using large databases could easily incorporate organizational level factors and utilize random effects or other cluster adjustments. The now more widespread appreciation of statistical methods for handling clustered data and improved computer power means the more sophisticated statistical methods utilized by a few studies in this review can be replicated. Furthermore, numerous structural and performance measures are available from existing surveys. Additionally, factors measured at the zip code level, like poverty or availability of primary care, are easily attainable and provide information on neighborhood effects and area resources. Again, these factors can be incorporated into models given the appropriate choice of statistical technique.
Variance in definitions makes drawing on the existing literature difficult
This paper has focused on preventable readmissions, but this is a term of convenience because the underlying possibility of prevention is variable across different readmissions. Unfortunately, it is frequently difficult to decide just how preventable the readmissions truly are due to numerous timeframes, the pervasive lack of conceptual clarity, and the varying use of terminology. Synthesizing results is thus hampered. These definitional difficulties call for a clear, shared vocabulary; for the choice of term makes a difference, as it not only indicates the degree to which the readmission is preventable, but also suggests by what mechanism prevention may be achieved.
This review contributes to that effort as some of the studies reviewed make strong efforts at conceptual and definitional clarity. From those studies, we can start to apply some common definitions and order to these terms. First, the term 'early' stresses the temporal association between the index and subsequent admissions. However, causality is not definite, because both elective and non-elective readmissions can occur shortly after discharge [8
]. 'Unplanned or non-elective' readmissions are not scheduled occurrences part of the medical process and undesired returns to the hospital [24
]. These labels are more descriptive and restrictive than simply 'early' because they eliminate some obviously non-preventable readmissions from consideration. Additionally, the word 'unplanned' sounds more like an aberrant event in the medical intervention initiated at the hospital, which ties the readmission to the care received during the index hospitalization. Finally, two terms clearly indicate a belief that intervention could effectively reduce the probability of readmission and employ more causal-type language. 'Potentially avoidable' draws upon the language of ambulatory care sensitive conditions, signifying appropriate, quality primary care can prevent readmission [25
]. By utilizing this established literature base, this label indicates a general strategy to reduce readmissions by improving the quality of, and access to, post-discharge care and patient management. 'Potentially preventable' was used by Goldfield et al.
] to describe clinically related, needless readmissions that quality care, discharge planning, follow up, or improved coordination would avert; this terminology not only claims a high expectation of preventability, but also implies broader opportunities for intervention inside and outside the hospital. Descriptions of readmissions adhering to the above terms and concepts would greatly facilitate comparisons between studies and simplify the national conversation on reform.
Methodological challenges make applying the existing literature to local practice difficult
Researchers, administrators, and clinicians have over many years pursued identification of readmission cases through predictive models with intentions of effectively intervening to extend or support a patient's care after discharge. While this review identified some consistent factors for such a model, it also catalogued a great deal of variety. For every reasonably consistent factor, like increasing co-morbidity scores, older age, or race/ethnicity disparities, there appeared to be multiple, detailed factors specific to the index and readmitting condition, like type of cardiovascular treatments, intraoperative measurements, surgical approaches, or specific existing conditions. This suggests a statistical model of just preventable readmissions may prove to be too elusive and that we should focus on condition specific preventable readmissions, either through stratified models or categorical dependent variables. While more complicated, that approach may prove more effective. Studies that do not restrict analysis to a single set of clinically-related index and readmitting conditions are most likely limited to effectively modeling only general risk factors, because the distinctive risks for various conditions may be may be lost in, or overpowered by, variables that apply to all conditions. Unfortunately, it is probably the condition-specific risks that provide the most opportunity for effective intervention within the hospital and in post-discharge settings. However, as much of the organizational and environmental factors are yet untapped, more information in the future may allow the question to be reexamined.
Four practical methodological challenges also hinder application of results in local practice. First, the studies in this review included both analyses of secondary linked datasets and those that relied on primary data collection and chart review. There is a difficulty in rectifying these two methods. Because primary data collection allows for many more detailed factors that may not be available in administrative databases, some findings may not be able to be utilized by those working in secondary data. In addition, the large sample sizes of the linked datasets may have indentified factors that will not be detectable in single-site studies. If it takes statewide or nationwide databases to identify statistically significant predictors because their effects are so small, it is difficult to assume any single facility will be able to generate the same level of precision in their own models. This is particularly true if we are going to have to stratify predictive models by specific condition or procedure. Third, the ability to adequately identify patients' previous and subsequent admissions may be very difficult for some facilities. The majority of studies relied on linked databases to ensure that all admissions to other facilities were being captured. Otherwise, extensive primary data collection was required. If facilities opt not to invest in primary data collection and patient follow-up, the ability of any single organization to identify their facility-specific risk factors for preventable readmissions may have to wait for fully developed local heath information exchange to follow patients between providers. Alternatively, the agency responsible for aggregating discharge claims within each state may have to take on the burden of patient matching. Finally, while in this review we have already advocated for more appropriate statistical techniques to account for the clustered nature of readmission, we recognize this type of modeling is not easy. Random-effects modeling requires expertise, specialized software, and sufficient computing power. Some organizations, like academic medical centers or VA facilities with access to health service research postdoctoral fellows, may be better positioned to engage in this type of predictive modeling. For other organizations, these approaches may be beyond their in-house capabilities.
Strategies for hospitals
If the conventional wisdom is to be believed, the cost of preventable readmissions will be borne principally by hospitals. However, as suggested in the introduction and as the existing literature has borne out, preventable readmissions are influenced by factors at the patient, encounter, organizational, and environmental levels. Which of these factors are actually in the hospitals' control or even amenable to direct influence?
Obviously, individual patient characteristics require significant consideration for those planning any interventions. It is an interesting contradiction that patient-level characteristics were the dominant area of inquiry for the reviewed studies, but most of these characteristics seem to be out of the hospitals' direct control. As O'Brien noted, 'unfortunately, many of these patient characteristics cannot be altered' [[37
] p2142]; a somewhat fatalistic comment, suggesting that research will need to increasingly identify behaviors and or contexts that can be targeted by interventions and evaluations. Furthermore, the increased risk for a preventable readmission for patients discharged against medical advice [23
] does not particularly bode well for any ideas that the hospital will be able to effectively influence subsequent health behaviors or even monitor resource utilization [50
]. However, more than one study in the earlier review by Benbassat and Taragin [3
] found interventions to provide post-discharge support or assistance reduced readmissions, and more recently, some systems such as Geisinger [60
] report success with patient-follow up after discharge.
Several encounter-level risk factors identified in this review, particularly those pertaining to specific procedures and medical interventions, are changeable by hospitals. In fact, the reviewed literature makes a few explicit recommendations, but these changes or improvements to clinical care while in the hospital are very condition-specific [21
]. The fact that there are so few specific recommendations for providers of care is not surprising because much of the literature was admittedly focused primarily on measurement methods [26
] and policy issues broader than intra-hospital operations [5
]. Therefore, beyond the few clinically-specific recommendations, the bulk of the remaining encounter-level risk factors hospitals either actually cannot change (such as who pays for the encounter or if the patient leaves against medical advice) or a simple, all-encompassing recommendation that is much more difficult (as in the case of length of stay, which is subject to a host of condition-specific clinical and payer influences). Similarly, hospitals may have limited or no effect on the supply or quality of primary care providers or home health, rehabilitation, or skilled nursing programs or facilities that may impact readmissions.
As deterministic actors, hospitals can make changes to their structure and processes and push back against environmental forces. Although hospitals can clearly change themselves and at least try to change the environment of their patients, the existing literature gives little guidance. As noted, the reviewed studies did not identify any organizational-level factors that can be easily targeted for change. Environmental-level determinants were also infrequently examined, but at least there we have some ideas of plausible interventions, mostly in the arena of changing patients' immediate support network. For example, Weaver et al.
] advised coordination with social workers or case managers during the discharge of cancer patients, and Timms et al.
] advocated for more qualitative information gathering through interviews with the patients, family members, and caregivers about the needs of elderly patients. These recommendations can be empirically tested in highly variable settings by multi-hospital systems or independent hospitals working on a joint program of research using quasi-experimental designs.
So what should hospitals do? Multiple options are available, but the choice of approach, in part, reflects the organization's underlying assumptions about the causes of readmissions, the applicability of predictive models, and the forthcoming financial policies. One viewpoint is that preventable readmissions are clearly a measure of overall hospital quality and that all preventable readmissions, regardless of causes, have some underlying driving factors [1
]. While this view would allow for statistical modeling as an effective means of performance measurement [26
], philosophically it implies that the search for individual risk factors or single interventions is too narrow in scope. If one accepts that preventable readmissions are failures at multiple processes, levels, and structures of healthcare, then these readmissions stand as a global indicator, not a single data point the organization tries to move; the potential changes in reimbursement are not intended to change a targeted practice or behavior, but to spur overall quality. That viewpoint suggests the solution to preventable readmissions is improvement in overall quality. That is definitely a hospital-centric view, where the efforts of the hospital are paramount in affecting preventable readmissions. In support of this view is that evidence indicates some hospitals are both better than expected and better than their peers in terms preventable readmission rates [26
]. Maybe these are the higher quality hospitals, or simply those who care for patients with lower severity conditions, or are located near more higher quality primary care and post-discharge care providers. For those organizations performing poorly on preventable readmissions, the implication is the need for organization-wide transformation. Transformation is not the adoption of a single technology or approach, but a profound change in the entire organization's culture and processes that improves quality [61
]. Unfortunately, the transformation in healthcare organizations has not been easily or widely achieved [65
A second general viewpoint is that preventable readmissions are not about the quality of care [33
]. Preventable readmissions are more about the person receiving care [24
] and the viewpoint is marked by phrases like 'unpredictable sequel' [33
] and 'cannot be predicted' [24
]. While not as dismissive of preventable readmissions as a marker of quality as the preceding quotations, those focusing on patients' post-discharge experiences, contexts, and resources [2
] could also be considered as sharing this extra-hospital viewpoint. This view is in stark contrast to the hospital-centric viewpoint, because whether preventable readmissions occur from pre-existing co-morbidities, health behaviors, or access to primary care, these things are all beyond the scope of services provided by the traditional inpatient setting. Reimbursement reform, therefore becomes an unfair financial penalty [67
] that hospitals try to avoid through various targeted initiatives like improved information systems [18
], case managers [52
], and post-discharges follow up [60
]. The underlying theme of these approaches and this extra-hospital view is that patients in some fashion have to be actively managed, because the negative financial outcomes are too great to take a passive role. For example, Ferraris et al.
] offered a practical, but an admittedly untested solution to the risk posed by patient co-morbidities: treat co-morbidities that raise the risk of readmission preoperatively. While intuitively a logical approach, this suggestion is more plausible under certain scenarios than others. A sufficient structure has to be in place to deliver that treatment. In case of infections, that care can exist within the hospital, but for chronic conditions, hospitals would need to possess an ambulatory care service line or have a strong connection to ambulatory care providers.
The concern over factors not modifiable by the hospital and the perceived need for continued, active post-discharge management are the types of reasons that justify integrated delivery systems and, now, the push toward accountable care organizations. Through vertical integration, integrated delivery systems are (theoretically) poised to facilitate transitions between different levels of care, and the care between inpatient, outpatient, and ambulatory care are better aligned. Accountable care organizations are to achieve the same alignment of effort toward the care of a population of patients [68
]. Becoming an integrated delivery system is not exactly a fast or necessarily feasible response. Accountable care organizations function under a variety of structures, possibly tied together only through a joint financial arrangement like a bundled payment or shared information system, which is at least somewhat more feasible to develop. Alternatively, those with the extra-hospital view will undoubtedly continue to look for more effective interventions for patients they rarely see.
First, as is the case with all reviews, even though we searched six databases for this review it is possible we omitted some studies. One of the included databases does include grey literature, but we would assume that is the source area in which this review may be lacking. However, because we were not attempting to quantify any effect sizes, this deficit probably does not dramatically alter any of our conclusions. Second, because we are concerned with the effects of organization and environment as well as the individual- and encounter-level determinants of readmissions, we limited our investigation to US based studies. However, significant and high quality work in defining and modeling predictive readmissions has been done internationally. A cursory look at this literature concurs with our earlier assumption of consistency of patient-level and encounter characteristics internationally. For example, older age [69
], ill health [70
], longer length of stays [71
], and prior utilization [69
] also appear as risks in other countries. Although potentially technically challenging, cross-national comparisons may prove to be very informative.