Readmissions to hospital are increasingly being used as a quality-of-care measure. They can indicate quality of care, however, only if an important proportion of them are deemed avoidable. In our systematic review, we identified 34 studies that measured the proportion of readmissions deemed avoidable. Subjective criteria and variable methods were used in every study. The proportions of readmissions deemed avoidable varied widely between the studies. This variability makes it difficult to state with any certainty how often readmissions are preventable. Nevertheless, the median proportion of readmissions deemed avoidable (27.1%) is certainly lower than the 76% reported in 2007 by the Medicare Payment Advisory Commission to the US Congress.39
Although the variation seen in these studies could reflect true differences in quality of patient care, it also reflects the subjectivity of the outcome itself as well as differences in study characteristics, including patient and hospital types included; factors considered in determining avoidability of readmissions; sources of information used to judge avoidable status; and the minimum number of reviewers per case.
Although subjectivity will always exist when determining whether readmissions are avoidable, steps can be taken to minimize resulting error. First, parameters required for reviewing readmissions — such as which factors responsible for a readmission (e.g., physician, nurse, patient) are classified as avoidable — need to be clarified. Second, the use of multiple reviewers is essential when dealing with subjective outcomes such as avoidable readmissions. Because the accuracy of reviews is never perfect, the use of multiple reviewers helps ensure that patient classifications are as accurate as possible. Finally, latent class models can be used to analyze multiple reviews and generate the probability that each patient truly had an avoidable readmission.40–42
We believe that such models may be useful to classify avoidable readmissions more reliably.
Our study has limitations. First, although we used a clear and sensible search strategy that identified a large number of studies, we may have missed relevant publications. In addition, we limited studies to those published in English. However, given the large number of studies included in our review, it is unlikely that our overall conclusions would change meaningfully if any missed studies were included.
Second, we used transparent meta-regression modelling to identify the most important sources of heterogeneity between studies. Although we limited this model to three covariates to avoid overfitting of the model, significant heterogeneity remained. This finding is not unexpected given the extensive amount of heterogeneity between the studies (). In addition, the model’s outcome (proportion of readmissions deemed avoidable) will have notable error in it because of the subjectivity involved in the classification of readmissions as avoidable or not. This error will not be captured by the study-level factors in our regression model.
Third, we combined studies from different health care systems. Although some factors contributing to the proportion of avoidable readmissions are likely universal (e.g., incorrect diagnosis), other factors influencing readmission rates that are unique to particular health care systems (e.g., health insurance coverage) will not be captured in our model.
Finally, we were unable to summarize disease-specific proportions of avoidable readmissions, because they were rarely reported in studies that included a broad assortment of diseases. Future studies would need to address this issue to identify possible diseases that could be targeted for interventions to decrease the risk of avoidable readmissions.
Our study showed that the proportion of hospital readmissions deemed avoidable has yet to be reliably determined. Furthermore, we found a lack of consensus regarding the methods necessary to judge whether readmissions are avoidable. Given the large variation in the proportion of avoidable readmissions between studies using primary data, “avoidability” cannot accurately be inferred based on diagnostic codes for the index admission and the readmission. Instead, it needs to be determined through a peer-review process in which readmissions are classified as avoidable or not based on expert opinion.
Criteria used in future studies need to focus on determining whether the readmission was preceded by an adverse event (i.e., a bad medical outcome due to medical care rather than the natural history of disease or bad luck); whether the adverse event could have been prevented; and whether the readmission would have occurred even without the adverse event or whether other factors were involved. In addition, future studies need to include a large number of readmissions in a broad spectrum of patients from multiple teaching and community hospitals; multiple sources of patient information between index admission and readmission on which decisions regarding avoidabililty are based; an explicit statement about which groups or factors contributing to readmissions are considered avoidable; at least three reviewers per readmission to judge avoidability; and the use of structural modelling methods such as the latent class model to measure the probability that patients truly had an avoidable readmission based on the judgments of reviewers.